diff --git a/build/make/Android.mk b/build/make/Android.mk
index 826ff2f..816334e 100644
--- a/build/make/Android.mk
+++ b/build/make/Android.mk
@@ -38,8 +38,9 @@
 # For this we import the 'cpufeatures' module from the NDK sources.
 # libvpx can also be configured without this runtime detection method.
 # Configuring with --disable-runtime-cpu-detect will assume presence of NEON.
-# Configuring with --disable-runtime-cpu-detect --disable-neon will remove any
-# NEON dependency.
+# Configuring with --disable-runtime-cpu-detect --disable-neon \
+#     --disable-neon-asm
+# will remove any NEON dependency.
 
 # To change to building armeabi, run ./libvpx/configure again, but with
 # --target=arm5te-android-gcc and modify the Application.mk file to
diff --git a/build/make/Makefile b/build/make/Makefile
index c4d53f1..63ec271 100644
--- a/build/make/Makefile
+++ b/build/make/Makefile
@@ -19,6 +19,7 @@
         done
 all: .DEFAULT
 clean:: .DEFAULT
+exampletest: .DEFAULT
 install:: .DEFAULT
 test:: .DEFAULT
 testdata:: .DEFAULT
@@ -105,6 +106,8 @@
 
 .PHONY: dist
 dist:
+.PHONY: exampletest
+exampletest:
 .PHONY: install
 install::
 .PHONY: test
diff --git a/build/make/configure.sh b/build/make/configure.sh
index a2a64c2..ad7dc82 100755
--- a/build/make/configure.sh
+++ b/build/make/configure.sh
@@ -518,7 +518,7 @@
         --enable-?*|--disable-?*)
         eval `echo "$opt" | sed 's/--/action=/;s/-/ option=/;s/-/_/g'`
         if echo "${ARCH_EXT_LIST}" | grep "^ *$option\$" >/dev/null; then
-            [ $action = "disable" ] && RTCD_OPTIONS="${RTCD_OPTIONS}${opt} "
+            [ $action = "disable" ] && RTCD_OPTIONS="${RTCD_OPTIONS}--disable-${option} "
         elif [ $action = "disable" ] && ! disabled $option ; then
           echo "${CMDLINE_SELECT}" | grep "^ *$option\$" >/dev/null ||
             die_unknown $opt
@@ -835,7 +835,7 @@
                 check_add_cflags  -march=armv7-a -mfloat-abi=${float_abi}
                 check_add_asflags -march=armv7-a -mfloat-abi=${float_abi}
 
-                if enabled neon
+                if enabled neon || enabled neon_asm
                 then
                     check_add_cflags -mfpu=neon #-ftree-vectorize
                     check_add_asflags -mfpu=neon
@@ -882,7 +882,7 @@
             tune_asflags="--cpu="
             if [ -z "${tune_cpu}" ]; then
                 if [ ${tgt_isa} = "armv7" ]; then
-                    if enabled neon
+                    if enabled neon || enabled neon_asm
                     then
                         check_add_cflags --fpu=softvfp+vfpv3
                         check_add_asflags --fpu=softvfp+vfpv3
@@ -1284,8 +1284,8 @@
     local makefile=$2
     shift 2
     for cfg; do
-        upname="`toupper $cfg`"
         if enabled $cfg; then
+            upname="`toupper $cfg`"
             echo "${prefix}_${upname}=yes" >> $makefile
         fi
     done
diff --git a/configure b/configure
index bd95056..b6d645a 100755
--- a/configure
+++ b/configure
@@ -52,6 +52,7 @@
   ${toggle_multi_res_encoding}    enable multiple-resolution encoding
   ${toggle_temporal_denoising}    enable temporal denoising and disable the spatial denoiser
   ${toggle_webm_io}               enable input from and output to WebM container
+  ${toggle_libyuv}                enable libyuv
 
 Codecs:
   Codecs can be selectively enabled or disabled individually, or by family:
@@ -265,8 +266,9 @@
     unistd_h
 "
 EXPERIMENT_LIST="
-    multiple_arf
     alpha
+    multiple_arf
+    spatial_svc
 "
 CONFIG_LIST="
     external_build
@@ -314,6 +316,7 @@
     os_support
     unit_tests
     webm_io
+    libyuv
     decode_perf_tests
     multi_res_encoding
     temporal_denoising
@@ -367,6 +370,7 @@
     postproc_visualizer
     unit_tests
     webm_io
+    libyuv
     decode_perf_tests
     multi_res_encoding
     temporal_denoising
@@ -708,9 +712,11 @@
         *-vs*)
             soft_enable unit_tests
             soft_enable webm_io
+            soft_enable libyuv
         ;;
         *-android-*)
             soft_enable webm_io
+            soft_enable libyuv
             # GTestLog must be modified to use Android logging utilities.
         ;;
         *-darwin-*)
@@ -727,6 +733,9 @@
             check_cxx "$@" <<EOF && soft_enable webm_io
 int z;
 EOF
+            check_cxx "$@" <<EOF && soft_enable libyuv
+int z;
+EOF
         ;;
         *)
             enabled pthread_h && check_cxx "$@" <<EOF && soft_enable unit_tests
@@ -735,6 +744,9 @@
             check_cxx "$@" <<EOF && soft_enable webm_io
 int z;
 EOF
+            check_cxx "$@" <<EOF && soft_enable libyuv
+int z;
+EOF
         ;;
     esac
     # libwebm needs to be linked with C++ standard library
diff --git a/examples.mk b/examples.mk
index c36159f..946c030 100644
--- a/examples.mk
+++ b/examples.mk
@@ -10,10 +10,24 @@
 
 LIBYUV_SRCS +=  third_party/libyuv/include/libyuv/basic_types.h  \
                 third_party/libyuv/include/libyuv/cpu_id.h  \
+                third_party/libyuv/include/libyuv/planar_functions.h  \
+                third_party/libyuv/include/libyuv/row.h  \
                 third_party/libyuv/include/libyuv/scale.h  \
-                third_party/libyuv/source/row.h \
-                third_party/libyuv/source/scale.c  \
-                third_party/libyuv/source/cpu_id.c
+                third_party/libyuv/include/libyuv/scale_row.h  \
+                third_party/libyuv/source/cpu_id.cc \
+                third_party/libyuv/source/planar_functions.cc \
+                third_party/libyuv/source/row_any.cc \
+                third_party/libyuv/source/row_common.cc \
+                third_party/libyuv/source/row_mips.cc \
+                third_party/libyuv/source/row_neon.cc \
+                third_party/libyuv/source/row_posix.cc \
+                third_party/libyuv/source/row_win.cc \
+                third_party/libyuv/source/scale.cc  \
+                third_party/libyuv/source/scale_common.cc \
+                third_party/libyuv/source/scale_mips.cc \
+                third_party/libyuv/source/scale_neon.cc \
+                third_party/libyuv/source/scale_posix.cc \
+                third_party/libyuv/source/scale_win.cc
 
 LIBWEBM_MUXER_SRCS += third_party/libwebm/mkvmuxer.cpp \
                       third_party/libwebm/mkvmuxerutil.cpp \
@@ -42,7 +56,9 @@
 vpxdec.SRCS                 += ivfdec.c ivfdec.h
 vpxdec.SRCS                 += tools_common.c tools_common.h
 vpxdec.SRCS                 += y4menc.c y4menc.h
-vpxdec.SRCS                 += $(LIBYUV_SRCS)
+ifeq ($(CONFIG_LIBYUV),yes)
+  vpxdec.SRCS                 += $(LIBYUV_SRCS)
+endif
 ifeq ($(CONFIG_WEBM_IO),yes)
   vpxdec.SRCS                 += $(LIBWEBM_PARSER_SRCS)
   vpxdec.SRCS                 += webmdec.cc webmdec.h
@@ -60,34 +76,38 @@
 vpxenc.SRCS                 += vpx_ports/mem_ops_aligned.h
 vpxenc.SRCS                 += vpx_ports/vpx_timer.h
 vpxenc.SRCS                 += vpxstats.c vpxstats.h
-vpxenc.SRCS                 += $(LIBYUV_SRCS)
+ifeq ($(CONFIG_LIBYUV),yes)
+  vpxenc.SRCS                 += $(LIBYUV_SRCS)
+endif
 ifeq ($(CONFIG_WEBM_IO),yes)
   vpxenc.SRCS                 += $(LIBWEBM_MUXER_SRCS)
   vpxenc.SRCS                 += webmenc.cc webmenc.h
 endif
 vpxenc.GUID                  = 548DEC74-7A15-4B2B-AFC3-AA102E7C25C1
 vpxenc.DESCRIPTION           = Full featured encoder
-EXAMPLES-$(CONFIG_VP9_ENCODER)    += vp9_spatial_scalable_encoder.c
-vp9_spatial_scalable_encoder.SRCS += args.c args.h
-vp9_spatial_scalable_encoder.SRCS += ivfenc.c ivfenc.h
-vp9_spatial_scalable_encoder.SRCS += tools_common.c tools_common.h
-vp9_spatial_scalable_encoder.SRCS += video_common.h
-vp9_spatial_scalable_encoder.SRCS += video_writer.h video_writer.c
-vp9_spatial_scalable_encoder.SRCS += vpxstats.c vpxstats.h
-vp9_spatial_scalable_encoder.GUID   = 4A38598D-627D-4505-9C7B-D4020C84100D
-vp9_spatial_scalable_encoder.DESCRIPTION = Spatial Scalable Encoder
+ifeq ($(CONFIG_SPATIAL_SVC),yes)
+  EXAMPLES-$(CONFIG_VP9_ENCODER)      += vp9_spatial_svc_encoder.c
+  vp9_spatial_svc_encoder.SRCS        += args.c args.h
+  vp9_spatial_svc_encoder.SRCS        += ivfenc.c ivfenc.h
+  vp9_spatial_svc_encoder.SRCS        += tools_common.c tools_common.h
+  vp9_spatial_svc_encoder.SRCS        += video_common.h
+  vp9_spatial_svc_encoder.SRCS        += video_writer.h video_writer.c
+  vp9_spatial_svc_encoder.SRCS        += vpxstats.c vpxstats.h
+  vp9_spatial_svc_encoder.GUID        = 4A38598D-627D-4505-9C7B-D4020C84100D
+  vp9_spatial_svc_encoder.DESCRIPTION = VP9 Spatial SVC Encoder
+endif
 
 ifneq ($(CONFIG_SHARED),yes)
 EXAMPLES-$(CONFIG_VP9_ENCODER)    += resize_util.c
 endif
 
-EXAMPLES-$(CONFIG_ENCODERS)         += vpx_temporal_scalable_patterns.c
-vpx_temporal_scalable_patterns.SRCS += ivfenc.c ivfenc.h
-vpx_temporal_scalable_patterns.SRCS += tools_common.c tools_common.h
-vpx_temporal_scalable_patterns.SRCS += video_common.h
-vpx_temporal_scalable_patterns.SRCS += video_writer.h video_writer.c
-vpx_temporal_scalable_patterns.GUID  = B18C08F2-A439-4502-A78E-849BE3D60947
-vpx_temporal_scalable_patterns.DESCRIPTION = Temporal Scalability Encoder
+EXAMPLES-$(CONFIG_ENCODERS)          += vpx_temporal_svc_encoder.c
+vpx_temporal_svc_encoder.SRCS        += ivfenc.c ivfenc.h
+vpx_temporal_svc_encoder.SRCS        += tools_common.c tools_common.h
+vpx_temporal_svc_encoder.SRCS        += video_common.h
+vpx_temporal_svc_encoder.SRCS        += video_writer.h video_writer.c
+vpx_temporal_svc_encoder.GUID        = B18C08F2-A439-4502-A78E-849BE3D60947
+vpx_temporal_svc_encoder.DESCRIPTION = Temporal SVC Encoder
 EXAMPLES-$(CONFIG_VP8_DECODER)     += simple_decoder.c
 simple_decoder.GUID                 = D3BBF1E9-2427-450D-BBFF-B2843C1D44CC
 simple_decoder.SRCS                += ivfdec.h ivfdec.c
@@ -158,11 +178,13 @@
 
 
 ifeq ($(CONFIG_MULTI_RES_ENCODING),yes)
+ifeq ($(CONFIG_LIBYUV),yes)
 EXAMPLES-$(CONFIG_VP8_DECODER)          += vp8_multi_resolution_encoder.c
 vp8_multi_resolution_encoder.SRCS       += $(LIBYUV_SRCS)
 vp8_multi_resolution_encoder.GUID        = 04f8738e-63c8-423b-90fa-7c2703a374de
 vp8_multi_resolution_encoder.DESCRIPTION = VP8 Multiple-resolution Encoding
 endif
+endif
 
 # Handle extra library flags depending on codec configuration
 
diff --git a/examples/set_maps.c b/examples/set_maps.c
index 4343832..4ba38ee 100644
--- a/examples/set_maps.c
+++ b/examples/set_maps.c
@@ -64,7 +64,8 @@
 static void set_roi_map(const vpx_codec_enc_cfg_t *cfg,
                         vpx_codec_ctx_t *codec) {
   unsigned int i;
-  vpx_roi_map_t roi = {0};
+  vpx_roi_map_t roi;
+  memset(&roi, 0, sizeof(roi));
 
   roi.rows = (cfg->g_h + 15) / 16;
   roi.cols = (cfg->g_w + 15) / 16;
@@ -97,7 +98,7 @@
 static void set_active_map(const vpx_codec_enc_cfg_t *cfg,
                            vpx_codec_ctx_t *codec) {
   unsigned int i;
-  vpx_active_map_t map = {0};
+  vpx_active_map_t map = {0, 0, 0};
 
   map.rows = (cfg->g_h + 15) / 16;
   map.cols = (cfg->g_w + 15) / 16;
@@ -114,7 +115,7 @@
 
 static void unset_active_map(const vpx_codec_enc_cfg_t *cfg,
                              vpx_codec_ctx_t *codec) {
-  vpx_active_map_t map = {0};
+  vpx_active_map_t map = {0, 0, 0};
 
   map.rows = (cfg->g_h + 15) / 16;
   map.cols = (cfg->g_w + 15) / 16;
@@ -153,22 +154,23 @@
 
 int main(int argc, char **argv) {
   FILE *infile = NULL;
-  vpx_codec_ctx_t codec = {0};
-  vpx_codec_enc_cfg_t cfg = {0};
+  vpx_codec_ctx_t codec;
+  vpx_codec_enc_cfg_t cfg;
   int frame_count = 0;
-  vpx_image_t raw = {0};
+  vpx_image_t raw;
   vpx_codec_err_t res;
-  VpxVideoInfo info = {0};
+  VpxVideoInfo info;
   VpxVideoWriter *writer = NULL;
   const VpxInterface *encoder = NULL;
   const int fps = 2;        // TODO(dkovalev) add command line argument
   const double bits_per_pixel_per_frame = 0.067;
 
   exec_name = argv[0];
-
   if (argc != 6)
     die("Invalid number of arguments");
 
+  memset(&info, 0, sizeof(info));
+
   encoder = get_vpx_encoder_by_name(argv[1]);
   if (!encoder)
     die("Unsupported codec.");
diff --git a/examples/vp9_spatial_scalable_encoder.c b/examples/vp9_spatial_svc_encoder.c
similarity index 100%
rename from examples/vp9_spatial_scalable_encoder.c
rename to examples/vp9_spatial_svc_encoder.c
diff --git a/examples/vpx_temporal_scalable_patterns.c b/examples/vpx_temporal_svc_encoder.c
similarity index 99%
rename from examples/vpx_temporal_scalable_patterns.c
rename to examples/vpx_temporal_svc_encoder.c
index 07dd318..e45b50c 100644
--- a/examples/vpx_temporal_scalable_patterns.c
+++ b/examples/vpx_temporal_svc_encoder.c
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-//  This is an example demonstrating how to implement a multi-layer VP9
+//  This is an example demonstrating how to implement a multi-layer VPx
 //  encoding scheme based on temporal scalability for video applications
 //  that benefit from a scalable bitstream.
 
diff --git a/libs.mk b/libs.mk
index 85c5f8a..d02e9bc 100644
--- a/libs.mk
+++ b/libs.mk
@@ -115,7 +115,7 @@
   CODEC_EXPORTS-yes += $(addprefix $(VP9_PREFIX),$(VP9_CX_EXPORTS))
   CODEC_SRCS-yes += $(VP9_PREFIX)vp9cx.mk vpx/vp8.h vpx/vp8cx.h
   INSTALL-LIBS-yes += include/vpx/vp8.h include/vpx/vp8cx.h
-  INSTALL-LIBS-yes += include/vpx/svc_context.h
+  INSTALL-LIBS-$(CONFIG_SPATIAL_SVC) += include/vpx/svc_context.h
   INSTALL_MAPS += include/vpx/% $(SRC_PATH_BARE)/$(VP9_PREFIX)/%
   CODEC_DOC_SRCS += vpx/vp8.h vpx/vp8cx.h
   CODEC_DOC_SECTIONS += vp9 vp9_encoder
@@ -556,3 +556,26 @@
 utiltest:
 	@echo Unit tests must be enabled to make the utiltest target.
 endif
+
+##
+## Example tests.
+##
+ifeq ($(CONFIG_UNIT_TESTS),yes)
+# All non-MSVC targets output example targets in a sub dir named examples.
+EXAMPLES_BIN_PATH = examples
+ifeq ($(CONFIG_MSVS),yes)
+# MSVC will build both Debug and Release configurations of the examples in a
+# sub directory named for the current target. Assume the user wants to
+# run the Release tools, and assign EXAMPLES_BIN_PATH accordingly.
+# TODO(tomfinegan): Is this adequate for ARM?
+# TODO(tomfinegan): Support running the debug versions of tools?
+EXAMPLES_BIN_PATH := $(TGT_OS:win64=x64)/Release
+endif
+exampletest: examples testdata
+	$(qexec)$(SRC_PATH_BARE)/test/examples.sh \
+		--test-data-path $(LIBVPX_TEST_DATA_PATH) \
+		--bin-path $(EXAMPLES_BIN_PATH)
+else
+exampletest:
+	@echo Unit tests must be enabled to make the exampletest target.
+endif
diff --git a/test/convolve_test.cc b/test/convolve_test.cc
index cbb4036..a6dcc98 100644
--- a/test/convolve_test.cc
+++ b/test/convolve_test.cc
@@ -341,6 +341,9 @@
   for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
     const InterpKernel *filters =
         vp9_get_interp_kernel(static_cast<INTERP_FILTER>(filter_bank));
+    const InterpKernel *const eighttap_smooth =
+        vp9_get_interp_kernel(EIGHTTAP_SMOOTH);
+
     for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
       for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
         filter_block2d_8_c(in, kInputStride,
@@ -348,7 +351,7 @@
                            ref, kOutputStride,
                            Width(), Height());
 
-        if (filters == vp9_sub_pel_filters_8lp || (filter_x && filter_y))
+        if (filters == eighttap_smooth || (filter_x && filter_y))
           REGISTER_STATE_CHECK(
               UUT_->hv8_(in, kInputStride, out, kOutputStride,
                          filters[filter_x], 16, filters[filter_y], 16,
@@ -396,6 +399,8 @@
   for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
     const InterpKernel *filters =
         vp9_get_interp_kernel(static_cast<INTERP_FILTER>(filter_bank));
+    const InterpKernel *const eighttap_smooth =
+        vp9_get_interp_kernel(EIGHTTAP_SMOOTH);
 
     for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
       for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
@@ -404,7 +409,7 @@
                                    ref, kOutputStride,
                                    Width(), Height());
 
-        if (filters == vp9_sub_pel_filters_8lp || (filter_x && filter_y))
+        if (filters == eighttap_smooth || (filter_x && filter_y))
           REGISTER_STATE_CHECK(
               UUT_->hv8_avg_(in, kInputStride, out, kOutputStride,
                              filters[filter_x], 16, filters[filter_y], 16,
@@ -544,6 +549,7 @@
 TEST_P(ConvolveTest, CheckScalingFiltering) {
   uint8_t* const in = input();
   uint8_t* const out = output();
+  const InterpKernel *const eighttap = vp9_get_interp_kernel(EIGHTTAP);
 
   SetConstantInput(127);
 
@@ -551,8 +557,8 @@
     for (int step = 1; step <= 32; ++step) {
       /* Test the horizontal and vertical filters in combination. */
       REGISTER_STATE_CHECK(UUT_->hv8_(in, kInputStride, out, kOutputStride,
-                                      vp9_sub_pel_filters_8[frac], step,
-                                      vp9_sub_pel_filters_8[frac], step,
+                                      eighttap[frac], step,
+                                      eighttap[frac], step,
                                       Width(), Height()));
 
       CheckGuardBlocks();
diff --git a/test/dct16x16_test.cc b/test/dct16x16_test.cc
index 143a267..7900bcf 100644
--- a/test/dct16x16_test.cc
+++ b/test/dct16x16_test.cc
@@ -512,7 +512,9 @@
         make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 2),
         make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 3)));
 
-#if HAVE_NEON_ASM
+// FIXME (jingning, fgalligan): need to simplify the corresponding steps
+// in neov version accordingly, and re-enable the unit test
+#if HAVE_NEON_ASM && 0
 INSTANTIATE_TEST_CASE_P(
     NEON, Trans16x16DCT,
     ::testing::Values(
diff --git a/test/examples.sh b/test/examples.sh
index ac2a18c..7ba9cce 100755
--- a/test/examples.sh
+++ b/test/examples.sh
@@ -24,5 +24,6 @@
 
 for test in ${example_tests}; do
   # Source each test script so that exporting variables can be avoided.
+  VPX_TEST_NAME="$(basename ${test%.*})"
   . "${test}"
 done
diff --git a/test/resize_util.sh b/test/resize_util.sh
new file mode 100755
index 0000000..2a8e3fb
--- /dev/null
+++ b/test/resize_util.sh
@@ -0,0 +1,66 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx resize_util example code. To add new tests to
+##  this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to resize_util_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+resize_util_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Resizes $YUV_RAW_INPUT using the resize_util example. $1 is the output
+# dimensions that will be passed to resize_util.
+resize_util() {
+  local resizer="${LIBVPX_BIN_PATH}/resize_util${VPX_TEST_EXE_SUFFIX}"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/resize_util.raw"
+  local frames_to_resize="10"
+  local target_dimensions="$1"
+
+  # resize_util is available only when CONFIG_SHARED is disabled.
+  if [ -z "$(vpx_config_option_enabled CONFIG_SHARED)" ]; then
+    [ -x "${resizer}" ] || return 1
+
+    eval "${resizer}" "${YUV_RAW_INPUT}" \
+        "${YUV_RAW_INPUT_WIDTH}x${YUV_RAW_INPUT_HEIGHT}" \
+        "${target_dimensions}" "${output_file}" ${frames_to_resize} \
+        ${devnull}
+
+    [ -e "${output_file}" ] || return 1
+  fi
+}
+
+# Halves each dimension of $YUV_RAW_INPUT using resize_util().
+resize_down() {
+  local target_width=$((${YUV_RAW_INPUT_WIDTH} / 2))
+  local target_height=$((${YUV_RAW_INPUT_HEIGHT} / 2))
+
+  resize_util "${target_width}x${target_height}"
+}
+
+# Doubles each dimension of $YUV_RAW_INPUT using resize_util().
+resize_up() {
+  local target_width=$((${YUV_RAW_INPUT_WIDTH} * 2))
+  local target_height=$((${YUV_RAW_INPUT_HEIGHT} * 2))
+
+  resize_util "${target_width}x${target_height}"
+}
+
+resize_util_tests="resize_down
+                   resize_up"
+
+run_tests resize_util_verify_environment "${resize_util_tests}"
diff --git a/test/test.mk b/test/test.mk
index 44d2f9c..f0a27c7 100644
--- a/test/test.mk
+++ b/test/test.mk
@@ -113,10 +113,13 @@
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += dct32x32_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += fdct4x4_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += fdct8x8_test.cc
-LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += svc_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += variance_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_subtract_test.cc
 
+ifeq ($(CONFIG_VP9_ENCODER),yes)
+LIBVPX_TEST_SRCS-$(CONFIG_SPATIAL_SVC) += svc_test.cc
+endif
+
 endif # VP9
 
 LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += sad_test.cc
diff --git a/test/tools_common.sh b/test/tools_common.sh
index 9c10d48..472111c 100755
--- a/test/tools_common.sh
+++ b/test/tools_common.sh
@@ -302,8 +302,13 @@
 # functions and are run unconditionally. Functions in positional parameter two
 # are run according to the rules specified in vpx_test_usage().
 run_tests() {
-  env_tests="verify_vpx_test_environment ${1}"
-  tests_to_filter="${2}"
+  local env_tests="verify_vpx_test_environment $1"
+  local tests_to_filter="$2"
+  local test_name="${VPX_TEST_NAME}"
+
+  if [ -z "${test_name}" ]; then
+    test_name="$(basename \"${0%.*}\")"
+  fi
 
   if [ "${VPX_TEST_RUN_DISABLED_TESTS}" != "yes" ]; then
     # Filter out DISABLED tests.
@@ -315,7 +320,7 @@
     tests_to_filter=$(filter_strings "${tests_to_filter}" ${VPX_TEST_FILTER})
   fi
 
-  tests_to_run="${env_tests} ${tests_to_filter}"
+  local tests_to_run="${env_tests} ${tests_to_filter}"
 
   check_git_hashes
 
@@ -328,8 +333,8 @@
     test_end "${test}"
   done
 
-  tested_config="$(test_configuration_target) @ $(current_hash)"
-  echo $(basename "${0%.*}"): Done, all tests pass for ${tested_config}.
+  local tested_config="$(test_configuration_target) @ $(current_hash)"
+  echo "${test_name}: Done, all tests pass for ${tested_config}."
 }
 
 vpx_test_usage() {
diff --git a/test/twopass_encoder.sh b/test/twopass_encoder.sh
new file mode 100755
index 0000000..fe3cbbb
--- /dev/null
+++ b/test/twopass_encoder.sh
@@ -0,0 +1,59 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx twopass_encoder example. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to twopass_encoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+twopass_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs twopass_encoder using the codec specified by $1.
+twopass_encoder() {
+  local encoder="${LIBVPX_BIN_PATH}/twopass_encoder${VPX_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/twopass_encoder_${codec}.ivf"
+
+  [ -x "${encoder}" ] || return 1
+
+  eval "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+twopass_encoder_vp8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    twopass_encoder vp8 || return 1
+  fi
+}
+
+# TODO(tomfinegan): Add a frame limit param to twopass_encoder and enable this
+# test. VP9 is just too slow right now: This test takes 31m16s+ on a fast
+# machine.
+DISABLED_twopass_encoder_vp9() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    twopass_encoder vp9 || return 1
+  fi
+}
+
+twopass_encoder_tests="twopass_encoder_vp8
+                       DISABLED_twopass_encoder_vp9"
+
+run_tests twopass_encoder_verify_environment "${twopass_encoder_tests}"
diff --git a/test/vp9_spatial_svc_encoder.sh b/test/vp9_spatial_svc_encoder.sh
new file mode 100755
index 0000000..635cfa2
--- /dev/null
+++ b/test/vp9_spatial_svc_encoder.sh
@@ -0,0 +1,89 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx vp9_spatial_svc_encoder example. To add new
+##  tests to to this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vp9_spatial_svc_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+vp9_spatial_svc_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs vp9_spatial_svc_encoder. $1 is the test name.
+vp9_spatial_svc_encoder() {
+  local encoder="${LIBVPX_BIN_PATH}/vp9_spatial_svc_encoder"
+  encoder="${encoder}${VPX_TEST_EXE_SUFFIX}"
+  local test_name="$1"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/vp9_ssvc_encoder${test_name}.ivf"
+  local frames_to_encode="10"
+  local max_kf="9999"
+
+  shift
+
+  [ -x "${encoder}" ] || return 1
+
+  eval "${encoder}" -w "${YUV_RAW_INPUT_WIDTH}" -h "${YUV_RAW_INPUT_HEIGHT}" \
+      -k "${max_kf}" -f "${frames_to_encode}" "$@" "${YUV_RAW_INPUT}" \
+      "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+# Each mode is run with layer count 1-$vp9_ssvc_test_layers.
+vp9_ssvc_test_layers=5
+
+vp9_spatial_svc_mode_i() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    local test_name="${FUNCNAME}"
+    for layers in $(seq 1 ${vp9_ssvc_test_layers}); do
+      vp9_spatial_svc_encoder "${test_name}" -m i -l ${layers}
+    done
+  fi
+}
+
+vp9_spatial_svc_mode_altip() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    local test_name="${FUNCNAME}"
+    for layers in $(seq 1 ${vp9_ssvc_test_layers}); do
+      vp9_spatial_svc_encoder "${test_name}" -m "alt-ip" -l ${layers}
+    done
+  fi
+}
+
+vp9_spatial_svc_mode_ip() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    local test_name="${FUNCNAME}"
+    vp9_spatial_svc_encoder "${test_name}" -m ip -l 1
+  fi
+}
+
+vp9_spatial_svc_mode_gf() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    local test_name="${FUNCNAME}"
+    for layers in $(seq 1 ${vp9_ssvc_test_layers}); do
+      vp9_spatial_svc_encoder "${test_name}" -m gf -l ${layers}
+    done
+  fi
+}
+
+vp9_spatial_svc_tests="vp9_spatial_svc_mode_i
+                       vp9_spatial_svc_mode_altip
+                       vp9_spatial_svc_mode_ip
+                       vp9_spatial_svc_mode_gf"
+
+run_tests vp9_spatial_svc_encoder_verify_environment "${vp9_spatial_svc_tests}"
diff --git a/test/vpx_temporal_svc_encoder.sh b/test/vpx_temporal_svc_encoder.sh
new file mode 100755
index 0000000..ff64740
--- /dev/null
+++ b/test/vpx_temporal_svc_encoder.sh
@@ -0,0 +1,283 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx vpx_temporal_svc_encoder example. To add new
+##  tests to this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vpx_tsvc_encoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+vpx_tsvc_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs vpx_temporal_svc_encoder using the codec specified by $1 and output file
+# name by $2. Additional positional parameters are passed directly to
+# vpx_temporal_svc_encoder.
+vpx_tsvc_encoder() {
+  local encoder="${LIBVPX_BIN_PATH}/vpx_temporal_svc_encoder"
+  encoder="${encoder}${VPX_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file_base="$2"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/${output_file_base}"
+  local timebase_num="1"
+  local timebase_den="1000"
+  local speed="6"
+  local frame_drop_thresh="30"
+
+  shift 2
+
+  [ -x "${encoder}" ] || return 1
+
+  eval "${encoder}" "${YUV_RAW_INPUT}" "${output_file}" "${codec}" \
+      "${YUV_RAW_INPUT_WIDTH}" "${YUV_RAW_INPUT_HEIGHT}" \
+      "${timebase_num}" "${timebase_den}" "${speed}" "${frame_drop_thresh}" \
+      "$@" \
+      ${devnull}
+}
+
+# Confirms that all expected output files exist given the output file name
+# passed to vpx_temporal_svc_encoder.
+# The file name passed to vpx_temporal_svc_encoder is joined with the stream
+# number and the extension .ivf to produce per stream output files.  Here $1 is
+# file name, and $2 is expected number of files.
+files_exist() {
+  local file_name="${VPX_TEST_OUTPUT_DIR}/$1"
+  local num_files="$(($2 - 1))"
+  for stream_num in $(seq 0 ${num_files}); do
+    [ -e "${file_name}_${stream_num}.ivf" ] || return 1
+  done
+}
+
+# Run vpx_temporal_svc_encoder in all supported modes for vp8 and vp9.
+
+vpx_tsvc_encoder_vp8_mode_0() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 0 200 || return 1
+    # Mode 0 produces 1 stream
+    files_exist "${FUNCNAME}" 1 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_1() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 1 200 400 || return 1
+    # Mode 1 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_2() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 2 200 400 || return 1
+    # Mode 2 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_3() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 3 200 400 600 || return 1
+    # Mode 3 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_4() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 4 200 400 600 || return 1
+    # Mode 4 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_5() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 5 200 400 600 || return 1
+    # Mode 5 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_6() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 6 200 400 600 || return 1
+    # Mode 6 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_7() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 7 200 400 600 800 1000 || return 1
+    # Mode 7 produces 5 streams
+    files_exist "${FUNCNAME}" 5 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 8 200 400 || return 1
+    # Mode 8 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_9() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 9 200 400 600 || return 1
+    # Mode 9 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_10() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 10 200 400 600 || return 1
+    # Mode 10 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_11() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 11 200 400 600 || return 1
+    # Mode 11 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_0() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 0 200 || return 1
+    # Mode 0 produces 1 stream
+    files_exist "${FUNCNAME}" 1 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_1() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 1 200 400 || return 1
+    # Mode 1 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_2() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 2 200 400 || return 1
+    # Mode 2 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_3() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 3 200 400 600 || return 1
+    # Mode 3 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_4() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 4 200 400 600 || return 1
+    # Mode 4 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_5() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 5 200 400 600 || return 1
+    # Mode 5 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_6() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 6 200 400 600 || return 1
+    # Mode 6 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_7() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 7 200 400 600 800 1000 || return 1
+    # Mode 7 produces 5 streams
+    files_exist "${FUNCNAME}" 5 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_8() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 8 200 400 || return 1
+    # Mode 8 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_9() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 9 200 400 600 || return 1
+    # Mode 9 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_10() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 10 200 400 600 || return 1
+    # Mode 10 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_11() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 11 200 400 600 || return 1
+    # Mode 11 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_tests="vpx_tsvc_encoder_vp8_mode_0
+                        vpx_tsvc_encoder_vp8_mode_1
+                        vpx_tsvc_encoder_vp8_mode_2
+                        vpx_tsvc_encoder_vp8_mode_3
+                        vpx_tsvc_encoder_vp8_mode_4
+                        vpx_tsvc_encoder_vp8_mode_5
+                        vpx_tsvc_encoder_vp8_mode_6
+                        vpx_tsvc_encoder_vp8_mode_7
+                        vpx_tsvc_encoder_vp8_mode_8
+                        vpx_tsvc_encoder_vp8_mode_9
+                        vpx_tsvc_encoder_vp8_mode_10
+                        vpx_tsvc_encoder_vp8_mode_11
+                        vpx_tsvc_encoder_vp9_mode_0
+                        vpx_tsvc_encoder_vp9_mode_1
+                        vpx_tsvc_encoder_vp9_mode_2
+                        vpx_tsvc_encoder_vp9_mode_3
+                        vpx_tsvc_encoder_vp9_mode_4
+                        vpx_tsvc_encoder_vp9_mode_5
+                        vpx_tsvc_encoder_vp9_mode_6
+                        vpx_tsvc_encoder_vp9_mode_7
+                        vpx_tsvc_encoder_vp9_mode_8
+                        vpx_tsvc_encoder_vp9_mode_9
+                        vpx_tsvc_encoder_vp9_mode_10
+                        vpx_tsvc_encoder_vp9_mode_11"
+
+run_tests vpx_tsvc_encoder_verify_environment "${vpx_tsvc_encoder_tests}"
diff --git a/third_party/libyuv/README.libvpx b/third_party/libyuv/README.libvpx
index d3495ca..577b42d 100644
--- a/third_party/libyuv/README.libvpx
+++ b/third_party/libyuv/README.libvpx
@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 102
+Version: 1005
 License: BSD
 License File: LICENSE
 
@@ -13,5 +13,5 @@
 in order to encode multiple resolution bit streams.
 
 Local Modifications:
-Modified the original scaler code from C++ to C to fit in our current build
-system. This is a temporal solution, and will be improved later.
\ No newline at end of file
+Modified the original scaler code minimally with include file changes to fit
+in our current build system.
diff --git a/third_party/libyuv/include/libyuv/basic_types.h b/third_party/libyuv/include/libyuv/basic_types.h
index 30504ce..beb750b 100644
--- a/third_party/libyuv/include/libyuv/basic_types.h
+++ b/third_party/libyuv/include/libyuv/basic_types.h
@@ -1,22 +1,25 @@
 /*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
+ *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_BASIC_TYPES_H_
+#ifndef INCLUDE_LIBYUV_BASIC_TYPES_H_  // NOLINT
 #define INCLUDE_LIBYUV_BASIC_TYPES_H_
 
 #include <stddef.h>  // for NULL, size_t
 
-#if !(defined(_MSC_VER) && (_MSC_VER < 1600))
+#if defined(__ANDROID__) || (defined(_MSC_VER) && (_MSC_VER < 1600))
+#include <sys/types.h>  // for uintptr_t on x86
+#else
 #include <stdint.h>  // for uintptr_t
 #endif
 
+#ifndef GG_LONGLONG
 #ifndef INT_TYPES_DEFINED
 #define INT_TYPES_DEFINED
 #ifdef COMPILER_MSVC
@@ -30,9 +33,9 @@
 #endif
 #define INT64_F "I64"
 #else  // COMPILER_MSVC
-#ifdef __LP64__
-typedef unsigned long uint64;
-typedef long int64;
+#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long uint64;  // NOLINT
+typedef long int64;  // NOLINT
 #ifndef INT64_C
 #define INT64_C(x) x ## L
 #endif
@@ -40,9 +43,9 @@
 #define UINT64_C(x) x ## UL
 #endif
 #define INT64_F "l"
-#else  // __LP64__
-typedef unsigned long long uint64;
-typedef long long int64;
+#else  // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long long uint64;  // NOLINT
+typedef long long int64;  // NOLINT
 #ifndef INT64_C
 #define INT64_C(x) x ## LL
 #endif
@@ -54,20 +57,62 @@
 #endif  // COMPILER_MSVC
 typedef unsigned int uint32;
 typedef int int32;
-typedef unsigned short uint16;
-typedef short int16;
+typedef unsigned short uint16;  // NOLINT
+typedef short int16;  // NOLINT
 typedef unsigned char uint8;
-typedef char int8;
+typedef signed char int8;
 #endif  // INT_TYPES_DEFINED
+#endif  // GG_LONGLONG
 
 // Detect compiler is for x86 or x64.
 #if defined(__x86_64__) || defined(_M_X64) || \
     defined(__i386__) || defined(_M_IX86)
 #define CPU_X86 1
 #endif
+// Detect compiler is for ARM.
+#if defined(__arm__) || defined(_M_ARM)
+#define CPU_ARM 1
+#endif
 
+#ifndef ALIGNP
+#ifdef __cplusplus
 #define ALIGNP(p, t) \
-  ((uint8*)((((uintptr_t)(p) + \
-  ((t)-1)) & ~((t)-1))))
+    (reinterpret_cast<uint8*>(((reinterpret_cast<uintptr_t>(p) + \
+    ((t) - 1)) & ~((t) - 1))))
+#else
+#define ALIGNP(p, t) \
+    ((uint8*)((((uintptr_t)(p) + ((t) - 1)) & ~((t) - 1))))  /* NOLINT */
+#endif
+#endif
 
-#endif // INCLUDE_LIBYUV_BASIC_TYPES_H_
+#if !defined(LIBYUV_API)
+#if defined(_WIN32) || defined(__CYGWIN__)
+#if defined(LIBYUV_BUILDING_SHARED_LIBRARY)
+#define LIBYUV_API __declspec(dllexport)
+#elif defined(LIBYUV_USING_SHARED_LIBRARY)
+#define LIBYUV_API __declspec(dllimport)
+#else
+#define LIBYUV_API
+#endif  // LIBYUV_BUILDING_SHARED_LIBRARY
+#elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__APPLE__) && \
+    (defined(LIBYUV_BUILDING_SHARED_LIBRARY) || \
+    defined(LIBYUV_USING_SHARED_LIBRARY))
+#define LIBYUV_API __attribute__ ((visibility ("default")))
+#else
+#define LIBYUV_API
+#endif  // __GNUC__
+#endif  // LIBYUV_API
+
+#define LIBYUV_BOOL int
+#define LIBYUV_FALSE 0
+#define LIBYUV_TRUE 1
+
+// Visual C x86 or GCC little endian.
+#if defined(__x86_64__) || defined(_M_X64) || \
+  defined(__i386__) || defined(_M_IX86) || \
+  defined(__arm__) || defined(_M_ARM) || \
+  (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#define LIBYUV_LITTLE_ENDIAN
+#endif
+
+#endif  // INCLUDE_LIBYUV_BASIC_TYPES_H_  NOLINT
diff --git a/third_party/libyuv/include/libyuv/cpu_id.h b/third_party/libyuv/include/libyuv/cpu_id.h
index 4a53b5b..fd6276b 100644
--- a/third_party/libyuv/include/libyuv/cpu_id.h
+++ b/third_party/libyuv/include/libyuv/cpu_id.h
@@ -1,49 +1,81 @@
 /*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
+ *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_CPU_ID_H_
+#ifndef INCLUDE_LIBYUV_CPU_ID_H_  // NOLINT
 #define INCLUDE_LIBYUV_CPU_ID_H_
 
+#include "basic_types.h"
+
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
-// These flags are only valid on x86 processors
-static const int kCpuHasSSE2 = 1;
-static const int kCpuHasSSSE3 = 2;
+// TODO(fbarchard): Consider overlapping bits for different architectures.
+// Internal flag to indicate cpuid requires initialization.
+#define kCpuInit 0x1
 
-// These flags are only valid on ARM processors
-static const int kCpuHasNEON = 4;
+// These flags are only valid on ARM processors.
+static const int kCpuHasARM = 0x2;
+static const int kCpuHasNEON = 0x4;
+// 0x8 reserved for future ARM flag.
 
-// Internal flag to indicate cpuid is initialized.
-static const int kCpuInitialized = 8;
+// These flags are only valid on x86 processors.
+static const int kCpuHasX86 = 0x10;
+static const int kCpuHasSSE2 = 0x20;
+static const int kCpuHasSSSE3 = 0x40;
+static const int kCpuHasSSE41 = 0x80;
+static const int kCpuHasSSE42 = 0x100;
+static const int kCpuHasAVX = 0x200;
+static const int kCpuHasAVX2 = 0x400;
+static const int kCpuHasERMS = 0x800;
+static const int kCpuHasFMA3 = 0x1000;
+// 0x2000, 0x4000, 0x8000 reserved for future X86 flags.
+
+// These flags are only valid on MIPS processors.
+static const int kCpuHasMIPS = 0x10000;
+static const int kCpuHasMIPS_DSP = 0x20000;
+static const int kCpuHasMIPS_DSPR2 = 0x40000;
+
+// Internal function used to auto-init.
+LIBYUV_API
+int InitCpuFlags(void);
+
+// Internal function for parsing /proc/cpuinfo.
+LIBYUV_API
+int ArmCpuCaps(const char* cpuinfo_name);
 
 // Detect CPU has SSE2 etc.
-// test_flag parameter should be one of kCpuHas constants above
+// Test_flag parameter should be one of kCpuHas constants above.
 // returns non-zero if instruction set is detected
 static __inline int TestCpuFlag(int test_flag) {
-  extern int cpu_info_;
-  extern int InitCpuFlags();
-  return (cpu_info_ ? cpu_info_ : InitCpuFlags()) & test_flag;
+  LIBYUV_API extern int cpu_info_;
+  return (cpu_info_ == kCpuInit ? InitCpuFlags() : cpu_info_) & test_flag;
 }
 
 // For testing, allow CPU flags to be disabled.
 // ie MaskCpuFlags(~kCpuHasSSSE3) to disable SSSE3.
-// -1 to enable all cpu specific optimizations.
-// 0 to disable all cpu specific optimizations.
+// MaskCpuFlags(-1) to enable all cpu specific optimizations.
+// MaskCpuFlags(0) to disable all cpu specific optimizations.
+LIBYUV_API
 void MaskCpuFlags(int enable_flags);
 
+// Low level cpuid for X86. Returns zeros on other CPUs.
+// eax is the info type that you want.
+// ecx is typically the cpu number, and should normally be zero.
+LIBYUV_API
+void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info);
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_CPU_ID_H_
+#endif  // INCLUDE_LIBYUV_CPU_ID_H_  NOLINT
diff --git a/third_party/libyuv/include/libyuv/planar_functions.h b/third_party/libyuv/include/libyuv/planar_functions.h
new file mode 100644
index 0000000..43f8df3
--- /dev/null
+++ b/third_party/libyuv/include/libyuv/planar_functions.h
@@ -0,0 +1,439 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  // NOLINT
+#define INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_
+
+#include "basic_types.h"
+
+// TODO(fbarchard): Remove the following headers includes.
+// #include "convert.h"
+// #include "convert_argb.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy a plane of data.
+LIBYUV_API
+void CopyPlane(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+LIBYUV_API
+void CopyPlane_16(const uint16* src_y, int src_stride_y,
+                  uint16* dst_y, int dst_stride_y,
+                  int width, int height);
+
+// Set a plane of data to a 32 bit value.
+LIBYUV_API
+void SetPlane(uint8* dst_y, int dst_stride_y,
+              int width, int height,
+              uint32 value);
+
+// Copy I400.  Supports inverting.
+LIBYUV_API
+int I400ToI400(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+
+// Copy I422 to I422.
+#define I422ToI422 I422Copy
+LIBYUV_API
+int I422Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Copy I444 to I444.
+#define I444ToI444 I444Copy
+LIBYUV_API
+int I444Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Convert YUY2 to I422.
+LIBYUV_API
+int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert UYVY to I422.
+LIBYUV_API
+int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I420 to I400. (calls CopyPlane ignoring u/v).
+LIBYUV_API
+int I420ToI400(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Alias
+#define I420ToI420Mirror I420Mirror
+
+// I420 mirror.
+LIBYUV_API
+int I420Mirror(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Alias
+#define I400ToI400Mirror I400Mirror
+
+// I400 mirror.  A single plane is mirrored horizontally.
+// Pass negative height to achieve 180 degree rotation.
+LIBYUV_API
+int I400Mirror(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Alias
+#define ARGBToARGBMirror ARGBMirror
+
+// ARGB mirror.
+LIBYUV_API
+int ARGBMirror(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// Convert NV21 to RGB565.
+LIBYUV_API
+int NV21ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// I422ToARGB is in convert_argb.h
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height);
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height);
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+// Draw a rectangle into I420.
+LIBYUV_API
+int I420Rect(uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int x, int y, int width, int height,
+             int value_y, int value_u, int value_v);
+
+// Draw a rectangle into ARGB.
+LIBYUV_API
+int ARGBRect(uint8* dst_argb, int dst_stride_argb,
+             int x, int y, int width, int height, uint32 value);
+
+// Convert ARGB to gray scale ARGB.
+LIBYUV_API
+int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Make a rectangle of ARGB gray scale.
+LIBYUV_API
+int ARGBGray(uint8* dst_argb, int dst_stride_argb,
+             int x, int y, int width, int height);
+
+// Make a rectangle of ARGB Sepia tone.
+LIBYUV_API
+int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
+              int x, int y, int width, int height);
+
+// Apply a matrix rotation to each ARGB pixel.
+// matrix_argb is 4 signed ARGB values. -128 to 127 representing -2 to 2.
+// The first 4 coefficients apply to B, G, R, A and produce B of the output.
+// The next 4 coefficients apply to B, G, R, A and produce G of the output.
+// The next 4 coefficients apply to B, G, R, A and produce R of the output.
+// The last 4 coefficients apply to B, G, R, A and produce A of the output.
+LIBYUV_API
+int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    const int8* matrix_argb,
+                    int width, int height);
+
+// Deprecated. Use ARGBColorMatrix instead.
+// Apply a matrix rotation to each ARGB pixel.
+// matrix_argb is 3 signed ARGB values. -128 to 127 representing -1 to 1.
+// The first 4 coefficients apply to B, G, R, A and produce B of the output.
+// The next 4 coefficients apply to B, G, R, A and produce G of the output.
+// The last 4 coefficients apply to B, G, R, A and produce R of the output.
+LIBYUV_API
+int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
+                   const int8* matrix_rgb,
+                   int x, int y, int width, int height);
+
+// Apply a color table each ARGB pixel.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                   const uint8* table_argb,
+                   int x, int y, int width, int height);
+
+// Apply a color table each ARGB pixel but preserve destination alpha.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                  const uint8* table_argb,
+                  int x, int y, int width, int height);
+
+// Apply a luma/color table each ARGB pixel but preserve destination alpha.
+// Table contains 32768 values indexed by [Y][C] where 7 it 7 bit luma from
+// RGB (YJ style) and C is an 8 bit color component (R, G or B).
+LIBYUV_API
+int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_argb, int dst_stride_argb,
+                       const uint8* luma_rgb_table,
+                       int width, int height);
+
+// Apply a 3 term polynomial to ARGB values.
+// poly points to a 4x4 matrix.  The first row is constants.  The 2nd row is
+// coefficients for b, g, r and a.  The 3rd row is coefficients for b squared,
+// g squared, r squared and a squared.  The 4rd row is coefficients for b to
+// the 3, g to the 3, r to the 3 and a to the 3.  The values are summed and
+// result clamped to 0 to 255.
+// A polynomial approximation can be dirived using software such as 'R'.
+
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height);
+
+// Quantize a rectangle of ARGB. Alpha unaffected.
+// scale is a 16 bit fractional fixed point scaler between 0 and 65535.
+// interval_size should be a value between 1 and 255.
+// interval_offset should be a value between 0 and 255.
+LIBYUV_API
+int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
+                 int scale, int interval_size, int interval_offset,
+                 int x, int y, int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
+                     uint8* dst_argb, int dst_stride_argb,
+                     int width, int height);
+
+typedef void (*ARGBBlendRow)(const uint8* src_argb0, const uint8* src_argb1,
+                             uint8* dst_argb, int width);
+
+// Get function to Alpha Blend ARGB pixels and store to destination.
+LIBYUV_API
+ARGBBlendRow GetARGBBlend();
+
+// Alpha Blend ARGB images and store to destination.
+// Alpha of destination is set to 255.
+LIBYUV_API
+int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
+              const uint8* src_argb1, int src_stride_argb1,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// Multiply ARGB image by ARGB image. Shifted down by 8. Saturates to 255.
+LIBYUV_API
+int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// Add ARGB image with ARGB image. Saturates to 255.
+LIBYUV_API
+int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
+            const uint8* src_argb1, int src_stride_argb1,
+            uint8* dst_argb, int dst_stride_argb,
+            int width, int height);
+
+// Subtract ARGB image (argb1) from ARGB image (argb0). Saturates to 0.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// Convert I422 to YUY2.
+LIBYUV_API
+int I422ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+// Convert I422 to UYVY.
+LIBYUV_API
+int I422ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+// Convert unattentuated ARGB to preattenuated ARGB.
+LIBYUV_API
+int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
+// Convert preattentuated ARGB to unattenuated ARGB.
+LIBYUV_API
+int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+// Convert MJPG to ARGB.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample, size_t sample_size,
+               uint8* argb, int argb_stride,
+               int w, int h, int dw, int dh);
+
+// Internal function - do not call directly.
+// Computes table of cumulative sum for image where the value is the sum
+// of all values above and to the left of the entry. Used by ARGBBlur.
+LIBYUV_API
+int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
+                             int32* dst_cumsum, int dst_stride32_cumsum,
+                             int width, int height);
+
+// Blur ARGB image.
+// dst_cumsum table of width * (height + 1) * 16 bytes aligned to
+//   16 byte boundary.
+// dst_stride32_cumsum is number of ints in a row (width * 4).
+// radius is number of pixels around the center.  e.g. 1 = 3x3. 2=5x5.
+// Blur is optimized for radius of 5 (11x11) or less.
+LIBYUV_API
+int ARGBBlur(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int32* dst_cumsum, int dst_stride32_cumsum,
+             int width, int height, int radius);
+
+// Multiply ARGB image by ARGB value.
+LIBYUV_API
+int ARGBShade(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height, uint32 value);
+
+// Interpolate between two ARGB images using specified amount of interpolation
+// (0 to 255) and store to destination.
+// 'interpolation' is specified as 8 bit fraction where 0 means 100% src_argb0
+// and 255 means 1% src_argb0 and 99% src_argb1.
+// Internally uses ARGBScale bilinear filtering.
+// Caveat: This function will write up to 16 bytes beyond the end of dst_argb.
+LIBYUV_API
+int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
+                    const uint8* src_argb1, int src_stride_argb1,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height, int interpolation);
+
+#if defined(__pnacl__) || defined(__CLR_VER) || defined(COVERAGE_ENABLED) || \
+    defined(TARGET_IPHONE_SIMULATOR)
+#define LIBYUV_DISABLE_X86
+#endif
+
+// Row functions for copying a pixels from a source with a slope to a row
+// of destination. Useful for scaling, rotation, mirror, texture mapping.
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width);
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width);
+#define HAS_ARGBAFFINEROW_SSE2
+#endif  // LIBYUV_DISABLE_X86
+
+// Shuffle ARGB channel order.  e.g. BGRA to ARGB.
+// shuffler is 16 bytes and must be aligned.
+LIBYUV_API
+int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
+                uint8* dst_argb, int dst_stride_argb,
+                const uint8* shuffler, int width, int height);
+
+// Sobel ARGB effect with planar output.
+LIBYUV_API
+int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_y, int dst_stride_y,
+                     int width, int height);
+
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// Sobel ARGB effect w/ Sobel X, Sobel, Sobel Y in ARGB.
+LIBYUV_API
+int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  NOLINT
diff --git a/third_party/libyuv/include/libyuv/row.h b/third_party/libyuv/include/libyuv/row.h
new file mode 100644
index 0000000..daf5a45
--- /dev/null
+++ b/third_party/libyuv/include/libyuv/row.h
@@ -0,0 +1,1704 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROW_H_
+
+#include <stdlib.h>  // For malloc.
+
+#include "basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a) - 1)))
+
+#ifdef __cplusplus
+#define align_buffer_64(var, size)                                             \
+  uint8* var##_mem = reinterpret_cast<uint8*>(malloc((size) + 63));            \
+  uint8* var = reinterpret_cast<uint8*>                                        \
+      ((reinterpret_cast<intptr_t>(var##_mem) + 63) & ~63)
+#else
+#define align_buffer_64(var, size)                                             \
+  uint8* var##_mem = (uint8*)(malloc((size) + 63));               /* NOLINT */ \
+  uint8* var = (uint8*)(((intptr_t)(var##_mem) + 63) & ~63)       /* NOLINT */
+#endif
+
+#define free_aligned_buffer_64(var) \
+  free(var##_mem);  \
+  var = 0
+
+#if defined(__pnacl__) || defined(__CLR_VER) || defined(COVERAGE_ENABLED) || \
+    defined(TARGET_IPHONE_SIMULATOR)
+#define LIBYUV_DISABLE_X86
+#endif
+// True if compiling for SSSE3 as a requirement.
+#if defined(__SSSE3__) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 3))
+#define LIBYUV_SSSE3_ONLY
+#endif
+
+// Enable for NaCL pepper 33 for bundle and AVX2 support.
+//  #define NEW_BINUTILS
+
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+// Effects:
+#define HAS_ARGBADDROW_SSE2
+#define HAS_ARGBAFFINEROW_SSE2
+#define HAS_ARGBATTENUATEROW_SSSE3
+#define HAS_ARGBBLENDROW_SSSE3
+#define HAS_ARGBCOLORMATRIXROW_SSSE3
+#define HAS_ARGBCOLORTABLEROW_X86
+#define HAS_ARGBCOPYALPHAROW_SSE2
+#define HAS_ARGBCOPYYTOALPHAROW_SSE2
+#define HAS_ARGBGRAYROW_SSSE3
+#define HAS_ARGBLUMACOLORTABLEROW_SSSE3
+#define HAS_ARGBMIRRORROW_SSSE3
+#define HAS_ARGBMULTIPLYROW_SSE2
+#define HAS_ARGBPOLYNOMIALROW_SSE2
+#define HAS_ARGBQUANTIZEROW_SSE2
+#define HAS_ARGBSEPIAROW_SSSE3
+#define HAS_ARGBSHADEROW_SSE2
+#define HAS_ARGBSUBTRACTROW_SSE2
+#define HAS_ARGBTOUVROW_SSSE3
+#define HAS_ARGBUNATTENUATEROW_SSE2
+#define HAS_COMPUTECUMULATIVESUMROW_SSE2
+#define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+#define HAS_INTERPOLATEROW_SSE2
+#define HAS_INTERPOLATEROW_SSSE3
+#define HAS_RGBCOLORTABLEROW_X86
+#define HAS_SOBELROW_SSE2
+#define HAS_SOBELTOPLANEROW_SSE2
+#define HAS_SOBELXROW_SSE2
+#define HAS_SOBELXYROW_SSE2
+#define HAS_SOBELYROW_SSE2
+
+// Conversions:
+#define HAS_ABGRTOUVROW_SSSE3
+#define HAS_ABGRTOYROW_SSSE3
+#define HAS_ARGB1555TOARGBROW_SSE2
+#define HAS_ARGB4444TOARGBROW_SSE2
+#define HAS_ARGBSHUFFLEROW_SSE2
+#define HAS_ARGBSHUFFLEROW_SSSE3
+#define HAS_ARGBTOARGB1555ROW_SSE2
+#define HAS_ARGBTOARGB4444ROW_SSE2
+#define HAS_ARGBTOBAYERGGROW_SSE2
+#define HAS_ARGBTOBAYERROW_SSSE3
+#define HAS_ARGBTORAWROW_SSSE3
+#define HAS_ARGBTORGB24ROW_SSSE3
+#define HAS_ARGBTORGB565ROW_SSE2
+#define HAS_ARGBTOUV422ROW_SSSE3
+#define HAS_ARGBTOUV444ROW_SSSE3
+#define HAS_ARGBTOUVJROW_SSSE3
+#define HAS_ARGBTOYJROW_SSSE3
+#define HAS_ARGBTOYROW_SSSE3
+#define HAS_BGRATOUVROW_SSSE3
+#define HAS_BGRATOYROW_SSSE3
+#define HAS_COPYROW_ERMS
+#define HAS_COPYROW_SSE2
+#define HAS_COPYROW_X86
+#define HAS_HALFROW_SSE2
+#define HAS_I400TOARGBROW_SSE2
+#define HAS_I411TOARGBROW_SSSE3
+#define HAS_I422TOARGB1555ROW_SSSE3
+#define HAS_I422TOABGRROW_SSSE3
+#define HAS_I422TOARGB1555ROW_SSSE3
+#define HAS_I422TOARGB4444ROW_SSSE3
+#define HAS_I422TOARGBROW_SSSE3
+#define HAS_I422TOBGRAROW_SSSE3
+#define HAS_I422TORAWROW_SSSE3
+#define HAS_I422TORGB24ROW_SSSE3
+#define HAS_I422TORGB565ROW_SSSE3
+#define HAS_I422TORGBAROW_SSSE3
+#define HAS_I422TOUYVYROW_SSE2
+#define HAS_I422TOYUY2ROW_SSE2
+#define HAS_I444TOARGBROW_SSSE3
+#define HAS_MERGEUVROW_SSE2
+#define HAS_MIRRORROW_SSE2
+#define HAS_MIRRORROW_SSSE3
+#define HAS_MIRRORROW_UV_SSSE3
+#define HAS_MIRRORUVROW_SSSE3
+#define HAS_NV12TOARGBROW_SSSE3
+#define HAS_NV12TORGB565ROW_SSSE3
+#define HAS_NV21TOARGBROW_SSSE3
+#define HAS_NV21TORGB565ROW_SSSE3
+#define HAS_RAWTOARGBROW_SSSE3
+#define HAS_RAWTOYROW_SSSE3
+#define HAS_RGB24TOARGBROW_SSSE3
+#define HAS_RGB24TOYROW_SSSE3
+#define HAS_RGB565TOARGBROW_SSE2
+#define HAS_RGBATOUVROW_SSSE3
+#define HAS_RGBATOYROW_SSSE3
+#define HAS_SETROW_X86
+#define HAS_SPLITUVROW_SSE2
+#define HAS_UYVYTOARGBROW_SSSE3
+#define HAS_UYVYTOUV422ROW_SSE2
+#define HAS_UYVYTOUVROW_SSE2
+#define HAS_UYVYTOYROW_SSE2
+#define HAS_YTOARGBROW_SSE2
+#define HAS_YUY2TOARGBROW_SSSE3
+#define HAS_YUY2TOUV422ROW_SSE2
+#define HAS_YUY2TOUVROW_SSE2
+#define HAS_YUY2TOYROW_SSE2
+#endif
+
+// GCC >= 4.7.0 required for AVX2.
+#if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
+#if (__GNUC__ > 4) || (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7))
+#define GCC_HAS_AVX2 1
+#endif  // GNUC >= 4.7
+#endif  // __GNUC__
+
+// clang >= 3.4.0 required for AVX2.
+#if defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
+#if (__clang_major__ > 3) || (__clang_major__ == 3 && (__clang_minor__ >= 4))
+#define CLANG_HAS_AVX2 1
+#endif  // clang >= 3.4
+#endif  // __clang__
+
+// Visual C 2012 required for AVX2.
+#if defined(_M_IX86) && defined(_MSC_VER) && _MSC_VER >= 1700
+#define VISUALC_HAS_AVX2 1
+#endif  // VisualStudio >= 2012
+
+// The following are available on all x86 platforms, but
+// require VS2012, clang 3.4 or gcc 4.7.
+// The code supports NaCL but requires a new compiler and validator.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(VISUALC_HAS_AVX2) || \
+    defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
+// Effects:
+#define HAS_ARGBPOLYNOMIALROW_AVX2
+#define HAS_ARGBSHUFFLEROW_AVX2
+#define HAS_ARGBCOPYALPHAROW_AVX2
+#define HAS_ARGBCOPYYTOALPHAROW_AVX2
+#endif
+
+// The following are require VS2012.
+// TODO(fbarchard): Port to gcc.
+#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
+#define HAS_ARGBTOUVROW_AVX2
+#define HAS_ARGBTOYJROW_AVX2
+#define HAS_ARGBTOYROW_AVX2
+#define HAS_HALFROW_AVX2
+#define HAS_I422TOARGBROW_AVX2
+#define HAS_INTERPOLATEROW_AVX2
+#define HAS_MERGEUVROW_AVX2
+#define HAS_MIRRORROW_AVX2
+#define HAS_SPLITUVROW_AVX2
+#define HAS_UYVYTOUV422ROW_AVX2
+#define HAS_UYVYTOUVROW_AVX2
+#define HAS_UYVYTOYROW_AVX2
+#define HAS_YUY2TOUV422ROW_AVX2
+#define HAS_YUY2TOUVROW_AVX2
+#define HAS_YUY2TOYROW_AVX2
+
+// Effects:
+#define HAS_ARGBADDROW_AVX2
+#define HAS_ARGBATTENUATEROW_AVX2
+#define HAS_ARGBMIRRORROW_AVX2
+#define HAS_ARGBMULTIPLYROW_AVX2
+#define HAS_ARGBSUBTRACTROW_AVX2
+#define HAS_ARGBUNATTENUATEROW_AVX2
+#endif  // defined(VISUALC_HAS_AVX2)
+
+// The following are Yasm x86 only:
+// TODO(fbarchard): Port AVX2 to inline.
+#if !defined(LIBYUV_DISABLE_X86) && defined(HAVE_YASM)
+    (defined(_M_IX86) || defined(_M_X64) || \
+    defined(__x86_64__) || defined(__i386__))
+#define HAS_MERGEUVROW_AVX2
+#define HAS_MERGEUVROW_MMX
+#define HAS_SPLITUVROW_AVX2
+#define HAS_SPLITUVROW_MMX
+#define HAS_UYVYTOYROW_AVX2
+#define HAS_UYVYTOYROW_MMX
+#define HAS_YUY2TOYROW_AVX2
+#define HAS_YUY2TOYROW_MMX
+#endif
+
+// The following are disabled when SSSE3 is available:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) && \
+    !defined(LIBYUV_SSSE3_ONLY)
+#define HAS_ARGBBLENDROW_SSE2
+#define HAS_ARGBATTENUATEROW_SSE2
+#define HAS_MIRRORROW_SSE2
+#endif
+
+// The following are available on Neon platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_ABGRTOUVROW_NEON
+#define HAS_ABGRTOYROW_NEON
+#define HAS_ARGB1555TOARGBROW_NEON
+#define HAS_ARGB1555TOUVROW_NEON
+#define HAS_ARGB1555TOYROW_NEON
+#define HAS_ARGB4444TOARGBROW_NEON
+#define HAS_ARGB4444TOUVROW_NEON
+#define HAS_ARGB4444TOYROW_NEON
+#define HAS_ARGBTOARGB1555ROW_NEON
+#define HAS_ARGBTOARGB4444ROW_NEON
+#define HAS_ARGBTOBAYERROW_NEON
+#define HAS_ARGBTOBAYERGGROW_NEON
+#define HAS_ARGBTORAWROW_NEON
+#define HAS_ARGBTORGB24ROW_NEON
+#define HAS_ARGBTORGB565ROW_NEON
+#define HAS_ARGBTOUV411ROW_NEON
+#define HAS_ARGBTOUV422ROW_NEON
+#define HAS_ARGBTOUV444ROW_NEON
+#define HAS_ARGBTOUVROW_NEON
+#define HAS_ARGBTOUVJROW_NEON
+#define HAS_ARGBTOYROW_NEON
+#define HAS_ARGBTOYJROW_NEON
+#define HAS_BGRATOUVROW_NEON
+#define HAS_BGRATOYROW_NEON
+#define HAS_COPYROW_NEON
+#define HAS_HALFROW_NEON
+#define HAS_I400TOARGBROW_NEON
+#define HAS_I411TOARGBROW_NEON
+#define HAS_I422TOABGRROW_NEON
+#define HAS_I422TOARGB1555ROW_NEON
+#define HAS_I422TOARGB4444ROW_NEON
+#define HAS_I422TOARGBROW_NEON
+#define HAS_I422TOBGRAROW_NEON
+#define HAS_I422TORAWROW_NEON
+#define HAS_I422TORGB24ROW_NEON
+#define HAS_I422TORGB565ROW_NEON
+#define HAS_I422TORGBAROW_NEON
+#define HAS_I422TOUYVYROW_NEON
+#define HAS_I422TOYUY2ROW_NEON
+#define HAS_I444TOARGBROW_NEON
+#define HAS_MERGEUVROW_NEON
+#define HAS_MIRRORROW_NEON
+#define HAS_MIRRORUVROW_NEON
+#define HAS_NV12TOARGBROW_NEON
+#define HAS_NV12TORGB565ROW_NEON
+#define HAS_NV21TOARGBROW_NEON
+#define HAS_NV21TORGB565ROW_NEON
+#define HAS_RAWTOARGBROW_NEON
+#define HAS_RAWTOUVROW_NEON
+#define HAS_RAWTOYROW_NEON
+#define HAS_RGB24TOARGBROW_NEON
+#define HAS_RGB24TOUVROW_NEON
+#define HAS_RGB24TOYROW_NEON
+#define HAS_RGB565TOARGBROW_NEON
+#define HAS_RGB565TOUVROW_NEON
+#define HAS_RGB565TOYROW_NEON
+#define HAS_RGBATOUVROW_NEON
+#define HAS_RGBATOYROW_NEON
+#define HAS_SETROW_NEON
+#define HAS_SPLITUVROW_NEON
+#define HAS_UYVYTOARGBROW_NEON
+#define HAS_UYVYTOUV422ROW_NEON
+#define HAS_UYVYTOUVROW_NEON
+#define HAS_UYVYTOYROW_NEON
+#define HAS_YTOARGBROW_NEON
+#define HAS_YUY2TOARGBROW_NEON
+#define HAS_YUY2TOUV422ROW_NEON
+#define HAS_YUY2TOUVROW_NEON
+#define HAS_YUY2TOYROW_NEON
+
+// Effects:
+#define HAS_ARGBADDROW_NEON
+#define HAS_ARGBATTENUATEROW_NEON
+#define HAS_ARGBBLENDROW_NEON
+#define HAS_ARGBGRAYROW_NEON
+#define HAS_ARGBMIRRORROW_NEON
+#define HAS_ARGBMULTIPLYROW_NEON
+#define HAS_ARGBQUANTIZEROW_NEON
+#define HAS_ARGBSEPIAROW_NEON
+#define HAS_ARGBSHADEROW_NEON
+#define HAS_ARGBSUBTRACTROW_NEON
+#define HAS_SOBELROW_NEON
+#define HAS_SOBELTOPLANEROW_NEON
+#define HAS_SOBELXYROW_NEON
+#define HAS_SOBELXROW_NEON
+#define HAS_SOBELYROW_NEON
+#define HAS_INTERPOLATEROW_NEON
+// TODO(fbarchard): Investigate neon unittest failure.
+// #define HAS_ARGBCOLORMATRIXROW_NEON
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__)
+#define HAS_COPYROW_MIPS
+#if defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_I422TOABGRROW_MIPS_DSPR2
+#define HAS_I422TOARGBROW_MIPS_DSPR2
+#define HAS_I422TOBGRAROW_MIPS_DSPR2
+#define HAS_INTERPOLATEROWS_MIPS_DSPR2
+#define HAS_MIRRORROW_MIPS_DSPR2
+#define HAS_MIRRORUVROW_MIPS_DSPR2
+#define HAS_SPLITUVROW_MIPS_DSPR2
+#endif
+#endif
+
+#if defined(_MSC_VER) && !defined(__CLR_VER)
+#define SIMD_ALIGNED(var) __declspec(align(16)) var
+typedef __declspec(align(16)) int16 vec16[8];
+typedef __declspec(align(16)) int32 vec32[4];
+typedef __declspec(align(16)) int8 vec8[16];
+typedef __declspec(align(16)) uint16 uvec16[8];
+typedef __declspec(align(16)) uint32 uvec32[4];
+typedef __declspec(align(16)) uint8 uvec8[16];
+typedef __declspec(align(32)) int16 lvec16[16];
+typedef __declspec(align(32)) int32 lvec32[8];
+typedef __declspec(align(32)) int8 lvec8[32];
+typedef __declspec(align(32)) uint16 ulvec16[16];
+typedef __declspec(align(32)) uint32 ulvec32[8];
+typedef __declspec(align(32)) uint8 ulvec8[32];
+
+#elif defined(__GNUC__)
+// Caveat GCC 4.2 to 4.7 have a known issue using vectors with const.
+#define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
+typedef int16 __attribute__((vector_size(16))) vec16;
+typedef int32 __attribute__((vector_size(16))) vec32;
+typedef int8 __attribute__((vector_size(16))) vec8;
+typedef uint16 __attribute__((vector_size(16))) uvec16;
+typedef uint32 __attribute__((vector_size(16))) uvec32;
+typedef uint8 __attribute__((vector_size(16))) uvec8;
+#else
+#define SIMD_ALIGNED(var) var
+typedef int16 vec16[8];
+typedef int32 vec32[4];
+typedef int8 vec8[16];
+typedef uint16 uvec16[8];
+typedef uint32 uvec32[4];
+typedef uint8 uvec8[16];
+#endif
+
+#if defined(__APPLE__) || defined(__x86_64__) || defined(__llvm__)
+#define OMITFP
+#else
+#define OMITFP __attribute__((optimize("omit-frame-pointer")))
+#endif
+
+// NaCL macros for GCC x86 and x64.
+
+// TODO(nfullagar): When pepper_33 toolchain is distributed, default to
+// NEW_BINUTILS and remove all BUNDLEALIGN occurances.
+#if defined(__native_client__)
+#define LABELALIGN ".p2align 5\n"
+#else
+#define LABELALIGN ".p2align 2\n"
+#endif
+#if defined(__native_client__) && defined(__x86_64__)
+#if defined(NEW_BINUTILS)
+#define BUNDLELOCK ".bundle_lock\n"
+#define BUNDLEUNLOCK ".bundle_unlock\n"
+#define BUNDLEALIGN "\n"
+#else
+#define BUNDLELOCK "\n"
+#define BUNDLEUNLOCK "\n"
+#define BUNDLEALIGN ".p2align 5\n"
+#endif
+#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
+#define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
+#define MEMLEA(offset, base) #offset "(%q" #base ")"
+#define MEMLEA3(offset, index, scale) \
+    #offset "(,%q" #index "," #scale ")"
+#define MEMLEA4(offset, base, index, scale) \
+    #offset "(%q" #base ",%q" #index "," #scale ")"
+#define MEMMOVESTRING(s, d) "%%nacl:(%q" #s "),%%nacl:(%q" #d "), %%r15"
+#define MEMSTORESTRING(reg, d) "%%" #reg ",%%nacl:(%q" #d "), %%r15"
+#define MEMOPREG(opcode, offset, base, index, scale, reg) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%%" #reg "\n" \
+    BUNDLEUNLOCK
+#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " %%" #reg ",(%%r15,%%r14)\n" \
+    BUNDLEUNLOCK
+#define MEMOPARG(opcode, offset, base, index, scale, arg) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%" #arg "\n" \
+    BUNDLEUNLOCK
+#else
+#define BUNDLEALIGN "\n"
+#define MEMACCESS(base) "(%" #base ")"
+#define MEMACCESS2(offset, base) #offset "(%" #base ")"
+#define MEMLEA(offset, base) #offset "(%" #base ")"
+#define MEMLEA3(offset, index, scale) \
+    #offset "(,%" #index "," #scale ")"
+#define MEMLEA4(offset, base, index, scale) \
+    #offset "(%" #base ",%" #index "," #scale ")"
+#define MEMMOVESTRING(s, d)
+#define MEMSTORESTRING(reg, d)
+#define MEMOPREG(opcode, offset, base, index, scale, reg) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
+#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
+    #opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
+#define MEMOPARG(opcode, offset, base, index, scale, arg) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%" #arg "\n"
+#endif
+
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width);
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width);
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width);
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width);
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width);
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width);
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width);
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width);
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width);
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_vu,
+                        uint8* dst_argb,
+                        int width);
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width);
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_vu,
+                          uint8* dst_rgb565,
+                          int width);
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width);
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width);
+
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Unaligned_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Unaligned_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Unaligned_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Unaligned_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Unaligned_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix);
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix);
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix);
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
+void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_C(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_C(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_C(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_C(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_C(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_C(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int pix);
+void ARGBToYRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Any_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Any_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Any_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Any_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Any_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Any_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Any_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Any_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Any_NEON(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_Any_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_Any_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_Any_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
+
+void ARGBToUVRow_AVX2(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Any_AVX2(const uint8* src_argb, int src_stride_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_SSSE3(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_SSSE3(const uint8* src_argb, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                       uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb, int src_stride_argb,
+                                  uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
+                            uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_Any_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_Any_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                           uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_Any_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUV422Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUV411Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUVRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVJRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
+                           uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToUVRow_Any_NEON(const uint8* src_bgra, int src_stride_bgra,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ABGRToUVRow_Any_NEON(const uint8* src_abgr, int src_stride_abgr,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void RGBAToUVRow_Any_NEON(const uint8* src_rgba, int src_stride_rgba,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void RGB24ToUVRow_Any_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                           uint8* dst_u, uint8* dst_v, int pix);
+void RAWToUVRow_Any_NEON(const uint8* src_raw, int src_stride_raw,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void RGB565ToUVRow_Any_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                            uint8* dst_u, uint8* dst_v, int pix);
+void ARGB1555ToUVRow_Any_NEON(const uint8* src_argb1555,
+                              int src_stride_argb1555,
+                              uint8* dst_u, uint8* dst_v, int pix);
+void ARGB4444ToUVRow_Any_NEON(const uint8* src_argb4444,
+                              int src_stride_argb4444,
+                              uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVRow_C(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_C(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_C(const uint8* src_bgra, int src_stride_bgra,
+                   uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_C(const uint8* src_abgr, int src_stride_abgr,
+                   uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_C(const uint8* src_rgba, int src_stride_rgba,
+                   uint8* dst_u, uint8* dst_v, int width);
+void RGB24ToUVRow_C(const uint8* src_rgb24, int src_stride_rgb24,
+                    uint8* dst_u, uint8* dst_v, int width);
+void RAWToUVRow_C(const uint8* src_raw, int src_stride_raw,
+                  uint8* dst_u, uint8* dst_v, int width);
+void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
+                     uint8* dst_u, uint8* dst_v, int width);
+void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
+                       uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV444Row_SSSE3(const uint8* src_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb,
+                                    uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Any_SSSE3(const uint8* src_argb,
+                              uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV422Row_SSSE3(const uint8* src_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb,
+                                    uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_Any_SSSE3(const uint8* src_argb,
+                              uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV444Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV411Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width);
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width);
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
+void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width);
+void MirrorRow_C(const uint8* src, uint8* dst, int width);
+
+void MirrorUVRow_SSSE3(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                       int width);
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width);
+void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                            int width);
+void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                   int width);
+
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width);
+
+void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                           int pix);
+void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix);
+void SplitUVRow_Unaligned_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u,
+                                     uint8* dst_v, int pix);
+void SplitUVRow_Any_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix);
+
+void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                  int width);
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
+                               uint8* dst_uv, int width);
+void MergeUVRow_Any_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+void MergeUVRow_Any_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+void MergeUVRow_Any_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count);
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count);
+void CopyRow_X86(const uint8* src, uint8* dst, int count);
+void CopyRow_NEON(const uint8* src, uint8* dst, int count);
+void CopyRow_MIPS(const uint8* src, uint8* dst, int count);
+void CopyRow_C(const uint8* src, uint8* dst, int count);
+
+void CopyRow_16_C(const uint16* src, uint16* dst, int count);
+
+void ARGBCopyAlphaRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBCopyAlphaRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBCopyAlphaRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+
+void ARGBCopyYToAlphaRow_C(const uint8* src_y, uint8* dst_argb, int width);
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
+
+void SetRow_X86(uint8* dst, uint32 v32, int count);
+void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
+                     int dst_stride, int height);
+void SetRow_NEON(uint8* dst, uint32 v32, int count);
+void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
+                      int dst_stride, int height);
+void SetRow_C(uint8* dst, uint32 v32, int count);
+void ARGBSetRows_C(uint8* dst, uint32 v32, int width, int dst_stride,
+                   int height);
+
+// ARGBShufflers for BGRAToARGB etc.
+void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
+                      const uint8* shuffler, int pix);
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix);
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                    const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_C(const uint8* src_rgb, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
+void ARGB4444ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
+void RGB24ToARGBRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_Any_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_Any_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+void RGB24ToARGBRow_Any_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_Any_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_Any_NEON(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+
+void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGBARow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int pix);
+
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I411ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void NV12ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_uv,
+                     uint8* dst_argb,
+                     int width);
+void NV21ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_vu,
+                       uint8* dst_argb,
+                       int width);
+void NV12ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_uv,
+                       uint8* dst_argb,
+                       int width);
+void NV21ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_vu,
+                     uint8* dst_argb,
+                     int width);
+void YUY2ToARGBRow_C(const uint8* src_yuy2,
+                     uint8* dst_argb,
+                     int width);
+void UYVYToARGBRow_C(const uint8* src_uyvy,
+                     uint8* dst_argb,
+                     int width);
+void I422ToBGRARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_bgra,
+                     int width);
+void I422ToABGRRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_abgr,
+                     int width);
+void I422ToRGBARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_rgba,
+                     int width);
+void I422ToRGB24Row_C(const uint8* src_y,
+                      const uint8* src_u,
+                      const uint8* src_v,
+                      uint8* dst_rgb24,
+                      int width);
+void I422ToRAWRow_C(const uint8* src_y,
+                    const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* dst_raw,
+                    int width);
+void I422ToARGB4444Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width);
+void I422ToARGB1555Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width);
+void I422ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_rgb565,
+                       int width);
+void YToARGBRow_C(const uint8* src_y,
+                  uint8* dst_argb,
+                  int width);
+void I422ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I444ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I422ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I411ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void NV12ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_uv,
+                         uint8* dst_argb,
+                         int width);
+void NV21ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_vu,
+                         uint8* dst_argb,
+                         int width);
+void NV12ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_uv,
+                           uint8* dst_argb,
+                           int width);
+void NV21ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_vu,
+                           uint8* dst_argb,
+                           int width);
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
+                         uint8* dst_argb,
+                         int width);
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
+                         uint8* dst_argb,
+                         int width);
+void I422ToBGRARow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_bgra,
+                         int width);
+void I422ToABGRRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_abgr,
+                         int width);
+void I422ToRGBARow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgba,
+                         int width);
+void I422ToARGB4444Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToARGB1555Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+// RGB24/RAW are unaligned.
+void I422ToRGB24Row_SSSE3(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb24,
+                          int width);
+void I422ToRAWRow_SSSE3(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_raw,
+                        int width);
+
+void I444ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_argb,
+                                   int width);
+void I422ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_argb,
+                                   int width);
+void I411ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_argb,
+                                   int width);
+void NV12ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_uv,
+                                   uint8* dst_argb,
+                                   int width);
+void NV21ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_vu,
+                                   uint8* dst_argb,
+                                   int width);
+void YUY2ToARGBRow_Unaligned_SSSE3(const uint8* src_yuy2,
+                                   uint8* dst_argb,
+                                   int width);
+void UYVYToARGBRow_Unaligned_SSSE3(const uint8* src_uyvy,
+                                   uint8* dst_argb,
+                                   int width);
+void I422ToBGRARow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_bgra,
+                                   int width);
+void I422ToABGRRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_abgr,
+                                   int width);
+void I422ToRGBARow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_rgba,
+                                   int width);
+void I422ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I444ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I411ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void NV12ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_uv,
+                             uint8* dst_argb,
+                             int width);
+void NV21ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_vu,
+                             uint8* dst_argb,
+                             int width);
+void NV12ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_uv,
+                               uint8* dst_argb,
+                               int width);
+void NV21ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_vu,
+                               uint8* dst_argb,
+                               int width);
+void YUY2ToARGBRow_Any_SSSE3(const uint8* src_yuy2,
+                             uint8* dst_argb,
+                             int width);
+void UYVYToARGBRow_Any_SSSE3(const uint8* src_uyvy,
+                             uint8* dst_argb,
+                             int width);
+void I422ToBGRARow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_bgra,
+                             int width);
+void I422ToABGRRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_abgr,
+                             int width);
+void I422ToRGBARow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_rgba,
+                             int width);
+void I422ToARGB4444Row_Any_SSSE3(const uint8* src_y,
+                                 const uint8* src_u,
+                                 const uint8* src_v,
+                                 uint8* dst_rgba,
+                                 int width);
+void I422ToARGB1555Row_Any_SSSE3(const uint8* src_y,
+                                 const uint8* src_u,
+                                 const uint8* src_v,
+                                 uint8* dst_rgba,
+                                 int width);
+void I422ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_u,
+                               const uint8* src_v,
+                               uint8* dst_rgba,
+                               int width);
+// RGB24/RAW are unaligned.
+void I422ToRGB24Row_Any_SSSE3(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToRAWRow_Any_SSSE3(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void YToARGBRow_SSE2(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width);
+void YToARGBRow_NEON(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width);
+void YToARGBRow_Any_SSE2(const uint8* src_y,
+                         uint8* dst_argb,
+                         int width);
+void YToARGBRow_Any_NEON(const uint8* src_y,
+                         uint8* dst_argb,
+                         int width);
+
+// ARGB preattenuated alpha blend.
+void ARGBBlendRow_SSSE3(const uint8* src_argb, const uint8* src_argb1,
+                        uint8* dst_argb, int width);
+void ARGBBlendRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBBlendRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1,
+                    uint8* dst_argb, int width);
+
+// ARGB multiply images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBMultiplyRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBMultiplyRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
+// ARGB add images.
+void ARGBAddRow_C(const uint8* src_argb, const uint8* src_argb1,
+                  uint8* dst_argb, int width);
+void ARGBAddRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+void ARGBAddRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+void ARGBAddRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+
+// ARGB subtract images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBSubtractRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBSubtractRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
+void ARGBToRGB24Row_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB24Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void I444ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I411ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToBGRARow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToABGRRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGBARow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGB24Row_Any_NEON(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToRAWRow_Any_NEON(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+void I422ToARGB4444Row_Any_NEON(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_argb,
+                                int width);
+void I422ToARGB1555Row_Any_NEON(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_argb,
+                                int width);
+void I422ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void NV12ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV21ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV12ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void NV21ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void YUY2ToARGBRow_Any_NEON(const uint8* src_yuy2,
+                            uint8* dst_argb,
+                            int width);
+void UYVYToARGBRow_Any_NEON(const uint8* src_uyvy,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+
+void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
+                               uint8* dst_y, int pix);
+void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                                uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
+                                   uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_C(const uint8* src_yuy2, int stride_yuy2,
+                   uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_C(const uint8* src_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_AVX2(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_SSE2(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_NEON(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_NEON(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
+                               uint8* dst_y, int pix);
+void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                                uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
+                                   uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_NEON(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+
+void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_C(const uint8* src_uyvy, int stride_uyvy,
+                   uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_C(const uint8* src_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_AVX2(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_SSE2(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_NEON(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_NEON(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+
+void HalfRow_C(const uint8* src_uv, int src_uv_stride,
+               uint8* dst_uv, int pix);
+void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix);
+void HalfRow_AVX2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix);
+void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix);
+
+void HalfRow_16_C(const uint16* src_uv, int src_uv_stride,
+                  uint16* dst_uv, int pix);
+
+void ARGBToBayerRow_C(const uint8* src_argb, uint8* dst_bayer,
+                      uint32 selector, int pix);
+void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                          uint32 selector, int pix);
+void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix);
+void ARGBToBayerRow_Any_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                              uint32 selector, int pix);
+void ARGBToBayerRow_Any_NEON(const uint8* src_argb, uint8* dst_bayer,
+                             uint32 selector, int pix);
+void ARGBToBayerGGRow_C(const uint8* src_argb, uint8* dst_bayer,
+                        uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_Any_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                               uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_Any_NEON(const uint8* src_argb, uint8* dst_bayer,
+                               uint32 /* selector */, int pix);
+
+void I422ToYUY2Row_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_yuy2, int width);
+void I422ToUYVYRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_uyvy, int width);
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width);
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width);
+void I422ToYUY2Row_Any_SSE2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_yuy2, int width);
+void I422ToUYVYRow_Any_SSE2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_uyvy, int width);
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width);
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width);
+void I422ToYUY2Row_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_yuy2, int width);
+void I422ToUYVYRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_uyvy, int width);
+
+// Effects related row functions.
+void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+void ARGBAttenuateRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                int width);
+void ARGBAttenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+void ARGBAttenuateRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+
+// Inverse table for unattenuate, shared by C and SSE2.
+extern const uint32 fixed_invtbl8[256];
+void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+void ARGBUnattenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+
+void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
+
+void ARGBSepiaRow_C(uint8* dst_argb, int width);
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width);
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width);
+
+void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
+                          const int8* matrix_argb, int width);
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width);
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width);
+
+void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
+
+void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
+
+void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
+                       int interval_offset, int width);
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
+
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value);
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+
+// Used for blur.
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst, int count);
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width);
+
+void CumulativeSumToAverageRow_C(const int32* topleft, const int32* botleft,
+                                 int width, int area, uint8* dst, int count);
+void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
+                               const int32* previous_cumsum, int width);
+
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width);
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width);
+
+// Used for I420Scale, ARGBScale, and ARGBInterpolate.
+void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+                      ptrdiff_t src_stride_ptr,
+                      int width, int source_y_fraction);
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride_ptr, int width,
+                          int source_y_fraction);
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRows_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                ptrdiff_t src_stride_ptr, int width,
+                                int source_y_fraction);
+void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride_ptr, int width,
+                                   int source_y_fraction);
+void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride_ptr, int width,
+                                    int source_y_fraction);
+void InterpolateRow_Any_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                              ptrdiff_t src_stride_ptr, int width,
+                              int source_y_fraction);
+void InterpolateRow_Any_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRows_Any_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride_ptr, int width,
+                                    int source_y_fraction);
+
+void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         ptrdiff_t src_stride_ptr,
+                         int width, int source_y_fraction);
+
+// Sobel images.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width);
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width);
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width);
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width);
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width);
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width);
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width);
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
+void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_y, int width);
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width);
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width);
+void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                  uint8* dst_argb, int width);
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width);
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width);
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width);
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+
+void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                             const uint8* luma, uint32 lumacoeff);
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROW_H_  NOLINT
diff --git a/third_party/libyuv/include/libyuv/scale.h b/third_party/libyuv/include/libyuv/scale.h
index 35d0ff5..973d464 100644
--- a/third_party/libyuv/include/libyuv/scale.h
+++ b/third_party/libyuv/include/libyuv/scale.h
@@ -1,30 +1,45 @@
 /*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
+ *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_SCALE_H_
+#ifndef INCLUDE_LIBYUV_SCALE_H_  // NOLINT
 #define INCLUDE_LIBYUV_SCALE_H_
 
-#include "third_party/libyuv/include/libyuv/basic_types.h"
+#include "basic_types.h"
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
-// Supported filtering
-typedef enum {
-  kFilterNone = 0,  // Point sample; Fastest
-  kFilterBilinear = 1,  // Faster than box, but lower quality scaling down.
-  kFilterBox = 2  // Highest quality
+// Supported filtering.
+typedef enum FilterMode {
+  kFilterNone = 0,  // Point sample; Fastest.
+  kFilterLinear = 1,  // Filter horizontally only.
+  kFilterBilinear = 2,  // Faster than box, but lower quality scaling down.
+  kFilterBox = 3  // Highest quality.
 } FilterModeEnum;
 
+// Scale a YUV plane.
+LIBYUV_API
+void ScalePlane(const uint8* src, int src_stride,
+                int src_width, int src_height,
+                uint8* dst, int dst_stride,
+                int dst_width, int dst_height,
+                enum FilterMode filtering);
+
+void ScalePlane_16(const uint16* src, int src_stride,
+                   int src_width, int src_height,
+                   uint16* dst, int dst_stride,
+                   int dst_width, int dst_height,
+                   enum FilterMode filtering);
+
 // Scales a YUV 4:2:0 image from the src width and height to the
 // dst width and height.
 // If filtering is kFilterNone, a simple nearest-neighbor algorithm is
@@ -35,6 +50,7 @@
 // quality image, at further expense of speed.
 // Returns 0 if successful.
 
+LIBYUV_API
 int I420Scale(const uint8* src_y, int src_stride_y,
               const uint8* src_u, int src_stride_u,
               const uint8* src_v, int src_stride_v,
@@ -43,28 +59,44 @@
               uint8* dst_u, int dst_stride_u,
               uint8* dst_v, int dst_stride_v,
               int dst_width, int dst_height,
-              FilterModeEnum filtering);
+              enum FilterMode filtering);
 
-// Legacy API.  Deprecated
+LIBYUV_API
+int I420Scale_16(const uint16* src_y, int src_stride_y,
+                 const uint16* src_u, int src_stride_u,
+                 const uint16* src_v, int src_stride_v,
+                 int src_width, int src_height,
+                 uint16* dst_y, int dst_stride_y,
+                 uint16* dst_u, int dst_stride_u,
+                 uint16* dst_v, int dst_stride_v,
+                 int dst_width, int dst_height,
+                 enum FilterMode filtering);
+
+#ifdef __cplusplus
+// Legacy API.  Deprecated.
+LIBYUV_API
 int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
           int src_stride_y, int src_stride_u, int src_stride_v,
           int src_width, int src_height,
           uint8* dst_y, uint8* dst_u, uint8* dst_v,
           int dst_stride_y, int dst_stride_u, int dst_stride_v,
           int dst_width, int dst_height,
-          int interpolate);
+          LIBYUV_BOOL interpolate);
 
-// Legacy API.  Deprecated
-int ScaleOffset(const uint8* src, int src_width, int src_height,
-                uint8* dst, int dst_width, int dst_height, int dst_yoffset,
-                int interpolate);
+// Legacy API.  Deprecated.
+LIBYUV_API
+int ScaleOffset(const uint8* src_i420, int src_width, int src_height,
+                uint8* dst_i420, int dst_width, int dst_height, int dst_yoffset,
+                LIBYUV_BOOL interpolate);
 
-// For testing, allow disabling of optimizations.
-void SetUseReferenceImpl(int use);
+// For testing, allow disabling of specialized scalers.
+LIBYUV_API
+void SetUseReferenceImpl(LIBYUV_BOOL use);
+#endif  // __cplusplus
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif // INCLUDE_LIBYUV_SCALE_H_
+#endif  // INCLUDE_LIBYUV_SCALE_H_  NOLINT
diff --git a/third_party/libyuv/include/libyuv/scale_row.h b/third_party/libyuv/include/libyuv/scale_row.h
new file mode 100644
index 0000000..5d91f8f
--- /dev/null
+++ b/third_party/libyuv/include/libyuv/scale_row.h
@@ -0,0 +1,341 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_ROW_H_
+
+#include "basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if defined(__pnacl__) || defined(__CLR_VER) || defined(COVERAGE_ENABLED) || \
+    defined(TARGET_IPHONE_SIMULATOR)
+#define LIBYUV_DISABLE_X86
+#endif
+
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_SCALEROWDOWN2_SSE2
+#define HAS_SCALEROWDOWN4_SSE2
+#define HAS_SCALEROWDOWN34_SSSE3
+#define HAS_SCALEROWDOWN38_SSSE3
+#define HAS_SCALEADDROWS_SSE2
+#define HAS_SCALEFILTERCOLS_SSSE3
+#define HAS_SCALECOLSUP2_SSE2
+#define HAS_SCALEARGBROWDOWN2_SSE2
+#define HAS_SCALEARGBROWDOWNEVEN_SSE2
+#define HAS_SCALEARGBCOLS_SSE2
+#define HAS_SCALEARGBFILTERCOLS_SSSE3
+#define HAS_SCALEARGBCOLSUP2_SSE2
+#define HAS_FIXEDDIV_X86
+#define HAS_FIXEDDIV1_X86
+#endif
+
+// The following are available on Neon platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_SCALEROWDOWN2_NEON
+#define HAS_SCALEROWDOWN4_NEON
+#define HAS_SCALEROWDOWN34_NEON
+#define HAS_SCALEROWDOWN38_NEON
+#define HAS_SCALEARGBROWDOWNEVEN_NEON
+#define HAS_SCALEARGBROWDOWN2_NEON
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
+    defined(__mips__) && defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_SCALEROWDOWN2_MIPS_DSPR2
+#define HAS_SCALEROWDOWN4_MIPS_DSPR2
+#define HAS_SCALEROWDOWN34_MIPS_DSPR2
+#define HAS_SCALEROWDOWN38_MIPS_DSPR2
+#endif
+
+// Scale ARGB vertically with bilinear interpolation.
+void ScalePlaneVertical(int src_height,
+                        int dst_width, int dst_height,
+                        int src_stride, int dst_stride,
+                        const uint8* src_argb, uint8* dst_argb,
+                        int x, int y, int dy,
+                        int bpp, enum FilterMode filtering);
+
+void ScalePlaneVertical_16(int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint16* src_argb, uint16* dst_argb,
+                           int x, int y, int dy,
+                           int wpp, enum FilterMode filtering);
+
+// Simplify the filtering based on scale factors.
+enum FilterMode ScaleFilterReduce(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  enum FilterMode filtering);
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_C(int num, int div);
+int FixedDiv_X86(int num, int div);
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
+int FixedDiv1_C(int num, int div);
+int FixedDiv1_X86(int num, int div);
+#ifdef HAS_FIXEDDIV_X86
+#define FixedDiv FixedDiv_X86
+#define FixedDiv1 FixedDiv1_X86
+#else
+#define FixedDiv FixedDiv_C
+#define FixedDiv1 FixedDiv1_C
+#endif
+
+// Compute slope values for stepping.
+void ScaleSlope(int src_width, int src_height,
+                int dst_width, int dst_height,
+                enum FilterMode filtering,
+                int* x, int* y, int* dx, int* dy);
+
+void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width);
+void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width);
+void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width);
+void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                              uint16* dst, int dst_width);
+void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width);
+void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width);
+void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width);
+void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width);
+void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width);
+void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width);
+void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width);
+void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width);
+void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width);
+void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width);
+void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                 int dst_width, int x, int dx);
+void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                    int dst_width, int x, int dx);
+void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
+                    int dst_width, int, int);
+void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int, int);
+void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx);
+void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                          int dst_width, int x, int dx);
+void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
+                         int dst_width, int x, int dx);
+void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                            int dst_width, int x, int dx);
+void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width);
+void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width);
+void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width);
+void ScaleAddRows_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height);
+void ScaleAddRows_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                       uint32* dst_ptr, int src_width, int src_height);
+void ScaleARGBRowDown2_C(const uint8* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            int src_stepx,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
+                     int dst_width, int x, int dx);
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x, int dx);
+void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int, int);
+void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx);
+void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
+                             int dst_width, int x, int dx);
+
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width,
+                       int src_height);
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx);
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx);
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx);
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx);
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx);
+// Row functions.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb, int src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, int src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width);
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+
+// ScaleRowDown2Box also used by planar functions
+// NEON downscalers with interpolation.
+
+// Note - not static due to reuse in convert for 444 to 420.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width);
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+//  to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width);
+// 32x3 -> 12x1
+void ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width);
+void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width);
+void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width);
+void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width);
+void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_ROW_H_  NOLINT
diff --git a/third_party/libyuv/source/cpu_id.c b/third_party/libyuv/source/cpu_id.c
deleted file mode 100644
index fccf3dd..0000000
--- a/third_party/libyuv/source/cpu_id.c
+++ /dev/null
@@ -1,81 +0,0 @@
-/*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "third_party/libyuv/include/libyuv/cpu_id.h"
-
-#ifdef _MSC_VER
-#include <intrin.h>
-#endif
-#ifdef __ANDROID__
-#include <cpu-features.h>
-#endif
-
-#include "third_party/libyuv/include/libyuv/basic_types.h"  // for CPU_X86
-
-// TODO(fbarchard): Use cpuid.h when gcc 4.4 is used on OSX and Linux.
-#if (defined(__pic__) || defined(__APPLE__)) && defined(__i386__)
-static inline void __cpuid(int cpu_info[4], int info_type) {
-  asm volatile (
-    "mov %%ebx, %%edi                          \n"
-    "cpuid                                     \n"
-    "xchg %%edi, %%ebx                         \n"
-    : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
-    : "a"(info_type)
-  );
-}
-#elif defined(__i386__) || defined(__x86_64__)
-static inline void __cpuid(int cpu_info[4], int info_type) {
-  asm volatile (
-    "cpuid                                     \n"
-    : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
-    : "a"(info_type)
-  );
-}
-#endif
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-// CPU detect function for SIMD instruction sets.
-int cpu_info_ = 0;
-
-int InitCpuFlags() {
-#ifdef CPU_X86
-  int cpu_info[4];
-  __cpuid(cpu_info, 1);
-  cpu_info_ = (cpu_info[3] & 0x04000000 ? kCpuHasSSE2 : 0) |
-              (cpu_info[2] & 0x00000200 ? kCpuHasSSSE3 : 0) |
-              kCpuInitialized;
-#elif defined(__ANDROID__) && defined(__ARM_NEON__)
-  uint64_t features = android_getCpuFeatures();
-  cpu_info_ = ((features & ANDROID_CPU_ARM_FEATURE_NEON) ? kCpuHasNEON : 0) |
-              kCpuInitialized;
-#elif defined(__ARM_NEON__)
-  // gcc -mfpu=neon defines __ARM_NEON__
-  // Enable Neon if you want support for Neon and Arm, and use MaskCpuFlags
-  // to disable Neon on devices that do not have it.
-  cpu_info_ = kCpuHasNEON | kCpuInitialized;
-#else
-  cpu_info_ = kCpuInitialized;
-#endif
-  return cpu_info_;
-}
-
-void MaskCpuFlags(int enable_flags) {
-  InitCpuFlags();
-  cpu_info_ = (cpu_info_ & enable_flags) | kCpuInitialized;
-}
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
diff --git a/third_party/libyuv/source/cpu_id.cc b/third_party/libyuv/source/cpu_id.cc
new file mode 100644
index 0000000..520cfe5
--- /dev/null
+++ b/third_party/libyuv/source/cpu_id.cc
@@ -0,0 +1,283 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/cpu_id.h"
+
+#ifdef _MSC_VER
+#include <intrin.h>  // For __cpuidex()
+#endif
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    !defined(__native_client__) && defined(_M_X64) && \
+    defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
+#include <immintrin.h>  // For _xgetbv()
+#endif
+
+#if !defined(__native_client__)
+#include <stdlib.h>  // For getenv()
+#endif
+
+// For ArmCpuCaps() but unittested on all platforms
+#include <stdio.h>
+#include <string.h>
+
+#include "third_party/libyuv/include/libyuv/basic_types.h"  // For CPU_X86
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// For functions that use the stack and have runtime checks for overflow,
+// use SAFEBUFFERS to avoid additional check.
+#if defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
+#define SAFEBUFFERS __declspec(safebuffers)
+#else
+#define SAFEBUFFERS
+#endif
+
+// Low level cpuid for X86. Returns zeros on other CPUs.
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    (defined(_M_IX86) || defined(_M_X64) || \
+    defined(__i386__) || defined(__x86_64__))
+LIBYUV_API
+void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
+#if defined(_MSC_VER)
+#if (_MSC_FULL_VER >= 160040219)
+  __cpuidex((int*)(cpu_info), info_eax, info_ecx);
+#elif defined(_M_IX86)
+  __asm {
+    mov        eax, info_eax
+    mov        ecx, info_ecx
+    mov        edi, cpu_info
+    cpuid
+    mov        [edi], eax
+    mov        [edi + 4], ebx
+    mov        [edi + 8], ecx
+    mov        [edi + 12], edx
+  }
+#else
+  if (info_ecx == 0) {
+    __cpuid((int*)(cpu_info), info_eax);
+  } else {
+    cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0;
+  }
+#endif
+#else  // defined(_MSC_VER)
+  uint32 info_ebx, info_edx;
+  asm volatile (  // NOLINT
+#if defined( __i386__) && defined(__PIC__)
+    // Preserve ebx for fpic 32 bit.
+    "mov %%ebx, %%edi                          \n"
+    "cpuid                                     \n"
+    "xchg %%edi, %%ebx                         \n"
+    : "=D" (info_ebx),
+#else
+    "cpuid                                     \n"
+    : "=b" (info_ebx),
+#endif  //  defined( __i386__) && defined(__PIC__)
+      "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx));
+  cpu_info[0] = info_eax;
+  cpu_info[1] = info_ebx;
+  cpu_info[2] = info_ecx;
+  cpu_info[3] = info_edx;
+#endif  // defined(_MSC_VER)
+}
+
+#if !defined(__native_client__)
+#define HAS_XGETBV
+// X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
+int TestOsSaveYmm() {
+  uint32 xcr0 = 0u;
+#if defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
+  xcr0 = (uint32)(_xgetbv(0));  // VS2010 SP1 required.
+#elif defined(_M_IX86)
+  __asm {
+    xor        ecx, ecx    // xcr 0
+    _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
+    mov        xcr0, eax
+  }
+#elif defined(__i386__) || defined(__x86_64__)
+  asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
+#endif  // defined(_MSC_VER)
+  return((xcr0 & 6) == 6);  // Is ymm saved?
+}
+#endif  // !defined(__native_client__)
+#else
+LIBYUV_API
+void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
+  cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
+}
+#endif
+
+// based on libvpx arm_cpudetect.c
+// For Arm, but public to allow testing on any CPU
+LIBYUV_API SAFEBUFFERS
+int ArmCpuCaps(const char* cpuinfo_name) {
+  char cpuinfo_line[512];
+  FILE* f = fopen(cpuinfo_name, "r");
+  if (!f) {
+    // Assume Neon if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasNEON;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
+    if (memcmp(cpuinfo_line, "Features", 8) == 0) {
+      char* p = strstr(cpuinfo_line, " neon");
+      if (p && (p[5] == ' ' || p[5] == '\n')) {
+        fclose(f);
+        return kCpuHasNEON;
+      }
+    }
+  }
+  fclose(f);
+  return 0;
+}
+
+#if defined(__mips__) && defined(__linux__)
+static int MipsCpuCaps(const char* search_string) {
+  char cpuinfo_line[512];
+  const char* file_name = "/proc/cpuinfo";
+  FILE* f = fopen(file_name, "r");
+  if (!f) {
+    // Assume DSP if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasMIPS_DSP;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f) != NULL) {
+    if (strstr(cpuinfo_line, search_string) != NULL) {
+      fclose(f);
+      return kCpuHasMIPS_DSP;
+    }
+  }
+  fclose(f);
+  return 0;
+}
+#endif
+
+// CPU detect function for SIMD instruction sets.
+LIBYUV_API
+int cpu_info_ = kCpuInit;  // cpu_info is not initialized yet.
+
+// Test environment variable for disabling CPU features. Any non-zero value
+// to disable. Zero ignored to make it easy to set the variable on/off.
+#if !defined(__native_client__) && !defined(_M_ARM)
+
+static LIBYUV_BOOL TestEnv(const char* name) {
+  const char* var = getenv(name);
+  if (var) {
+    if (var[0] != '0') {
+      return LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_FALSE;
+}
+#else  // nacl does not support getenv().
+static LIBYUV_BOOL TestEnv(const char*) {
+  return LIBYUV_FALSE;
+}
+#endif
+
+LIBYUV_API SAFEBUFFERS
+int InitCpuFlags(void) {
+#if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86)
+
+  uint32 cpu_info1[4] = { 0, 0, 0, 0 };
+  uint32 cpu_info7[4] = { 0, 0, 0, 0 };
+  CpuId(1, 0, cpu_info1);
+  CpuId(7, 0, cpu_info7);
+  cpu_info_ = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
+              ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
+              ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
+              ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
+              ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) |
+              ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
+              kCpuHasX86;
+#ifdef HAS_XGETBV
+  if ((cpu_info1[2] & 0x18000000) == 0x18000000 &&  // AVX and OSSave
+      TestOsSaveYmm()) {  // Saves YMM.
+    cpu_info_ |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) |
+                 kCpuHasAVX;
+  }
+#endif
+  // Environment variable overrides for testing.
+  if (TestEnv("LIBYUV_DISABLE_X86")) {
+    cpu_info_ &= ~kCpuHasX86;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE2")) {
+    cpu_info_ &= ~kCpuHasSSE2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
+    cpu_info_ &= ~kCpuHasSSSE3;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE41")) {
+    cpu_info_ &= ~kCpuHasSSE41;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE42")) {
+    cpu_info_ &= ~kCpuHasSSE42;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX")) {
+    cpu_info_ &= ~kCpuHasAVX;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX2")) {
+    cpu_info_ &= ~kCpuHasAVX2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_ERMS")) {
+    cpu_info_ &= ~kCpuHasERMS;
+  }
+  if (TestEnv("LIBYUV_DISABLE_FMA3")) {
+    cpu_info_ &= ~kCpuHasFMA3;
+  }
+#elif defined(__mips__) && defined(__linux__)
+  // Linux mips parse text file for dsp detect.
+  cpu_info_ = MipsCpuCaps("dsp");  // set kCpuHasMIPS_DSP.
+#if defined(__mips_dspr2)
+  cpu_info_ |= kCpuHasMIPS_DSPR2;
+#endif
+  cpu_info_ |= kCpuHasMIPS;
+
+  if (getenv("LIBYUV_DISABLE_MIPS")) {
+    cpu_info_ &= ~kCpuHasMIPS;
+  }
+  if (getenv("LIBYUV_DISABLE_MIPS_DSP")) {
+    cpu_info_ &= ~kCpuHasMIPS_DSP;
+  }
+  if (getenv("LIBYUV_DISABLE_MIPS_DSPR2")) {
+    cpu_info_ &= ~kCpuHasMIPS_DSPR2;
+  }
+#elif defined(__arm__)
+// gcc -mfpu=neon defines __ARM_NEON__
+// __ARM_NEON__ generates code that requires Neon.  NaCL also requires Neon.
+// For Linux, /proc/cpuinfo can be tested but without that assume Neon.
+#if defined(__ARM_NEON__) || defined(__native_client__) || !defined(__linux__)
+  cpu_info_ = kCpuHasNEON;
+#else
+  // Linux arm parse text file for neon detect.
+  cpu_info_ = ArmCpuCaps("/proc/cpuinfo");
+#endif
+  cpu_info_ |= kCpuHasARM;
+  if (TestEnv("LIBYUV_DISABLE_NEON")) {
+    cpu_info_ &= ~kCpuHasNEON;
+  }
+#endif  // __arm__
+  if (TestEnv("LIBYUV_DISABLE_ASM")) {
+    cpu_info_ = 0;
+  }
+  return cpu_info_;
+}
+
+LIBYUV_API
+void MaskCpuFlags(int enable_flags) {
+  cpu_info_ = InitCpuFlags() & enable_flags;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/planar_functions.cc b/third_party/libyuv/source/planar_functions.cc
new file mode 100644
index 0000000..68b8f46
--- /dev/null
+++ b/third_party/libyuv/source/planar_functions.cc
@@ -0,0 +1,2287 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/planar_functions.h"
+
+#include <string.h>  // for memset()
+
+#include "third_party/libyuv/include/libyuv/cpu_id.h"
+#ifdef HAVE_JPEG
+#include "third_party/libyuv/include/libyuv/mjpeg_decoder.h"
+#endif
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy a plane of data
+LIBYUV_API
+void CopyPlane(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  int y;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+#if defined(HAS_COPYROW_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height; ++y) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+LIBYUV_API
+void CopyPlane_16(const uint16* src_y, int src_stride_y,
+                  uint16* dst_y, int dst_stride_y,
+                  int width, int height) {
+  int y;
+  void (*CopyRow)(const uint16* src, uint16* dst, int width) = CopyRow_16_C;
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+#if defined(HAS_COPYROW_16_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_16_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    CopyRow = CopyRow_16_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_16_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_16_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_16_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_16_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_16_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height; ++y) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Copy I422.
+LIBYUV_API
+int I422Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+  return 0;
+}
+
+// Copy I444.
+LIBYUV_API
+int I444Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+  return 0;
+}
+
+// Copy I400.
+LIBYUV_API
+int I400ToI400(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Convert I420 to I400.
+LIBYUV_API
+int I420ToI400(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Mirror a plane of data.
+void MirrorPlane(const uint8* src_y, int src_stride_y,
+                 uint8* dst_y, int dst_stride_y,
+                 int width, int height) {
+  int y;
+  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    MirrorRow = MirrorRow_NEON;
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16)) {
+    MirrorRow = MirrorRow_SSE2;
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    MirrorRow = MirrorRow_SSSE3;
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 32)) {
+    MirrorRow = MirrorRow_AVX2;
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    MirrorRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Convert YUY2 to I422.
+LIBYUV_API
+int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*YUY2ToUV422Row)(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) =
+      YUY2ToUV422Row_C;
+  void (*YUY2ToYRow)(const uint8* src_yuy2, uint8* dst_y, int pix) =
+      YUY2ToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_yuy2 == width * 2 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_yuy2 = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Unaligned_SSE2;
+      YUY2ToYRow = YUY2ToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16)) {
+        YUY2ToUV422Row = YUY2ToUV422Row_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          YUY2ToYRow = YUY2ToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    if (width >= 16) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUV422Row = YUY2ToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Convert UYVY to I422.
+LIBYUV_API
+int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*UYVYToUV422Row)(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) =
+      UYVYToUV422Row_C;
+  void (*UYVYToYRow)(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) = UYVYToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    UYVYToUV422Row = UYVYToUV422Row_Any_SSE2;
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToUV422Row = UYVYToUV422Row_Unaligned_SSE2;
+      UYVYToYRow = UYVYToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_uyvy, 16) && IS_ALIGNED(src_stride_uyvy, 16)) {
+        UYVYToUV422Row = UYVYToUV422Row_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          UYVYToYRow = UYVYToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    UYVYToUV422Row = UYVYToUV422Row_Any_AVX2;
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToUV422Row = UYVYToUV422Row_AVX2;
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    if (width >= 16) {
+      UYVYToUV422Row = UYVYToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+      UYVYToUV422Row = UYVYToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    UYVYToUV422Row(src_uyvy, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+    src_uyvy += src_stride_uyvy;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Mirror I400 with optional flipping
+LIBYUV_API
+int I400Mirror(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+
+  MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Mirror I420 with optional flipping
+LIBYUV_API
+int I420Mirror(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  MirrorPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
+  MirrorPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
+  return 0;
+}
+
+// ARGB mirror.
+LIBYUV_API
+int ARGBMirror(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+      ARGBMirrorRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+#if defined(HAS_ARGBMIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBMirrorRow = ARGBMirrorRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
+    ARGBMirrorRow = ARGBMirrorRow_AVX2;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 4)) {
+    ARGBMirrorRow = ARGBMirrorRow_NEON;
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    ARGBMirrorRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Get a blender that optimized for the CPU, alignment and pixel count.
+// As there are 6 blenders to choose from, the caller should try to use
+// the same blend function for all pixels if possible.
+LIBYUV_API
+ARGBBlendRow GetARGBBlend() {
+  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width) = ARGBBlendRow_C;
+#if defined(HAS_ARGBBLENDROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBBlendRow = ARGBBlendRow_SSSE3;
+    return ARGBBlendRow;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBBlendRow = ARGBBlendRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBBlendRow = ARGBBlendRow_NEON;
+  }
+#endif
+  return ARGBBlendRow;
+}
+
+// Alpha Blend 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
+              const uint8* src_argb1, int src_stride_argb1,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  int y;
+  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width) = GetARGBBlend();
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+
+  for (y = 0; y < height; ++y) {
+    ARGBBlendRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Multiply 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*ARGBMultiplyRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBMultiplyRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBMULTIPLYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_NEON;
+    }
+  }
+#endif
+
+  // Multiply plane
+  for (y = 0; y < height; ++y) {
+    ARGBMultiplyRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Add 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
+            const uint8* src_argb1, int src_stride_argb1,
+            uint8* dst_argb, int dst_stride_argb,
+            int width, int height) {
+  int y;
+  void (*ARGBAddRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                     int width) = ARGBAddRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBADDROW_SSE2) && defined(_MSC_VER)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBAddRow = ARGBAddRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBADDROW_SSE2) && !defined(_MSC_VER)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBAddRow = ARGBAddRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAddRow = ARGBAddRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBAddRow = ARGBAddRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBAddRow = ARGBAddRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_NEON;
+    }
+  }
+#endif
+
+  // Add plane
+  for (y = 0; y < height; ++y) {
+    ARGBAddRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Subtract 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*ARGBSubtractRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBSubtractRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSUBTRACTROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSubtractRow = ARGBSubtractRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSUBTRACTROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSUBTRACTROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_NEON;
+    }
+  }
+#endif
+
+  // Subtract plane
+  for (y = 0; y < height; ++y) {
+    ARGBSubtractRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height) {
+  int y;
+  void (*I422ToBGRARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToBGRARow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_bgra ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra;
+    dst_stride_bgra = -dst_stride_bgra;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_bgra == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_bgra = 0;
+  }
+#if defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToBGRARow = I422ToBGRARow_NEON;
+    }
+  }
+#elif defined(HAS_I422TOBGRAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_bgra, 16) && IS_ALIGNED(dst_stride_bgra, 16)) {
+        I422ToBGRARow = I422ToBGRARow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) {
+    I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width);
+    dst_bgra += dst_stride_bgra;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height) {
+  int y;
+  void (*I422ToABGRRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToABGRRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_abgr ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr;
+    dst_stride_abgr = -dst_stride_abgr;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_abgr == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_abgr = 0;
+  }
+#if defined(HAS_I422TOABGRROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToABGRRow = I422ToABGRRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_NEON;
+    }
+  }
+#elif defined(HAS_I422TOABGRROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_abgr, 16) && IS_ALIGNED(dst_stride_abgr, 16)) {
+        I422ToABGRRow = I422ToABGRRow_SSSE3;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width);
+    dst_abgr += dst_stride_abgr;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRGBARow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_rgba ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_rgba == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_rgba = 0;
+  }
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#elif defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_rgba, 16) && IS_ALIGNED(dst_stride_rgba, 16)) {
+        I422ToRGBARow = I422ToRGBARow_SSSE3;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*NV12ToRGB565Row)(const uint8* y_buf,
+                          const uint8* uv_buf,
+                          uint8* rgb_buf,
+                          int width) = NV12ToRGB565Row_C;
+  if (!src_y || !src_uv || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV12TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_SSSE3;
+    }
+  }
+#elif defined(HAS_NV12TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToRGB565Row(src_y, src_uv, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV21 to RGB565.
+LIBYUV_API
+int NV21ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_vu, int src_stride_vu,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*NV21ToRGB565Row)(const uint8* y_buf,
+                          const uint8* src_vu,
+                          uint8* rgb_buf,
+                          int width) = NV21ToRGB565Row_C;
+  if (!src_y || !src_vu || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV21TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_SSSE3;
+    }
+  }
+#elif defined(HAS_NV21TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV21ToRGB565Row(src_y, src_vu, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_vu += src_stride_vu;
+    }
+  }
+  return 0;
+}
+
+LIBYUV_API
+void SetPlane(uint8* dst_y, int dst_stride_y,
+              int width, int height,
+              uint32 value) {
+  int y;
+  uint32 v32 = value | (value << 8) | (value << 16) | (value << 24);
+  void (*SetRow)(uint8* dst, uint32 value, int pix) = SetRow_C;
+  // Coalesce rows.
+  if (dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    dst_stride_y = 0;
+  }
+#if defined(HAS_SETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) &&
+      IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    SetRow = SetRow_NEON;
+  }
+#endif
+#if defined(HAS_SETROW_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    SetRow = SetRow_X86;
+  }
+#endif
+
+  // Set plane
+  for (y = 0; y < height; ++y) {
+    SetRow(dst_y, v32, width);
+    dst_y += dst_stride_y;
+  }
+}
+
+// Draw a rectangle into I420
+LIBYUV_API
+int I420Rect(uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int x, int y,
+             int width, int height,
+             int value_y, int value_u, int value_v) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  uint8* start_y = dst_y + y * dst_stride_y + x;
+  uint8* start_u = dst_u + (y / 2) * dst_stride_u + (x / 2);
+  uint8* start_v = dst_v + (y / 2) * dst_stride_v + (x / 2);
+  if (!dst_y || !dst_u || !dst_v ||
+      width <= 0 || height <= 0 ||
+      x < 0 || y < 0 ||
+      value_y < 0 || value_y > 255 ||
+      value_u < 0 || value_u > 255 ||
+      value_v < 0 || value_v > 255) {
+    return -1;
+  }
+
+  SetPlane(start_y, dst_stride_y, width, height, value_y);
+  SetPlane(start_u, dst_stride_u, halfwidth, halfheight, value_u);
+  SetPlane(start_v, dst_stride_v, halfwidth, halfheight, value_v);
+  return 0;
+}
+
+// Draw a rectangle into ARGB
+LIBYUV_API
+int ARGBRect(uint8* dst_argb, int dst_stride_argb,
+             int dst_x, int dst_y,
+             int width, int height,
+             uint32 value) {
+  if (!dst_argb ||
+      width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  dst_argb += dst_y * dst_stride_argb + dst_x * 4;
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_SETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBSetRows_NEON(dst_argb, value, width, dst_stride_argb, height);
+    return 0;
+  }
+#endif
+#if defined(HAS_SETROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    ARGBSetRows_X86(dst_argb, value, width, dst_stride_argb, height);
+    return 0;
+  }
+#endif
+  ARGBSetRows_C(dst_argb, value, width, dst_stride_argb, height);
+  return 0;
+}
+
+// Convert unattentuated ARGB to preattenuated ARGB.
+// An unattenutated ARGB alpha blend uses the formula
+// p = a * f + (1 - a) * b
+// where
+//   p is output pixel
+//   f is foreground pixel
+//   b is background pixel
+//   a is alpha value from foreground pixel
+// An preattenutated ARGB alpha blend uses the formula
+// p = f + (1 - a) * b
+// where
+//   f is foreground pixel premultiplied by alpha
+
+LIBYUV_API
+int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  int y;
+  void (*ARGBAttenuateRow)(const uint8* src_argb, uint8* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBATTENUATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 4) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBAttenuateRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert preattentuated ARGB to unattenuated ARGB.
+LIBYUV_API
+int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height) {
+  int y;
+  void (*ARGBUnattenuateRow)(const uint8* src_argb, uint8* dst_argb,
+                             int width) = ARGBUnattenuateRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBUNATTENUATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBUNATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Neon version.
+
+  for (y = 0; y < height; ++y) {
+    ARGBUnattenuateRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB to Grayed ARGB.
+LIBYUV_API
+int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
+                      int width) = ARGBGrayRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBGRAYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBGrayRow = ARGBGrayRow_SSSE3;
+  }
+#elif defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBGrayRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Make a rectangle of ARGB gray scale.
+LIBYUV_API
+int ARGBGray(uint8* dst_argb, int dst_stride_argb,
+             int dst_x, int dst_y,
+             int width, int height) {
+  int y;
+  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
+                      int width) = ARGBGrayRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBGRAYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBGrayRow = ARGBGrayRow_SSSE3;
+  }
+#elif defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBGrayRow(dst, dst, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Make a rectangle of ARGB Sepia tone.
+LIBYUV_API
+int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
+              int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBSepiaRow)(uint8* dst_argb, int width) = ARGBSepiaRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSEPIAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBSepiaRow = ARGBSepiaRow_SSSE3;
+  }
+#elif defined(HAS_ARGBSEPIAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBSepiaRow = ARGBSepiaRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBSepiaRow(dst, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a 4x4 matrix to each ARGB pixel.
+// Note: Normally for shading, but can be used to swizzle or invert.
+LIBYUV_API
+int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    const int8* matrix_argb,
+                    int width, int height) {
+  int y;
+  void (*ARGBColorMatrixRow)(const uint8* src_argb, uint8* dst_argb,
+      const int8* matrix_argb, int width) = ARGBColorMatrixRow_C;
+  if (!src_argb || !dst_argb || !matrix_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOLORMATRIXROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_SSSE3;
+  }
+#elif defined(HAS_ARGBCOLORMATRIXROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBColorMatrixRow(src_argb, dst_argb, matrix_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a 4x3 matrix to each ARGB pixel.
+// Deprecated.
+LIBYUV_API
+int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
+                   const int8* matrix_rgb,
+                   int dst_x, int dst_y, int width, int height) {
+  SIMD_ALIGNED(int8 matrix_argb[16]);
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !matrix_rgb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+
+  // Convert 4x3 7 bit matrix to 4x4 6 bit matrix.
+  matrix_argb[0] = matrix_rgb[0] / 2;
+  matrix_argb[1] = matrix_rgb[1] / 2;
+  matrix_argb[2] = matrix_rgb[2] / 2;
+  matrix_argb[3] = matrix_rgb[3] / 2;
+  matrix_argb[4] = matrix_rgb[4] / 2;
+  matrix_argb[5] = matrix_rgb[5] / 2;
+  matrix_argb[6] = matrix_rgb[6] / 2;
+  matrix_argb[7] = matrix_rgb[7] / 2;
+  matrix_argb[8] = matrix_rgb[8] / 2;
+  matrix_argb[9] = matrix_rgb[9] / 2;
+  matrix_argb[10] = matrix_rgb[10] / 2;
+  matrix_argb[11] = matrix_rgb[11] / 2;
+  matrix_argb[14] = matrix_argb[13] = matrix_argb[12] = 0;
+  matrix_argb[15] = 64;  // 1.0
+
+  return ARGBColorMatrix((const uint8*)(dst), dst_stride_argb,
+                         dst, dst_stride_argb,
+                         &matrix_argb[0], width, height);
+}
+
+// Apply a color table each ARGB pixel.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                   const uint8* table_argb,
+                   int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+                            int width) = ARGBColorTableRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOLORTABLEROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    ARGBColorTableRow = ARGBColorTableRow_X86;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBColorTableRow(dst, table_argb, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a color table each ARGB pixel but preserve destination alpha.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                  const uint8* table_argb,
+                  int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*RGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+                           int width) = RGBColorTableRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_RGBCOLORTABLEROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    RGBColorTableRow = RGBColorTableRow_X86;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    RGBColorTableRow(dst, table_argb, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// ARGBQuantize is used to posterize art.
+// e.g. rgb / qvalue * qvalue + qvalue / 2
+// But the low levels implement efficiently with 3 parameters, and could be
+// used for other high level operations.
+// dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
+// where scale is 1 / interval_size as a fixed point value.
+// The divide is replaces with a multiply by reciprocal fixed point multiply.
+// Caveat - although SSE2 saturates, the C function does not and should be used
+// with care if doing anything but quantization.
+LIBYUV_API
+int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
+                 int scale, int interval_size, int interval_offset,
+                 int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBQuantizeRow)(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) = ARGBQuantizeRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0 ||
+      interval_size < 1 || interval_size > 255) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBQUANTIZEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_SSE2;
+  }
+#elif defined(HAS_ARGBQUANTIZEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBQuantizeRow(dst, scale, interval_size, interval_offset, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Computes table of cumulative sum for image where the value is the sum
+// of all values above and to the left of the entry. Used by ARGBBlur.
+LIBYUV_API
+int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
+                             int32* dst_cumsum, int dst_stride32_cumsum,
+                             int width, int height) {
+  int y;
+  void (*ComputeCumulativeSumRow)(const uint8* row, int32* cumsum,
+      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
+  int32* previous_cumsum = dst_cumsum;
+  if (!dst_cumsum || !src_argb || width <= 0 || height <= 0) {
+    return -1;
+  }
+#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
+  }
+#endif
+  memset(dst_cumsum, 0, width * sizeof(dst_cumsum[0]) * 4);  // 4 int per pixel.
+  for (y = 0; y < height; ++y) {
+    ComputeCumulativeSumRow(src_argb, dst_cumsum, previous_cumsum, width);
+    previous_cumsum = dst_cumsum;
+    dst_cumsum += dst_stride32_cumsum;
+    src_argb += src_stride_argb;
+  }
+  return 0;
+}
+
+// Blur ARGB image.
+// Caller should allocate CumulativeSum table of width * height * 16 bytes
+// aligned to 16 byte boundary. height can be radius * 2 + 2 to save memory
+// as the buffer is treated as circular.
+LIBYUV_API
+int ARGBBlur(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int32* dst_cumsum, int dst_stride32_cumsum,
+             int width, int height, int radius) {
+  int y;
+  void (*ComputeCumulativeSumRow)(const uint8 *row, int32 *cumsum,
+      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
+  void (*CumulativeSumToAverageRow)(const int32* topleft, const int32* botleft,
+      int width, int area, uint8* dst, int count) = CumulativeSumToAverageRow_C;
+  int32* cumsum_bot_row;
+  int32* max_cumsum_bot_row;
+  int32* cumsum_top_row;
+
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  if (radius > height) {
+    radius = height;
+  }
+  if (radius > (width / 2 - 1)) {
+    radius = width / 2 - 1;
+  }
+  if (radius <= 0) {
+    return -1;
+  }
+#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
+    CumulativeSumToAverageRow = CumulativeSumToAverageRow_SSE2;
+  }
+#endif
+  // Compute enough CumulativeSum for first row to be blurred. After this
+  // one row of CumulativeSum is updated at a time.
+  ARGBComputeCumulativeSum(src_argb, src_stride_argb,
+                           dst_cumsum, dst_stride32_cumsum,
+                           width, radius);
+
+  src_argb = src_argb + radius * src_stride_argb;
+  cumsum_bot_row = &dst_cumsum[(radius - 1) * dst_stride32_cumsum];
+
+  max_cumsum_bot_row = &dst_cumsum[(radius * 2 + 2) * dst_stride32_cumsum];
+  cumsum_top_row = &dst_cumsum[0];
+
+  for (y = 0; y < height; ++y) {
+    int top_y = ((y - radius - 1) >= 0) ? (y - radius - 1) : 0;
+    int bot_y = ((y + radius) < height) ? (y + radius) : (height - 1);
+    int area = radius * (bot_y - top_y);
+    int boxwidth = radius * 4;
+    int x;
+    int n;
+
+    // Increment cumsum_top_row pointer with circular buffer wrap around.
+    if (top_y) {
+      cumsum_top_row += dst_stride32_cumsum;
+      if (cumsum_top_row >= max_cumsum_bot_row) {
+        cumsum_top_row = dst_cumsum;
+      }
+    }
+    // Increment cumsum_bot_row pointer with circular buffer wrap around and
+    // then fill in a row of CumulativeSum.
+    if ((y + radius) < height) {
+      const int32* prev_cumsum_bot_row = cumsum_bot_row;
+      cumsum_bot_row += dst_stride32_cumsum;
+      if (cumsum_bot_row >= max_cumsum_bot_row) {
+        cumsum_bot_row = dst_cumsum;
+      }
+      ComputeCumulativeSumRow(src_argb, cumsum_bot_row, prev_cumsum_bot_row,
+                              width);
+      src_argb += src_stride_argb;
+    }
+
+    // Left clipped.
+    for (x = 0; x < radius + 1; ++x) {
+      CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
+                                boxwidth, area, &dst_argb[x * 4], 1);
+      area += (bot_y - top_y);
+      boxwidth += 4;
+    }
+
+    // Middle unclipped.
+    n = (width - 1) - radius - x + 1;
+    CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
+                              boxwidth, area, &dst_argb[x * 4], n);
+
+    // Right clipped.
+    for (x += n; x <= width - 1; ++x) {
+      area -= (bot_y - top_y);
+      boxwidth -= 4;
+      CumulativeSumToAverageRow(cumsum_top_row + (x - radius - 1) * 4,
+                                cumsum_bot_row + (x - radius - 1) * 4,
+                                boxwidth, area, &dst_argb[x * 4], 1);
+    }
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Multiply ARGB image by a specified ARGB value.
+LIBYUV_API
+int ARGBShade(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height, uint32 value) {
+  int y;
+  void (*ARGBShadeRow)(const uint8* src_argb, uint8* dst_argb,
+                       int width, uint32 value) = ARGBShadeRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0 || value == 0u) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSHADEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBShadeRow = ARGBShadeRow_SSE2;
+  }
+#elif defined(HAS_ARGBSHADEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBShadeRow = ARGBShadeRow_NEON;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBShadeRow(src_argb, dst_argb, width, value);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Interpolate 2 ARGB images by specified amount (0 to 255).
+LIBYUV_API
+int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
+                    const uint8* src_argb1, int src_stride_argb1,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height, int interpolation) {
+  int y;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_argb0, 16) && IS_ALIGNED(src_stride_argb0, 16) &&
+          IS_ALIGNED(src_argb1, 16) && IS_ALIGNED(src_stride_argb1, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb0, 16) && IS_ALIGNED(src_stride_argb0, 16) &&
+          IS_ALIGNED(src_argb1, 16) && IS_ALIGNED(src_stride_argb1, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 4) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && width >= 1 &&
+      IS_ALIGNED(src_argb0, 4) && IS_ALIGNED(src_stride_argb0, 4) &&
+      IS_ALIGNED(src_argb1, 4) && IS_ALIGNED(src_stride_argb1, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    ScaleARGBFilterRows = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    InterpolateRow(dst_argb, src_argb0, src_argb1 - src_argb0,
+                   width * 4, interpolation);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Shuffle ARGB channel order.  e.g. BGRA to ARGB.
+LIBYUV_API
+int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
+                uint8* dst_argb, int dst_stride_argb,
+                const uint8* shuffler, int width, int height) {
+  int y;
+  void (*ARGBShuffleRow)(const uint8* src_bgra, uint8* dst_argb,
+                         const uint8* shuffler, int pix) = ARGBShuffleRow_C;
+  if (!src_bgra || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_bgra = src_bgra + (height - 1) * src_stride_bgra;
+    src_stride_bgra = -src_stride_bgra;
+  }
+  // Coalesce rows.
+  if (src_stride_bgra == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_bgra = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSHUFFLEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBShuffleRow = ARGBShuffleRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBShuffleRow = ARGBShuffleRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_bgra, 16) && IS_ALIGNED(src_stride_bgra, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        ARGBShuffleRow = ARGBShuffleRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBShuffleRow = ARGBShuffleRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 4) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBShuffleRow = ARGBShuffleRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBShuffleRow(src_bgra, dst_argb, shuffler, width);
+    src_bgra += src_stride_bgra;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Sobel ARGB effect.
+static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
+                        uint8* dst_argb, int dst_stride_argb,
+                        int width, int height,
+                        void (*SobelRow)(const uint8* src_sobelx,
+                                         const uint8* src_sobely,
+                                         uint8* dst, int width)) {
+  int y;
+  void (*ARGBToBayerRow)(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) = ARGBToBayerGGRow_C;
+  void (*SobelYRow)(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) = SobelYRow_C;
+  void (*SobelXRow)(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobely, int width) =
+      SobelXRow_C;
+  const int kEdge = 16;  // Extra pixels at start of row for extrude/align.
+  if (!src_argb  || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // ARGBToBayer used to select G channel from ARGB.
+#if defined(HAS_ARGBTOBAYERGGROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToBayerRow = ARGBToBayerGGRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerGGRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOBAYERROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOBAYERGGROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToBayerRow = ARGBToBayerGGRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerGGRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SOBELYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelYRow = SobelYRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelYRow = SobelYRow_NEON;
+  }
+#endif
+#if defined(HAS_SOBELXROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelXRow = SobelXRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELXROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelXRow = SobelXRow_NEON;
+  }
+#endif
+  {
+    // 3 rows with edges before/after.
+    const int kRowSize = (width + kEdge + 15) & ~15;
+    align_buffer_64(rows, kRowSize * 2 + (kEdge + kRowSize * 3 + kEdge));
+    uint8* row_sobelx = rows;
+    uint8* row_sobely = rows + kRowSize;
+    uint8* row_y = rows + kRowSize * 2;
+
+    // Convert first row.
+    uint8* row_y0 = row_y + kEdge;
+    uint8* row_y1 = row_y0 + kRowSize;
+    uint8* row_y2 = row_y1 + kRowSize;
+    ARGBToBayerRow(src_argb, row_y0, 0x0d090501, width);
+    row_y0[-1] = row_y0[0];
+    memset(row_y0 + width, row_y0[width - 1], 16);  // Extrude 16 for valgrind.
+    ARGBToBayerRow(src_argb, row_y1, 0x0d090501, width);
+    row_y1[-1] = row_y1[0];
+    memset(row_y1 + width, row_y1[width - 1], 16);
+    memset(row_y2 + width, 0, 16);
+
+    for (y = 0; y < height; ++y) {
+      // Convert next row of ARGB to Y.
+      if (y < (height - 1)) {
+        src_argb += src_stride_argb;
+      }
+      ARGBToBayerRow(src_argb, row_y2, 0x0d090501, width);
+      row_y2[-1] = row_y2[0];
+      row_y2[width] = row_y2[width - 1];
+
+      SobelXRow(row_y0 - 1, row_y1 - 1, row_y2 - 1, row_sobelx, width);
+      SobelYRow(row_y0 - 1, row_y2 - 1, row_sobely, width);
+      SobelRow(row_sobelx, row_sobely, dst_argb, width);
+
+      // Cycle thru circular queue of 3 row_y buffers.
+      {
+        uint8* row_yt = row_y0;
+        row_y0 = row_y1;
+        row_y1 = row_y2;
+        row_y2 = row_yt;
+      }
+
+      dst_argb += dst_stride_argb;
+    }
+    free_aligned_buffer_64(rows);
+  }
+  return 0;
+}
+
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) = SobelRow_C;
+#if defined(HAS_SOBELROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    SobelRow = SobelRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    SobelRow = SobelRow_NEON;
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height, SobelRow);
+}
+
+// Sobel ARGB effect with planar output.
+LIBYUV_API
+int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_y, int dst_stride_y,
+                     int width, int height) {
+  void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_, int width) = SobelToPlaneRow_C;
+#if defined(HAS_SOBELTOPLANEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    SobelToPlaneRow = SobelToPlaneRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELTOPLANEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    SobelToPlaneRow = SobelToPlaneRow_NEON;
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_y, dst_stride_y,
+                      width, height, SobelToPlaneRow);
+}
+
+// SobelXY ARGB effect.
+// Similar to Sobel, but also stores Sobel X in R and Sobel Y in B.  G = Sobel.
+LIBYUV_API
+int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height) {
+  void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) = SobelXYRow_C;
+#if defined(HAS_SOBELXYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    SobelXYRow = SobelXYRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELXYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    SobelXYRow = SobelXYRow_NEON;
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height, SobelXYRow);
+}
+
+// Apply a 4x4 polynomial to each ARGB pixel.
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height) {
+  int y;
+  void (*ARGBPolynomialRow)(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) = ARGBPolynomialRow_C;
+  if (!src_argb || !dst_argb || !poly || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBPOLYNOMIALROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBPOLYNOMIALROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && TestCpuFlag(kCpuHasFMA3) &&
+      IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBPolynomialRow(src_argb, dst_argb, poly, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a lumacolortable to each ARGB pixel.
+LIBYUV_API
+int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_argb, int dst_stride_argb,
+                       const uint8* luma,
+                       int width, int height) {
+  int y;
+  void (*ARGBLumaColorTableRow)(const uint8* src_argb, uint8* dst_argb,
+      int width, const uint8* luma, const uint32 lumacoeff) =
+      ARGBLumaColorTableRow_C;
+  if (!src_argb || !dst_argb || !luma || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBLUMACOLORTABLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4)) {
+    ARGBLumaColorTableRow = ARGBLumaColorTableRow_SSSE3;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBLumaColorTableRow(src_argb, dst_argb, width, luma, 0x00264b0f);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Copy Alpha from one ARGB image to another.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  int y;
+  void (*ARGBCopyAlphaRow)(const uint8* src_argb, uint8* dst_argb, int width) =
+      ARGBCopyAlphaRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOPYALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16) &&
+      IS_ALIGNED(width, 8)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBCOPYALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBCopyAlphaRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Copy a planar Y channel to the alpha channel of a destination ARGB image.
+LIBYUV_API
+int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
+                     uint8* dst_argb, int dst_stride_argb,
+                     int width, int height) {
+  int y;
+  void (*ARGBCopyYToAlphaRow)(const uint8* src_y, uint8* dst_argb, int width) =
+      ARGBCopyYToAlphaRow_C;
+  if (!src_y || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOPYYTOALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16) &&
+      IS_ALIGNED(width, 8)) {
+    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBCOPYYTOALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
+    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBCopyYToAlphaRow(src_y, dst_argb, width);
+    src_y += src_stride_y;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/row.h b/third_party/libyuv/source/row.h
deleted file mode 100644
index eabe180..0000000
--- a/third_party/libyuv/source/row.h
+++ /dev/null
@@ -1,264 +0,0 @@
-/*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#ifndef LIBYUV_SOURCE_ROW_H_
-#define LIBYUV_SOURCE_ROW_H_
-
-#include "third_party/libyuv/include/libyuv/basic_types.h"
-
-#define kMaxStride (2048 * 4)
-#define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a) - 1)))
-
-#if defined(COVERAGE_ENABLED) || defined(TARGET_IPHONE_SIMULATOR)
-#define YUV_DISABLE_ASM
-#endif
-
-#if defined(__ARM_NEON__) && !defined(YUV_DISABLE_ASM)
-#define HAS_FASTCONVERTYUVTOARGBROW_NEON
-void FastConvertYUVToARGBRow_NEON(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-#define HAS_FASTCONVERTYUVTOBGRAROW_NEON
-void FastConvertYUVToBGRARow_NEON(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-#define HAS_FASTCONVERTYUVTOABGRROW_NEON
-void FastConvertYUVToABGRRow_NEON(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-#endif
-
-// The following are available on all x86 platforms
-#if (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) && \
-    !defined(YUV_DISABLE_ASM)
-#define HAS_ABGRTOARGBROW_SSSE3
-#define HAS_BGRATOARGBROW_SSSE3
-#define HAS_BG24TOARGBROW_SSSE3
-#define HAS_RAWTOARGBROW_SSSE3
-#define HAS_RGB24TOYROW_SSSE3
-#define HAS_RAWTOYROW_SSSE3
-#define HAS_RGB24TOUVROW_SSSE3
-#define HAS_RAWTOUVROW_SSSE3
-#define HAS_ARGBTOYROW_SSSE3
-#define HAS_BGRATOYROW_SSSE3
-#define HAS_ABGRTOYROW_SSSE3
-#define HAS_ARGBTOUVROW_SSSE3
-#define HAS_BGRATOUVROW_SSSE3
-#define HAS_ABGRTOUVROW_SSSE3
-#define HAS_I400TOARGBROW_SSE2
-#define HAS_FASTCONVERTYTOARGBROW_SSE2
-#define HAS_FASTCONVERTYUVTOARGBROW_SSSE3
-#define HAS_FASTCONVERTYUVTOBGRAROW_SSSE3
-#define HAS_FASTCONVERTYUVTOABGRROW_SSSE3
-#define HAS_FASTCONVERTYUV444TOARGBROW_SSSE3
-#define HAS_REVERSE_ROW_SSSE3
-#endif
-
-// The following are available on Neon platforms
-#if defined(__ARM_NEON__) && !defined(YUV_DISABLE_ASM)
-#define HAS_REVERSE_ROW_NEON
-#endif
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-#ifdef HAS_ARGBTOYROW_SSSE3
-void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
-void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
-void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
-void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width);
-void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width);
-void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width);
-#endif
-#if defined(HAS_BG24TOARGBROW_SSSE3) && defined(HAS_ARGBTOYROW_SSSE3)
-#define HASRGB24TOYROW_SSSE3
-#endif
-#ifdef HASRGB24TOYROW_SSSE3
-void RGB24ToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
-void RAWToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
-void RGB24ToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                        uint8* dst_u, uint8* dst_v, int width);
-void RAWToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width);
-#endif
-#ifdef HAS_REVERSE_ROW_SSSE3
-void ReverseRow_SSSE3(const uint8* src, uint8* dst, int width);
-#endif
-#ifdef HAS_REVERSE_ROW_NEON
-void ReverseRow_NEON(const uint8* src, uint8* dst, int width);
-#endif
-void ReverseRow_C(const uint8* src, uint8* dst, int width);
-
-void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
-void BGRAToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
-void ABGRToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
-void RGB24ToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
-void RAWToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
-void ARGBToUVRow_C(const uint8* src_argb0, int src_stride_argb,
-                   uint8* dst_u, uint8* dst_v, int width);
-void BGRAToUVRow_C(const uint8* src_argb0, int src_stride_argb,
-                   uint8* dst_u, uint8* dst_v, int width);
-void ABGRToUVRow_C(const uint8* src_argb0, int src_stride_argb,
-                   uint8* dst_u, uint8* dst_v, int width);
-void RGB24ToUVRow_C(const uint8* src_argb0, int src_stride_argb,
-                    uint8* dst_u, uint8* dst_v, int width);
-void RAWToUVRow_C(const uint8* src_argb0, int src_stride_argb,
-                  uint8* dst_u, uint8* dst_v, int width);
-
-#ifdef HAS_BG24TOARGBROW_SSSE3
-void ABGRToARGBRow_SSSE3(const uint8* src_abgr, uint8* dst_argb, int pix);
-void BGRAToARGBRow_SSSE3(const uint8* src_bgra, uint8* dst_argb, int pix);
-void BG24ToARGBRow_SSSE3(const uint8* src_bg24, uint8* dst_argb, int pix);
-void RAWToARGBRow_SSSE3(const uint8* src_bg24, uint8* dst_argb, int pix);
-#endif
-void ABGRToARGBRow_C(const uint8* src_abgr, uint8* dst_argb, int pix);
-void BGRAToARGBRow_C(const uint8* src_bgra, uint8* dst_argb, int pix);
-void BG24ToARGBRow_C(const uint8* src_bg24, uint8* dst_argb, int pix);
-void RAWToARGBRow_C(const uint8* src_bg24, uint8* dst_argb, int pix);
-
-#ifdef HAS_I400TOARGBROW_SSE2
-void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
-#endif
-void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix);
-
-#if defined(_MSC_VER)
-#define SIMD_ALIGNED(var) __declspec(align(16)) var
-typedef __declspec(align(16)) signed char vec8[16];
-typedef __declspec(align(16)) unsigned char uvec8[16];
-typedef __declspec(align(16)) signed short vec16[8];
-#else // __GNUC__
-#define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
-typedef signed char __attribute__((vector_size(16))) vec8;
-typedef unsigned char __attribute__((vector_size(16))) uvec8;
-typedef signed short __attribute__((vector_size(16))) vec16;
-#endif
-
-//extern "C"
-SIMD_ALIGNED(const int16 kCoefficientsRgbY[768][4]);
-//extern "C"
-SIMD_ALIGNED(const int16 kCoefficientsBgraY[768][4]);
-//extern "C"
-SIMD_ALIGNED(const int16 kCoefficientsAbgrY[768][4]);
-
-void FastConvertYUVToARGBRow_C(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               uint8* rgb_buf,
-                               int width);
-
-void FastConvertYUVToBGRARow_C(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               uint8* rgb_buf,
-                               int width);
-
-void FastConvertYUVToABGRRow_C(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               uint8* rgb_buf,
-                               int width);
-
-void FastConvertYUV444ToARGBRow_C(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-
-void FastConvertYToARGBRow_C(const uint8* y_buf,
-                             uint8* rgb_buf,
-                             int width);
-
-#ifdef HAS_FASTCONVERTYUVTOARGBROW_SSE2
-void FastConvertYUVToARGBRow_SSE2(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-
-void FastConvertYUVToARGBRow4_SSE2(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* rgb_buf,
-                                   int width);
-
-void FastConvertYUVToBGRARow_SSE2(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-
-void FastConvertYUVToABGRRow_SSE2(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-
-void FastConvertYUV444ToARGBRow_SSE2(const uint8* y_buf,
-                                     const uint8* u_buf,
-                                     const uint8* v_buf,
-                                     uint8* rgb_buf,
-                                     int width);
-
-void FastConvertYToARGBRow_SSE2(const uint8* y_buf,
-                                uint8* rgb_buf,
-                                int width);
-#endif
-
-#ifdef HAS_FASTCONVERTYUVTOARGBROW_SSSE3
-void FastConvertYUVToARGBRow_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* rgb_buf,
-                                   int width);
-
-void FastConvertYUVToBGRARow_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* rgb_buf,
-                                   int width);
-
-void FastConvertYUVToABGRRow_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* rgb_buf,
-                                   int width);
-
-void FastConvertYUV444ToARGBRow_SSSE3(const uint8* y_buf,
-                                      const uint8* u_buf,
-                                      const uint8* v_buf,
-                                      uint8* rgb_buf,
-                                      int width);
-
-#endif
-
-#ifdef HAS_FASTCONVERTYTOARGBROW_SSE2
-void FastConvertYToARGBRow_SSE2(const uint8* y_buf,
-                                uint8* rgb_buf,
-                                int width);
-
-#endif
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
-
-#endif  // LIBYUV_SOURCE_ROW_H_
diff --git a/third_party/libyuv/source/row_any.cc b/third_party/libyuv/source/row_any.cc
new file mode 100644
index 0000000..27a0de1
--- /dev/null
+++ b/third_party/libyuv/source/row_any.cc
@@ -0,0 +1,542 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#include "third_party/libyuv/include/libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// TODO(fbarchard): Consider 'any' functions handling any quantity of pixels.
+// TODO(fbarchard): Consider 'any' functions handling odd alignment.
+// YUV to RGB does multiple of 8 with SIMD and remainder with C.
+#define YANY(NAMEANY, I420TORGB_SIMD, I420TORGB_C, UV_SHIFT, BPP, MASK)        \
+    void NAMEANY(const uint8* y_buf,                                           \
+                 const uint8* u_buf,                                           \
+                 const uint8* v_buf,                                           \
+                 uint8* rgb_buf,                                               \
+                 int width) {                                                  \
+      int n = width & ~MASK;                                                   \
+      I420TORGB_SIMD(y_buf, u_buf, v_buf, rgb_buf, n);                         \
+      I420TORGB_C(y_buf + n,                                                   \
+                  u_buf + (n >> UV_SHIFT),                                     \
+                  v_buf + (n >> UV_SHIFT),                                     \
+                  rgb_buf + n * BPP, width & MASK);                            \
+    }
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+YANY(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_Unaligned_SSSE3, I444ToARGBRow_C,
+     0, 4, 7)
+YANY(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_Unaligned_SSSE3, I422ToARGBRow_C,
+     1, 4, 7)
+YANY(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_Unaligned_SSSE3, I411ToARGBRow_C,
+     2, 4, 7)
+YANY(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_Unaligned_SSSE3, I422ToBGRARow_C,
+     1, 4, 7)
+YANY(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_Unaligned_SSSE3, I422ToABGRRow_C,
+     1, 4, 7)
+YANY(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_Unaligned_SSSE3, I422ToRGBARow_C,
+     1, 4, 7)
+// I422ToRGB565Row_SSSE3 is unaligned.
+YANY(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, I422ToARGB4444Row_C,
+     1, 2, 7)
+YANY(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, I422ToARGB1555Row_C,
+     1, 2, 7)
+YANY(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, I422ToRGB565Row_C,
+     1, 2, 7)
+// I422ToRGB24Row_SSSE3 is unaligned.
+YANY(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, I422ToRGB24Row_C, 1, 3, 7)
+YANY(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, I422ToRAWRow_C, 1, 3, 7)
+YANY(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, I422ToYUY2Row_C, 1, 2, 15)
+YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15)
+#endif  // HAS_I422TOARGBROW_SSSE3
+#ifdef HAS_I422TOARGBROW_AVX2
+YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15)
+#endif  // HAS_I422TOARGBROW_AVX2
+#ifdef HAS_I422TOARGBROW_NEON
+YANY(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, I444ToARGBRow_C, 0, 4, 7)
+YANY(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, I422ToARGBRow_C, 1, 4, 7)
+YANY(I411ToARGBRow_Any_NEON, I411ToARGBRow_NEON, I411ToARGBRow_C, 2, 4, 7)
+YANY(I422ToBGRARow_Any_NEON, I422ToBGRARow_NEON, I422ToBGRARow_C, 1, 4, 7)
+YANY(I422ToABGRRow_Any_NEON, I422ToABGRRow_NEON, I422ToABGRRow_C, 1, 4, 7)
+YANY(I422ToRGBARow_Any_NEON, I422ToRGBARow_NEON, I422ToRGBARow_C, 1, 4, 7)
+YANY(I422ToRGB24Row_Any_NEON, I422ToRGB24Row_NEON, I422ToRGB24Row_C, 1, 3, 7)
+YANY(I422ToRAWRow_Any_NEON, I422ToRAWRow_NEON, I422ToRAWRow_C, 1, 3, 7)
+YANY(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, I422ToARGB4444Row_C,
+     1, 2, 7)
+YANY(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, I422ToARGB1555Row_C,
+     1, 2, 7)
+YANY(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, I422ToRGB565Row_C, 1, 2, 7)
+YANY(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, I422ToYUY2Row_C, 1, 2, 15)
+YANY(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, I422ToUYVYRow_C, 1, 2, 15)
+#endif  // HAS_I422TOARGBROW_NEON
+#undef YANY
+
+// Wrappers to handle odd width
+#define NV2NY(NAMEANY, NV12TORGB_SIMD, NV12TORGB_C, UV_SHIFT, BPP)             \
+    void NAMEANY(const uint8* y_buf,                                           \
+                 const uint8* uv_buf,                                          \
+                 uint8* rgb_buf,                                               \
+                 int width) {                                                  \
+      int n = width & ~7;                                                      \
+      NV12TORGB_SIMD(y_buf, uv_buf, rgb_buf, n);                               \
+      NV12TORGB_C(y_buf + n,                                                   \
+                  uv_buf + (n >> UV_SHIFT),                                    \
+                  rgb_buf + n * BPP, width & 7);                               \
+    }
+
+#ifdef HAS_NV12TOARGBROW_SSSE3
+NV2NY(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_Unaligned_SSSE3, NV12ToARGBRow_C,
+      0, 4)
+NV2NY(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_Unaligned_SSSE3, NV21ToARGBRow_C,
+      0, 4)
+#endif  // HAS_NV12TOARGBROW_SSSE3
+#ifdef HAS_NV12TOARGBROW_NEON
+NV2NY(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, NV12ToARGBRow_C, 0, 4)
+NV2NY(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, NV21ToARGBRow_C, 0, 4)
+#endif  // HAS_NV12TOARGBROW_NEON
+#ifdef HAS_NV12TORGB565ROW_SSSE3
+NV2NY(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, NV12ToRGB565Row_C,
+      0, 2)
+NV2NY(NV21ToRGB565Row_Any_SSSE3, NV21ToRGB565Row_SSSE3, NV21ToRGB565Row_C,
+      0, 2)
+#endif  // HAS_NV12TORGB565ROW_SSSE3
+#ifdef HAS_NV12TORGB565ROW_NEON
+NV2NY(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, NV12ToRGB565Row_C, 0, 2)
+NV2NY(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, NV21ToRGB565Row_C, 0, 2)
+#endif  // HAS_NV12TORGB565ROW_NEON
+#undef NVANY
+
+#define RGBANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP)          \
+    void NAMEANY(const uint8* src,                                             \
+                 uint8* dst,                                                   \
+                 int width) {                                                  \
+      int n = width & ~MASK;                                                   \
+      ARGBTORGB_SIMD(src, dst, n);                                             \
+      ARGBTORGB_C(src + n * SBPP, dst + n * BPP, width & MASK);                \
+    }
+
+#if defined(HAS_ARGBTORGB24ROW_SSSE3)
+RGBANY(ARGBToRGB24Row_Any_SSSE3, ARGBToRGB24Row_SSSE3, ARGBToRGB24Row_C,
+       15, 4, 3)
+RGBANY(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, ARGBToRAWRow_C,
+       15, 4, 3)
+RGBANY(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, ARGBToRGB565Row_C,
+       3, 4, 2)
+RGBANY(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, ARGBToARGB1555Row_C,
+       3, 4, 2)
+RGBANY(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, ARGBToARGB4444Row_C,
+       3, 4, 2)
+#endif
+#if defined(HAS_I400TOARGBROW_SSE2)
+RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_Unaligned_SSE2, I400ToARGBRow_C,
+       7, 1, 4)
+#endif
+#if defined(HAS_YTOARGBROW_SSE2)
+RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C,
+       7, 1, 4)
+RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_Unaligned_SSSE3, YUY2ToARGBRow_C,
+       15, 2, 4)
+RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_Unaligned_SSSE3, UYVYToARGBRow_C,
+       15, 2, 4)
+// These require alignment on ARGB, so C is used for remainder.
+RGBANY(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, RGB24ToARGBRow_C,
+       15, 3, 4)
+RGBANY(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, RAWToARGBRow_C,
+       15, 3, 4)
+RGBANY(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, RGB565ToARGBRow_C,
+       7, 2, 4)
+RGBANY(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, ARGB1555ToARGBRow_C,
+       7, 2, 4)
+RGBANY(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, ARGB4444ToARGBRow_C,
+       7, 2, 4)
+#endif
+#if defined(HAS_ARGBTORGB24ROW_NEON)
+RGBANY(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, ARGBToRGB24Row_C, 7, 4, 3)
+RGBANY(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, ARGBToRAWRow_C, 7, 4, 3)
+RGBANY(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, ARGBToRGB565Row_C,
+       7, 4, 2)
+RGBANY(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, ARGBToARGB1555Row_C,
+       7, 4, 2)
+RGBANY(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, ARGBToARGB4444Row_C,
+       7, 4, 2)
+RGBANY(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, I400ToARGBRow_C,
+       7, 1, 4)
+RGBANY(YToARGBRow_Any_NEON, YToARGBRow_NEON, YToARGBRow_C,
+       7, 1, 4)
+RGBANY(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, YUY2ToARGBRow_C,
+       7, 2, 4)
+RGBANY(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, UYVYToARGBRow_C,
+       7, 2, 4)
+#endif
+#undef RGBANY
+
+// ARGB to Bayer does multiple of 4 pixels, SSSE3 aligned src, unaligned dst.
+#define BAYERANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP)        \
+    void NAMEANY(const uint8* src,                                             \
+                 uint8* dst, uint32 selector,                                  \
+                 int width) {                                                  \
+      int n = width & ~MASK;                                                   \
+      ARGBTORGB_SIMD(src, dst, selector, n);                                   \
+      ARGBTORGB_C(src + n * SBPP, dst + n * BPP, selector, width & MASK);      \
+    }
+
+#if defined(HAS_ARGBTOBAYERROW_SSSE3)
+BAYERANY(ARGBToBayerRow_Any_SSSE3, ARGBToBayerRow_SSSE3, ARGBToBayerRow_C,
+         7, 4, 1)
+#endif
+#if defined(HAS_ARGBTOBAYERROW_NEON)
+BAYERANY(ARGBToBayerRow_Any_NEON, ARGBToBayerRow_NEON, ARGBToBayerRow_C,
+         7, 4, 1)
+#endif
+#if defined(HAS_ARGBTOBAYERGGROW_SSE2)
+BAYERANY(ARGBToBayerGGRow_Any_SSE2, ARGBToBayerGGRow_SSE2, ARGBToBayerGGRow_C,
+         7, 4, 1)
+#endif
+#if defined(HAS_ARGBTOBAYERGGROW_NEON)
+BAYERANY(ARGBToBayerGGRow_Any_NEON, ARGBToBayerGGRow_NEON, ARGBToBayerGGRow_C,
+         7, 4, 1)
+#endif
+
+#undef BAYERANY
+
+// RGB/YUV to Y does multiple of 16 with SIMD and last 16 with SIMD.
+#define YANY(NAMEANY, ARGBTOY_SIMD, SBPP, BPP, NUM)                            \
+    void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) {             \
+      ARGBTOY_SIMD(src_argb, dst_y, width - NUM);                              \
+      ARGBTOY_SIMD(src_argb + (width - NUM) * SBPP,                            \
+                   dst_y + (width - NUM) * BPP, NUM);                          \
+    }
+
+#ifdef HAS_ARGBTOYROW_AVX2
+YANY(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 4, 1, 32)
+YANY(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 4, 1, 32)
+YANY(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 2, 1, 32)
+YANY(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 2, 1, 32)
+#endif
+#ifdef HAS_ARGBTOYROW_SSSE3
+YANY(ARGBToYRow_Any_SSSE3, ARGBToYRow_Unaligned_SSSE3, 4, 1, 16)
+#endif
+#ifdef HAS_BGRATOYROW_SSSE3
+YANY(BGRAToYRow_Any_SSSE3, BGRAToYRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(ABGRToYRow_Any_SSSE3, ABGRToYRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(RGBAToYRow_Any_SSSE3, RGBAToYRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(YUY2ToYRow_Any_SSE2, YUY2ToYRow_Unaligned_SSE2, 2, 1, 16)
+YANY(UYVYToYRow_Any_SSE2, UYVYToYRow_Unaligned_SSE2, 2, 1, 16)
+#endif
+#ifdef HAS_ARGBTOYJROW_SSSE3
+YANY(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_Unaligned_SSSE3, 4, 1, 16)
+#endif
+#ifdef HAS_ARGBTOYROW_NEON
+YANY(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 4, 1, 8)
+YANY(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 4, 1, 8)
+YANY(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 4, 1, 8)
+YANY(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 4, 1, 8)
+YANY(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 4, 1, 8)
+YANY(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 3, 1, 8)
+YANY(RAWToYRow_Any_NEON, RAWToYRow_NEON, 3, 1, 8)
+YANY(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 2, 1, 8)
+YANY(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 2, 1, 8)
+YANY(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 2, 1, 8)
+YANY(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 2, 1, 16)
+YANY(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 2, 1, 16)
+YANY(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 3, 4, 8)
+YANY(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 3, 4, 8)
+YANY(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 2, 4, 8)
+YANY(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 2, 4, 8)
+YANY(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 2, 4, 8)
+#endif
+#undef YANY
+
+#define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK)                \
+    void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) {             \
+      int n = width & ~MASK;                                                   \
+      ARGBTOY_SIMD(src_argb, dst_y, n);                                        \
+      ARGBTOY_C(src_argb + n * SBPP,                                           \
+                dst_y  + n * BPP, width & MASK);                               \
+    }
+
+// Attenuate is destructive so last16 method can not be used due to overlap.
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+YANY(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, ARGBAttenuateRow_C,
+     4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+YANY(ARGBAttenuateRow_Any_SSE2, ARGBAttenuateRow_SSE2, ARGBAttenuateRow_C,
+     4, 4, 3)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+YANY(ARGBUnattenuateRow_Any_SSE2, ARGBUnattenuateRow_SSE2, ARGBUnattenuateRow_C,
+     4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+YANY(ARGBAttenuateRow_Any_AVX2, ARGBAttenuateRow_AVX2, ARGBAttenuateRow_C,
+     4, 4, 7)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+YANY(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, ARGBUnattenuateRow_C,
+     4, 4, 7)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_NEON
+YANY(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, ARGBAttenuateRow_C,
+     4, 4, 7)
+#endif
+#undef YANY
+
+// RGB/YUV to UV does multiple of 16 with SIMD and remainder with C.
+#define UVANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK)                     \
+    void NAMEANY(const uint8* src_argb, int src_stride_argb,                   \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_argb, src_stride_argb, dst_u, dst_v, n);                \
+      ANYTOUV_C(src_argb  + n * BPP, src_stride_argb,                          \
+                dst_u + (n >> 1),                                              \
+                dst_v + (n >> 1),                                              \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+UVANY(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, ARGBToUVRow_C, 4, 31)
+UVANY(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, YUY2ToUVRow_C, 2, 31)
+UVANY(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, UYVYToUVRow_C, 2, 31)
+#endif
+#ifdef HAS_ARGBTOUVROW_SSSE3
+UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_Unaligned_SSSE3, ARGBToUVRow_C, 4, 15)
+UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_Unaligned_SSSE3, ARGBToUVJRow_C,
+      4, 15)
+UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_Unaligned_SSSE3, BGRAToUVRow_C, 4, 15)
+UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_Unaligned_SSSE3, ABGRToUVRow_C, 4, 15)
+UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_Unaligned_SSSE3, RGBAToUVRow_C, 4, 15)
+UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_Unaligned_SSE2, YUY2ToUVRow_C, 2, 15)
+UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_Unaligned_SSE2, UYVYToUVRow_C, 2, 15)
+#endif
+#ifdef HAS_ARGBTOUVROW_NEON
+UVANY(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, ARGBToUVRow_C, 4, 15)
+UVANY(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, ARGBToUVJRow_C, 4, 15)
+UVANY(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, BGRAToUVRow_C, 4, 15)
+UVANY(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, ABGRToUVRow_C, 4, 15)
+UVANY(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, RGBAToUVRow_C, 4, 15)
+UVANY(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, RGB24ToUVRow_C, 3, 15)
+UVANY(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, RAWToUVRow_C, 3, 15)
+UVANY(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, RGB565ToUVRow_C, 2, 15)
+UVANY(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, ARGB1555ToUVRow_C, 2, 15)
+UVANY(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, ARGB4444ToUVRow_C, 2, 15)
+UVANY(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, YUY2ToUVRow_C, 2, 15)
+UVANY(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, UYVYToUVRow_C, 2, 15)
+#endif
+#undef UVANY
+
+#define UV422ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK, SHIFT)           \
+    void NAMEANY(const uint8* src_uv,                                          \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_uv, dst_u, dst_v, n);                                   \
+      ANYTOUV_C(src_uv  + n * BPP,                                             \
+                dst_u + (n >> SHIFT),                                          \
+                dst_v + (n >> SHIFT),                                          \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_ARGBTOUV444ROW_SSSE3
+UV422ANY(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_Unaligned_SSSE3,
+         ARGBToUV444Row_C, 4, 15, 0)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_AVX2
+UV422ANY(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2,
+         YUY2ToUV422Row_C, 2, 31, 1)
+UV422ANY(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2,
+         UYVYToUV422Row_C, 2, 31, 1)
+#endif
+#ifdef HAS_ARGBTOUVROW_SSSE3
+UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_Unaligned_SSSE3,
+         ARGBToUV422Row_C, 4, 15, 1)
+UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_Unaligned_SSE2,
+         YUY2ToUV422Row_C, 2, 15, 1)
+UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_Unaligned_SSE2,
+         UYVYToUV422Row_C, 2, 15, 1)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_NEON
+UV422ANY(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON,
+         ARGBToUV444Row_C, 4, 7, 0)
+UV422ANY(ARGBToUV422Row_Any_NEON, ARGBToUV422Row_NEON,
+         ARGBToUV422Row_C, 4, 15, 1)
+UV422ANY(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON,
+         ARGBToUV411Row_C, 4, 31, 2)
+UV422ANY(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON,
+         YUY2ToUV422Row_C, 2, 15, 1)
+UV422ANY(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON,
+         UYVYToUV422Row_C, 2, 15, 1)
+#endif
+#undef UV422ANY
+
+#define SPLITUVROWANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK)                  \
+    void NAMEANY(const uint8* src_uv,                                          \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_uv, dst_u, dst_v, n);                                   \
+      ANYTOUV_C(src_uv + n * 2,                                                \
+                dst_u + n,                                                     \
+                dst_v + n,                                                     \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_SPLITUVROW_SSE2
+SPLITUVROWANY(SplitUVRow_Any_SSE2, SplitUVRow_Unaligned_SSE2, SplitUVRow_C, 15)
+#endif
+#ifdef HAS_SPLITUVROW_AVX2
+SPLITUVROWANY(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, SplitUVRow_C, 31)
+#endif
+#ifdef HAS_SPLITUVROW_NEON
+SPLITUVROWANY(SplitUVRow_Any_NEON, SplitUVRow_NEON, SplitUVRow_C, 15)
+#endif
+#ifdef HAS_SPLITUVROW_MIPS_DSPR2
+SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_Unaligned_MIPS_DSPR2,
+              SplitUVRow_C, 15)
+#endif
+#undef SPLITUVROWANY
+
+#define MERGEUVROW_ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK)                 \
+    void NAMEANY(const uint8* src_u, const uint8* src_v,                       \
+                 uint8* dst_uv, int width) {                                   \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_u, src_v, dst_uv, n);                                   \
+      ANYTOUV_C(src_u + n,                                                     \
+                src_v + n,                                                     \
+                dst_uv + n * 2,                                                \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_MERGEUVROW_SSE2
+MERGEUVROW_ANY(MergeUVRow_Any_SSE2, MergeUVRow_Unaligned_SSE2, MergeUVRow_C, 15)
+#endif
+#ifdef HAS_MERGEUVROW_AVX2
+MERGEUVROW_ANY(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, MergeUVRow_C, 31)
+#endif
+#ifdef HAS_MERGEUVROW_NEON
+MERGEUVROW_ANY(MergeUVRow_Any_NEON, MergeUVRow_NEON, MergeUVRow_C, 15)
+#endif
+#undef MERGEUVROW_ANY
+
+#define MATHROW_ANY(NAMEANY, ARGBMATH_SIMD, ARGBMATH_C, MASK)                  \
+    void NAMEANY(const uint8* src_argb0, const uint8* src_argb1,               \
+                 uint8* dst_argb, int width) {                                 \
+      int n = width & ~MASK;                                                   \
+      ARGBMATH_SIMD(src_argb0, src_argb1, dst_argb, n);                        \
+      ARGBMATH_C(src_argb0 + n * 4,                                            \
+                 src_argb1 + n * 4,                                            \
+                 dst_argb + n * 4,                                             \
+                 width & MASK);                                                \
+    }
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+MATHROW_ANY(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, ARGBMultiplyRow_C,
+            3)
+#endif
+#ifdef HAS_ARGBADDROW_SSE2
+MATHROW_ANY(ARGBAddRow_Any_SSE2, ARGBAddRow_SSE2, ARGBAddRow_C, 3)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+MATHROW_ANY(ARGBSubtractRow_Any_SSE2, ARGBSubtractRow_SSE2, ARGBSubtractRow_C,
+            3)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+MATHROW_ANY(ARGBMultiplyRow_Any_AVX2, ARGBMultiplyRow_AVX2, ARGBMultiplyRow_C,
+            7)
+#endif
+#ifdef HAS_ARGBADDROW_AVX2
+MATHROW_ANY(ARGBAddRow_Any_AVX2, ARGBAddRow_AVX2, ARGBAddRow_C, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+MATHROW_ANY(ARGBSubtractRow_Any_AVX2, ARGBSubtractRow_AVX2, ARGBSubtractRow_C,
+            7)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+MATHROW_ANY(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, ARGBMultiplyRow_C,
+            7)
+#endif
+#ifdef HAS_ARGBADDROW_NEON
+MATHROW_ANY(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, ARGBAddRow_C, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_NEON
+MATHROW_ANY(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, ARGBSubtractRow_C,
+            7)
+#endif
+#undef MATHROW_ANY
+
+// Shuffle may want to work in place, so last16 method can not be used.
+#define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK)                \
+    void NAMEANY(const uint8* src_argb, uint8* dst_argb,                       \
+                 const uint8* shuffler, int width) {                           \
+      int n = width & ~MASK;                                                   \
+      ARGBTOY_SIMD(src_argb, dst_argb, shuffler, n);                           \
+      ARGBTOY_C(src_argb + n * SBPP,                                           \
+                dst_argb  + n * BPP, shuffler, width & MASK);                  \
+    }
+
+#ifdef HAS_ARGBSHUFFLEROW_SSE2
+YANY(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2,
+     ARGBShuffleRow_C, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_SSSE3
+YANY(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_Unaligned_SSSE3,
+     ARGBShuffleRow_C, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+YANY(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2,
+     ARGBShuffleRow_C, 4, 4, 15)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_NEON
+YANY(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON,
+     ARGBShuffleRow_C, 4, 4, 3)
+#endif
+#undef YANY
+
+// Interpolate may want to work in place, so last16 method can not be used.
+#define NANY(NAMEANY, TERP_SIMD, TERP_C, SBPP, BPP, MASK)                      \
+    void NAMEANY(uint8* dst_ptr, const uint8* src_ptr,                         \
+                 ptrdiff_t src_stride_ptr, int width,                          \
+                 int source_y_fraction) {                                      \
+      int n = width & ~MASK;                                                   \
+      TERP_SIMD(dst_ptr, src_ptr, src_stride_ptr,                              \
+                n, source_y_fraction);                                         \
+      TERP_C(dst_ptr + n * BPP,                                                \
+             src_ptr + n * SBPP, src_stride_ptr,                               \
+             width & MASK, source_y_fraction);                                 \
+    }
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+NANY(InterpolateRow_Any_AVX2, InterpolateRow_AVX2,
+     InterpolateRow_C, 1, 1, 32)
+#endif
+#ifdef HAS_INTERPOLATEROW_SSSE3
+NANY(InterpolateRow_Any_SSSE3, InterpolateRow_Unaligned_SSSE3,
+     InterpolateRow_C, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_SSE2
+NANY(InterpolateRow_Any_SSE2, InterpolateRow_Unaligned_SSE2,
+     InterpolateRow_C, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_NEON
+NANY(InterpolateRow_Any_NEON, InterpolateRow_NEON,
+     InterpolateRow_C, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_MIPS_DSPR2
+NANY(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2,
+     InterpolateRow_C, 1, 1, 3)
+#endif
+#undef NANY
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/row_common.cc b/third_party/libyuv/source/row_common.cc
new file mode 100644
index 0000000..ceb3836
--- /dev/null
+++ b/third_party/libyuv/source/row_common.cc
@@ -0,0 +1,2286 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#include <string.h>  // For memcpy and memset.
+
+#include "third_party/libyuv/include/libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// llvm x86 is poor at ternary operator, so use branchless min/max.
+
+#define USE_BRANCHLESS 1
+#if USE_BRANCHLESS
+static __inline int32 clamp0(int32 v) {
+  return ((-(v) >> 31) & (v));
+}
+
+static __inline int32 clamp255(int32 v) {
+  return (((255 - (v)) >> 31) | (v)) & 255;
+}
+
+static __inline uint32 Clamp(int32 val) {
+  int v = clamp0(val);
+  return (uint32)(clamp255(v));
+}
+
+static __inline uint32 Abs(int32 v) {
+  int m = v >> 31;
+  return (v + m) ^ m;
+}
+#else  // USE_BRANCHLESS
+static __inline int32 clamp0(int32 v) {
+  return (v < 0) ? 0 : v;
+}
+
+static __inline int32 clamp255(int32 v) {
+  return (v > 255) ? 255 : v;
+}
+
+static __inline uint32 Clamp(int32 val) {
+  int v = clamp0(val);
+  return (uint32)(clamp255(v));
+}
+
+static __inline uint32 Abs(int32 v) {
+  return (v < 0) ? -v : v;
+}
+#endif  // USE_BRANCHLESS
+
+#ifdef LIBYUV_LITTLE_ENDIAN
+#define WRITEWORD(p, v) *(uint32*)(p) = v
+#else
+static inline void WRITEWORD(uint8* p, uint32 v) {
+  p[0] = (uint8)(v & 255);
+  p[1] = (uint8)((v >> 8) & 255);
+  p[2] = (uint8)((v >> 16) & 255);
+  p[3] = (uint8)((v >> 24) & 255);
+}
+#endif
+
+void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb24[0];
+    uint8 g = src_rgb24[1];
+    uint8 r = src_rgb24[2];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_rgb24 += 3;
+  }
+}
+
+void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 r = src_raw[0];
+    uint8 g = src_raw[1];
+    uint8 b = src_raw[2];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_raw += 3;
+  }
+}
+
+void RGB565ToARGBRow_C(const uint8* src_rgb565, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb565[0] & 0x1f;
+    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r = src_rgb565[1] >> 3;
+    dst_argb[0] = (b << 3) | (b >> 2);
+    dst_argb[1] = (g << 2) | (g >> 4);
+    dst_argb[2] = (r << 3) | (r >> 2);
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_rgb565 += 2;
+  }
+}
+
+void ARGB1555ToARGBRow_C(const uint8* src_argb1555, uint8* dst_argb,
+                         int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb1555[0] & 0x1f;
+    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 a = src_argb1555[1] >> 7;
+    dst_argb[0] = (b << 3) | (b >> 2);
+    dst_argb[1] = (g << 3) | (g >> 2);
+    dst_argb[2] = (r << 3) | (r >> 2);
+    dst_argb[3] = -a;
+    dst_argb += 4;
+    src_argb1555 += 2;
+  }
+}
+
+void ARGB4444ToARGBRow_C(const uint8* src_argb4444, uint8* dst_argb,
+                         int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb4444[0] & 0x0f;
+    uint8 g = src_argb4444[0] >> 4;
+    uint8 r = src_argb4444[1] & 0x0f;
+    uint8 a = src_argb4444[1] >> 4;
+    dst_argb[0] = (b << 4) | b;
+    dst_argb[1] = (g << 4) | g;
+    dst_argb[2] = (r << 4) | r;
+    dst_argb[3] = (a << 4) | a;
+    dst_argb += 4;
+    src_argb4444 += 2;
+  }
+}
+
+void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb[0];
+    uint8 g = src_argb[1];
+    uint8 r = src_argb[2];
+    dst_rgb[0] = b;
+    dst_rgb[1] = g;
+    dst_rgb[2] = r;
+    dst_rgb += 3;
+    src_argb += 4;
+  }
+}
+
+void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb[0];
+    uint8 g = src_argb[1];
+    uint8 r = src_argb[2];
+    dst_rgb[0] = r;
+    dst_rgb[1] = g;
+    dst_rgb[2] = b;
+    dst_rgb += 3;
+    src_argb += 4;
+  }
+}
+
+void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 2;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 b1 = src_argb[4] >> 3;
+    uint8 g1 = src_argb[5] >> 2;
+    uint8 r1 = src_argb[6] >> 3;
+    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) |
+              (b1 << 16) | (g1 << 21) | (r1 << 27));
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 2;
+    uint8 r0 = src_argb[2] >> 3;
+    *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 3;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 a0 = src_argb[3] >> 7;
+    uint8 b1 = src_argb[4] >> 3;
+    uint8 g1 = src_argb[5] >> 3;
+    uint8 r1 = src_argb[6] >> 3;
+    uint8 a1 = src_argb[7] >> 7;
+    *(uint32*)(dst_rgb) =
+        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15) |
+        (b1 << 16) | (g1 << 21) | (r1 << 26) | (a1 << 31);
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 3;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 a0 = src_argb[3] >> 7;
+    *(uint16*)(dst_rgb) =
+        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15);
+  }
+}
+
+void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 4;
+    uint8 g0 = src_argb[1] >> 4;
+    uint8 r0 = src_argb[2] >> 4;
+    uint8 a0 = src_argb[3] >> 4;
+    uint8 b1 = src_argb[4] >> 4;
+    uint8 g1 = src_argb[5] >> 4;
+    uint8 r1 = src_argb[6] >> 4;
+    uint8 a1 = src_argb[7] >> 4;
+    *(uint32*)(dst_rgb) =
+        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12) |
+        (b1 << 16) | (g1 << 20) | (r1 << 24) | (a1 << 28);
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 4;
+    uint8 g0 = src_argb[1] >> 4;
+    uint8 r0 = src_argb[2] >> 4;
+    uint8 a0 = src_argb[3] >> 4;
+    *(uint16*)(dst_rgb) =
+        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12);
+  }
+}
+
+static __inline int RGBToY(uint8 r, uint8 g, uint8 b) {
+  return (66 * r + 129 * g +  25 * b + 0x1080) >> 8;
+}
+
+static __inline int RGBToU(uint8 r, uint8 g, uint8 b) {
+  return (112 * b - 74 * g - 38 * r + 0x8080) >> 8;
+}
+static __inline int RGBToV(uint8 r, uint8 g, uint8 b) {
+  return (112 * r - 94 * g - 18 * b + 0x8080) >> 8;
+}
+
+#define MAKEROWY(NAME, R, G, B, BPP) \
+void NAME ## ToYRow_C(const uint8* src_argb0, uint8* dst_y, int width) {       \
+  int x;                                                                       \
+  for (x = 0; x < width; ++x) {                                                \
+    dst_y[0] = RGBToY(src_argb0[R], src_argb0[G], src_argb0[B]);               \
+    src_argb0 += BPP;                                                          \
+    dst_y += 1;                                                                \
+  }                                                                            \
+}                                                                              \
+void NAME ## ToUVRow_C(const uint8* src_rgb0, int src_stride_rgb,              \
+                       uint8* dst_u, uint8* dst_v, int width) {                \
+  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
+  int x;                                                                       \
+  for (x = 0; x < width - 1; x += 2) {                                         \
+    uint8 ab = (src_rgb0[B] + src_rgb0[B + BPP] +                              \
+               src_rgb1[B] + src_rgb1[B + BPP]) >> 2;                          \
+    uint8 ag = (src_rgb0[G] + src_rgb0[G + BPP] +                              \
+               src_rgb1[G] + src_rgb1[G + BPP]) >> 2;                          \
+    uint8 ar = (src_rgb0[R] + src_rgb0[R + BPP] +                              \
+               src_rgb1[R] + src_rgb1[R + BPP]) >> 2;                          \
+    dst_u[0] = RGBToU(ar, ag, ab);                                             \
+    dst_v[0] = RGBToV(ar, ag, ab);                                             \
+    src_rgb0 += BPP * 2;                                                       \
+    src_rgb1 += BPP * 2;                                                       \
+    dst_u += 1;                                                                \
+    dst_v += 1;                                                                \
+  }                                                                            \
+  if (width & 1) {                                                             \
+    uint8 ab = (src_rgb0[B] + src_rgb1[B]) >> 1;                               \
+    uint8 ag = (src_rgb0[G] + src_rgb1[G]) >> 1;                               \
+    uint8 ar = (src_rgb0[R] + src_rgb1[R]) >> 1;                               \
+    dst_u[0] = RGBToU(ar, ag, ab);                                             \
+    dst_v[0] = RGBToV(ar, ag, ab);                                             \
+  }                                                                            \
+}
+
+MAKEROWY(ARGB, 2, 1, 0, 4)
+MAKEROWY(BGRA, 1, 2, 3, 4)
+MAKEROWY(ABGR, 0, 1, 2, 4)
+MAKEROWY(RGBA, 3, 2, 1, 4)
+MAKEROWY(RGB24, 2, 1, 0, 3)
+MAKEROWY(RAW, 0, 1, 2, 3)
+#undef MAKEROWY
+
+// JPeg uses a variation on BT.601-1 full range
+// y =  0.29900 * r + 0.58700 * g + 0.11400 * b
+// u = -0.16874 * r - 0.33126 * g + 0.50000 * b  + center
+// v =  0.50000 * r - 0.41869 * g - 0.08131 * b  + center
+// BT.601 Mpeg range uses:
+// b 0.1016 * 255 = 25.908 = 25
+// g 0.5078 * 255 = 129.489 = 129
+// r 0.2578 * 255 = 65.739 = 66
+// JPeg 8 bit Y (not used):
+// b 0.11400 * 256 = 29.184 = 29
+// g 0.58700 * 256 = 150.272 = 150
+// r 0.29900 * 256 = 76.544 = 77
+// JPeg 7 bit Y:
+// b 0.11400 * 128 = 14.592 = 15
+// g 0.58700 * 128 = 75.136 = 75
+// r 0.29900 * 128 = 38.272 = 38
+// JPeg 8 bit U:
+// b  0.50000 * 255 = 127.5 = 127
+// g -0.33126 * 255 = -84.4713 = -84
+// r -0.16874 * 255 = -43.0287 = -43
+// JPeg 8 bit V:
+// b -0.08131 * 255 = -20.73405 = -20
+// g -0.41869 * 255 = -106.76595 = -107
+// r  0.50000 * 255 = 127.5 = 127
+
+static __inline int RGBToYJ(uint8 r, uint8 g, uint8 b) {
+  return (38 * r + 75 * g +  15 * b + 64) >> 7;
+}
+
+static __inline int RGBToUJ(uint8 r, uint8 g, uint8 b) {
+  return (127 * b - 84 * g - 43 * r + 0x8080) >> 8;
+}
+static __inline int RGBToVJ(uint8 r, uint8 g, uint8 b) {
+  return (127 * r - 107 * g - 20 * b + 0x8080) >> 8;
+}
+
+#define AVGB(a, b) (((a) + (b) + 1) >> 1)
+
+#define MAKEROWYJ(NAME, R, G, B, BPP) \
+void NAME ## ToYJRow_C(const uint8* src_argb0, uint8* dst_y, int width) {      \
+  int x;                                                                       \
+  for (x = 0; x < width; ++x) {                                                \
+    dst_y[0] = RGBToYJ(src_argb0[R], src_argb0[G], src_argb0[B]);              \
+    src_argb0 += BPP;                                                          \
+    dst_y += 1;                                                                \
+  }                                                                            \
+}                                                                              \
+void NAME ## ToUVJRow_C(const uint8* src_rgb0, int src_stride_rgb,             \
+                        uint8* dst_u, uint8* dst_v, int width) {               \
+  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
+  int x;                                                                       \
+  for (x = 0; x < width - 1; x += 2) {                                         \
+    uint8 ab = AVGB(AVGB(src_rgb0[B], src_rgb1[B]),                            \
+                    AVGB(src_rgb0[B + BPP], src_rgb1[B + BPP]));               \
+    uint8 ag = AVGB(AVGB(src_rgb0[G], src_rgb1[G]),                            \
+                    AVGB(src_rgb0[G + BPP], src_rgb1[G + BPP]));               \
+    uint8 ar = AVGB(AVGB(src_rgb0[R], src_rgb1[R]),                            \
+                    AVGB(src_rgb0[R + BPP], src_rgb1[R + BPP]));               \
+    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
+    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
+    src_rgb0 += BPP * 2;                                                       \
+    src_rgb1 += BPP * 2;                                                       \
+    dst_u += 1;                                                                \
+    dst_v += 1;                                                                \
+  }                                                                            \
+  if (width & 1) {                                                             \
+    uint8 ab = AVGB(src_rgb0[B], src_rgb1[B]);                                 \
+    uint8 ag = AVGB(src_rgb0[G], src_rgb1[G]);                                 \
+    uint8 ar = AVGB(src_rgb0[R], src_rgb1[R]);                                 \
+    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
+    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
+  }                                                                            \
+}
+
+MAKEROWYJ(ARGB, 2, 1, 0, 4)
+#undef MAKEROWYJ
+
+void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb565[0] & 0x1f;
+    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r = src_rgb565[1] >> 3;
+    b = (b << 3) | (b >> 2);
+    g = (g << 2) | (g >> 4);
+    r = (r << 3) | (r >> 2);
+    dst_y[0] = RGBToY(r, g, b);
+    src_rgb565 += 2;
+    dst_y += 1;
+  }
+}
+
+void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb1555[0] & 0x1f;
+    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
+    b = (b << 3) | (b >> 2);
+    g = (g << 3) | (g >> 2);
+    r = (r << 3) | (r >> 2);
+    dst_y[0] = RGBToY(r, g, b);
+    src_argb1555 += 2;
+    dst_y += 1;
+  }
+}
+
+void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb4444[0] & 0x0f;
+    uint8 g = src_argb4444[0] >> 4;
+    uint8 r = src_argb4444[1] & 0x0f;
+    b = (b << 4) | b;
+    g = (g << 4) | g;
+    r = (r << 4) | r;
+    dst_y[0] = RGBToY(r, g, b);
+    src_argb4444 += 2;
+    dst_y += 1;
+  }
+}
+
+void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
+                     uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_rgb565 = src_rgb565 + src_stride_rgb565;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_rgb565[0] & 0x1f;
+    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r0 = src_rgb565[1] >> 3;
+    uint8 b1 = src_rgb565[2] & 0x1f;
+    uint8 g1 = (src_rgb565[2] >> 5) | ((src_rgb565[3] & 0x07) << 3);
+    uint8 r1 = src_rgb565[3] >> 3;
+    uint8 b2 = next_rgb565[0] & 0x1f;
+    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
+    uint8 r2 = next_rgb565[1] >> 3;
+    uint8 b3 = next_rgb565[2] & 0x1f;
+    uint8 g3 = (next_rgb565[2] >> 5) | ((next_rgb565[3] & 0x07) << 3);
+    uint8 r3 = next_rgb565[3] >> 3;
+    uint8 b = (b0 + b1 + b2 + b3);  // 565 * 4 = 787.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 1) | (b >> 6);  // 787 -> 888.
+    r = (r << 1) | (r >> 6);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_rgb565 += 4;
+    next_rgb565 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_rgb565[0] & 0x1f;
+    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r0 = src_rgb565[1] >> 3;
+    uint8 b2 = next_rgb565[0] & 0x1f;
+    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
+    uint8 r2 = next_rgb565[1] >> 3;
+    uint8 b = (b0 + b2);  // 565 * 2 = 676.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 2) | (b >> 4);  // 676 -> 888
+    g = (g << 1) | (g >> 6);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_argb1555 = src_argb1555 + src_stride_argb1555;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb1555[0] & 0x1f;
+    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 b1 = src_argb1555[2] & 0x1f;
+    uint8 g1 = (src_argb1555[2] >> 5) | ((src_argb1555[3] & 0x03) << 3);
+    uint8 r1 = (src_argb1555[3] & 0x7c) >> 2;
+    uint8 b2 = next_argb1555[0] & 0x1f;
+    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
+    uint8 r2 = (next_argb1555[1] & 0x7c) >> 2;
+    uint8 b3 = next_argb1555[2] & 0x1f;
+    uint8 g3 = (next_argb1555[2] >> 5) | ((next_argb1555[3] & 0x03) << 3);
+    uint8 r3 = (next_argb1555[3] & 0x7c) >> 2;
+    uint8 b = (b0 + b1 + b2 + b3);  // 555 * 4 = 777.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 1) | (b >> 6);  // 777 -> 888.
+    g = (g << 1) | (g >> 6);
+    r = (r << 1) | (r >> 6);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_argb1555 += 4;
+    next_argb1555 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb1555[0] & 0x1f;
+    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 b2 = next_argb1555[0] & 0x1f;
+    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
+    uint8 r2 = next_argb1555[1] >> 3;
+    uint8 b = (b0 + b2);  // 555 * 2 = 666.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 2) | (b >> 4);  // 666 -> 888.
+    g = (g << 2) | (g >> 4);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_argb4444 = src_argb4444 + src_stride_argb4444;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb4444[0] & 0x0f;
+    uint8 g0 = src_argb4444[0] >> 4;
+    uint8 r0 = src_argb4444[1] & 0x0f;
+    uint8 b1 = src_argb4444[2] & 0x0f;
+    uint8 g1 = src_argb4444[2] >> 4;
+    uint8 r1 = src_argb4444[3] & 0x0f;
+    uint8 b2 = next_argb4444[0] & 0x0f;
+    uint8 g2 = next_argb4444[0] >> 4;
+    uint8 r2 = next_argb4444[1] & 0x0f;
+    uint8 b3 = next_argb4444[2] & 0x0f;
+    uint8 g3 = next_argb4444[2] >> 4;
+    uint8 r3 = next_argb4444[3] & 0x0f;
+    uint8 b = (b0 + b1 + b2 + b3);  // 444 * 4 = 666.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 2) | (b >> 4);  // 666 -> 888.
+    g = (g << 2) | (g >> 4);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_argb4444 += 4;
+    next_argb4444 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb4444[0] & 0x0f;
+    uint8 g0 = src_argb4444[0] >> 4;
+    uint8 r0 = src_argb4444[1] & 0x0f;
+    uint8 b2 = next_argb4444[0] & 0x0f;
+    uint8 g2 = next_argb4444[0] >> 4;
+    uint8 r2 = next_argb4444[1] & 0x0f;
+    uint8 b = (b0 + b2);  // 444 * 2 = 555.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 3) | (b >> 2);  // 555 -> 888.
+    g = (g << 3) | (g >> 2);
+    r = (r << 3) | (r >> 2);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGBToUV444Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+void ARGBToUV422Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 8;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  }
+}
+
+void ARGBToUV411Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 3; x += 4) {
+    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8] + src_argb[12]) >> 2;
+    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9] + src_argb[13]) >> 2;
+    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10] + src_argb[14]) >> 2;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 16;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if ((width & 3) == 3) {
+    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8]) / 3;
+    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9]) / 3;
+    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10]) / 3;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  } else if ((width & 3) == 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  } else if ((width & 3) == 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  }
+}
+
+void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 y = RGBToYJ(src_argb[2], src_argb[1], src_argb[0]);
+    dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
+    dst_argb[3] = src_argb[3];
+    dst_argb += 4;
+    src_argb += 4;
+  }
+}
+
+// Convert a row of image to Sepia tone.
+void ARGBSepiaRow_C(uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    int sb = (b * 17 + g * 68 + r * 35) >> 7;
+    int sg = (b * 22 + g * 88 + r * 45) >> 7;
+    int sr = (b * 24 + g * 98 + r * 50) >> 7;
+    // b does not over flow. a is preserved from original.
+    dst_argb[0] = sb;
+    dst_argb[1] = clamp255(sg);
+    dst_argb[2] = clamp255(sr);
+    dst_argb += 4;
+  }
+}
+
+// Apply color matrix to a row of image. Matrix is signed.
+// TODO(fbarchard): Consider adding rounding (+32).
+void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
+                          const int8* matrix_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = src_argb[0];
+    int g = src_argb[1];
+    int r = src_argb[2];
+    int a = src_argb[3];
+    int sb = (b * matrix_argb[0] + g * matrix_argb[1] +
+              r * matrix_argb[2] + a * matrix_argb[3]) >> 6;
+    int sg = (b * matrix_argb[4] + g * matrix_argb[5] +
+              r * matrix_argb[6] + a * matrix_argb[7]) >> 6;
+    int sr = (b * matrix_argb[8] + g * matrix_argb[9] +
+              r * matrix_argb[10] + a * matrix_argb[11]) >> 6;
+    int sa = (b * matrix_argb[12] + g * matrix_argb[13] +
+              r * matrix_argb[14] + a * matrix_argb[15]) >> 6;
+    dst_argb[0] = Clamp(sb);
+    dst_argb[1] = Clamp(sg);
+    dst_argb[2] = Clamp(sr);
+    dst_argb[3] = Clamp(sa);
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+// Apply color table to a row of image.
+void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    int a = dst_argb[3];
+    dst_argb[0] = table_argb[b * 4 + 0];
+    dst_argb[1] = table_argb[g * 4 + 1];
+    dst_argb[2] = table_argb[r * 4 + 2];
+    dst_argb[3] = table_argb[a * 4 + 3];
+    dst_argb += 4;
+  }
+}
+
+// Apply color table to a row of image.
+void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    dst_argb[0] = table_argb[b * 4 + 0];
+    dst_argb[1] = table_argb[g * 4 + 1];
+    dst_argb[2] = table_argb[r * 4 + 2];
+    dst_argb += 4;
+  }
+}
+
+void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
+                       int interval_offset, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
+    dst_argb[1] = (g * scale >> 16) * interval_size + interval_offset;
+    dst_argb[2] = (r * scale >> 16) * interval_size + interval_offset;
+    dst_argb += 4;
+  }
+}
+
+#define REPEAT8(v) (v) | ((v) << 8)
+#define SHADE(f, v) v * f >> 24
+
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value) {
+  const uint32 b_scale = REPEAT8(value & 0xff);
+  const uint32 g_scale = REPEAT8((value >> 8) & 0xff);
+  const uint32 r_scale = REPEAT8((value >> 16) & 0xff);
+  const uint32 a_scale = REPEAT8(value >> 24);
+
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32 b = REPEAT8(src_argb[0]);
+    const uint32 g = REPEAT8(src_argb[1]);
+    const uint32 r = REPEAT8(src_argb[2]);
+    const uint32 a = REPEAT8(src_argb[3]);
+    dst_argb[0] = SHADE(b, b_scale);
+    dst_argb[1] = SHADE(g, g_scale);
+    dst_argb[2] = SHADE(r, r_scale);
+    dst_argb[3] = SHADE(a, a_scale);
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+#undef REPEAT8
+#undef SHADE
+
+#define REPEAT8(v) (v) | ((v) << 8)
+#define SHADE(f, v) v * f >> 16
+
+void ARGBMultiplyRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32 b = REPEAT8(src_argb0[0]);
+    const uint32 g = REPEAT8(src_argb0[1]);
+    const uint32 r = REPEAT8(src_argb0[2]);
+    const uint32 a = REPEAT8(src_argb0[3]);
+    const uint32 b_scale = src_argb1[0];
+    const uint32 g_scale = src_argb1[1];
+    const uint32 r_scale = src_argb1[2];
+    const uint32 a_scale = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_scale);
+    dst_argb[1] = SHADE(g, g_scale);
+    dst_argb[2] = SHADE(r, r_scale);
+    dst_argb[3] = SHADE(a, a_scale);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef REPEAT8
+#undef SHADE
+
+#define SHADE(f, v) clamp255(v + f)
+
+void ARGBAddRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                  uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const int b = src_argb0[0];
+    const int g = src_argb0[1];
+    const int r = src_argb0[2];
+    const int a = src_argb0[3];
+    const int b_add = src_argb1[0];
+    const int g_add = src_argb1[1];
+    const int r_add = src_argb1[2];
+    const int a_add = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_add);
+    dst_argb[1] = SHADE(g, g_add);
+    dst_argb[2] = SHADE(r, r_add);
+    dst_argb[3] = SHADE(a, a_add);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
+#define SHADE(f, v) clamp0(f - v)
+
+void ARGBSubtractRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const int b = src_argb0[0];
+    const int g = src_argb0[1];
+    const int r = src_argb0[2];
+    const int a = src_argb0[3];
+    const int b_sub = src_argb1[0];
+    const int g_sub = src_argb1[1];
+    const int r_sub = src_argb1[2];
+    const int a_sub = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_sub);
+    dst_argb[1] = SHADE(g, g_sub);
+    dst_argb[2] = SHADE(r, r_sub);
+    dst_argb[3] = SHADE(a, a_sub);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
+// Sobel functions which mimics SSSE3.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int a = src_y0[i];
+    int b = src_y1[i];
+    int c = src_y2[i];
+    int a_sub = src_y0[i + 2];
+    int b_sub = src_y1[i + 2];
+    int c_sub = src_y2[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = Abs(a_diff + b_diff * 2 + c_diff);
+    dst_sobelx[i] = (uint8)(clamp255(sobel));
+  }
+}
+
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int a = src_y0[i + 0];
+    int b = src_y0[i + 1];
+    int c = src_y0[i + 2];
+    int a_sub = src_y1[i + 0];
+    int b_sub = src_y1[i + 1];
+    int c_sub = src_y1[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = Abs(a_diff + b_diff * 2 + c_diff);
+    dst_sobely[i] = (uint8)(clamp255(sobel));
+  }
+}
+
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = clamp255(r + b);
+    dst_argb[0] = (uint8)(s);
+    dst_argb[1] = (uint8)(s);
+    dst_argb[2] = (uint8)(s);
+    dst_argb[3] = (uint8)(255u);
+    dst_argb += 4;
+  }
+}
+
+void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_y, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = clamp255(r + b);
+    dst_y[i] = (uint8)(s);
+  }
+}
+
+void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                  uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int g = clamp255(r + b);
+    dst_argb[0] = (uint8)(b);
+    dst_argb[1] = (uint8)(g);
+    dst_argb[2] = (uint8)(r);
+    dst_argb[3] = (uint8)(255u);
+    dst_argb += 4;
+  }
+}
+
+void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) {
+  // Copy a Y to RGB.
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 y = src_y[0];
+    dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    ++src_y;
+  }
+}
+
+// C reference code that mimics the YUV assembly.
+
+#define YG 74 /* (int8)(1.164 * 64 + 0.5) */
+
+#define UB 127 /* min(63,(int8)(2.018 * 64)) */
+#define UG -25 /* (int8)(-0.391 * 64 - 0.5) */
+#define UR 0
+
+#define VB 0
+#define VG -52 /* (int8)(-0.813 * 64 - 0.5) */
+#define VR 102 /* (int8)(1.596 * 64 + 0.5) */
+
+// Bias
+#define BB UB * 128 + VB * 128
+#define BG UG * 128 + VG * 128
+#define BR UR * 128 + VR * 128
+
+static __inline void YuvPixel(uint8 y, uint8 u, uint8 v,
+                              uint8* b, uint8* g, uint8* r) {
+  int32 y1 = ((int32)(y) - 16) * YG;
+  *b = Clamp((int32)((u * UB + v * VB) - (BB) + y1) >> 6);
+  *g = Clamp((int32)((u * UG + v * VG) - (BG) + y1) >> 6);
+  *r = Clamp((int32)((u * UR + v * VR) - (BR) + y1) >> 6);
+}
+
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+// C mimic assembly.
+// TODO(fbarchard): Remove subsampling from Neon.
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 u = (src_u[0] + src_u[1] + 1) >> 1;
+    uint8 v = (src_v[0] + src_v[1] + 1) >> 1;
+    YuvPixel(src_y[0], u, v, rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], u, v, rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 2;
+    src_v += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+  }
+}
+#else
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 4;  // Advance 1 pixel.
+  }
+}
+#endif
+// Also used for 420
+void I422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToRGB24Row_C(const uint8* src_y,
+                      const uint8* src_u,
+                      const uint8* src_v,
+                      uint8* rgb_buf,
+                      int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 4, rgb_buf + 5);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+  }
+}
+
+void I422ToRAWRow_C(const uint8* src_y,
+                    const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* rgb_buf,
+                    int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 5, rgb_buf + 4, rgb_buf + 3);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+  }
+}
+
+void I422ToARGB4444Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 4;
+    g0 = g0 >> 4;
+    r0 = r0 >> 4;
+    b1 = b1 >> 4;
+    g1 = g1 >> 4;
+    r1 = r1 >> 4;
+    *(uint32*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
+        (b1 << 16) | (g1 << 20) | (r1 << 24) | 0xf000f000;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_argb4444 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 4;
+    g0 = g0 >> 4;
+    r0 = r0 >> 4;
+    *(uint16*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
+        0xf000;
+  }
+}
+
+void I422ToARGB1555Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb1555,
+                         int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 3;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 3;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
+        (b1 << 16) | (g1 << 21) | (r1 << 26) | 0x80008000;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_argb1555 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 3;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
+        0x8000;
+  }
+}
+
+void I422ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void I411ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 3; x += 4) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    YuvPixel(src_y[2], src_u[0], src_v[0],
+             rgb_buf + 8, rgb_buf + 9, rgb_buf + 10);
+    rgb_buf[11] = 255;
+    YuvPixel(src_y[3], src_u[0], src_v[0],
+             rgb_buf + 12, rgb_buf + 13, rgb_buf + 14);
+    rgb_buf[15] = 255;
+    src_y += 4;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 16;  // Advance 4 pixels.
+  }
+  if (width & 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV12ToARGBRow_C(const uint8* src_y,
+                     const uint8* usrc_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], usrc_v[0], usrc_v[1],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    usrc_v += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV21ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_vu,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+
+    YuvPixel(src_y[1], src_vu[1], src_vu[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+
+    src_y += 2;
+    src_vu += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV12ToRGB565Row_C(const uint8* src_y,
+                       const uint8* usrc_v,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1], &b0, &g0, &r0);
+    YuvPixel(src_y[1], usrc_v[0], usrc_v[1], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    usrc_v += 2;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void NV21ToRGB565Row_C(const uint8* src_y,
+                       const uint8* vsrc_u,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], vsrc_u[1], vsrc_u[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    vsrc_u += 2;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void YUY2ToARGBRow_C(const uint8* src_yuy2,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_yuy2[2], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_yuy2 += 4;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void UYVYToARGBRow_C(const uint8* src_uyvy,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_uyvy[3], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_uyvy += 4;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToBGRARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 2, rgb_buf + 1);
+    rgb_buf[0] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 7, rgb_buf + 6, rgb_buf + 5);
+    rgb_buf[4] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 2, rgb_buf + 1);
+    rgb_buf[0] = 255;
+  }
+}
+
+void I422ToABGRRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 6, rgb_buf + 5, rgb_buf + 4);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToRGBARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3);
+    rgb_buf[0] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 5, rgb_buf + 6, rgb_buf + 7);
+    rgb_buf[4] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3);
+    rgb_buf[0] = 255;
+  }
+}
+
+void YToARGBRow_C(const uint8* src_y, uint8* rgb_buf, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], 128, 128,
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], 128, 128,
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], 128, 128,
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void MirrorRow_C(const uint8* src, uint8* dst, int width) {
+  int x;
+  src += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst[x] = src[0];
+    dst[x + 1] = src[-1];
+    src -= 2;
+  }
+  if (width & 1) {
+    dst[width - 1] = src[0];
+  }
+}
+
+void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  src_uv += (width - 1) << 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_u[x] = src_uv[0];
+    dst_u[x + 1] = src_uv[-2];
+    dst_v[x] = src_uv[1];
+    dst_v[x + 1] = src_uv[-2 + 1];
+    src_uv -= 4;
+  }
+  if (width & 1) {
+    dst_u[width - 1] = src_uv[0];
+    dst_v[width - 1] = src_uv[1];
+  }
+}
+
+void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width) {
+  int x;
+  const uint32* src32 = (const uint32*)(src);
+  uint32* dst32 = (uint32*)(dst);
+  src32 += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst32[x] = src32[0];
+    dst32[x + 1] = src32[-1];
+    src32 -= 2;
+  }
+  if (width & 1) {
+    dst32[width - 1] = src32[0];
+  }
+}
+
+void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_u[x] = src_uv[0];
+    dst_u[x + 1] = src_uv[2];
+    dst_v[x] = src_uv[1];
+    dst_v[x + 1] = src_uv[3];
+    src_uv += 4;
+  }
+  if (width & 1) {
+    dst_u[width - 1] = src_uv[0];
+    dst_v[width - 1] = src_uv[1];
+  }
+}
+
+void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                  int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_uv[0] = src_u[x];
+    dst_uv[1] = src_v[x];
+    dst_uv[2] = src_u[x + 1];
+    dst_uv[3] = src_v[x + 1];
+    dst_uv += 4;
+  }
+  if (width & 1) {
+    dst_uv[0] = src_u[width - 1];
+    dst_uv[1] = src_v[width - 1];
+  }
+}
+
+void CopyRow_C(const uint8* src, uint8* dst, int count) {
+  memcpy(dst, src, count);
+}
+
+void CopyRow_16_C(const uint16* src, uint16* dst, int count) {
+  memcpy(dst, src, count * 2);
+}
+
+void SetRow_C(uint8* dst, uint32 v8, int count) {
+#ifdef _MSC_VER
+  // VC will generate rep stosb.
+  int x;
+  for (x = 0; x < count; ++x) {
+    dst[x] = v8;
+  }
+#else
+  memset(dst, v8, count);
+#endif
+}
+
+void ARGBSetRows_C(uint8* dst, uint32 v32, int width,
+                 int dst_stride, int height) {
+  int y;
+  for (y = 0; y < height; ++y) {
+    uint32* d = (uint32*)(dst);
+    int x;
+    for (x = 0; x < width; ++x) {
+      d[x] = v32;
+    }
+    dst += dst_stride;
+  }
+}
+
+// Filter 2 rows of YUY2 UV's (422) into U and V (420).
+void YUY2ToUVRow_C(const uint8* src_yuy2, int src_stride_yuy2,
+                   uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values, filtering 2 rows of YUY2.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = (src_yuy2[1] + src_yuy2[src_stride_yuy2 + 1] + 1) >> 1;
+    dst_v[0] = (src_yuy2[3] + src_yuy2[src_stride_yuy2 + 3] + 1) >> 1;
+    src_yuy2 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of YUY2 UV's (422) into U and V (422).
+void YUY2ToUV422Row_C(const uint8* src_yuy2,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = src_yuy2[1];
+    dst_v[0] = src_yuy2[3];
+    src_yuy2 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of YUY2 Y's (422) into Y (420/422).
+void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_y[x] = src_yuy2[0];
+    dst_y[x + 1] = src_yuy2[2];
+    src_yuy2 += 4;
+  }
+  if (width & 1) {
+    dst_y[width - 1] = src_yuy2[0];
+  }
+}
+
+// Filter 2 rows of UYVY UV's (422) into U and V (420).
+void UYVYToUVRow_C(const uint8* src_uyvy, int src_stride_uyvy,
+                   uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = (src_uyvy[0] + src_uyvy[src_stride_uyvy + 0] + 1) >> 1;
+    dst_v[0] = (src_uyvy[2] + src_uyvy[src_stride_uyvy + 2] + 1) >> 1;
+    src_uyvy += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of UYVY UV's (422) into U and V (422).
+void UYVYToUV422Row_C(const uint8* src_uyvy,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = src_uyvy[0];
+    dst_v[0] = src_uyvy[2];
+    src_uyvy += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of UYVY Y's (422) into Y (420/422).
+void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_y[x] = src_uyvy[1];
+    dst_y[x + 1] = src_uyvy[3];
+    src_uyvy += 4;
+  }
+  if (width & 1) {
+    dst_y[width - 1] = src_uyvy[1];
+  }
+}
+
+#define BLEND(f, b, a) (((256 - a) * b) >> 8) + f
+
+// Blend src_argb0 over src_argb1 and store to dst_argb.
+// dst_argb may be src_argb0 or src_argb1.
+// This code mimics the SSSE3 version for better testability.
+void ARGBBlendRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                    uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint32 fb = src_argb0[0];
+    uint32 fg = src_argb0[1];
+    uint32 fr = src_argb0[2];
+    uint32 a = src_argb0[3];
+    uint32 bb = src_argb1[0];
+    uint32 bg = src_argb1[1];
+    uint32 br = src_argb1[2];
+    dst_argb[0] = BLEND(fb, bb, a);
+    dst_argb[1] = BLEND(fg, bg, a);
+    dst_argb[2] = BLEND(fr, br, a);
+    dst_argb[3] = 255u;
+
+    fb = src_argb0[4 + 0];
+    fg = src_argb0[4 + 1];
+    fr = src_argb0[4 + 2];
+    a = src_argb0[4 + 3];
+    bb = src_argb1[4 + 0];
+    bg = src_argb1[4 + 1];
+    br = src_argb1[4 + 2];
+    dst_argb[4 + 0] = BLEND(fb, bb, a);
+    dst_argb[4 + 1] = BLEND(fg, bg, a);
+    dst_argb[4 + 2] = BLEND(fr, br, a);
+    dst_argb[4 + 3] = 255u;
+    src_argb0 += 8;
+    src_argb1 += 8;
+    dst_argb += 8;
+  }
+
+  if (width & 1) {
+    uint32 fb = src_argb0[0];
+    uint32 fg = src_argb0[1];
+    uint32 fr = src_argb0[2];
+    uint32 a = src_argb0[3];
+    uint32 bb = src_argb1[0];
+    uint32 bg = src_argb1[1];
+    uint32 br = src_argb1[2];
+    dst_argb[0] = BLEND(fb, bb, a);
+    dst_argb[1] = BLEND(fg, bg, a);
+    dst_argb[2] = BLEND(fr, br, a);
+    dst_argb[3] = 255u;
+  }
+}
+#undef BLEND
+#define ATTENUATE(f, a) (a | (a << 8)) * (f | (f << 8)) >> 24
+
+// Multiply source RGB by alpha and store to destination.
+// This code mimics the SSSE3 version for better testability.
+void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    uint32 b = src_argb[0];
+    uint32 g = src_argb[1];
+    uint32 r = src_argb[2];
+    uint32 a = src_argb[3];
+    dst_argb[0] = ATTENUATE(b, a);
+    dst_argb[1] = ATTENUATE(g, a);
+    dst_argb[2] = ATTENUATE(r, a);
+    dst_argb[3] = a;
+    b = src_argb[4];
+    g = src_argb[5];
+    r = src_argb[6];
+    a = src_argb[7];
+    dst_argb[4] = ATTENUATE(b, a);
+    dst_argb[5] = ATTENUATE(g, a);
+    dst_argb[6] = ATTENUATE(r, a);
+    dst_argb[7] = a;
+    src_argb += 8;
+    dst_argb += 8;
+  }
+
+  if (width & 1) {
+    const uint32 b = src_argb[0];
+    const uint32 g = src_argb[1];
+    const uint32 r = src_argb[2];
+    const uint32 a = src_argb[3];
+    dst_argb[0] = ATTENUATE(b, a);
+    dst_argb[1] = ATTENUATE(g, a);
+    dst_argb[2] = ATTENUATE(r, a);
+    dst_argb[3] = a;
+  }
+}
+#undef ATTENUATE
+
+// Divide source RGB by alpha and store to destination.
+// b = (b * 255 + (a / 2)) / a;
+// g = (g * 255 + (a / 2)) / a;
+// r = (r * 255 + (a / 2)) / a;
+// Reciprocal method is off by 1 on some values. ie 125
+// 8.8 fixed point inverse table with 1.0 in upper short and 1 / a in lower.
+#define T(a) 0x01000000 + (0x10000 / a)
+const uint32 fixed_invtbl8[256] = {
+  0x01000000, 0x0100ffff, T(0x02), T(0x03), T(0x04), T(0x05), T(0x06), T(0x07),
+  T(0x08), T(0x09), T(0x0a), T(0x0b), T(0x0c), T(0x0d), T(0x0e), T(0x0f),
+  T(0x10), T(0x11), T(0x12), T(0x13), T(0x14), T(0x15), T(0x16), T(0x17),
+  T(0x18), T(0x19), T(0x1a), T(0x1b), T(0x1c), T(0x1d), T(0x1e), T(0x1f),
+  T(0x20), T(0x21), T(0x22), T(0x23), T(0x24), T(0x25), T(0x26), T(0x27),
+  T(0x28), T(0x29), T(0x2a), T(0x2b), T(0x2c), T(0x2d), T(0x2e), T(0x2f),
+  T(0x30), T(0x31), T(0x32), T(0x33), T(0x34), T(0x35), T(0x36), T(0x37),
+  T(0x38), T(0x39), T(0x3a), T(0x3b), T(0x3c), T(0x3d), T(0x3e), T(0x3f),
+  T(0x40), T(0x41), T(0x42), T(0x43), T(0x44), T(0x45), T(0x46), T(0x47),
+  T(0x48), T(0x49), T(0x4a), T(0x4b), T(0x4c), T(0x4d), T(0x4e), T(0x4f),
+  T(0x50), T(0x51), T(0x52), T(0x53), T(0x54), T(0x55), T(0x56), T(0x57),
+  T(0x58), T(0x59), T(0x5a), T(0x5b), T(0x5c), T(0x5d), T(0x5e), T(0x5f),
+  T(0x60), T(0x61), T(0x62), T(0x63), T(0x64), T(0x65), T(0x66), T(0x67),
+  T(0x68), T(0x69), T(0x6a), T(0x6b), T(0x6c), T(0x6d), T(0x6e), T(0x6f),
+  T(0x70), T(0x71), T(0x72), T(0x73), T(0x74), T(0x75), T(0x76), T(0x77),
+  T(0x78), T(0x79), T(0x7a), T(0x7b), T(0x7c), T(0x7d), T(0x7e), T(0x7f),
+  T(0x80), T(0x81), T(0x82), T(0x83), T(0x84), T(0x85), T(0x86), T(0x87),
+  T(0x88), T(0x89), T(0x8a), T(0x8b), T(0x8c), T(0x8d), T(0x8e), T(0x8f),
+  T(0x90), T(0x91), T(0x92), T(0x93), T(0x94), T(0x95), T(0x96), T(0x97),
+  T(0x98), T(0x99), T(0x9a), T(0x9b), T(0x9c), T(0x9d), T(0x9e), T(0x9f),
+  T(0xa0), T(0xa1), T(0xa2), T(0xa3), T(0xa4), T(0xa5), T(0xa6), T(0xa7),
+  T(0xa8), T(0xa9), T(0xaa), T(0xab), T(0xac), T(0xad), T(0xae), T(0xaf),
+  T(0xb0), T(0xb1), T(0xb2), T(0xb3), T(0xb4), T(0xb5), T(0xb6), T(0xb7),
+  T(0xb8), T(0xb9), T(0xba), T(0xbb), T(0xbc), T(0xbd), T(0xbe), T(0xbf),
+  T(0xc0), T(0xc1), T(0xc2), T(0xc3), T(0xc4), T(0xc5), T(0xc6), T(0xc7),
+  T(0xc8), T(0xc9), T(0xca), T(0xcb), T(0xcc), T(0xcd), T(0xce), T(0xcf),
+  T(0xd0), T(0xd1), T(0xd2), T(0xd3), T(0xd4), T(0xd5), T(0xd6), T(0xd7),
+  T(0xd8), T(0xd9), T(0xda), T(0xdb), T(0xdc), T(0xdd), T(0xde), T(0xdf),
+  T(0xe0), T(0xe1), T(0xe2), T(0xe3), T(0xe4), T(0xe5), T(0xe6), T(0xe7),
+  T(0xe8), T(0xe9), T(0xea), T(0xeb), T(0xec), T(0xed), T(0xee), T(0xef),
+  T(0xf0), T(0xf1), T(0xf2), T(0xf3), T(0xf4), T(0xf5), T(0xf6), T(0xf7),
+  T(0xf8), T(0xf9), T(0xfa), T(0xfb), T(0xfc), T(0xfd), T(0xfe), 0x01000100 };
+#undef T
+
+void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    uint32 b = src_argb[0];
+    uint32 g = src_argb[1];
+    uint32 r = src_argb[2];
+    const uint32 a = src_argb[3];
+    const uint32 ia = fixed_invtbl8[a] & 0xffff;  // 8.8 fixed point
+    b = (b * ia) >> 8;
+    g = (g * ia) >> 8;
+    r = (r * ia) >> 8;
+    // Clamping should not be necessary but is free in assembly.
+    dst_argb[0] = clamp255(b);
+    dst_argb[1] = clamp255(g);
+    dst_argb[2] = clamp255(r);
+    dst_argb[3] = a;
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
+                               const int32* previous_cumsum, int width) {
+  int32 row_sum[4] = {0, 0, 0, 0};
+  int x;
+  for (x = 0; x < width; ++x) {
+    row_sum[0] += row[x * 4 + 0];
+    row_sum[1] += row[x * 4 + 1];
+    row_sum[2] += row[x * 4 + 2];
+    row_sum[3] += row[x * 4 + 3];
+    cumsum[x * 4 + 0] = row_sum[0]  + previous_cumsum[x * 4 + 0];
+    cumsum[x * 4 + 1] = row_sum[1]  + previous_cumsum[x * 4 + 1];
+    cumsum[x * 4 + 2] = row_sum[2]  + previous_cumsum[x * 4 + 2];
+    cumsum[x * 4 + 3] = row_sum[3]  + previous_cumsum[x * 4 + 3];
+  }
+}
+
+void CumulativeSumToAverageRow_C(const int32* tl, const int32* bl,
+                                int w, int area, uint8* dst, int count) {
+  float ooa = 1.0f / area;
+  int i;
+  for (i = 0; i < count; ++i) {
+    dst[0] = (uint8)((bl[w + 0] + tl[0] - bl[0] - tl[w + 0]) * ooa);
+    dst[1] = (uint8)((bl[w + 1] + tl[1] - bl[1] - tl[w + 1]) * ooa);
+    dst[2] = (uint8)((bl[w + 2] + tl[2] - bl[2] - tl[w + 2]) * ooa);
+    dst[3] = (uint8)((bl[w + 3] + tl[3] - bl[3] - tl[w + 3]) * ooa);
+    dst += 4;
+    tl += 4;
+    bl += 4;
+  }
+}
+
+// Copy pixels from rotated source to destination row with a slope.
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width) {
+  int i;
+  // Render a row of pixels from source into a buffer.
+  float uv[2];
+  uv[0] = uv_dudv[0];
+  uv[1] = uv_dudv[1];
+  for (i = 0; i < width; ++i) {
+    int x = (int)(uv[0]);
+    int y = (int)(uv[1]);
+    *(uint32*)(dst_argb) =
+        *(const uint32*)(src_argb + y * src_argb_stride +
+                                         x * 4);
+    dst_argb += 4;
+    uv[0] += uv_dudv[2];
+    uv[1] += uv_dudv[3];
+  }
+}
+
+// Blend 2 rows into 1 for conversions such as I422ToI420.
+void HalfRow_C(const uint8* src_uv, int src_uv_stride,
+               uint8* dst_uv, int pix) {
+  int x;
+  for (x = 0; x < pix; ++x) {
+    dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
+  }
+}
+
+void HalfRow_16_C(const uint16* src_uv, int src_uv_stride,
+                  uint16* dst_uv, int pix) {
+  int x;
+  for (x = 0; x < pix; ++x) {
+    dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
+  }
+}
+
+// C version 2x2 -> 2x1.
+void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+                      ptrdiff_t src_stride,
+                      int width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  int x;
+  if (source_y_fraction == 0) {
+    memcpy(dst_ptr, src_ptr, width);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_C(src_ptr, (int)(src_stride), dst_ptr, width);
+    return;
+  }
+  for (x = 0; x < width - 1; x += 2) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
+    src_ptr += 2;
+    src_ptr1 += 2;
+    dst_ptr += 2;
+  }
+  if (width & 1) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+  }
+}
+
+void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         ptrdiff_t src_stride,
+                         int width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16* src_ptr1 = src_ptr + src_stride;
+  int x;
+  if (source_y_fraction == 0) {
+    memcpy(dst_ptr, src_ptr, width * 2);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_16_C(src_ptr, (int)(src_stride), dst_ptr, width);
+    return;
+  }
+  for (x = 0; x < width - 1; x += 2) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
+    src_ptr += 2;
+    src_ptr1 += 2;
+    dst_ptr += 2;
+  }
+  if (width & 1) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+  }
+}
+
+// Select 2 channels from ARGB on alternating pixels.  e.g.  BGBGBGBG
+void ARGBToBayerRow_C(const uint8* src_argb,
+                      uint8* dst_bayer, uint32 selector, int pix) {
+  int index0 = selector & 0xff;
+  int index1 = (selector >> 8) & 0xff;
+  // Copy a row of Bayer.
+  int x;
+  for (x = 0; x < pix - 1; x += 2) {
+    dst_bayer[0] = src_argb[index0];
+    dst_bayer[1] = src_argb[index1];
+    src_argb += 8;
+    dst_bayer += 2;
+  }
+  if (pix & 1) {
+    dst_bayer[0] = src_argb[index0];
+  }
+}
+
+// Select G channel from ARGB.  e.g.  GGGGGGGG
+void ARGBToBayerGGRow_C(const uint8* src_argb,
+                        uint8* dst_bayer, uint32 selector, int pix) {
+  // Copy a row of G.
+  int x;
+  for (x = 0; x < pix - 1; x += 2) {
+    dst_bayer[0] = src_argb[1];
+    dst_bayer[1] = src_argb[5];
+    src_argb += 8;
+    dst_bayer += 2;
+  }
+  if (pix & 1) {
+    dst_bayer[0] = src_argb[1];
+  }
+}
+
+// Use first 4 shuffler values to reorder ARGB channels.
+void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
+                      const uint8* shuffler, int pix) {
+  int index0 = shuffler[0];
+  int index1 = shuffler[1];
+  int index2 = shuffler[2];
+  int index3 = shuffler[3];
+  // Shuffle a row of ARGB.
+  int x;
+  for (x = 0; x < pix; ++x) {
+    // To support in-place conversion.
+    uint8 b = src_argb[index0];
+    uint8 g = src_argb[index1];
+    uint8 r = src_argb[index2];
+    uint8 a = src_argb[index3];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = a;
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void I422ToYUY2Row_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_frame, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_frame[0] = src_y[0];
+    dst_frame[1] = src_u[0];
+    dst_frame[2] = src_y[1];
+    dst_frame[3] = src_v[0];
+    dst_frame += 4;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+  }
+  if (width & 1) {
+    dst_frame[0] = src_y[0];
+    dst_frame[1] = src_u[0];
+    dst_frame[2] = src_y[0];  // duplicate last y
+    dst_frame[3] = src_v[0];
+  }
+}
+
+void I422ToUYVYRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_frame, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_frame[0] = src_u[0];
+    dst_frame[1] = src_y[0];
+    dst_frame[2] = src_v[0];
+    dst_frame[3] = src_y[1];
+    dst_frame += 4;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+  }
+  if (width & 1) {
+    dst_frame[0] = src_u[0];
+    dst_frame[1] = src_y[0];
+    dst_frame[2] = src_v[0];
+    dst_frame[3] = src_y[0];  // duplicate last y
+  }
+}
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(HAS_I422TOARGBROW_SSSE3)
+// row_win.cc has asm version, but GCC uses 2 step wrapper.
+#if !defined(_MSC_VER) && (defined(__x86_64__) || defined(__i386__))
+void I422ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* rgb_buf,
+                           int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
+  ARGBToRGB565Row_SSE2(row, rgb_buf, width);
+  free_aligned_buffer_64(row);
+}
+#endif  // !defined(_MSC_VER) && (defined(__x86_64__) || defined(__i386__))
+
+#if defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
+void I422ToARGB1555Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* rgb_buf,
+                             int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
+  ARGBToARGB1555Row_SSE2(row, rgb_buf, width);
+  free_aligned_buffer_64(row);
+}
+
+void I422ToARGB4444Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* rgb_buf,
+                             int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
+  ARGBToARGB4444Row_SSE2(row, rgb_buf, width);
+  free_aligned_buffer_64(row);
+}
+
+void NV12ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_uv,
+                           uint8* dst_rgb565,
+                           int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  NV12ToARGBRow_SSSE3(src_y, src_uv, row, width);
+  ARGBToRGB565Row_SSE2(row, dst_rgb565, width);
+  free_aligned_buffer_64(row);
+}
+
+void NV21ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_vu,
+                           uint8* dst_rgb565,
+                           int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  NV21ToARGBRow_SSSE3(src_y, src_vu, row, width);
+  ARGBToRGB565Row_SSE2(row, dst_rgb565, width);
+  free_aligned_buffer_64(row);
+}
+
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
+                         uint8* dst_argb,
+                         int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  YUY2ToUV422Row_SSE2(src_yuy2, row_u, row_v, width);
+  YUY2ToYRow_SSE2(src_yuy2, row_y, width);
+  I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+void YUY2ToARGBRow_Unaligned_SSSE3(const uint8* src_yuy2,
+                                   uint8* dst_argb,
+                                   int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  YUY2ToUV422Row_Unaligned_SSE2(src_yuy2, row_u, row_v, width);
+  YUY2ToYRow_Unaligned_SSE2(src_yuy2, row_y, width);
+  I422ToARGBRow_Unaligned_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
+                         uint8* dst_argb,
+                         int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  UYVYToUV422Row_SSE2(src_uyvy, row_u, row_v, width);
+  UYVYToYRow_SSE2(src_uyvy, row_y, width);
+  I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+void UYVYToARGBRow_Unaligned_SSSE3(const uint8* src_uyvy,
+                                   uint8* dst_argb,
+                                   int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  UYVYToUV422Row_Unaligned_SSE2(src_uyvy, row_u, row_v, width);
+  UYVYToYRow_Unaligned_SSE2(src_uyvy, row_y, width);
+  I422ToARGBRow_Unaligned_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+#endif  // defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
+#endif  // !defined(LIBYUV_DISABLE_X86)
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    float b = (float)(src_argb[0]);
+    float g = (float)(src_argb[1]);
+    float r = (float)(src_argb[2]);
+    float a = (float)(src_argb[3]);
+    float b2 = b * b;
+    float g2 = g * g;
+    float r2 = r * r;
+    float a2 = a * a;
+    float db = poly[0] + poly[4] * b;
+    float dg = poly[1] + poly[5] * g;
+    float dr = poly[2] + poly[6] * r;
+    float da = poly[3] + poly[7] * a;
+    float b3 = b2 * b;
+    float g3 = g2 * g;
+    float r3 = r2 * r;
+    float a3 = a2 * a;
+    db += poly[8] * b2;
+    dg += poly[9] * g2;
+    dr += poly[10] * r2;
+    da += poly[11] * a2;
+    db += poly[12] * b3;
+    dg += poly[13] * g3;
+    dr += poly[14] * r3;
+    da += poly[15] * a3;
+
+    dst_argb[0] = Clamp((int32)(db));
+    dst_argb[1] = Clamp((int32)(dg));
+    dst_argb[2] = Clamp((int32)(dr));
+    dst_argb[3] = Clamp((int32)(da));
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                             const uint8* luma, uint32 lumacoeff) {
+  uint32 bc = lumacoeff & 0xff;
+  uint32 gc = (lumacoeff >> 8) & 0xff;
+  uint32 rc = (lumacoeff >> 16) & 0xff;
+
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
+                           src_argb[2] * rc) & 0x7F00u) + luma;
+    const uint8* luma1;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+    luma1 = ((src_argb[4] * bc + src_argb[5] * gc +
+              src_argb[6] * rc) & 0x7F00u) + luma;
+    dst_argb[4] = luma1[src_argb[4]];
+    dst_argb[5] = luma1[src_argb[5]];
+    dst_argb[6] = luma1[src_argb[6]];
+    dst_argb[7] = src_argb[7];
+    src_argb += 8;
+    dst_argb += 8;
+  }
+  if (width & 1) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
+                           src_argb[2] * rc) & 0x7F00u) + luma;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+  }
+}
+
+void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst[3] = src[3];
+    dst[7] = src[7];
+    dst += 8;
+    src += 8;
+  }
+  if (width & 1) {
+    dst[3] = src[3];
+  }
+}
+
+void ARGBCopyYToAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst[3] = src[0];
+    dst[7] = src[1];
+    dst += 8;
+    src += 2;
+  }
+  if (width & 1) {
+    dst[3] = src[0];
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/row_mips.cc b/third_party/libyuv/source/row_mips.cc
new file mode 100644
index 0000000..a804670
--- /dev/null
+++ b/third_party/libyuv/source/row_mips.cc
@@ -0,0 +1,991 @@
+/*
+ *  Copyright (c) 2012 The LibYuv project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__)
+
+#ifdef HAS_COPYROW_MIPS
+void CopyRow_MIPS(const uint8* src, uint8* dst, int count) {
+  __asm__ __volatile__ (
+    ".set      noreorder                         \n"
+    ".set      noat                              \n"
+    "slti      $at, %[count], 8                  \n"
+    "bne       $at ,$zero, $last8                \n"
+    "xor       $t8, %[src], %[dst]               \n"
+    "andi      $t8, $t8, 0x3                     \n"
+
+    "bne       $t8, $zero, unaligned             \n"
+    "negu      $a3, %[dst]                       \n"
+    // make dst/src aligned
+    "andi      $a3, $a3, 0x3                     \n"
+    "beq       $a3, $zero, $chk16w               \n"
+    // word-aligned now count is the remining bytes count
+    "subu     %[count], %[count], $a3            \n"
+
+    "lwr       $t8, 0(%[src])                    \n"
+    "addu      %[src], %[src], $a3               \n"
+    "swr       $t8, 0(%[dst])                    \n"
+    "addu      %[dst], %[dst], $a3               \n"
+
+    // Now the dst/src are mutually word-aligned with word-aligned addresses
+    "$chk16w:                                    \n"
+    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
+    // t8 is the byte count after 64-byte chunks
+    "beq       %[count], $t8, chk8w              \n"
+    // There will be at most 1 32-byte chunk after it
+    "subu      $a3, %[count], $t8                \n"  // the reminder
+    // Here a3 counts bytes in 16w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // Now a3 is the final dst after 64-byte chunks
+    "addu      $t0, %[dst], %[count]             \n"
+    // t0 is the "past the end" address
+
+    // When in the loop we exercise "pref 30,x(a1)", the a1+x should not be past
+    // the "t0-32" address
+    // This means: for x=128 the last "safe" a1 address is "t0-160"
+    // Alternatively, for x=64 the last "safe" a1 address is "t0-96"
+    // we will use "pref 30,128(a1)", so "t0-160" is the limit
+    "subu      $t9, $t0, 160                     \n"
+    // t9 is the "last safe pref 30,128(a1)" address
+    "pref      0, 0(%[src])                      \n"  // first line of src
+    "pref      0, 32(%[src])                     \n"  // second line of src
+    "pref      0, 64(%[src])                     \n"
+    "pref      30, 32(%[dst])                    \n"
+    // In case the a1 > t9 don't use "pref 30" at all
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bgtz      $v1, $loop16w                     \n"
+    "nop                                         \n"
+    // otherwise, start with using pref30
+    "pref      30, 64(%[dst])                    \n"
+    "$loop16w:                                    \n"
+    "pref      0, 96(%[src])                     \n"
+    "lw        $t0, 0(%[src])                    \n"
+    "bgtz      $v1, $skip_pref30_96              \n"  // skip
+    "lw        $t1, 4(%[src])                    \n"
+    "pref      30, 96(%[dst])                    \n"  // continue
+    "$skip_pref30_96:                            \n"
+    "lw        $t2, 8(%[src])                    \n"
+    "lw        $t3, 12(%[src])                   \n"
+    "lw        $t4, 16(%[src])                   \n"
+    "lw        $t5, 20(%[src])                   \n"
+    "lw        $t6, 24(%[src])                   \n"
+    "lw        $t7, 28(%[src])                   \n"
+    "pref      0, 128(%[src])                    \n"
+    //  bring the next lines of src, addr 128
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "lw        $t0, 32(%[src])                   \n"
+    "bgtz      $v1, $skip_pref30_128             \n"  // skip pref 30,128(a1)
+    "lw        $t1, 36(%[src])                   \n"
+    "pref      30, 128(%[dst])                   \n"  // set dest, addr 128
+    "$skip_pref30_128:                           \n"
+    "lw        $t2, 40(%[src])                   \n"
+    "lw        $t3, 44(%[src])                   \n"
+    "lw        $t4, 48(%[src])                   \n"
+    "lw        $t5, 52(%[src])                   \n"
+    "lw        $t6, 56(%[src])                   \n"
+    "lw        $t7, 60(%[src])                   \n"
+    "pref      0, 160(%[src])                    \n"
+    // bring the next lines of src, addr 160
+    "sw        $t0, 32(%[dst])                   \n"
+    "sw        $t1, 36(%[dst])                   \n"
+    "sw        $t2, 40(%[dst])                   \n"
+    "sw        $t3, 44(%[dst])                   \n"
+    "sw        $t4, 48(%[dst])                   \n"
+    "sw        $t5, 52(%[dst])                   \n"
+    "sw        $t6, 56(%[dst])                   \n"
+    "sw        $t7, 60(%[dst])                   \n"
+
+    "addiu     %[dst], %[dst], 64                \n"  // adding 64 to dest
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bne       %[dst], $a3, $loop16w             \n"
+    " addiu    %[src], %[src], 64                \n"  // adding 64 to src
+    "move      %[count], $t8                     \n"
+
+    // Here we have src and dest word-aligned but less than 64-bytes to go
+
+    "chk8w:                                      \n"
+    "pref      0, 0x0(%[src])                    \n"
+    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
+    // the t8 is the reminder count past 32-bytes
+    "beq       %[count], $t8, chk1w              \n"
+    // count=t8,no 32-byte chunk
+    " nop                                        \n"
+
+    "lw        $t0, 0(%[src])                    \n"
+    "lw        $t1, 4(%[src])                    \n"
+    "lw        $t2, 8(%[src])                    \n"
+    "lw        $t3, 12(%[src])                   \n"
+    "lw        $t4, 16(%[src])                   \n"
+    "lw        $t5, 20(%[src])                   \n"
+    "lw        $t6, 24(%[src])                   \n"
+    "lw        $t7, 28(%[src])                   \n"
+    "addiu     %[src], %[src], 32                \n"
+
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "addiu     %[dst], %[dst], 32                \n"
+
+    "chk1w:                                      \n"
+    "andi      %[count], $t8, 0x3                \n"
+    // now count is the reminder past 1w chunks
+    "beq       %[count], $t8, $last8             \n"
+    " subu     $a3, $t8, %[count]                \n"
+    // a3 is count of bytes in 1w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // now a3 is the dst address past the 1w chunks
+    // copying in words (4-byte chunks)
+    "$wordCopy_loop:                             \n"
+    "lw        $t3, 0(%[src])                    \n"
+    // the first t3 may be equal t0 ... optimize?
+    "addiu     %[src], %[src],4                  \n"
+    "addiu     %[dst], %[dst],4                  \n"
+    "bne       %[dst], $a3,$wordCopy_loop        \n"
+    " sw       $t3, -4(%[dst])                   \n"
+
+    // For the last (<8) bytes
+    "$last8:                                     \n"
+    "blez      %[count], leave                   \n"
+    " addu     $a3, %[dst], %[count]             \n"  // a3 -last dst address
+    "$last8loop:                                 \n"
+    "lb        $v1, 0(%[src])                    \n"
+    "addiu     %[src], %[src], 1                 \n"
+    "addiu     %[dst], %[dst], 1                 \n"
+    "bne       %[dst], $a3, $last8loop           \n"
+    " sb       $v1, -1(%[dst])                   \n"
+
+    "leave:                                      \n"
+    "  j       $ra                               \n"
+    "  nop                                       \n"
+
+    //
+    // UNALIGNED case
+    //
+
+    "unaligned:                                  \n"
+    // got here with a3="negu a1"
+    "andi      $a3, $a3, 0x3                     \n"  // a1 is word aligned?
+    "beqz      $a3, $ua_chk16w                   \n"
+    " subu     %[count], %[count], $a3           \n"
+    // bytes left after initial a3 bytes
+    "lwr       $v1, 0(%[src])                    \n"
+    "lwl       $v1, 3(%[src])                    \n"
+    "addu      %[src], %[src], $a3               \n"  // a3 may be 1, 2 or 3
+    "swr       $v1, 0(%[dst])                    \n"
+    "addu      %[dst], %[dst], $a3               \n"
+    // below the dst will be word aligned (NOTE1)
+    "$ua_chk16w:                                 \n"
+    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
+    // t8 is the byte count after 64-byte chunks
+    "beq       %[count], $t8, ua_chk8w           \n"
+    // if a2==t8, no 64-byte chunks
+    // There will be at most 1 32-byte chunk after it
+    "subu      $a3, %[count], $t8                \n"  // the reminder
+    // Here a3 counts bytes in 16w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // Now a3 is the final dst after 64-byte chunks
+    "addu      $t0, %[dst], %[count]             \n"  // t0 "past the end"
+    "subu      $t9, $t0, 160                     \n"
+    // t9 is the "last safe pref 30,128(a1)" address
+    "pref      0, 0(%[src])                      \n"  // first line of src
+    "pref      0, 32(%[src])                     \n"  // second line  addr 32
+    "pref      0, 64(%[src])                     \n"
+    "pref      30, 32(%[dst])                    \n"
+    // safe, as we have at least 64 bytes ahead
+    // In case the a1 > t9 don't use "pref 30" at all
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bgtz      $v1, $ua_loop16w                  \n"
+    // skip "pref 30,64(a1)" for too short arrays
+    " nop                                        \n"
+    // otherwise, start with using pref30
+    "pref      30, 64(%[dst])                    \n"
+    "$ua_loop16w:                                \n"
+    "pref      0, 96(%[src])                     \n"
+    "lwr       $t0, 0(%[src])                    \n"
+    "lwl       $t0, 3(%[src])                    \n"
+    "lwr       $t1, 4(%[src])                    \n"
+    "bgtz      $v1, $ua_skip_pref30_96           \n"
+    " lwl      $t1, 7(%[src])                    \n"
+    "pref      30, 96(%[dst])                    \n"
+    // continue setting up the dest, addr 96
+    "$ua_skip_pref30_96:                         \n"
+    "lwr       $t2, 8(%[src])                    \n"
+    "lwl       $t2, 11(%[src])                   \n"
+    "lwr       $t3, 12(%[src])                   \n"
+    "lwl       $t3, 15(%[src])                   \n"
+    "lwr       $t4, 16(%[src])                   \n"
+    "lwl       $t4, 19(%[src])                   \n"
+    "lwr       $t5, 20(%[src])                   \n"
+    "lwl       $t5, 23(%[src])                   \n"
+    "lwr       $t6, 24(%[src])                   \n"
+    "lwl       $t6, 27(%[src])                   \n"
+    "lwr       $t7, 28(%[src])                   \n"
+    "lwl       $t7, 31(%[src])                   \n"
+    "pref      0, 128(%[src])                    \n"
+    // bring the next lines of src, addr 128
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "lwr       $t0, 32(%[src])                   \n"
+    "lwl       $t0, 35(%[src])                   \n"
+    "lwr       $t1, 36(%[src])                   \n"
+    "bgtz      $v1, ua_skip_pref30_128           \n"
+    " lwl      $t1, 39(%[src])                   \n"
+    "pref      30, 128(%[dst])                   \n"
+    // continue setting up the dest, addr 128
+    "ua_skip_pref30_128:                         \n"
+
+    "lwr       $t2, 40(%[src])                   \n"
+    "lwl       $t2, 43(%[src])                   \n"
+    "lwr       $t3, 44(%[src])                   \n"
+    "lwl       $t3, 47(%[src])                   \n"
+    "lwr       $t4, 48(%[src])                   \n"
+    "lwl       $t4, 51(%[src])                   \n"
+    "lwr       $t5, 52(%[src])                   \n"
+    "lwl       $t5, 55(%[src])                   \n"
+    "lwr       $t6, 56(%[src])                   \n"
+    "lwl       $t6, 59(%[src])                   \n"
+    "lwr       $t7, 60(%[src])                   \n"
+    "lwl       $t7, 63(%[src])                   \n"
+    "pref      0, 160(%[src])                    \n"
+    // bring the next lines of src, addr 160
+    "sw        $t0, 32(%[dst])                   \n"
+    "sw        $t1, 36(%[dst])                   \n"
+    "sw        $t2, 40(%[dst])                   \n"
+    "sw        $t3, 44(%[dst])                   \n"
+    "sw        $t4, 48(%[dst])                   \n"
+    "sw        $t5, 52(%[dst])                   \n"
+    "sw        $t6, 56(%[dst])                   \n"
+    "sw        $t7, 60(%[dst])                   \n"
+
+    "addiu     %[dst],%[dst],64                  \n"  // adding 64 to dest
+    "sgtu      $v1,%[dst],$t9                    \n"
+    "bne       %[dst],$a3,$ua_loop16w            \n"
+    " addiu    %[src],%[src],64                  \n"  // adding 64 to src
+    "move      %[count],$t8                      \n"
+
+    // Here we have src and dest word-aligned but less than 64-bytes to go
+
+    "ua_chk8w:                                   \n"
+    "pref      0, 0x0(%[src])                    \n"
+    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
+    // the t8 is the reminder count
+    "beq       %[count], $t8, $ua_chk1w          \n"
+    // when count==t8, no 32-byte chunk
+
+    "lwr       $t0, 0(%[src])                    \n"
+    "lwl       $t0, 3(%[src])                    \n"
+    "lwr       $t1, 4(%[src])                    \n"
+    "lwl       $t1, 7(%[src])                    \n"
+    "lwr       $t2, 8(%[src])                    \n"
+    "lwl       $t2, 11(%[src])                   \n"
+    "lwr       $t3, 12(%[src])                   \n"
+    "lwl       $t3, 15(%[src])                   \n"
+    "lwr       $t4, 16(%[src])                   \n"
+    "lwl       $t4, 19(%[src])                   \n"
+    "lwr       $t5, 20(%[src])                   \n"
+    "lwl       $t5, 23(%[src])                   \n"
+    "lwr       $t6, 24(%[src])                   \n"
+    "lwl       $t6, 27(%[src])                   \n"
+    "lwr       $t7, 28(%[src])                   \n"
+    "lwl       $t7, 31(%[src])                   \n"
+    "addiu     %[src], %[src], 32                \n"
+
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "addiu     %[dst], %[dst], 32                \n"
+
+    "$ua_chk1w:                                  \n"
+    "andi      %[count], $t8, 0x3                \n"
+    // now count is the reminder past 1w chunks
+    "beq       %[count], $t8, ua_smallCopy       \n"
+    "subu      $a3, $t8, %[count]                \n"
+    // a3 is count of bytes in 1w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // now a3 is the dst address past the 1w chunks
+
+    // copying in words (4-byte chunks)
+    "$ua_wordCopy_loop:                          \n"
+    "lwr       $v1, 0(%[src])                    \n"
+    "lwl       $v1, 3(%[src])                    \n"
+    "addiu     %[src], %[src], 4                 \n"
+    "addiu     %[dst], %[dst], 4                 \n"
+    // note: dst=a1 is word aligned here, see NOTE1
+    "bne       %[dst], $a3, $ua_wordCopy_loop    \n"
+    " sw       $v1,-4(%[dst])                    \n"
+
+    // Now less than 4 bytes (value in count) left to copy
+    "ua_smallCopy:                               \n"
+    "beqz      %[count], leave                   \n"
+    " addu     $a3, %[dst], %[count]             \n" // a3 = last dst address
+    "$ua_smallCopy_loop:                         \n"
+    "lb        $v1, 0(%[src])                    \n"
+    "addiu     %[src], %[src], 1                 \n"
+    "addiu     %[dst], %[dst], 1                 \n"
+    "bne       %[dst],$a3,$ua_smallCopy_loop     \n"
+    " sb       $v1, -1(%[dst])                   \n"
+
+    "j         $ra                               \n"
+    " nop                                        \n"
+    ".set      at                                \n"
+    ".set      reorder                           \n"
+       : [dst] "+r" (dst), [src] "+r" (src)
+       : [count] "r" (count)
+       : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7",
+       "t8", "t9", "a3", "v1", "at"
+  );
+}
+#endif  // HAS_COPYROW_MIPS
+
+// MIPS DSPR2 functions
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_dsp) && \
+    (__mips_dsp_rev >= 2)
+void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                           int width) {
+  __asm__ __volatile__ (
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+    "srl             $t4, %[width], 4              \n"  // multiplies of 16
+    "blez            $t4, 2f                       \n"
+    " andi           %[width], %[width], 0xf       \n"  // residual
+
+    ".p2align        2                             \n"
+  "1:                                              \n"
+    "addiu           $t4, $t4, -1                  \n"
+    "lw              $t0, 0(%[src_uv])             \n"  // V1 | U1 | V0 | U0
+    "lw              $t1, 4(%[src_uv])             \n"  // V3 | U3 | V2 | U2
+    "lw              $t2, 8(%[src_uv])             \n"  // V5 | U5 | V4 | U4
+    "lw              $t3, 12(%[src_uv])            \n"  // V7 | U7 | V6 | U6
+    "lw              $t5, 16(%[src_uv])            \n"  // V9 | U9 | V8 | U8
+    "lw              $t6, 20(%[src_uv])            \n"  // V11 | U11 | V10 | U10
+    "lw              $t7, 24(%[src_uv])            \n"  // V13 | U13 | V12 | U12
+    "lw              $t8, 28(%[src_uv])            \n"  // V15 | U15 | V14 | U14
+    "addiu           %[src_uv], %[src_uv], 32      \n"
+    "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
+    "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
+    "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
+    "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
+    "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
+    "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
+    "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 | V12
+    "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 | U12
+    "sw              $t9, 0(%[dst_v])              \n"
+    "sw              $t0, 0(%[dst_u])              \n"
+    "sw              $t1, 4(%[dst_v])              \n"
+    "sw              $t2, 4(%[dst_u])              \n"
+    "sw              $t3, 8(%[dst_v])              \n"
+    "sw              $t5, 8(%[dst_u])              \n"
+    "sw              $t6, 12(%[dst_v])             \n"
+    "sw              $t7, 12(%[dst_u])             \n"
+    "addiu           %[dst_v], %[dst_v], 16        \n"
+    "bgtz            $t4, 1b                       \n"
+    " addiu          %[dst_u], %[dst_u], 16        \n"
+
+    "beqz            %[width], 3f                  \n"
+    " nop                                          \n"
+
+  "2:                                              \n"
+    "lbu             $t0, 0(%[src_uv])             \n"
+    "lbu             $t1, 1(%[src_uv])             \n"
+    "addiu           %[src_uv], %[src_uv], 2       \n"
+    "addiu           %[width], %[width], -1        \n"
+    "sb              $t0, 0(%[dst_u])              \n"
+    "sb              $t1, 0(%[dst_v])              \n"
+    "addiu           %[dst_u], %[dst_u], 1         \n"
+    "bgtz            %[width], 2b                  \n"
+    " addiu          %[dst_v], %[dst_v], 1         \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+     : [src_uv] "+r" (src_uv),
+       [width] "+r" (width),
+       [dst_u] "+r" (dst_u),
+       [dst_v] "+r" (dst_v)
+     :
+     : "t0", "t1", "t2", "t3",
+     "t4", "t5", "t6", "t7", "t8", "t9"
+  );
+}
+
+void SplitUVRow_Unaligned_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u,
+                                     uint8* dst_v, int width) {
+  __asm__ __volatile__ (
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+    "srl             $t4, %[width], 4              \n"  // multiplies of 16
+    "blez            $t4, 2f                       \n"
+    " andi           %[width], %[width], 0xf       \n"  // residual
+
+    ".p2align        2                             \n"
+  "1:                                              \n"
+    "addiu           $t4, $t4, -1                  \n"
+    "lwr             $t0, 0(%[src_uv])             \n"
+    "lwl             $t0, 3(%[src_uv])             \n"  // V1 | U1 | V0 | U0
+    "lwr             $t1, 4(%[src_uv])             \n"
+    "lwl             $t1, 7(%[src_uv])             \n"  // V3 | U3 | V2 | U2
+    "lwr             $t2, 8(%[src_uv])             \n"
+    "lwl             $t2, 11(%[src_uv])            \n"  // V5 | U5 | V4 | U4
+    "lwr             $t3, 12(%[src_uv])            \n"
+    "lwl             $t3, 15(%[src_uv])            \n"  // V7 | U7 | V6 | U6
+    "lwr             $t5, 16(%[src_uv])            \n"
+    "lwl             $t5, 19(%[src_uv])            \n"  // V9 | U9 | V8 | U8
+    "lwr             $t6, 20(%[src_uv])            \n"
+    "lwl             $t6, 23(%[src_uv])            \n"  // V11 | U11 | V10 | U10
+    "lwr             $t7, 24(%[src_uv])            \n"
+    "lwl             $t7, 27(%[src_uv])            \n"  // V13 | U13 | V12 | U12
+    "lwr             $t8, 28(%[src_uv])            \n"
+    "lwl             $t8, 31(%[src_uv])            \n"  // V15 | U15 | V14 | U14
+    "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
+    "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
+    "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
+    "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
+    "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
+    "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
+    "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 | V12
+    "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 | U12
+    "addiu           %[src_uv], %[src_uv], 32      \n"
+    "swr             $t9, 0(%[dst_v])              \n"
+    "swl             $t9, 3(%[dst_v])              \n"
+    "swr             $t0, 0(%[dst_u])              \n"
+    "swl             $t0, 3(%[dst_u])              \n"
+    "swr             $t1, 4(%[dst_v])              \n"
+    "swl             $t1, 7(%[dst_v])              \n"
+    "swr             $t2, 4(%[dst_u])              \n"
+    "swl             $t2, 7(%[dst_u])              \n"
+    "swr             $t3, 8(%[dst_v])              \n"
+    "swl             $t3, 11(%[dst_v])             \n"
+    "swr             $t5, 8(%[dst_u])              \n"
+    "swl             $t5, 11(%[dst_u])             \n"
+    "swr             $t6, 12(%[dst_v])             \n"
+    "swl             $t6, 15(%[dst_v])             \n"
+    "swr             $t7, 12(%[dst_u])             \n"
+    "swl             $t7, 15(%[dst_u])             \n"
+    "addiu           %[dst_u], %[dst_u], 16        \n"
+    "bgtz            $t4, 1b                       \n"
+    " addiu          %[dst_v], %[dst_v], 16        \n"
+
+    "beqz            %[width], 3f                  \n"
+    " nop                                          \n"
+
+  "2:                                              \n"
+    "lbu             $t0, 0(%[src_uv])             \n"
+    "lbu             $t1, 1(%[src_uv])             \n"
+    "addiu           %[src_uv], %[src_uv], 2       \n"
+    "addiu           %[width], %[width], -1        \n"
+    "sb              $t0, 0(%[dst_u])              \n"
+    "sb              $t1, 0(%[dst_v])              \n"
+    "addiu           %[dst_u], %[dst_u], 1         \n"
+    "bgtz            %[width], 2b                  \n"
+    " addiu          %[dst_v], %[dst_v], 1         \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+     : [src_uv] "+r" (src_uv),
+       [width] "+r" (width),
+       [dst_u] "+r" (dst_u),
+       [dst_v] "+r" (dst_v)
+     :
+     : "t0", "t1", "t2", "t3",
+     "t4", "t5", "t6", "t7", "t8", "t9"
+  );
+}
+
+void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width) {
+  __asm__ __volatile__ (
+    ".set push                             \n"
+    ".set noreorder                        \n"
+
+    "srl       $t4, %[width], 4            \n"  // multiplies of 16
+    "andi      $t5, %[width], 0xf          \n"
+    "blez      $t4, 2f                     \n"
+    " addu     %[src], %[src], %[width]    \n"  // src += width
+
+    ".p2align  2                           \n"
+   "1:                                     \n"
+    "lw        $t0, -16(%[src])            \n"  // |3|2|1|0|
+    "lw        $t1, -12(%[src])            \n"  // |7|6|5|4|
+    "lw        $t2, -8(%[src])             \n"  // |11|10|9|8|
+    "lw        $t3, -4(%[src])             \n"  // |15|14|13|12|
+    "wsbh      $t0, $t0                    \n"  // |2|3|0|1|
+    "wsbh      $t1, $t1                    \n"  // |6|7|4|5|
+    "wsbh      $t2, $t2                    \n"  // |10|11|8|9|
+    "wsbh      $t3, $t3                    \n"  // |14|15|12|13|
+    "rotr      $t0, $t0, 16                \n"  // |0|1|2|3|
+    "rotr      $t1, $t1, 16                \n"  // |4|5|6|7|
+    "rotr      $t2, $t2, 16                \n"  // |8|9|10|11|
+    "rotr      $t3, $t3, 16                \n"  // |12|13|14|15|
+    "addiu     %[src], %[src], -16         \n"
+    "addiu     $t4, $t4, -1                \n"
+    "sw        $t3, 0(%[dst])              \n"  // |15|14|13|12|
+    "sw        $t2, 4(%[dst])              \n"  // |11|10|9|8|
+    "sw        $t1, 8(%[dst])              \n"  // |7|6|5|4|
+    "sw        $t0, 12(%[dst])             \n"  // |3|2|1|0|
+    "bgtz      $t4, 1b                     \n"
+    " addiu    %[dst], %[dst], 16          \n"
+    "beqz      $t5, 3f                     \n"
+    " nop                                  \n"
+
+   "2:                                     \n"
+    "lbu       $t0, -1(%[src])             \n"
+    "addiu     $t5, $t5, -1                \n"
+    "addiu     %[src], %[src], -1          \n"
+    "sb        $t0, 0(%[dst])              \n"
+    "bgez      $t5, 2b                     \n"
+    " addiu    %[dst], %[dst], 1           \n"
+
+   "3:                                     \n"
+    ".set pop                              \n"
+      : [src] "+r" (src), [dst] "+r" (dst)
+      : [width] "r" (width)
+      : "t0", "t1", "t2", "t3", "t4", "t5"
+  );
+}
+
+void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                            int width) {
+  int x = 0;
+  int y = 0;
+  __asm__ __volatile__ (
+    ".set push                                    \n"
+    ".set noreorder                               \n"
+
+    "addu            $t4, %[width], %[width]      \n"
+    "srl             %[x], %[width], 4            \n"
+    "andi            %[y], %[width], 0xf          \n"
+    "blez            %[x], 2f                     \n"
+    " addu           %[src_uv], %[src_uv], $t4    \n"
+
+    ".p2align        2                            \n"
+   "1:                                            \n"
+    "lw              $t0, -32(%[src_uv])          \n"  // |3|2|1|0|
+    "lw              $t1, -28(%[src_uv])          \n"  // |7|6|5|4|
+    "lw              $t2, -24(%[src_uv])          \n"  // |11|10|9|8|
+    "lw              $t3, -20(%[src_uv])          \n"  // |15|14|13|12|
+    "lw              $t4, -16(%[src_uv])          \n"  // |19|18|17|16|
+    "lw              $t6, -12(%[src_uv])          \n"  // |23|22|21|20|
+    "lw              $t7, -8(%[src_uv])           \n"  // |27|26|25|24|
+    "lw              $t8, -4(%[src_uv])           \n"  // |31|30|29|28|
+
+    "rotr            $t0, $t0, 16                 \n"  // |1|0|3|2|
+    "rotr            $t1, $t1, 16                 \n"  // |5|4|7|6|
+    "rotr            $t2, $t2, 16                 \n"  // |9|8|11|10|
+    "rotr            $t3, $t3, 16                 \n"  // |13|12|15|14|
+    "rotr            $t4, $t4, 16                 \n"  // |17|16|19|18|
+    "rotr            $t6, $t6, 16                 \n"  // |21|20|23|22|
+    "rotr            $t7, $t7, 16                 \n"  // |25|24|27|26|
+    "rotr            $t8, $t8, 16                 \n"  // |29|28|31|30|
+    "precr.qb.ph     $t9, $t0, $t1                \n"  // |0|2|4|6|
+    "precrq.qb.ph    $t5, $t0, $t1                \n"  // |1|3|5|7|
+    "precr.qb.ph     $t0, $t2, $t3                \n"  // |8|10|12|14|
+    "precrq.qb.ph    $t1, $t2, $t3                \n"  // |9|11|13|15|
+    "precr.qb.ph     $t2, $t4, $t6                \n"  // |16|18|20|22|
+    "precrq.qb.ph    $t3, $t4, $t6                \n"  // |17|19|21|23|
+    "precr.qb.ph     $t4, $t7, $t8                \n"  // |24|26|28|30|
+    "precrq.qb.ph    $t6, $t7, $t8                \n"  // |25|27|29|31|
+    "addiu           %[src_uv], %[src_uv], -32    \n"
+    "addiu           %[x], %[x], -1               \n"
+    "swr             $t4, 0(%[dst_u])             \n"
+    "swl             $t4, 3(%[dst_u])             \n"  // |30|28|26|24|
+    "swr             $t6, 0(%[dst_v])             \n"
+    "swl             $t6, 3(%[dst_v])             \n"  // |31|29|27|25|
+    "swr             $t2, 4(%[dst_u])             \n"
+    "swl             $t2, 7(%[dst_u])             \n"  // |22|20|18|16|
+    "swr             $t3, 4(%[dst_v])             \n"
+    "swl             $t3, 7(%[dst_v])             \n"  // |23|21|19|17|
+    "swr             $t0, 8(%[dst_u])             \n"
+    "swl             $t0, 11(%[dst_u])            \n"  // |14|12|10|8|
+    "swr             $t1, 8(%[dst_v])             \n"
+    "swl             $t1, 11(%[dst_v])            \n"  // |15|13|11|9|
+    "swr             $t9, 12(%[dst_u])            \n"
+    "swl             $t9, 15(%[dst_u])            \n"  // |6|4|2|0|
+    "swr             $t5, 12(%[dst_v])            \n"
+    "swl             $t5, 15(%[dst_v])            \n"  // |7|5|3|1|
+    "addiu           %[dst_v], %[dst_v], 16       \n"
+    "bgtz            %[x], 1b                     \n"
+    " addiu          %[dst_u], %[dst_u], 16       \n"
+    "beqz            %[y], 3f                     \n"
+    " nop                                         \n"
+    "b               2f                           \n"
+    " nop                                         \n"
+
+   "2:                                            \n"
+    "lbu             $t0, -2(%[src_uv])           \n"
+    "lbu             $t1, -1(%[src_uv])           \n"
+    "addiu           %[src_uv], %[src_uv], -2     \n"
+    "addiu           %[y], %[y], -1               \n"
+    "sb              $t0, 0(%[dst_u])             \n"
+    "sb              $t1, 0(%[dst_v])             \n"
+    "addiu           %[dst_u], %[dst_u], 1        \n"
+    "bgtz            %[y], 2b                     \n"
+    " addiu          %[dst_v], %[dst_v], 1        \n"
+
+   "3:                                            \n"
+    ".set pop                                     \n"
+      : [src_uv] "+r" (src_uv),
+        [dst_u] "+r" (dst_u),
+        [dst_v] "+r" (dst_v),
+        [x] "=&r" (x),
+        [y] "+r" (y)
+      : [width] "r" (width)
+      : "t0", "t1", "t2", "t3", "t4",
+      "t5", "t7", "t8", "t9"
+  );
+}
+
+// Convert (4 Y and 2 VU) I422 and arrange RGB values into
+// t5 = | 0 | B0 | 0 | b0 |
+// t4 = | 0 | B1 | 0 | b1 |
+// t9 = | 0 | G0 | 0 | g0 |
+// t8 = | 0 | G1 | 0 | g1 |
+// t2 = | 0 | R0 | 0 | r0 |
+// t1 = | 0 | R1 | 0 | r1 |
+#define I422ToTransientMipsRGB                                                 \
+      "lw                $t0, 0(%[y_buf])       \n"                            \
+      "lhu               $t1, 0(%[u_buf])       \n"                            \
+      "lhu               $t2, 0(%[v_buf])       \n"                            \
+      "preceu.ph.qbr     $t1, $t1               \n"                            \
+      "preceu.ph.qbr     $t2, $t2               \n"                            \
+      "preceu.ph.qbra    $t3, $t0               \n"                            \
+      "preceu.ph.qbla    $t0, $t0               \n"                            \
+      "subu.ph           $t1, $t1, $s5          \n"                            \
+      "subu.ph           $t2, $t2, $s5          \n"                            \
+      "subu.ph           $t3, $t3, $s4          \n"                            \
+      "subu.ph           $t0, $t0, $s4          \n"                            \
+      "mul.ph            $t3, $t3, $s0          \n"                            \
+      "mul.ph            $t0, $t0, $s0          \n"                            \
+      "shll.ph           $t4, $t1, 0x7          \n"                            \
+      "subu.ph           $t4, $t4, $t1          \n"                            \
+      "mul.ph            $t6, $t1, $s1          \n"                            \
+      "mul.ph            $t1, $t2, $s2          \n"                            \
+      "addq_s.ph         $t5, $t4, $t3          \n"                            \
+      "addq_s.ph         $t4, $t4, $t0          \n"                            \
+      "shra.ph           $t5, $t5, 6            \n"                            \
+      "shra.ph           $t4, $t4, 6            \n"                            \
+      "addiu             %[u_buf], 2            \n"                            \
+      "addiu             %[v_buf], 2            \n"                            \
+      "addu.ph           $t6, $t6, $t1          \n"                            \
+      "mul.ph            $t1, $t2, $s3          \n"                            \
+      "addu.ph           $t9, $t6, $t3          \n"                            \
+      "addu.ph           $t8, $t6, $t0          \n"                            \
+      "shra.ph           $t9, $t9, 6            \n"                            \
+      "shra.ph           $t8, $t8, 6            \n"                            \
+      "addu.ph           $t2, $t1, $t3          \n"                            \
+      "addu.ph           $t1, $t1, $t0          \n"                            \
+      "shra.ph           $t2, $t2, 6            \n"                            \
+      "shra.ph           $t1, $t1, 6            \n"                            \
+      "subu.ph           $t5, $t5, $s5          \n"                            \
+      "subu.ph           $t4, $t4, $s5          \n"                            \
+      "subu.ph           $t9, $t9, $s5          \n"                            \
+      "subu.ph           $t8, $t8, $s5          \n"                            \
+      "subu.ph           $t2, $t2, $s5          \n"                            \
+      "subu.ph           $t1, $t1, $s5          \n"                            \
+      "shll_s.ph         $t5, $t5, 8            \n"                            \
+      "shll_s.ph         $t4, $t4, 8            \n"                            \
+      "shll_s.ph         $t9, $t9, 8            \n"                            \
+      "shll_s.ph         $t8, $t8, 8            \n"                            \
+      "shll_s.ph         $t2, $t2, 8            \n"                            \
+      "shll_s.ph         $t1, $t1, 8            \n"                            \
+      "shra.ph           $t5, $t5, 8            \n"                            \
+      "shra.ph           $t4, $t4, 8            \n"                            \
+      "shra.ph           $t9, $t9, 8            \n"                            \
+      "shra.ph           $t8, $t8, 8            \n"                            \
+      "shra.ph           $t2, $t2, 8            \n"                            \
+      "shra.ph           $t1, $t1, 8            \n"                            \
+      "addu.ph           $t5, $t5, $s5          \n"                            \
+      "addu.ph           $t4, $t4, $s5          \n"                            \
+      "addu.ph           $t9, $t9, $s5          \n"                            \
+      "addu.ph           $t8, $t8, $s5          \n"                            \
+      "addu.ph           $t2, $t2, $s5          \n"                            \
+      "addu.ph           $t1, $t1, $s5          \n"
+
+void I422ToARGBRow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128| // clipping
+    "lui               $s6, 0xff00            \n"
+    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|ff|
+
+    ".p2align          2                      \n"
+   "1:                                        \n"
+      I422ToTransientMipsRGB
+// Arranging into argb format
+    "precr.qb.ph       $t4, $t8, $t4          \n"  // |G1|g1|B1|b1|
+    "precr.qb.ph       $t5, $t9, $t5          \n"  // |G0|g0|B0|b0|
+    "addiu             %[width], -4           \n"
+    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |G1|B1|G0|B0|
+    "precr.qb.ph       $t9, $t4, $t5          \n"  // |g1|b1|g0|b0|
+    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
+
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
+    "or                $t1, $t1, $s6          \n"  // |ff|R1|ff|R0|
+    "or                $t2, $t2, $s6          \n"  // |ff|r1|ff|r0|
+    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|r1|g1|b1|
+    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|R1|G1|B1|
+    "sll               $t9, $t9, 16           \n"
+    "sll               $t8, $t8, 16           \n"
+    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|r0|g0|b0|
+    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|R0|G0|B0|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+void I422ToABGRRow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128|
+    "lui               $s6, 0xff00            \n"
+    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|
+
+    ".p2align          2                       \n"
+   "1:                                         \n"
+      I422ToTransientMipsRGB
+// Arranging into abgr format
+    "precr.qb.ph      $t0, $t8, $t1           \n"  // |G1|g1|R1|r1|
+    "precr.qb.ph      $t3, $t9, $t2           \n"  // |G0|g0|R0|r0|
+    "precrq.qb.ph     $t8, $t0, $t3           \n"  // |G1|R1|G0|R0|
+    "precr.qb.ph      $t9, $t0, $t3           \n"  // |g1|r1|g0|r0|
+
+    "precr.qb.ph       $t2, $t4, $t5          \n"  // |B1|b1|B0|b0|
+    "addiu             %[width], -4           \n"
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |B1|0 |B0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |b1|0 |b0|
+    "or                $t1, $t1, $s6          \n"  // |ff|B1|ff|B0|
+    "or                $t2, $t2, $s6          \n"  // |ff|b1|ff|b0|
+    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|b1|g1|r1|
+    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|B1|G1|R1|
+    "sll               $t9, $t9, 16           \n"
+    "sll               $t8, $t8, 16           \n"
+    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|b0|g0|r0|
+    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|B0|G0|R0|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+void I422ToBGRARow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74 |74 |
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128|
+    "lui               $s6, 0xff              \n"
+    "ori               $s6, 0xff              \n"  // |00|ff|00|ff|
+
+    ".p2align          2                      \n"
+   "1:                                        \n"
+      I422ToTransientMipsRGB
+      // Arranging into bgra format
+    "precr.qb.ph       $t4, $t4, $t8          \n"  // |B1|b1|G1|g1|
+    "precr.qb.ph       $t5, $t5, $t9          \n"  // |B0|b0|G0|g0|
+    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |B1|G1|B0|G0|
+    "precr.qb.ph       $t9, $t4, $t5          \n"  // |b1|g1|b0|g0|
+
+    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
+    "addiu             %[width], -4           \n"
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
+    "sll               $t1, $t1, 8            \n"  // |R1|0 |R0|0 |
+    "sll               $t2, $t2, 8            \n"  // |r1|0 |r0|0 |
+    "or                $t1, $t1, $s6          \n"  // |R1|ff|R0|ff|
+    "or                $t2, $t2, $s6          \n"  // |r1|ff|r0|ff|
+    "precrq.ph.w       $t0, $t9, $t2          \n"  // |b1|g1|r1|ff|
+    "precrq.ph.w       $t3, $t8, $t1          \n"  // |B1|G1|R1|ff|
+    "sll               $t1, $t1, 16           \n"
+    "sll               $t2, $t2, 16           \n"
+    "packrl.ph         $t2, $t9, $t2          \n"  // |b0|g0|r0|ff|
+    "packrl.ph         $t1, $t8, $t1          \n"  // |B0|G0|R0|ff|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+// Bilinear filter 8x2 -> 8x1
+void InterpolateRows_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                ptrdiff_t src_stride, int dst_width,
+                                int source_y_fraction) {
+    int y0_fraction = 256 - source_y_fraction;
+    const uint8* src_ptr1 = src_ptr + src_stride;
+
+  __asm__ __volatile__ (
+     ".set push                                           \n"
+     ".set noreorder                                      \n"
+
+     "replv.ph          $t0, %[y0_fraction]               \n"
+     "replv.ph          $t1, %[source_y_fraction]         \n"
+
+    ".p2align           2                                 \n"
+   "1:                                                    \n"
+     "lw                $t2, 0(%[src_ptr])                \n"
+     "lw                $t3, 0(%[src_ptr1])               \n"
+     "lw                $t4, 4(%[src_ptr])                \n"
+     "lw                $t5, 4(%[src_ptr1])               \n"
+     "muleu_s.ph.qbl    $t6, $t2, $t0                     \n"
+     "muleu_s.ph.qbr    $t7, $t2, $t0                     \n"
+     "muleu_s.ph.qbl    $t8, $t3, $t1                     \n"
+     "muleu_s.ph.qbr    $t9, $t3, $t1                     \n"
+     "muleu_s.ph.qbl    $t2, $t4, $t0                     \n"
+     "muleu_s.ph.qbr    $t3, $t4, $t0                     \n"
+     "muleu_s.ph.qbl    $t4, $t5, $t1                     \n"
+     "muleu_s.ph.qbr    $t5, $t5, $t1                     \n"
+     "addq.ph           $t6, $t6, $t8                     \n"
+     "addq.ph           $t7, $t7, $t9                     \n"
+     "addq.ph           $t2, $t2, $t4                     \n"
+     "addq.ph           $t3, $t3, $t5                     \n"
+     "shra.ph           $t6, $t6, 8                       \n"
+     "shra.ph           $t7, $t7, 8                       \n"
+     "shra.ph           $t2, $t2, 8                       \n"
+     "shra.ph           $t3, $t3, 8                       \n"
+     "precr.qb.ph       $t6, $t6, $t7                     \n"
+     "precr.qb.ph       $t2, $t2, $t3                     \n"
+     "addiu             %[src_ptr], %[src_ptr], 8         \n"
+     "addiu             %[src_ptr1], %[src_ptr1], 8       \n"
+     "addiu             %[dst_width], %[dst_width], -8    \n"
+     "sw                $t6, 0(%[dst_ptr])                \n"
+     "sw                $t2, 4(%[dst_ptr])                \n"
+     "bgtz              %[dst_width], 1b                  \n"
+     " addiu            %[dst_ptr], %[dst_ptr], 8         \n"
+
+     ".set pop                                            \n"
+  : [dst_ptr] "+r" (dst_ptr),
+    [src_ptr1] "+r" (src_ptr1),
+    [src_ptr] "+r" (src_ptr),
+    [dst_width] "+r" (dst_width)
+  : [source_y_fraction] "r" (source_y_fraction),
+    [y0_fraction] "r" (y0_fraction),
+    [src_stride] "r" (src_stride)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+#endif  // __mips_dsp_rev >= 2
+
+#endif  // defined(__mips__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/row_neon.cc b/third_party/libyuv/source/row_neon.cc
new file mode 100644
index 0000000..c5ae2c5
--- /dev/null
+++ b/third_party/libyuv/source/row_neon.cc
@@ -0,0 +1,2844 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+
+// Read 8 Y, 4 U and 4 V from 422
+#define READYUV422                                                             \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vld1.32    {d2[0]}, [%1]!                 \n"                             \
+    "vld1.32    {d2[1]}, [%2]!                 \n"
+
+// Read 8 Y, 2 U and 2 V from 422
+#define READYUV411                                                             \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vld1.16    {d2[0]}, [%1]!                 \n"                             \
+    "vld1.16    {d2[1]}, [%2]!                 \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vzip.u8    d2, d3                         \n"
+
+// Read 8 Y, 8 U and 8 V from 444
+#define READYUV444                                                             \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vld1.8     {d3}, [%2]!                    \n"                             \
+    "vpaddl.u8  q1, q1                         \n"                             \
+    "vrshrn.u16 d2, q1, #1                     \n"
+
+// Read 8 Y, and set 4 U and 4 V to 128
+#define READYUV400                                                             \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vmov.u8    d2, #128                       \n"
+
+// Read 8 Y and 4 UV from NV12
+#define READNV12                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 Y and 4 VU from NV21
+#define READNV21                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d3, d2                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 YUY2
+#define READYUY2                                                               \
+    "vld2.8     {d0, d2}, [%0]!                \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 UYVY
+#define READUYVY                                                               \
+    "vld2.8     {d2, d3}, [%0]!                \n"                             \
+    "vmov.u8    d0, d3                         \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+#define YUV422TORGB                                                            \
+    "veor.u8    d2, d26                        \n"/*subtract 128 from u and v*/\
+    "vmull.s8   q8, d2, d24                    \n"/*  u/v B/R component      */\
+    "vmull.s8   q9, d2, d25                    \n"/*  u/v G component        */\
+    "vmov.u8    d1, #0                         \n"/*  split odd/even y apart */\
+    "vtrn.u8    d0, d1                         \n"                             \
+    "vsub.s16   q0, q0, q15                    \n"/*  offset y               */\
+    "vmul.s16   q0, q0, q14                    \n"                             \
+    "vadd.s16   d18, d19                       \n"                             \
+    "vqadd.s16  d20, d0, d16                   \n" /* B */                     \
+    "vqadd.s16  d21, d1, d16                   \n"                             \
+    "vqadd.s16  d22, d0, d17                   \n" /* R */                     \
+    "vqadd.s16  d23, d1, d17                   \n"                             \
+    "vqadd.s16  d16, d0, d18                   \n" /* G */                     \
+    "vqadd.s16  d17, d1, d18                   \n"                             \
+    "vqshrun.s16 d0, q10, #6                   \n" /* B */                     \
+    "vqshrun.s16 d1, q11, #6                   \n" /* G */                     \
+    "vqshrun.s16 d2, q8, #6                    \n" /* R */                     \
+    "vmovl.u8   q10, d0                        \n"/*  set up for reinterleave*/\
+    "vmovl.u8   q11, d1                        \n"                             \
+    "vmovl.u8   q8, d2                         \n"                             \
+    "vtrn.u8    d20, d21                       \n"                             \
+    "vtrn.u8    d22, d23                       \n"                             \
+    "vtrn.u8    d16, d17                       \n"                             \
+    "vmov.u8    d21, d16                       \n"
+
+static vec8 kUVToRB  = { 127, 127, 127, 127, 102, 102, 102, 102,
+                         0, 0, 0, 0, 0, 0, 0, 0 };
+static vec8 kUVToG = { -25, -25, -25, -25, -52, -52, -52, -52,
+                       0, 0, 0, 0, 0, 0, 0, 0 };
+
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV444
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV411
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d19, #255                      \n"
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_bgra),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_abgr),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d19, #255                      \n"
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgba),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),      // %0
+      "+r"(src_u),      // %1
+      "+r"(src_v),      // %2
+      "+r"(dst_rgb24),  // %3
+      "+r"(width)       // %4
+    : "r"(&kUVToRB),    // %5
+      "r"(&kUVToG)      // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_raw),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTORGB565                                                           \
+    "vshr.u8    d20, d20, #3                   \n"  /* B                    */ \
+    "vshr.u8    d21, d21, #2                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #11                  \n"  /* R                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q0, q0, q10                    \n"  /* BGR                  */
+
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    ARGBTORGB565
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_rgb565),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTOARGB1555                                                         \
+    "vshr.u8    q10, q10, #3                   \n"  /* B                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vshr.u8    d23, d23, #7                   \n"  /* A                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vmovl.u8   q11, d23                       \n"  /* A                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #10                  \n"  /* R                    */ \
+    "vshl.u16   q11, q11, #15                  \n"  /* A                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q1, q10, q11                   \n"  /* RA                   */ \
+    "vorr       q0, q0, q1                     \n"  /* BGRA                 */
+
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB1555
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb1555),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTOARGB4444                                                         \
+    "vshr.u8    d20, d20, #4                   \n"  /* B                    */ \
+    "vbic.32    d21, d21, d4                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #4                   \n"  /* R                    */ \
+    "vbic.32    d23, d23, d4                   \n"  /* A                    */ \
+    "vorr       d0, d20, d21                   \n"  /* BG                   */ \
+    "vorr       d1, d22, d23                   \n"  /* RA                   */ \
+    "vzip.u8    d0, d1                         \n"  /* BGRA                 */
+
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB4444
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb4444),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void YToARGBRow_NEON(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%3]                    \n"
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV400
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+    "vmov.u8    d23, #255                      \n"
+  "1:                                          \n"
+    "vld1.8     {d20}, [%0]!                   \n"
+    "vmov       d21, d20                       \n"
+    "vmov       d22, d20                       \n"
+    "subs       %2, %2, #8                     \n"
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    :
+    : "cc", "memory", "d20", "d21", "d22", "d23"
+  );
+}
+
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%4]                    \n"
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%4]                    \n"
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%4]                    \n"
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%4]                    \n"
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%3]                    \n"
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUY2
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_yuy2),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%3]                    \n"
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READUYVY
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_uyvy),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pairs of UV
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vst1.8     {q0}, [%1]!                    \n"  // store U
+    "vst1.8     {q1}, [%2]!                    \n"  // store V
+    "bgt        1b                             \n"
+    : "+r"(src_uv),  // %0
+      "+r"(dst_u),   // %1
+      "+r"(dst_v),   // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load U
+    "vld1.8     {q1}, [%1]!                    \n"  // load V
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vst2.u8    {q0, q1}, [%2]!                \n"  // store 16 pairs of UV
+    "bgt        1b                             \n"
+    :
+      "+r"(src_u),   // %0
+      "+r"(src_v),   // %1
+      "+r"(dst_uv),  // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
+void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 32
+    "subs       %2, %2, #32                    \n"  // 32 processed per loop
+    "vst1.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 32
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2  // Output registers
+  :                     // Input registers
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SetRow8 writes 'count' bytes using a 32 bit value repeated.
+void SetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
+    "1:                                        \n"
+    "subs      %1, %1, #16                     \n"  // 16 bytes per loop
+    "vst1.8    {q0}, [%0]!                     \n"  // store
+    "bgt       1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "q0"
+  );
+}
+
+// TODO(fbarchard): Make fully assembler
+// SetRow32 writes 'count' words using a 32 bit value repeated.
+void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
+                      int dst_stride, int height) {
+  for (int y = 0; y < height; ++y) {
+    SetRow_NEON(dst, v32, width << 2);
+    dst += dst_stride;
+  }
+}
+
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2                     \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #16                        \n"  // 16 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r12, #-16                      \n"
+    "add        %0, %0, %3, lsl #1             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {d0, d1}, [%0], r12            \n"  // src -= 16
+    "subs       %3, #8                         \n"  // 8 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // dst += 8
+    "vst1.8     {d1}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_u),   // %1
+    "+r"(dst_v),   // %2
+    "+r"(width)    // %3
+  :
+  : "cc", "memory", "r12", "q0"
+  );
+}
+
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2, lsl #2             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #4                         \n"  // 4 pixels per loop.
+    "vrev64.32  q0, q0                         \n"
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),   // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+#define RGB565TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxGGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB RRRRRxxx */ \
+    "vshl.u8    d6, d6, #2                     \n"  /* G GGGGGG00 upper 6   */ \
+    "vshr.u8    d1, d1, #3                     \n"  /* R 000RRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #6                     \n"  /* G 000000GG lower 2   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB1555TOARGB                                                         \
+    "vshrn.u16  d7, q0, #8                     \n"  /* A Arrrrrxx           */ \
+    "vshr.u8    d6, d7, #2                     \n"  /* R xxxRRRRR           */ \
+    "vshrn.u16  d5, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vmovn.u16  d4, q0                         \n"  /* B xxxBBBBB           */ \
+    "vshr.u8    d7, d7, #7                     \n"  /* A 0000000A           */ \
+    "vneg.s8    d7, d7                         \n"  /* A AAAAAAAA upper 8   */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* R RRRRR000 upper 5   */ \
+    "vshr.u8    q1, q3, #5                     \n"  /* R,A 00000RRR lower 3 */ \
+    "vshl.u8    q0, q2, #3                     \n"  /* B,G BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,G 00000BBB lower 3 */ \
+    "vorr.u8    q1, q1, q3                     \n"  /* R,A                  */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,G                  */ \
+
+// RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
+#define RGB555TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB xRRRRRxx */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* G GGGGG000 upper 5   */ \
+    "vshr.u8    d1, d1, #2                     \n"  /* R 00xRRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #5                     \n"  /* G 00000GGG lower 3   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB4444TOARGB                                                         \
+    "vuzp.u8    d0, d1                         \n"  /* d0 BG, d1 RA         */ \
+    "vshl.u8    q2, q0, #4                     \n"  /* B,R BBBB0000         */ \
+    "vshr.u8    q1, q0, #4                     \n"  /* G,A 0000GGGG         */ \
+    "vshr.u8    q0, q2, #4                     \n"  /* B,R 0000BBBB         */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,R BBBBBBBB         */ \
+    "vshl.u8    q2, q1, #4                     \n"  /* G,A GGGG0000         */ \
+    "vorr.u8    q1, q1, q2                     \n"  /* G,A GGGGGGGG         */ \
+    "vswp.u8    d1, d2                         \n"  /* B,R,G,A -> B,G,R,A   */
+
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RGB24.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_rgb24),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RAW.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_raw),   // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of YUY2.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    "vst1.8     {q0}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of UYVY.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    "vst1.8     {q1}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    "vst1.8     {d1}, [%1]!                    \n"  // store 8 U.
+    "vst1.8     {d3}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 U.
+    "vst1.8     {d2}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_yuy2
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row YUY2.
+    "vrhadd.u8  d1, d1, d5                     \n"  // average rows of U
+    "vrhadd.u8  d3, d3, d7                     \n"  // average rows of V
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 U.
+    "vst1.8     {d3}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),     // %0
+    "+r"(stride_yuy2),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_uyvy
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row UYVY.
+    "vrhadd.u8  d0, d0, d4                     \n"  // average rows of U
+    "vrhadd.u8  d2, d2, d6                     \n"  // average rows of V
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 U.
+    "vst1.8     {d2}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),     // %0
+    "+r"(stride_uyvy),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+
+void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %0                         \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load row 1 16 pixels.
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vld1.8     {q1}, [%1]!                    \n"  // load row 2 16 pixels.
+    "vrhadd.u8  q0, q1                         \n"  // average row 1 and 2
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_uv),         // %0
+    "+r"(src_uv_stride),  // %1
+    "+r"(dst_uv),         // %2
+    "+r"(pix)             // %3
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Select 2 channels from ARGB on alternating pixels.  e.g.  BGBGBGBG
+void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) {
+  asm volatile (
+    "vmov.u32   d6[0], %3                      \n"  // selector
+  "1:                                          \n"
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 8 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vtbl.8     d4, {d0, d1}, d6               \n"  // look up 4 pixels
+    "vtbl.8     d5, {d2, d3}, d6               \n"  // look up 4 pixels
+    "vtrn.u32   d4, d5                         \n"  // combine 8 pixels
+    "vst1.8     {d4}, [%1]!                    \n"  // store 8.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_bayer),  // %1
+    "+r"(pix)         // %2
+  : "r"(selector)     // %3
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// Select G channels from ARGB.  e.g.  GGGGGGGG
+void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 /*selector*/, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load row 8 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vst1.8     {d1}, [%1]!                    \n"  // store 8 G's.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_bayer),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    "vld1.8     {q2}, [%3]                     \n"  // shuffler
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 4 pixels.
+    "subs       %2, %2, #4                     \n"  // 4 processed per loop
+    "vtbl.8     d2, {d0, d1}, d4               \n"  // look up 2 first pixels
+    "vtbl.8     d3, {d0, d1}, d5               \n"  // look up 2 next pixels
+    "vst1.8     {q1}, [%1]!                    \n"  // store 4.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {d0, d2}, [%0]!                \n"  // load 16 Ys
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 Us
+    "vld1.8     {d3}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 YUY2/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_yuy2),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {d1, d3}, [%0]!                \n"  // load 16 Ys
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 Us
+    "vld1.8     {d2}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 UYVY/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_uyvy),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTORGB565
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_rgb565),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
+                            int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB1555
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb1555),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB4444
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb4444),  // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit Y
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+// 8x1 pixels.
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.u8    d24, #112                      \n"  // UB / VR 0.875 coefficient
+    "vmov.u8    d25, #74                       \n"  // UG -0.5781 coefficient
+    "vmov.u8    d26, #38                       \n"  // UR -0.2969 coefficient
+    "vmov.u8    d27, #18                       \n"  // VB -0.1406 coefficient
+    "vmov.u8    d28, #94                       \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlsl.u8   q2, d1, d25                    \n"  // G
+    "vmlsl.u8   q2, d2, d26                    \n"  // R
+    "vadd.u16   q2, q2, q15                    \n"  // +128 -> unsigned
+
+    "vmull.u8   q3, d2, d24                    \n"  // R
+    "vmlsl.u8   q3, d1, d28                    \n"  // G
+    "vmlsl.u8   q3, d0, d27                    \n"  // B
+    "vadd.u16   q3, q3, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q2, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q3, #8                    \n"  // 16 bit to 8 bit V
+
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 32x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 32.
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%0]!     \n"  // load 8 more ARGB pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%0]!     \n"  // load last 8 ARGB pixels.
+    "vpaddl.u8  q4, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q5, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q6, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vpadd.u16  d0, d0, d1                     \n"  // B 16 shorts -> 8 shorts.
+    "vpadd.u16  d1, d8, d9                     \n"  // B
+    "vpadd.u16  d2, d2, d3                     \n"  // G 16 shorts -> 8 shorts.
+    "vpadd.u16  d3, d10, d11                   \n"  // G
+    "vpadd.u16  d4, d4, d5                     \n"  // R 16 shorts -> 8 shorts.
+    "vpadd.u16  d5, d12, d13                   \n"  // R
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %3, %3, #32                    \n"  // 32 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#define RGBTOUV(QB, QG, QR) \
+    "vmul.s16   q8, " #QB ", q10               \n"  /* B                    */ \
+    "vmls.s16   q8, " #QG ", q11               \n"  /* G                    */ \
+    "vmls.s16   q8, " #QR ", q12               \n"  /* R                    */ \
+    "vadd.u16   q8, q8, q15                    \n"  /* +128 -> unsigned     */ \
+    "vmul.s16   q9, " #QR ", q10               \n"  /* R                    */ \
+    "vmls.s16   q9, " #QG ", q14               \n"  /* G                    */ \
+    "vmls.s16   q9, " #QB ", q13               \n"  /* B                    */ \
+    "vadd.u16   q9, q9, q15                    \n"  /* +128 -> unsigned     */ \
+    "vqshrn.u16  d0, q8, #8                    \n"  /* 16 bit to 8 bit U    */ \
+    "vqshrn.u16  d1, q9, #8                    \n"  /* 16 bit to 8 bit V    */
+
+// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): Subsample match C code.
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #127 / 2                  \n"  // UB / VR 0.500 coefficient
+    "vmov.s16   q11, #84 / 2                   \n"  // UG -0.33126 coefficient
+    "vmov.s16   q12, #43 / 2                   \n"  // UR -0.16874 coefficient
+    "vmov.s16   q13, #20 / 2                   \n"  // VB -0.08131 coefficient
+    "vmov.s16   q14, #107 / 2                  \n"  // VG -0.41869 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_bgra
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 BGRA pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 BGRA pixels.
+    "vpaddl.u8  q3, q3                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more BGRA pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 BGRA pixels.
+    "vpadal.u8  q3, q7                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q1, q1, #1                     \n"  // 2x average
+    "vrshr.u16  q2, q2, #1                     \n"
+    "vrshr.u16  q3, q3, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q3, q2, q1)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(src_stride_bgra),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_abgr
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ABGR pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ABGR pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ABGR pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ABGR pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_stride_abgr),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgba
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 RGBA pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 RGBA pixels.
+    "vpaddl.u8  q0, q1                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q3                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more RGBA pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 RGBA pixels.
+    "vpadal.u8  q0, q5                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q7                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(src_stride_rgba),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgb24
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RGB24 pixels.
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RGB24 pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RGB24 pixels.
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RGB24 pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_stride_rgb24),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_raw
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RAW pixels.
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RAW pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RAW pixels.
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RAW pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_stride_raw),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(src_stride_rgb565),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(src_stride_argb1555),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(src_stride_argb4444),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_y),       // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of BGRA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // R
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ABGR.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // R
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of RGBA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // B
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_NEON(uint8* dst_ptr,
+                         const uint8* src_ptr, ptrdiff_t src_stride,
+                         int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp        %4, #0                         \n"
+    "beq        100f                           \n"
+    "add        %2, %1                         \n"
+    "cmp        %4, #64                        \n"
+    "beq        75f                            \n"
+    "cmp        %4, #128                       \n"
+    "beq        50f                            \n"
+    "cmp        %4, #192                       \n"
+    "beq        25f                            \n"
+
+    "vdup.8     d5, %4                         \n"
+    "rsb        %4, #256                       \n"
+    "vdup.8     d4, %4                         \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    "vld1.8     {q0}, [%1]!                    \n"
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vmull.u8   q13, d0, d4                    \n"
+    "vmull.u8   q14, d1, d4                    \n"
+    "vmlal.u8   q13, d2, d5                    \n"
+    "vmlal.u8   q14, d3, d5                    \n"
+    "vrshrn.u16 d0, q13, #8                    \n"
+    "vrshrn.u16 d1, q14, #8                    \n"
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        1b                             \n"
+    "b          99f                            \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    "vld1.8     {q0}, [%1]!                    \n"
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        25b                            \n"
+    "b          99f                            \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    "vld1.8     {q0}, [%1]!                    \n"
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        50b                            \n"
+    "b          99f                            \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    "vld1.8     {q1}, [%1]!                    \n"
+    "vld1.8     {q0}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        75b                            \n"
+    "b          99f                            \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    "vld1.8     {q0}, [%1]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        100b                           \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14"
+  );
+}
+
+// dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
+void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "subs       %3, #8                         \n"
+    "blt        89f                            \n"
+    // Blend 8 pixels.
+  "8:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB0.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 pixels of ARGB1.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 pixels of ARGB.
+    "bge        8b                             \n"
+
+  "89:                                         \n"
+    "adds       %3, #8-1                       \n"
+    "blt        99f                            \n"
+
+    // Blend 1 pixels.
+  "1:                                          \n"
+    "vld4.8     {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
+    "vld4.8     {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
+    "subs       %3, %3, #1                     \n"  // 1 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    "vst4.8     {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
+    "bge        1b                             \n"
+
+  "99:                                         \n"
+
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12"
+  );
+}
+
+// Attenuate 8 pixels at a time.
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    // Attenuate 8 pixels.
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d0, d3                    \n"  // b * a
+    "vmull.u8   q11, d1, d3                    \n"  // g * a
+    "vmull.u8   q12, d2, d3                    \n"  // r * a
+    "vqrshrn.u16 d0, q10, #8                   \n"  // b >>= 8
+    "vqrshrn.u16 d1, q11, #8                   \n"  // g >>= 8
+    "vqrshrn.u16 d2, q12, #8                   \n"  // r >>= 8
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  :
+  : "cc", "memory", "q0", "q1", "q10", "q11", "q12"
+  );
+}
+
+// Quantize 8 ARGB pixels (32 bytes).
+// dst = (dst * scale >> 16) * interval_size + interval_offset;
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "vdup.u16   q8, %2                         \n"
+    "vshr.u16   q8, q8, #1                     \n"  // scale >>= 1
+    "vdup.u16   q9, %3                         \n"  // interval multiply.
+    "vdup.u16   q10, %4                        \n"  // interval add
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]         \n"  // load 8 pixels of ARGB.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q0, d0                         \n"  // b (0 .. 255)
+    "vmovl.u8   q1, d2                         \n"
+    "vmovl.u8   q2, d4                         \n"
+    "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
+    "vqdmulh.s16 q1, q1, q8                    \n"  // g
+    "vqdmulh.s16 q2, q2, q8                    \n"  // r
+    "vmul.u16   q0, q0, q9                     \n"  // b * interval_size
+    "vmul.u16   q1, q1, q9                     \n"  // g
+    "vmul.u16   q2, q2, q9                     \n"  // r
+    "vadd.u16   q0, q0, q10                    \n"  // b + interval_offset
+    "vadd.u16   q1, q1, q10                    \n"  // g
+    "vadd.u16   q2, q2, q10                    \n"  // r
+    "vqmovn.u16 d0, q0                         \n"
+    "vqmovn.u16 d2, q1                         \n"
+    "vqmovn.u16 d4, q2                         \n"
+    "vst4.8     {d0, d2, d4, d6}, [%0]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10"
+  );
+}
+
+// Shade 8 pixels at a time by specified value.
+// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
+// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "vdup.u32   q0, %3                         \n"  // duplicate scale value.
+    "vzip.u8    d0, d1                         \n"  // d0 aarrggbb.
+    "vshr.u16   q0, q0, #1                     \n"  // scale / 2.
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d20, d22, d24, d26}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q10, d20                       \n"  // b (0 .. 255)
+    "vmovl.u8   q11, d22                       \n"
+    "vmovl.u8   q12, d24                       \n"
+    "vmovl.u8   q13, d26                       \n"
+    "vqrdmulh.s16 q10, q10, d0[0]              \n"  // b * scale * 2
+    "vqrdmulh.s16 q11, q11, d0[1]              \n"  // g
+    "vqrdmulh.s16 q12, q12, d0[2]              \n"  // r
+    "vqrdmulh.s16 q13, q13, d0[3]              \n"  // a
+    "vqmovn.u16 d20, q10                       \n"
+    "vqmovn.u16 d22, q11                       \n"
+    "vqmovn.u16 d24, q12                       \n"
+    "vqmovn.u16 d26, q13                       \n"
+    "vst4.8     {d20, d22, d24, d26}, [%1]!    \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),       // %0
+    "+r"(dst_argb),       // %1
+    "+r"(width)           // %2
+  : "r"(value)            // %3
+  : "cc", "memory", "q0", "q10", "q11", "q12", "q13"
+  );
+}
+
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+// Similar to ARGBToYJ but stores ARGB.
+// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit B
+    "vmov       d1, d0                         \n"  // G
+    "vmov       d2, d0                         \n"  // R
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d20, #17                       \n"  // BB coefficient
+    "vmov.u8    d21, #68                       \n"  // BG coefficient
+    "vmov.u8    d22, #35                       \n"  // BR coefficient
+    "vmov.u8    d24, #22                       \n"  // GB coefficient
+    "vmov.u8    d25, #88                       \n"  // GG coefficient
+    "vmov.u8    d26, #45                       \n"  // GR coefficient
+    "vmov.u8    d28, #24                       \n"  // BB coefficient
+    "vmov.u8    d29, #98                       \n"  // BG coefficient
+    "vmov.u8    d30, #50                       \n"  // BR coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]         \n"  // load 8 ARGB pixels.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d20                    \n"  // B to Sepia B
+    "vmlal.u8   q2, d1, d21                    \n"  // G
+    "vmlal.u8   q2, d2, d22                    \n"  // R
+    "vmull.u8   q3, d0, d24                    \n"  // B to Sepia G
+    "vmlal.u8   q3, d1, d25                    \n"  // G
+    "vmlal.u8   q3, d2, d26                    \n"  // R
+    "vmull.u8   q8, d0, d28                    \n"  // B to Sepia R
+    "vmlal.u8   q8, d1, d29                    \n"  // G
+    "vmlal.u8   q8, d2, d30                    \n"  // R
+    "vqshrn.u16 d0, q2, #7                     \n"  // 16 bit to 8 bit B
+    "vqshrn.u16 d1, q3, #7                     \n"  // 16 bit to 8 bit G
+    "vqshrn.u16 d2, q8, #7                     \n"  // 16 bit to 8 bit R
+    "vst4.8     {d0, d1, d2, d3}, [%0]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),  // %0
+    "+r"(width)      // %1
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// TODO(fbarchard): Was same as Sepia except matrix is provided.  This function
+// needs to saturate.  Consider doing a non-saturating version.
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width) {
+  asm volatile (
+    "vld1.8     {q2}, [%3]                     \n"  // load 3 ARGB vectors.
+    "vmovl.s8   q0, d4                         \n"  // B,G coefficients s16.
+    "vmovl.s8   q1, d5                         \n"  // R,A coefficients s16.
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d16, d18, d20, d22}, [%0]!    \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q8, d16                        \n"  // b (0 .. 255) 16 bit
+    "vmovl.u8   q9, d18                        \n"  // g
+    "vmovl.u8   q10, d20                       \n"  // r
+    "vmovl.u8   q15, d22                       \n"  // a
+    "vmul.s16   q12, q8, d0[0]                 \n"  // B = B * Matrix B
+    "vmul.s16   q13, q8, d1[0]                 \n"  // G = B * Matrix G
+    "vmul.s16   q14, q8, d2[0]                 \n"  // R = B * Matrix R
+    "vmul.s16   q15, q8, d3[0]                 \n"  // A = B * Matrix A
+    "vmul.s16   q4, q9, d0[1]                  \n"  // B += G * Matrix B
+    "vmul.s16   q5, q9, d1[1]                  \n"  // G += G * Matrix G
+    "vmul.s16   q6, q9, d2[1]                  \n"  // R += G * Matrix R
+    "vmul.s16   q7, q9, d3[1]                  \n"  // A += G * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q10, d0[2]                 \n"  // B += R * Matrix B
+    "vmul.s16   q5, q10, d1[2]                 \n"  // G += R * Matrix G
+    "vmul.s16   q6, q10, d2[2]                 \n"  // R += R * Matrix R
+    "vmul.s16   q7, q10, d3[2]                 \n"  // A += R * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q15, d0[3]                 \n"  // B += A * Matrix B
+    "vmul.s16   q5, q15, d1[3]                 \n"  // G += A * Matrix G
+    "vmul.s16   q6, q15, d2[3]                 \n"  // R += A * Matrix R
+    "vmul.s16   q7, q15, d3[3]                 \n"  // A += A * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vqshrun.s16 d16, q12, #6                  \n"  // 16 bit to 8 bit B
+    "vqshrun.s16 d18, q13, #6                  \n"  // 16 bit to 8 bit G
+    "vqshrun.s16 d20, q14, #6                  \n"  // 16 bit to 8 bit R
+    "vqshrun.s16 d22, q15, #6                  \n"  // 16 bit to 8 bit A
+    "vst4.8     {d16, d18, d20, d22}, [%1]!    \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "r"(matrix_argb)  // %3
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable.
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q0, d0, d1                     \n"  // multiply B
+    "vmull.u8   q1, d2, d3                     \n"  // multiply G
+    "vmull.u8   q2, d4, d5                     \n"  // multiply R
+    "vmull.u8   q3, d6, d7                     \n"  // multiply A
+    "vrshrn.u16 d0, q0, #8                     \n"  // 16 bit to 8 bit B
+    "vrshrn.u16 d1, q1, #8                     \n"  // 16 bit to 8 bit G
+    "vrshrn.u16 d2, q2, #8                     \n"  // 16 bit to 8 bit R
+    "vrshrn.u16 d3, q3, #8                     \n"  // 16 bit to 8 bit A
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_NEON
+
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   q0, q0, q2                     \n"  // add B, G
+    "vqadd.u8   q1, q1, q3                     \n"  // add R, A
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqsub.u8   q0, q0, q2                     \n"  // subtract B, G
+    "vqsub.u8   q1, q1, q3                     \n"  // subtract R, A
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0}, [%0]!                    \n"  // load 8 sobelx.
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d0, d0, d1                     \n"  // add
+    "vmov.u8    d1, d0                         \n"
+    "vmov.u8    d2, d0                         \n"
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// Adds Sobel X and Sobel Y and stores Sobel into plane.
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    // 16 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 16 sobelx.
+    "vld1.8     {q1}, [%1]!                    \n"  // load 16 sobely.
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vqadd.u8   q0, q0, q1                     \n"  // add
+    "vst1.8     {q0}, [%2]!                    \n"  // store 16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d2}, [%0]!                    \n"  // load 8 sobelx.
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d1, d0, d2                     \n"  // add
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0}, [%0],%5                  \n"  // top
+    "vld1.8     {d1}, [%0],%6                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    "vld1.8     {d2}, [%1],%5                  \n"  // center * 2
+    "vld1.8     {d3}, [%1],%6                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vld1.8     {d2}, [%2],%5                  \n"  // bottom
+    "vld1.8     {d3}, [%2],%6                  \n"
+    "subs       %4, %4, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    "vst1.8     {d0}, [%3]!                    \n"  // store 8 sobelx
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  : "r"(2),            // %5
+    "r"(6)             // %6
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0}, [%0],%4                  \n"  // left
+    "vld1.8     {d1}, [%1],%4                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    "vld1.8     {d2}, [%0],%4                  \n"  // center * 2
+    "vld1.8     {d3}, [%1],%4                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vld1.8     {d2}, [%0],%5                  \n"  // right
+    "vld1.8     {d3}, [%1],%5                  \n"
+    "subs       %3, %3, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 sobely
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  : "r"(1),            // %4
+    "r"(6)             // %5
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // __ARM_NEON__
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/row_posix.cc b/third_party/libyuv/source/row_posix.cc
new file mode 100644
index 0000000..e477088
--- /dev/null
+++ b/third_party/libyuv/source/row_posix.cc
@@ -0,0 +1,6443 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
+
+// Constants for ARGB
+static vec8 kARGBToY = {
+  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
+};
+
+// JPeg full range.
+static vec8 kARGBToYJ = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+#endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
+
+#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
+
+static vec8 kARGBToU = {
+  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
+};
+
+static vec8 kARGBToUJ = {
+  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
+};
+
+static vec8 kARGBToV = {
+  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+};
+
+static vec8 kARGBToVJ = {
+  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
+};
+
+// Constants for BGRA
+static vec8 kBGRAToY = {
+  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
+};
+
+static vec8 kBGRAToU = {
+  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
+};
+
+static vec8 kBGRAToV = {
+  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
+};
+
+// Constants for ABGR
+static vec8 kABGRToY = {
+  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
+};
+
+static vec8 kABGRToU = {
+  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
+};
+
+static vec8 kABGRToV = {
+  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
+};
+
+// Constants for RGBA.
+static vec8 kRGBAToY = {
+  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
+};
+
+static vec8 kRGBAToU = {
+  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
+};
+
+static vec8 kRGBAToV = {
+  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
+};
+
+static uvec8 kAddY16 = {
+  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
+};
+
+static vec16 kAddYJ64 = {
+  64, 64, 64, 64, 64, 64, 64, 64
+};
+
+static uvec8 kAddUV128 = {
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
+};
+
+static uvec16 kAddUVJ128 = {
+  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
+};
+#endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
+
+#ifdef HAS_RGB24TOARGBROW_SSSE3
+
+// Shuffle table for converting RGB24 to ARGB.
+static uvec8 kShuffleMaskRGB24ToARGB = {
+  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
+};
+
+// Shuffle table for converting RAW to ARGB.
+static uvec8 kShuffleMaskRAWToARGB = {
+  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
+};
+
+// Shuffle table for converting ARGB to RGB24.
+static uvec8 kShuffleMaskARGBToRGB24 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static uvec8 kShuffleMaskARGBToRAW = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
+static uvec8 kShuffleMaskARGBToRGB24_0 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static uvec8 kShuffleMaskARGBToRAW_0 = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u
+};
+#endif  // HAS_RGB24TOARGBROW_SSSE3
+
+#if defined(TESTING) && defined(__x86_64__)
+void TestRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  asm volatile (
+    ".p2align  5                               \n"
+    "mov       %%eax,%%eax                     \n"
+    "mov       %%ebx,%%ebx                     \n"
+    "mov       %%ecx,%%ecx                     \n"
+    "mov       %%edx,%%edx                     \n"
+    "mov       %%esi,%%esi                     \n"
+    "mov       %%edi,%%edi                     \n"
+    "mov       %%ebp,%%ebp                     \n"
+    "mov       %%esp,%%esp                     \n"
+    ".p2align  5                               \n"
+    "mov       %%r8d,%%r8d                     \n"
+    "mov       %%r9d,%%r9d                     \n"
+    "mov       %%r10d,%%r10d                   \n"
+    "mov       %%r11d,%%r11d                   \n"
+    "mov       %%r12d,%%r12d                   \n"
+    "mov       %%r13d,%%r13d                   \n"
+    "mov       %%r14d,%%r14d                   \n"
+    "mov       %%r15d,%%r15d                   \n"
+    ".p2align  5                               \n"
+    "lea       (%%rax),%%eax                   \n"
+    "lea       (%%rbx),%%ebx                   \n"
+    "lea       (%%rcx),%%ecx                   \n"
+    "lea       (%%rdx),%%edx                   \n"
+    "lea       (%%rsi),%%esi                   \n"
+    "lea       (%%rdi),%%edi                   \n"
+    "lea       (%%rbp),%%ebp                   \n"
+    "lea       (%%rsp),%%esp                   \n"
+    ".p2align  5                               \n"
+    "lea       (%%r8),%%r8d                    \n"
+    "lea       (%%r9),%%r9d                    \n"
+    "lea       (%%r10),%%r10d                  \n"
+    "lea       (%%r11),%%r11d                  \n"
+    "lea       (%%r12),%%r12d                  \n"
+    "lea       (%%r13),%%r13d                  \n"
+    "lea       (%%r14),%%r14d                  \n"
+    "lea       (%%r15),%%r15d                  \n"
+
+    ".p2align  5                               \n"
+    "lea       0x10(%%rax),%%eax               \n"
+    "lea       0x10(%%rbx),%%ebx               \n"
+    "lea       0x10(%%rcx),%%ecx               \n"
+    "lea       0x10(%%rdx),%%edx               \n"
+    "lea       0x10(%%rsi),%%esi               \n"
+    "lea       0x10(%%rdi),%%edi               \n"
+    "lea       0x10(%%rbp),%%ebp               \n"
+    "lea       0x10(%%rsp),%%esp               \n"
+    ".p2align  5                               \n"
+    "lea       0x10(%%r8),%%r8d                \n"
+    "lea       0x10(%%r9),%%r9d                \n"
+    "lea       0x10(%%r10),%%r10d              \n"
+    "lea       0x10(%%r11),%%r11d              \n"
+    "lea       0x10(%%r12),%%r12d              \n"
+    "lea       0x10(%%r13),%%r13d              \n"
+    "lea       0x10(%%r14),%%r14d              \n"
+    "lea       0x10(%%r15),%%r15d              \n"
+
+    ".p2align  5                               \n"
+    "add       0x10,%%eax                      \n"
+    "add       0x10,%%ebx                      \n"
+    "add       0x10,%%ecx                      \n"
+    "add       0x10,%%edx                      \n"
+    "add       0x10,%%esi                      \n"
+    "add       0x10,%%edi                      \n"
+    "add       0x10,%%ebp                      \n"
+    "add       0x10,%%esp                      \n"
+    ".p2align  5                               \n"
+    "add       0x10,%%r8d                      \n"
+    "add       0x10,%%r9d                      \n"
+    "add       0x10,%%r10d                     \n"
+    "add       0x10,%%r11d                     \n"
+    "add       0x10,%%r12d                     \n"
+    "add       0x10,%%r13d                     \n"
+    "add       0x10,%%r14d                     \n"
+    "add       0x10,%%r15d                     \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // TESTING
+
+#ifdef HAS_I400TOARGBROW_SSE2
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb,
+                                  int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_I400TOARGBROW_SSE2
+
+#ifdef HAS_RGB24TOARGBROW_SSSE3
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
+    "pslld     $0x18,%%xmm5                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x30,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm2                   \n"
+    "palignr   $0x8,%%xmm1,%%xmm2              \n"
+    "pshufb    %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "palignr   $0x4,%%xmm3,%%xmm3              \n"
+    "pshufb    %%xmm4,%%xmm3                   \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "m"(kShuffleMaskRGB24ToARGB)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
+    "pslld     $0x18,%%xmm5                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x30,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm2                   \n"
+    "palignr   $0x8,%%xmm1,%%xmm2              \n"
+    "pshufb    %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "palignr   $0x4,%%xmm3,%%xmm3              \n"
+    "pshufb    %%xmm4,%%xmm3                   \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "m"(kShuffleMaskRAWToARGB)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RGB565ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0x1080108,%%eax                \n"
+    "movd      %%eax,%%xmm5                    \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "mov       $0x20802080,%%eax               \n"
+    "movd      %%eax,%%xmm6                    \n"
+    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psllw     $0xb,%%xmm3                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0xa,%%xmm4                     \n"
+    "psrlw     $0x5,%%xmm4                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psllw     $0x8,%%xmm7                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "psllw     $0xb,%%xmm2                     \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "psllw     $0x8,%%xmm1                     \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "pmulhuw   %%xmm6,%%xmm0                   \n"
+    "por       %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpckhbw %%xmm0,%%xmm2                   \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm1,0x00,1,0,2)           //  movdqa  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqa,xmm2,0x10,1,0,2)           //  movdqa  %%xmm2,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGB1555ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0x1080108,%%eax                \n"
+    "movd      %%eax,%%xmm5                    \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "mov       $0x42004200,%%eax               \n"
+    "movd      %%eax,%%xmm6                    \n"
+    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psllw     $0xb,%%xmm3                     \n"
+    "movdqa    %%xmm3,%%xmm4                   \n"
+    "psrlw     $0x6,%%xmm4                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psllw     $0x8,%%xmm7                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psllw     $0x1,%%xmm1                     \n"
+    "psllw     $0xb,%%xmm2                     \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "psllw     $0x8,%%xmm1                     \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "pmulhuw   %%xmm6,%%xmm0                   \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpckhbw %%xmm0,%%xmm2                   \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm1,0x00,1,0,2)           //  movdqa  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqa,xmm2,0x10,1,0,2)           //  movdqa  %%xmm2,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGB4444ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0xf0f0f0f,%%eax                \n"
+    "movd      %%eax,%%xmm4                    \n"
+    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "pslld     $0x4,%%xmm5                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "psllw     $0x4,%%xmm1                     \n"
+    "psrlw     $0x4,%%xmm3                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,2)           //  movdqa  %%xmm0,(%1,%0,2)
+    MEMOPMEM(movdqa,xmm1,0x10,1,0,2)           //  movdqa  %%xmm1,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToRGB24Row_SSSE3(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "pshufb    %%xmm6,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm6,%%xmm2                   \n"
+    "pshufb    %%xmm6,%%xmm3                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "psrldq    $0x4,%%xmm1                     \n"
+    "pslldq    $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm2,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pslldq    $0x8,%%xmm5                     \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "psrldq    $0x8,%%xmm2                     \n"
+    "pslldq    $0x4,%%xmm3                     \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x30,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "m"(kShuffleMaskARGBToRGB24)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void ARGBToRAWRow_SSSE3(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "pshufb    %%xmm6,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm6,%%xmm2                   \n"
+    "pshufb    %%xmm6,%%xmm3                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "psrldq    $0x4,%%xmm1                     \n"
+    "pslldq    $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm2,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pslldq    $0x8,%%xmm5                     \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "psrldq    $0x8,%%xmm2                     \n"
+    "pslldq    $0x4,%%xmm3                     \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x30,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "m"(kShuffleMaskARGBToRAW)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void ARGBToRGB565Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psrld     $0x1b,%%xmm3                    \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psrld     $0x1a,%%xmm4                    \n"
+    "pslld     $0x5,%%xmm4                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0xb,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pslld     $0x8,%%xmm0                     \n"
+    "psrld     $0x3,%%xmm1                     \n"
+    "psrld     $0x5,%%xmm2                     \n"
+    "psrad     $0x10,%%xmm0                    \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm4,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psrld     $0x1b,%%xmm4                    \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "pslld     $0x5,%%xmm5                     \n"
+    "movdqa    %%xmm4,%%xmm6                   \n"
+    "pslld     $0xa,%%xmm6                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "pslld     $0xf,%%xmm7                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "psrad     $0x10,%%xmm0                    \n"
+    "psrld     $0x3,%%xmm1                     \n"
+    "psrld     $0x6,%%xmm2                     \n"
+    "psrld     $0x9,%%xmm3                     \n"
+    "pand      %%xmm7,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm6,%%xmm3                   \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMACCESS2(0x8,1) ",%1        \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToARGB4444Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm4,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm3,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm1                   \n"
+    "psrlq     $0x4,%%xmm0                     \n"
+    "psrlq     $0x8,%%xmm1                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+#endif  // HAS_RGB24TOARGBROW_SSSE3
+
+#ifdef HAS_ARGBTOYROW_SSSE3
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOYROW_SSSE3
+
+#ifdef HAS_ARGBTOYJROW_SSSE3
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),  // %3
+    "m"(kAddYJ64)    // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),  // %3
+    "m"(kAddYJ64)    // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOYJROW_SSSE3
+
+#ifdef HAS_ARGBTOUVROW_SSSE3
+// TODO(fbarchard): pass xmm constants to single block of assembly.
+// fpic on GCC 4.2 for OSX runs out of GPR registers. "m" effectively takes
+// 3 registers - ebx, ebp and eax. "m" can be passed with 3 normal registers,
+// or 4 if stack frame is disabled. Doing 2 assembly blocks is a work around
+// and considered unsafe.
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps    %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+// TODO(fbarchard): Share code with ARGBToUVRow_SSSE3.
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToUJ),  // %0
+    "m"(kARGBToVJ),  // %1
+    "m"(kAddUVJ128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),         // %0
+    "m"(kARGBToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                  uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToUJ),         // %0
+    "m"(kARGBToVJ),         // %1
+    "m"(kAddUVJ128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb))
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUV444Row_SSSE3(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                          int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,2,1)           //  movdqa  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),        // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6"
+#endif
+  );
+}
+
+void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_u,
+                                    uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,2,1)           //  movdqu  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),        // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6"
+#endif
+  );
+}
+
+void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0,
+                                    uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kBGRAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void BGRAToYRow_Unaligned_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kBGRAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void BGRAToUVRow_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kBGRAToU),         // %0
+    "m"(kBGRAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_bgra)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kBGRAToU),         // %0
+    "m"(kBGRAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_bgra)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kABGRToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ABGRToYRow_Unaligned_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kABGRToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kRGBAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RGBAToYRow_Unaligned_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kRGBAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ABGRToUVRow_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kABGRToU),         // %0
+    "m"(kABGRToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_abgr)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kABGRToU),         // %0
+    "m"(kABGRToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_abgr)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void RGBAToUVRow_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kRGBAToU),         // %0
+    "m"(kRGBAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_rgba))
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kRGBAToU),         // %0
+    "m"(kRGBAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_rgba)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOUVROW_SSSE3
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+#define UB 127 /* min(63,(int8)(2.018 * 64)) */
+#define UG -25 /* (int8)(-0.391 * 64 - 0.5) */
+#define UR 0
+
+#define VB 0
+#define VG -52 /* (int8)(-0.813 * 64 - 0.5) */
+#define VR 102 /* (int8)(1.596 * 64 + 0.5) */
+
+// Bias
+#define BB UB * 128 + VB * 128
+#define BG UG * 128 + VG * 128
+#define BR UR * 128 + VR * 128
+
+#define YG 74 /* (int8)(1.164 * 64 + 0.5) */
+
+struct {
+  vec8 kUVToB;  // 0
+  vec8 kUVToG;  // 16
+  vec8 kUVToR;  // 32
+  vec16 kUVBiasB;  // 48
+  vec16 kUVBiasG;  // 64
+  vec16 kUVBiasR;  // 80
+  vec16 kYSub16;  // 96
+  vec16 kYToRgb;  // 112
+  vec8 kVUToB;  // 128
+  vec8 kVUToG;  // 144
+  vec8 kVUToR;  // 160
+} static SIMD_ALIGNED(kYuvConstants) = {
+  { UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB },
+  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
+  { UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR },
+  { BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR },
+  { 16, 16, 16, 16, 16, 16, 16, 16 },
+  { YG, YG, YG, YG, YG, YG, YG, YG },
+  { VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB },
+  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
+  { VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR }
+};
+
+
+// Read 8 UV from 411
+#define READYUV444                                                             \
+    "movq       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    BUNDLEALIGN                                                                \
+    MEMOPREG(movq, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"
+
+// Read 4 UV from 422, upsample to 8 UV
+#define READYUV422                                                             \
+    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    BUNDLEALIGN                                                                \
+    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x4, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"
+
+// Read 2 UV from 411, upsample to 8 UV
+#define READYUV411                                                             \
+    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    BUNDLEALIGN                                                                \
+    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x2, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"            \
+    "punpckldq  %%xmm0,%%xmm0                                   \n"
+
+// Read 4 UV from NV12, upsample to 8 UV
+#define READNV12                                                               \
+    "movq       " MEMACCESS([uv_buf]) ",%%xmm0                  \n"            \
+    "lea        " MEMLEA(0x8, [uv_buf]) ",%[uv_buf]             \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"
+
+// Convert 8 pixels: 8 UV and 8 Y
+#define YUVTORGB                                                               \
+    "movdqa     %%xmm0,%%xmm1                                   \n"            \
+    "movdqa     %%xmm0,%%xmm2                                   \n"            \
+    "pmaddubsw  " MEMACCESS([kYuvConstants]) ",%%xmm0           \n"            \
+    "pmaddubsw  " MEMACCESS2(16, [kYuvConstants]) ",%%xmm1      \n"            \
+    "pmaddubsw  " MEMACCESS2(32, [kYuvConstants]) ",%%xmm2      \n"            \
+    "psubw      " MEMACCESS2(48, [kYuvConstants]) ",%%xmm0      \n"            \
+    "psubw      " MEMACCESS2(64, [kYuvConstants]) ",%%xmm1      \n"            \
+    "psubw      " MEMACCESS2(80, [kYuvConstants]) ",%%xmm2      \n"            \
+    "movq       " MEMACCESS([y_buf]) ",%%xmm3                   \n"            \
+    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
+    "punpcklbw  %%xmm4,%%xmm3                                   \n"            \
+    "psubsw     " MEMACCESS2(96, [kYuvConstants]) ",%%xmm3      \n"            \
+    "pmullw     " MEMACCESS2(112, [kYuvConstants]) ",%%xmm3     \n"            \
+    "paddsw     %%xmm3,%%xmm0                                   \n"            \
+    "paddsw     %%xmm3,%%xmm1                                   \n"            \
+    "paddsw     %%xmm3,%%xmm2                                   \n"            \
+    "psraw      $0x6,%%xmm0                                     \n"            \
+    "psraw      $0x6,%%xmm1                                     \n"            \
+    "psraw      $0x6,%%xmm2                                     \n"            \
+    "packuswb   %%xmm0,%%xmm0                                   \n"            \
+    "packuswb   %%xmm1,%%xmm1                                   \n"            \
+    "packuswb   %%xmm2,%%xmm2                                   \n"
+
+// Convert 8 pixels: 8 VU and 8 Y
+#define YVUTORGB                                                               \
+    "movdqa     %%xmm0,%%xmm1                                   \n"            \
+    "movdqa     %%xmm0,%%xmm2                                   \n"            \
+    "pmaddubsw  " MEMACCESS2(128, [kYuvConstants]) ",%%xmm0     \n"            \
+    "pmaddubsw  " MEMACCESS2(144, [kYuvConstants]) ",%%xmm1     \n"            \
+    "pmaddubsw  " MEMACCESS2(160, [kYuvConstants]) ",%%xmm2     \n"            \
+    "psubw      " MEMACCESS2(48, [kYuvConstants]) ",%%xmm0      \n"            \
+    "psubw      " MEMACCESS2(64, [kYuvConstants]) ",%%xmm1      \n"            \
+    "psubw      " MEMACCESS2(80, [kYuvConstants]) ",%%xmm2      \n"            \
+    "movq       " MEMACCESS([y_buf]) ",%%xmm3                   \n"            \
+    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
+    "punpcklbw  %%xmm4,%%xmm3                                   \n"            \
+    "psubsw     " MEMACCESS2(96, [kYuvConstants]) ",%%xmm3      \n"            \
+    "pmullw     " MEMACCESS2(112, [kYuvConstants]) ",%%xmm3     \n"            \
+    "paddsw     %%xmm3,%%xmm0                                   \n"            \
+    "paddsw     %%xmm3,%%xmm1                                   \n"            \
+    "paddsw     %%xmm3,%%xmm2                                   \n"            \
+    "psraw      $0x6,%%xmm0                                     \n"            \
+    "psraw      $0x6,%%xmm1                                     \n"            \
+    "psraw      $0x6,%%xmm2                                     \n"            \
+    "packuswb   %%xmm0,%%xmm0                                   \n"            \
+    "packuswb   %%xmm1,%%xmm1                                   \n"            \
+    "packuswb   %%xmm2,%%xmm2                                   \n"
+
+void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV444
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "   \n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb]  \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf,
+                                 const uint8* u_buf,
+                                 const uint8* v_buf,
+                                 uint8* dst_rgb24,
+                                 int width) {
+// fpic 32 bit gcc 4.2 on OSX runs out of GPR regs.
+#if defined(__i386__)
+  asm volatile (
+    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
+  :: [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
+    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24));
+#endif
+
+  asm volatile (
+#if !defined(__i386__)
+    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
+#endif
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "movq      %%xmm0," MEMACCESS([dst_rgb24]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_rgb24]) "\n"
+    "lea       " MEMLEA(0x18,[dst_rgb24]) ",%[dst_rgb24] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgb24]"+r"(dst_rgb24),  // %[dst_rgb24]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)
+#if !defined(__i386__)
+    , [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
+    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24)
+#endif
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void OMITFP I422ToRAWRow_SSSE3(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_raw,
+                               int width) {
+// fpic 32 bit gcc 4.2 on OSX runs out of GPR regs.
+#if defined(__i386__)
+  asm volatile (
+    "movdqa    %[kShuffleMaskARGBToRAW_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRAW],%%xmm6   \n"
+  :: [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
+    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW));
+#endif
+
+  asm volatile (
+#if !defined(__i386__)
+    "movdqa    %[kShuffleMaskARGBToRAW_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRAW],%%xmm6   \n"
+#endif
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "movq      %%xmm0," MEMACCESS([dst_raw]) " \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_raw]) "\n"
+    "lea       " MEMLEA(0x18,[dst_raw]) ",%[dst_raw] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_raw]"+r"(dst_raw),  // %[dst_raw]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)
+#if !defined(__i386__)
+    , [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
+    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW)
+#endif
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void OMITFP I422ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I411ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV411
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* uv_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV21ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* uv_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YVUTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I444ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV444
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I411ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV411
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV12ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* uv_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV21ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* uv_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YVUTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToBGRARow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_bgra,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm5," MEMACCESS([dst_bgra]) "\n"
+    "movdqa    %%xmm0," MEMACCESS2(0x10,[dst_bgra]) "\n"
+    "lea       " MEMLEA(0x20,[dst_bgra]) ",%[dst_bgra] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToABGRRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_abgr,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm2                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2," MEMACCESS([dst_abgr]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_abgr]) "\n"
+    "lea       " MEMLEA(0x20,[dst_abgr]) ",%[dst_abgr] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_rgba,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm2,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm5," MEMACCESS([dst_rgba]) "\n"
+    "movdqa    %%xmm0," MEMACCESS2(0x10,[dst_rgba]) "\n"
+    "lea       " MEMLEA(0x20,[dst_rgba]) ",%[dst_rgba] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToBGRARow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_bgra,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm5," MEMACCESS([dst_bgra]) "\n"
+    "movdqu    %%xmm0," MEMACCESS2(0x10,[dst_bgra]) "\n"
+    "lea       " MEMLEA(0x20,[dst_bgra]) ",%[dst_bgra] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToABGRRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_abgr,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm2                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2," MEMACCESS([dst_abgr]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_abgr]) "\n"
+    "lea       " MEMLEA(0x20,[dst_abgr]) ",%[dst_abgr] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToRGBARow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_rgba,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm2,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm5," MEMACCESS([dst_rgba]) "\n"
+    "movdqu    %%xmm0," MEMACCESS2(0x10,[dst_rgba]) "\n"
+    "lea       " MEMLEA(0x20,[dst_rgba]) ",%[dst_rgba] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+#ifdef HAS_YTOARGBROW_SSE2
+void YToARGBRow_SSE2(const uint8* y_buf,
+                     uint8* dst_argb,
+                     int width) {
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "mov       $0x00100010,%%eax               \n"
+    "movd      %%eax,%%xmm3                    \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "mov       $0x004a004a,%%eax               \n"
+    "movd      %%eax,%%xmm2                    \n"
+    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    LABELALIGN
+  "1:                                          \n"
+    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "psubusw   %%xmm3,%%xmm0                   \n"
+    "pmullw    %%xmm2,%%xmm0                   \n"
+    "psrlw     $6, %%xmm0                      \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+
+    // Step 2: Weave into ARGB
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "por       %%xmm4,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(y_buf),     // %0
+    "+r"(dst_argb),  // %1
+    "+rm"(width)     // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+#endif  // HAS_YTOARGBROW_SSE2
+
+#ifdef HAS_MIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static uvec8 kShuffleMirror = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "movdqa    %3,%%xmm5                       \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(movdqa,0x00,0,2,1,xmm0)           //  movdqa  (%0,%2),%%xmm0
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kShuffleMirror) // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_MIRRORROW_SSSE3
+
+#ifdef HAS_MIRRORROW_SSE2
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(movdqu,0x00,0,2,1,xmm0)           //  movdqu  (%0,%2),%%xmm0
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "psllw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufd    $0x4e,%%xmm0,%%xmm0             \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1)",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_MIRRORROW_SSE2
+
+#ifdef HAS_MIRRORROW_UV_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static uvec8 kShuffleMirrorUV = {
+  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
+};
+void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
+                       int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "movdqa    %4,%%xmm1                       \n"
+    "lea       " MEMLEA4(-0x10,0,3,2) ",%0       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(-0x10,0) ",%0            \n"
+    "pshufb    %%xmm1,%%xmm0                   \n"
+    "sub       $8,%3                           \n"
+    "movlpd    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhpd,xmm0,0x00,1,2,1)           //  movhpd    %%xmm0,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src),      // %0
+    "+r"(dst_u),    // %1
+    "+r"(dst_v),    // %2
+    "+r"(temp_width)  // %3
+  : "m"(kShuffleMirrorUV)  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_MIRRORROW_UV_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static uvec8 kARGBShuffleMirror = {
+  12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u, 4u, 5u, 6u, 7u, 0u, 1u, 2u, 3u
+};
+
+void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "lea       " MEMLEA4(-0x10,0,2,4) ",%0     \n"
+    "movdqa    %3,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kARGBShuffleMirror)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_SSSE3
+
+#ifdef HAS_SPLITUVROW_SSE2
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb    %%xmm5,%%xmm5                    \n"
+    "psrlw      $0x8,%%xmm5                      \n"
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqa     " MEMACCESS(0) ",%%xmm0          \n"
+    "movdqa     " MEMACCESS2(0x10,0) ",%%xmm1    \n"
+    "lea        " MEMLEA(0x20,0) ",%0            \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "pand       %%xmm5,%%xmm0                    \n"
+    "pand       %%xmm5,%%xmm1                    \n"
+    "packuswb   %%xmm1,%%xmm0                    \n"
+    "psrlw      $0x8,%%xmm2                      \n"
+    "psrlw      $0x8,%%xmm3                      \n"
+    "packuswb   %%xmm3,%%xmm2                    \n"
+    "movdqa     %%xmm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(movdqa,xmm2,0x00,1,2,1)             // movdqa     %%xmm2,(%1,%2)
+    "lea        " MEMLEA(0x10,1) ",%1            \n"
+    "sub        $0x10,%3                         \n"
+    "jg         1b                               \n"
+  : "+r"(src_uv),     // %0
+    "+r"(dst_u),      // %1
+    "+r"(dst_v),      // %2
+    "+r"(pix)         // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix) {
+  asm volatile (
+    "pcmpeqb    %%xmm5,%%xmm5                    \n"
+    "psrlw      $0x8,%%xmm5                      \n"
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqu     " MEMACCESS(0) ",%%xmm0          \n"
+    "movdqu     " MEMACCESS2(0x10,0) ",%%xmm1    \n"
+    "lea        " MEMLEA(0x20,0) ",%0            \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "pand       %%xmm5,%%xmm0                    \n"
+    "pand       %%xmm5,%%xmm1                    \n"
+    "packuswb   %%xmm1,%%xmm0                    \n"
+    "psrlw      $0x8,%%xmm2                      \n"
+    "psrlw      $0x8,%%xmm3                      \n"
+    "packuswb   %%xmm3,%%xmm2                    \n"
+    "movdqu     %%xmm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(movdqu,xmm2,0x00,1,2,1)             //  movdqu     %%xmm2,(%1,%2)
+    "lea        " MEMLEA(0x10,1) ",%1            \n"
+    "sub        $0x10,%3                         \n"
+    "jg         1b                               \n"
+  : "+r"(src_uv),     // %0
+    "+r"(dst_u),      // %1
+    "+r"(dst_v),      // %2
+    "+r"(pix)         // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SPLITUVROW_SSE2
+
+#ifdef HAS_MERGEUVROW_SSE2
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    "sub       %0,%1                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0           \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)             //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm2                     \n"
+    "punpcklbw %%xmm1,%%xmm0                     \n"
+    "punpckhbw %%xmm1,%%xmm2                     \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "           \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x10,2) "     \n"
+    "lea       " MEMLEA(0x20,2) ",%2             \n"
+    "sub       $0x10,%3                          \n"
+    "jg        1b                                \n"
+  : "+r"(src_u),     // %0
+    "+r"(src_v),     // %1
+    "+r"(dst_uv),    // %2
+    "+r"(width)      // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+
+void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
+                               uint8* dst_uv, int width) {
+  asm volatile (
+    "sub       %0,%1                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)             //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm2                     \n"
+    "punpcklbw %%xmm1,%%xmm0                     \n"
+    "punpckhbw %%xmm1,%%xmm2                     \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "           \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,2) "     \n"
+    "lea       " MEMLEA(0x20,2) ",%2             \n"
+    "sub       $0x10,%3                          \n"
+    "jg        1b                                \n"
+  : "+r"(src_u),     // %0
+    "+r"(src_v),     // %1
+    "+r"(dst_uv),    // %2
+    "+r"(width)      // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_MERGEUVROW_SSE2
+
+#ifdef HAS_COPYROW_SSE2
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_COPYROW_SSE2
+
+#ifdef HAS_COPYROW_X86
+void CopyRow_X86(const uint8* src, uint8* dst, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "shr       $0x2,%2                         \n"
+    "rep movsl " MEMMOVESTRING(0,1) "          \n"
+  : "+S"(src),  // %0
+    "+D"(dst),  // %1
+    "+c"(width_tmp) // %2
+  :
+  : "memory", "cc"
+  );
+}
+#endif  // HAS_COPYROW_X86
+
+#ifdef HAS_COPYROW_ERMS
+// Unaligned Multiple of 1.
+void CopyRow_ERMS(const uint8* src, uint8* dst, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep movsb " MEMMOVESTRING(0,1) "          \n"
+  : "+S"(src),  // %0
+    "+D"(dst),  // %1
+    "+c"(width_tmp) // %2
+  :
+  : "memory", "cc"
+  );
+}
+#endif  // HAS_COPYROW_ERMS
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm0,%%xmm0                   \n"
+    "pslld     $0x18,%%xmm0                    \n"
+    "pcmpeqb   %%xmm1,%%xmm1                   \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqa    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "pand      %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm0,%%xmm3                   \n"
+    "pand      %%xmm1,%%xmm4                   \n"
+    "pand      %%xmm1,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYALPHAROW_AVX2
+// width in pixels
+void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm1         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm2   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
+    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
+    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYALPHAROW_AVX2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
+// width in pixels
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm0,%%xmm0                   \n"
+    "pslld     $0x18,%%xmm0                    \n"
+    "pcmpeqb   %%xmm1,%%xmm1                   \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpckhwd %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm2,%%xmm2                   \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqa    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "pand      %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm0,%%xmm3                   \n"
+    "pand      %%xmm1,%%xmm4                   \n"
+    "pand      %%xmm1,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
+// width in pixels
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vpmovzxbd " MEMACCESS(0) ",%%ymm1         \n"
+    "vpmovzxbd " MEMACCESS2(0x8,0) ",%%ymm2    \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "vpslld    $0x18,%%ymm1,%%ymm1             \n"
+    "vpslld    $0x18,%%ymm2,%%ymm2             \n"
+    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
+    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
+    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
+
+#ifdef HAS_SETROW_X86
+void SetRow_X86(uint8* dst, uint32 v32, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "shr       $0x2,%1                         \n"
+    "rep stosl " MEMSTORESTRING(eax,0) "       \n"
+    : "+D"(dst),       // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v32)         // %2
+    : "memory", "cc");
+}
+
+void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
+                   int dst_stride, int height) {
+  for (int y = 0; y < height; ++y) {
+    size_t width_tmp = (size_t)(width);
+    uint32* d = (uint32*)(dst);
+    asm volatile (
+      "rep stosl " MEMSTORESTRING(eax,0) "     \n"
+      : "+D"(d),         // %0
+        "+c"(width_tmp)  // %1
+      : "a"(v32)         // %2
+      : "memory", "cc");
+    dst += dst_stride;
+  }
+}
+#endif  // HAS_SETROW_X86
+
+#ifdef HAS_YUY2TOYROW_SSE2
+void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
+                               uint8* dst_y, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2,
+                                int stride_yuy2,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
+                               uint8* dst_y, int pix) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_YUY2TOYROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSE2
+// Blend 8 pixels at a time.
+void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0xf,%%xmm7                     \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x8,%%xmm6                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psllw     $0x8,%%xmm5                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "sub       $0x1,%3                         \n"
+    "je        91f                             \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop until destination pointer is aligned.
+  "10:                                         \n"
+    "test      $0xf,%2                         \n"
+    "je        19f                             \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+    "add       $1-4,%3                         \n"
+    "jl        49f                             \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "41:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       41b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop.
+  "91:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       91b                             \n"
+  "99:                                         \n"
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBBLENDROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSSE3
+// Shuffle table for isolating alpha.
+static uvec8 kShuffleAlpha = {
+  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
+  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
+};
+
+// Blend 8 pixels at a time
+// Shuffle table for reversing the bytes.
+
+// Same as SSE2, but replaces
+//    psrlw      xmm3, 8          // alpha
+//    pshufhw    xmm3, xmm3,0F5h  // 8 alpha words
+//    pshuflw    xmm3, xmm3,0F5h
+// with..
+//    pshufb     xmm3, kShuffleAlpha // alpha
+
+void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
+                        uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0xf,%%xmm7                     \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x8,%%xmm6                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psllw     $0x8,%%xmm5                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "sub       $0x1,%3                         \n"
+    "je        91f                             \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop until destination pointer is aligned.
+  "10:                                         \n"
+    "test      $0xf,%2                         \n"
+    "je        19f                             \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+    "add       $1-4,%3                         \n"
+    "jl        49f                             \n"
+    "test      $0xf,%0                         \n"
+    "jne       41f                             \n"
+    "test      $0xf,%1                         \n"
+    "jne       41f                             \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "40:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       40b                             \n"
+    "jmp       49f                             \n"
+
+    // 4 pixel unaligned loop.
+    LABELALIGN
+  "41:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       41b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop.
+  "91:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       91b                             \n"
+  "99:                                         \n"
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  : "m"(kShuffleAlpha)  // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBBLENDROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+// Attenuate 4 pixels at a time.
+// aligned to 16 bytes
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x8,%%xmm5                     \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "pshufhw   $0xff,%%xmm0,%%xmm2             \n"
+    "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pshufhw   $0xff,%%xmm1,%%xmm2             \n"
+    "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "pand      %%xmm4,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+// Shuffle table duplicating alpha
+static uvec8 kShuffleAlpha0 = {
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+};
+static uvec8 kShuffleAlpha1 = {
+  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+};
+// Attenuate 4 pixels at a time.
+// aligned to 16 bytes
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "pslld     $0x18,%%xmm3                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpcklbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm1,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "punpckhbw %%xmm2,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "pand      %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  : "m"(kShuffleAlpha0),  // %3
+    "m"(kShuffleAlpha1)  // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_SSSE3
+
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+// Unattenuate 4 pixels at a time.
+// aligned to 16 bytes
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  uintptr_t alpha = 0;
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "movlhps   %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "movlhps   %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width),       // %2
+    "+r"(alpha)        // %3
+  : "r"(fixed_invtbl8)  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBUNATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBGRAYROW_SSSE3
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrld     $0x18,%%xmm2                    \n"
+    "psrld     $0x18,%%xmm3                    \n"
+    "packuswb  %%xmm3,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm3                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "m"(kARGBToYJ),   // %3
+    "m"(kAddYJ64)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBGRAYROW_SSSE3
+
+#ifdef HAS_ARGBSEPIAROW_SSSE3
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+// Constant for ARGB color to sepia tone
+static vec8 kARGBToSepiaB = {
+  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
+};
+
+static vec8 kARGBToSepiaG = {
+  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
+};
+
+static vec8 kARGBToSepiaR = {
+  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
+};
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+  asm volatile (
+    "movdqa    %2,%%xmm2                       \n"
+    "movdqa    %3,%%xmm3                       \n"
+    "movdqa    %4,%%xmm4                       \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "pmaddubsw %%xmm2,%%xmm6                   \n"
+    "phaddw    %%xmm6,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm3,%%xmm5                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm5                   \n"
+    "psrlw     $0x7,%%xmm5                     \n"
+    "packuswb  %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm4,%%xmm5                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm5                   \n"
+    "psrlw     $0x7,%%xmm5                     \n"
+    "packuswb  %%xmm5,%%xmm5                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "psrld     $0x18,%%xmm6                    \n"
+    "psrld     $0x18,%%xmm1                    \n"
+    "packuswb  %%xmm1,%%xmm6                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm5                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm5,%%xmm0                   \n"
+    "punpckhwd %%xmm5,%%xmm1                   \n"
+    "sub       $0x8,%1                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),      // %0
+    "+r"(width)          // %1
+  : "m"(kARGBToSepiaB),  // %2
+    "m"(kARGBToSepiaG),  // %3
+    "m"(kARGBToSepiaR)   // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+#endif  // HAS_ARGBSEPIAROW_SSSE3
+
+#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// Same as Sepia except matrix is provided.
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width) {
+  asm volatile (
+    "movdqu    " MEMACCESS(3) ",%%xmm5         \n"
+    "pshufd    $0x00,%%xmm5,%%xmm2             \n"
+    "pshufd    $0x55,%%xmm5,%%xmm3             \n"
+    "pshufd    $0xaa,%%xmm5,%%xmm4             \n"
+    "pshufd    $0xff,%%xmm5,%%xmm5             \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "pmaddubsw %%xmm2,%%xmm7                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "phaddsw   %%xmm7,%%xmm0                   \n"
+    "phaddsw   %%xmm1,%%xmm6                   \n"
+    "psraw     $0x6,%%xmm0                     \n"
+    "psraw     $0x6,%%xmm6                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm7                   \n"
+    "phaddsw   %%xmm7,%%xmm1                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm7                   \n"
+    "phaddsw   %%xmm7,%%xmm6                   \n"
+    "psraw     $0x6,%%xmm1                     \n"
+    "psraw     $0x6,%%xmm6                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "punpcklwd %%xmm1,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm6                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb),      // %1
+    "+r"(width)          // %2
+  : "r"(matrix_argb)     // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
+
+#ifdef HAS_ARGBQUANTIZEROW_SSE2
+// Quantize 4 ARGB pixels (16 bytes).
+// aligned to 16 bytes
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "movd      %2,%%xmm2                       \n"
+    "movd      %3,%%xmm3                       \n"
+    "movd      %4,%%xmm4                       \n"
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshufd    $0x44,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "pshufd    $0x44,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0x40,%%xmm4,%%xmm4             \n"
+    "pshufd    $0x44,%%xmm4,%%xmm4             \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "pslld     $0x18,%%xmm6                    \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "pmullw    %%xmm3,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm7         \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm6,%%xmm7                   \n"
+    "paddw     %%xmm4,%%xmm0                   \n"
+    "paddw     %%xmm4,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "por       %%xmm7,%%xmm0                   \n"
+    "sub       $0x4,%1                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBQUANTIZEROW_SSE2
+
+#ifdef HAS_ARGBSHADEROW_SSE2
+// Shade 4 pixels at a time by specified value.
+// Aligned to 16 bytes.
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "movd      %3,%%xmm2                       \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm2                  \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(value)       // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHADEROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+// Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpckhbw %%xmm5,%%xmm3                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_SSE2
+
+#ifdef HAS_ARGBADDROW_SSE2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_ARGBADDROW_SSE2
+
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels, 4 pixels at a time.
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psubusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
+#ifdef HAS_SOBELXROW_SSE2
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+    "sub       %0,%3                           \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "movq      " MEMACCESS2(0x2,0) ",%%xmm1    \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "psubw     %%xmm1,%%xmm0                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
+    MEMOPREG(movq,0x02,0,1,1,xmm2)             //  movq      0x2(%0,%1,1),%%xmm2
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "psubw     %%xmm2,%%xmm1                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movq,0x00,0,2,1,xmm2)             //  movq      (%0,%2,1),%%xmm2
+    MEMOPREG(movq,0x02,0,2,1,xmm3)             //  movq      0x2(%0,%2,1),%%xmm3
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "psubw     %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm1                   \n"
+    "pmaxsw    %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "sub       $0x8,%4                         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm0,0x00,0,3,1)             //  movq      %%xmm0,(%0,%3,1)
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "jg        1b                              \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SOBELXROW_SSE2
+
+#ifdef HAS_SOBELYROW_SSE2
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "psubw     %%xmm1,%%xmm0                   \n"
+    BUNDLEALIGN
+    "movq      " MEMACCESS2(0x1,0) ",%%xmm1    \n"
+    MEMOPREG(movq,0x01,0,1,1,xmm2)             //  movq      0x1(%0,%1,1),%%xmm2
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "psubw     %%xmm2,%%xmm1                   \n"
+    BUNDLEALIGN
+    "movq      " MEMACCESS2(0x2,0) ",%%xmm2    \n"
+    MEMOPREG(movq,0x02,0,1,1,xmm3)             //  movq      0x2(%0,%1,1),%%xmm3
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "psubw     %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm1                   \n"
+    "pmaxsw    %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "sub       $0x8,%3                         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm0,0x00,0,2,1)             //  movq      %%xmm0,(%0,%2,1)
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "jg        1b                              \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SOBELYROW_SSE2
+
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm2                   \n"
+    "punpckhbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm1                   \n"
+    "punpckhwd %%xmm2,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklwd %%xmm0,%%xmm3                   \n"
+    "punpckhwd %%xmm0,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm1," MEMACCESS(2) "         \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x10,2) "   \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x20,2) "   \n"
+    "movdqa    %%xmm0," MEMACCESS2(0x30,2) "   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SOBELROW_SSE2
+
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into a plane.
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_SOBELTOPLANEROW_SSE2
+
+#ifdef HAS_SOBELXYROW_SSE2
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "paddusb   %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "punpckhbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "punpcklbw %%xmm2,%%xmm4                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm6                   \n"
+    "punpcklwd %%xmm3,%%xmm6                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "movdqa    %%xmm1,%%xmm7                   \n"
+    "punpcklwd %%xmm0,%%xmm7                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm6," MEMACCESS(2) "         \n"
+    "movdqa    %%xmm4," MEMACCESS2(0x10,2) "   \n"
+    "movdqa    %%xmm7," MEMACCESS2(0x20,2) "   \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x30,2) "   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_SOBELXYROW_SSE2
+
+#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
+// Creates a table of cumulative sums where each value is a sum of all values
+// above and to the left of the value, inclusive of the value.
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width) {
+  asm volatile (
+    "pxor      %%xmm0,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+    "test      $0xf,%1                         \n"
+    "jne       49f                             \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm2,%%xmm4                   \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
+    "punpckhwd %%xmm1,%%xmm3                   \n"
+    "punpckhbw %%xmm1,%%xmm4                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "punpcklwd %%xmm1,%%xmm4                   \n"
+    "punpckhwd %%xmm1,%%xmm5                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(2) ",%%xmm2         \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "movdqa    " MEMACCESS2(0x10,2) ",%%xmm3   \n"
+    "paddd     %%xmm0,%%xmm3                   \n"
+    "paddd     %%xmm4,%%xmm0                   \n"
+    "movdqa    " MEMACCESS2(0x20,2) ",%%xmm4   \n"
+    "paddd     %%xmm0,%%xmm4                   \n"
+    "paddd     %%xmm5,%%xmm0                   \n"
+    "movdqa    " MEMACCESS2(0x30,2) ",%%xmm5   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "paddd     %%xmm0,%%xmm5                   \n"
+    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "movdqa    %%xmm4," MEMACCESS2(0x20,1) "   \n"
+    "movdqa    %%xmm5," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x1,%3                         \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+  : "+r"(row),  // %0
+    "+r"(cumsum),  // %1
+    "+r"(previous_cumsum),  // %2
+    "+r"(width)  // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
+
+#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst,
+                                    int count) {
+  asm volatile (
+    "movd      %5,%%xmm5                       \n"
+    "cvtdq2ps  %%xmm5,%%xmm5                   \n"
+    "rcpss     %%xmm5,%%xmm4                   \n"
+    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+    "cmpl      $0x80,%5                        \n"
+    "ja        40f                             \n"
+
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrld     $0x10,%%xmm6                    \n"
+    "cvtdq2ps  %%xmm6,%%xmm6                   \n"
+    "addps     %%xmm6,%%xmm5                   \n"
+    "mulps     %%xmm4,%%xmm5                   \n"
+    "cvtps2dq  %%xmm5,%%xmm5                   \n"
+    "packssdw  %%xmm5,%%xmm5                   \n"
+
+  // 4 pixel small loop                        \n"
+    LABELALIGN
+  "4:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
+    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
+    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
+    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
+    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
+    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
+    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm5,%%xmm0                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       4b                              \n"
+    "jmp       49f                             \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
+    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
+    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
+    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
+    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
+    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
+    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm1,%%xmm1                   \n"
+    "mulps     %%xmm4,%%xmm0                   \n"
+    "mulps     %%xmm4,%%xmm1                   \n"
+    "cvtdq2ps  %%xmm2,%%xmm2                   \n"
+    "cvtdq2ps  %%xmm3,%%xmm3                   \n"
+    "mulps     %%xmm4,%%xmm2                   \n"
+    "mulps     %%xmm4,%%xmm3                   \n"
+    "cvtps2dq  %%xmm0,%%xmm0                   \n"
+    "cvtps2dq  %%xmm1,%%xmm1                   \n"
+    "cvtps2dq  %%xmm2,%%xmm2                   \n"
+    "cvtps2dq  %%xmm3,%%xmm3                   \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm3,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    BUNDLEALIGN
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "mulps     %%xmm4,%%xmm0                   \n"
+    "cvtps2dq  %%xmm0,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
+    "jge       10b                             \n"
+  "19:                                         \n"
+  : "+r"(topleft),  // %0
+    "+r"(botleft),  // %1
+    "+r"(dst),      // %2
+    "+rm"(count)    // %3
+  : "r"((intptr_t)(width)),  // %4
+    "rm"(area)     // %5
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+#endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+
+#ifdef HAS_ARGBAFFINEROW_SSE2
+// Copy ARGB pixels from source image with slope to a row of destination.
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* src_dudv, int width) {
+  intptr_t src_argb_stride_temp = src_argb_stride;
+  intptr_t temp = 0;
+  asm volatile (
+    "movq      " MEMACCESS(3) ",%%xmm2         \n"
+    "movq      " MEMACCESS2(0x08,3) ",%%xmm7   \n"
+    "shl       $0x10,%1                        \n"
+    "add       $0x4,%1                         \n"
+    "movd      %1,%%xmm5                       \n"
+    "sub       $0x4,%4                         \n"
+    "jl        49f                             \n"
+
+    "pshufd    $0x44,%%xmm7,%%xmm7             \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm7,%%xmm0                   \n"
+    "movlhps   %%xmm0,%%xmm2                   \n"
+    "movdqa    %%xmm7,%%xmm4                   \n"
+    "addps     %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "addps     %%xmm4,%%xmm3                   \n"
+    "addps     %%xmm4,%%xmm4                   \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "cvttps2dq %%xmm2,%%xmm0                   \n"  // x, y float to int first 2
+    "cvttps2dq %%xmm3,%%xmm1                   \n"  // x, y float to int next 2
+    "packssdw  %%xmm1,%%xmm0                   \n"  // x, y as 8 shorts
+    "pmaddwd   %%xmm5,%%xmm0                   \n"  // off = x * 4 + y * stride
+    "movd      %%xmm0,%k1                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    "movd      %%xmm0,%k5                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
+    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
+    "punpckldq %%xmm6,%%xmm1                   \n"
+    "addps     %%xmm4,%%xmm2                   \n"
+    "movq      %%xmm1," MEMACCESS(2) "         \n"
+    "movd      %%xmm0,%k1                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    "movd      %%xmm0,%k5                      \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
+    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
+    "punpckldq %%xmm6,%%xmm0                   \n"
+    "addps     %%xmm4,%%xmm3                   \n"
+    "sub       $0x4,%4                         \n"
+    "movq      %%xmm0," MEMACCESS2(0x08,2) "   \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%4                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "cvttps2dq %%xmm2,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "pmaddwd   %%xmm5,%%xmm0                   \n"
+    "addps     %%xmm7,%%xmm2                   \n"
+    "movd      %%xmm0,%k1                      \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
+    "sub       $0x1,%4                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x04,2) ",%2           \n"
+    "jge       10b                             \n"
+  "19:                                         \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_argb_stride_temp),  // %1
+    "+r"(dst_argb),  // %2
+    "+r"(src_dudv),  // %3
+    "+rm"(width),    // %4
+    "+r"(temp)   // %5
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBAFFINEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm2)
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "pmaddubsw %%xmm5,%%xmm0                   \n"
+    "pmaddubsw %%xmm5,%%xmm1                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm0)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm2)           //  movdqa    (%1,%4,1),%%xmm2
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm2                   \n"
+    "psubw     %%xmm1,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm2                   \n"
+    "paddw     %%xmm3,%%xmm3                   \n"
+    "pmulhw    %%xmm5,%%xmm2                   \n"
+    "pmulhw    %%xmm5,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)           //  movdqa    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)           //  movdqa    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm0)           //  movdqa    (%1,%4,1),%%xmm0
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride, int dst_width,
+                                    int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "pmaddubsw %%xmm5,%%xmm0                   \n"
+    "pmaddubsw %%xmm5,%%xmm1                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm5"
+#endif
+  );
+}
+#endif   // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride, int dst_width,
+                                   int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)           //  movdqu    (%1,%4,1),%%xmm2
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm2                   \n"
+    "psubw     %%xmm1,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm2                   \n"
+    "paddw     %%xmm3,%%xmm3                   \n"
+    "pmulhw    %%xmm5,%%xmm2                   \n"
+    "pmulhw    %%xmm5,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)           //  movdqu    (%1,%4,1),%%xmm0
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+#ifdef HAS_HALFROW_SSE2
+void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(pavgb,0x00,0,3,1,xmm0)            //  pavgb     (%0,%3),%%xmm0
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqa,xmm0,0x00,0,1,1)           //  movdqa    %%xmm0,(%0,%1)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "jg        1b                              \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_uv),  // %1
+    "+r"(pix)      // %2
+  : "r"((intptr_t)(src_uv_stride))  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+      , "xmm0"
+#endif
+  );
+}
+#endif  // HAS_HALFROW_SSE2
+
+#ifdef HAS_ARGBTOBAYERROW_SSSE3
+void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                          uint32 selector, int pix) {
+  asm volatile (
+    // NaCL caveat - assumes movd is from GPR
+    "movd      %3,%%xmm5                       \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    "sub       $0x8,%2                         \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_bayer), // %1
+    "+r"(pix)        // %2
+  : "g"(selector)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOBAYERROW_SSSE3
+
+#ifdef HAS_ARGBTOBAYERGGROW_SSE2
+void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 selector, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrld     $0x8,%%xmm0                     \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x8,%2                         \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_bayer), // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOBAYERGGROW_SSE2
+
+#ifdef HAS_ARGBSHUFFLEROW_SSSE3
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix) {
+  asm volatile (
+    "movdqa    " MEMACCESS(3) ",%%xmm5         \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                    const uint8* shuffler, int pix) {
+  asm volatile (
+    "movdqa    " MEMACCESS(3) ",%%xmm5         \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_SSSE3
+
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    "vbroadcastf128 " MEMACCESS(3) ",%%ymm5    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpshufb   %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpshufb   %%ymm5,%%ymm1,%%ymm1            \n"
+    "sub       $0x10,%2                        \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_AVX2
+
+#ifdef HAS_ARGBSHUFFLEROW_SSE2
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "mov       " MEMACCESS(4) ",%k2            \n"
+    "cmp       $0x3000102,%k2                  \n"
+    "je        3012f                           \n"
+    "cmp       $0x10203,%k2                    \n"
+    "je        123f                            \n"
+    "cmp       $0x30201,%k2                    \n"
+    "je        321f                            \n"
+    "cmp       $0x2010003,%k2                  \n"
+    "je        2103f                           \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(4) ",%2             \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS(1) "            \n"
+    "movzb     " MEMACCESS2(0x1,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x1,1) "       \n"
+    BUNDLEALIGN
+    "movzb     " MEMACCESS2(0x2,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x2,1) "       \n"
+    "movzb     " MEMACCESS2(0x3,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x3,1) "       \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "sub       $0x1,%3                         \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "123:                                        \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x1b,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x1b,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        123b                            \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "321:                                        \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x39,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x39,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x39,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x39,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        321b                            \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "2103:                                       \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x93,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x93,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x93,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x93,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        2103b                           \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "3012:                                       \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0xc6,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0xc6,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0xc6,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0xc6,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        3012b                           \n"
+
+  "99:                                         \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+d"(pixel_temp),  // %2
+    "+r"(pix)         // %3
+  : "r"(shuffler)      // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_SSE2
+
+#ifdef HAS_I422TOYUY2ROW_SSE2
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+ asm volatile (
+    "sub       %1,%2                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movq      " MEMACCESS(1) ",%%xmm2           \n"
+    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
+    "lea       " MEMLEA(0x8,1) ",%1              \n"
+    "punpcklbw %%xmm3,%%xmm2                     \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm1                     \n"
+    "punpcklbw %%xmm2,%%xmm0                     \n"
+    "punpckhbw %%xmm2,%%xmm1                     \n"
+    "movdqu    %%xmm0," MEMACCESS(3) "           \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,3) "     \n"
+    "lea       " MEMLEA(0x20,3) ",%3             \n"
+    "sub       $0x10,%4                          \n"
+    "jg         1b                               \n"
+    : "+r"(src_y),  // %0
+      "+r"(src_u),  // %1
+      "+r"(src_v),  // %2
+      "+r"(dst_frame),  // %3
+      "+rm"(width)  // %4
+    :
+    : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+#endif  // HAS_I422TOYUY2ROW_SSE2
+
+#ifdef HAS_I422TOUYVYROW_SSE2
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+ asm volatile (
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movq      " MEMACCESS(1) ",%%xmm2           \n"
+    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
+    "lea       " MEMLEA(0x8,1) ",%1              \n"
+    "punpcklbw %%xmm3,%%xmm2                     \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    "movdqa    %%xmm2,%%xmm1                     \n"
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "punpcklbw %%xmm0,%%xmm1                     \n"
+    "punpckhbw %%xmm0,%%xmm2                     \n"
+    "movdqu    %%xmm1," MEMACCESS(3) "           \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,3) "     \n"
+    "lea       " MEMLEA(0x20,3) ",%3             \n"
+    "sub       $0x10,%4                          \n"
+    "jg         1b                               \n"
+    : "+r"(src_y),  // %0
+      "+r"(src_u),  // %1
+      "+r"(src_v),  // %2
+      "+r"(dst_frame),  // %3
+      "+rm"(width)  // %4
+    :
+    : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+#endif  // HAS_I422TOUYVYROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "pxor      %%xmm3,%%xmm3                   \n"
+
+    // 2 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm3,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm4                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "mulps     " MEMACCESS2(0x10,3) ",%%xmm0   \n"
+    "mulps     " MEMACCESS2(0x10,3) ",%%xmm4   \n"
+    "addps     " MEMACCESS(3) ",%%xmm0         \n"
+    "addps     " MEMACCESS(3) ",%%xmm4         \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm1,%%xmm2                   \n"
+    "mulps     %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm2,%%xmm1                   \n"
+    "mulps     %%xmm6,%%xmm5                   \n"
+    "mulps     " MEMACCESS2(0x20,3) ",%%xmm2   \n"
+    "mulps     " MEMACCESS2(0x20,3) ",%%xmm6   \n"
+    "mulps     " MEMACCESS2(0x30,3) ",%%xmm1   \n"
+    "mulps     " MEMACCESS2(0x30,3) ",%%xmm5   \n"
+    "addps     %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm6,%%xmm4                   \n"
+    "addps     %%xmm1,%%xmm0                   \n"
+    "addps     %%xmm5,%%xmm4                   \n"
+    "cvttps2dq %%xmm0,%%xmm0                   \n"
+    "cvttps2dq %%xmm4,%%xmm4                   \n"
+    "packuswb  %%xmm4,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "sub       $0x2,%2                         \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "vbroadcastf128 " MEMACCESS(3) ",%%ymm4     \n"
+    "vbroadcastf128 " MEMACCESS2(0x10,3) ",%%ymm5 \n"
+    "vbroadcastf128 " MEMACCESS2(0x20,3) ",%%ymm6 \n"
+    "vbroadcastf128 " MEMACCESS2(0x30,3) ",%%ymm7 \n"
+
+    // 2 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vpmovzxbd   " MEMACCESS(0) ",%%ymm0       \n"  // 2 ARGB pixels
+    "lea         " MEMLEA(0x8,0) ",%0          \n"
+    "vcvtdq2ps   %%ymm0,%%ymm0                 \n"  // X 8 floats
+    "vmulps      %%ymm0,%%ymm0,%%ymm2          \n"  // X * X
+    "vmulps      %%ymm7,%%ymm0,%%ymm3          \n"  // C3 * X
+    "vfmadd132ps %%ymm5,%%ymm4,%%ymm0          \n"  // result = C0 + C1 * X
+    "vfmadd231ps %%ymm6,%%ymm2,%%ymm0          \n"  // result += C2 * X * X
+    "vfmadd231ps %%ymm3,%%ymm2,%%ymm0          \n"  // result += C3 * X * X * X
+    "vcvttps2dq  %%ymm0,%%ymm0                 \n"
+    "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
+    "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+    "vpackuswb   %%xmm0,%%xmm0,%%xmm0          \n"
+    "sub         $0x2,%2                       \n"
+    "vmovq       %%xmm0," MEMACCESS(1) "       \n"
+    "lea         " MEMLEA(0x8,1) ",%1          \n"
+    "jg          1b                            \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+// TODO(fbarchard): declare ymm usage when applicable.
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
+#ifdef HAS_ARGBCOLORTABLEROW_X86
+// Tranform ARGB pixels with color table.
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+                           int width) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    // 1 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(0) ",%1             \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
+    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
+    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
+    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
+    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
+    "movzb     " MEMACCESS2(-0x1,0) ",%1       \n"
+    MEMOPARG(movzb,0x03,3,1,4,1) "             \n"  // movzb 0x3(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x1,0) "      \n"
+    "dec       %2                              \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),   // %0
+    "+d"(pixel_temp), // %1
+    "+r"(width)       // %2
+  : "r"(table_argb)   // %3
+  : "memory", "cc");
+}
+#endif  // HAS_ARGBCOLORTABLEROW_X86
+
+#ifdef HAS_RGBCOLORTABLEROW_X86
+// Tranform RGB pixels with color table.
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    // 1 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(0) ",%1             \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
+    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
+    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
+    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
+    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
+    "dec       %2                              \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),   // %0
+    "+d"(pixel_temp), // %1
+    "+r"(width)       // %2
+  : "r"(table_argb)   // %3
+  : "memory", "cc");
+}
+#endif  // HAS_RGBCOLORTABLEROW_X86
+
+#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// Tranform RGB pixels with luma table.
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff) {
+  uintptr_t pixel_temp = 0u;
+  uintptr_t table_temp = 0u;
+  asm volatile (
+    "movd      %6,%%xmm3                       \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0x8,%%xmm4                     \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm0         \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "phaddw    %%xmm0,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "punpcklwd %%xmm5,%%xmm0                   \n"
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS(2) ",%0             \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS(3) "            \n"
+    "movzb     " MEMACCESS2(0x1,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x1,3) "       \n"
+    "movzb     " MEMACCESS2(0x2,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x2,3) "       \n"
+    "movzb     " MEMACCESS2(0x3,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0x3,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS2(0x4,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x4,3) "       \n"
+    BUNDLEALIGN
+    "movzb     " MEMACCESS2(0x5,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x5,3) "       \n"
+    "movzb     " MEMACCESS2(0x6,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x6,3) "       \n"
+    "movzb     " MEMACCESS2(0x7,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0x7,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS2(0x8,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x8,3) "       \n"
+    "movzb     " MEMACCESS2(0x9,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x9,3) "       \n"
+    "movzb     " MEMACCESS2(0xa,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xa,3) "       \n"
+    "movzb     " MEMACCESS2(0xb,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0xb,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+
+    "movzb     " MEMACCESS2(0xc,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xc,3) "       \n"
+    "movzb     " MEMACCESS2(0xd,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xd,3) "       \n"
+    "movzb     " MEMACCESS2(0xe,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xe,3) "       \n"
+    "movzb     " MEMACCESS2(0xf,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0xf,3) "       \n"
+    "sub       $0x4,%4                         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "lea       " MEMLEA(0x10,3) ",%3           \n"
+    "jg        1b                              \n"
+  : "+d"(pixel_temp),  // %0
+    "+a"(table_temp),  // %1
+    "+r"(src_argb),    // %2
+    "+r"(dst_argb),    // %3
+    "+rm"(width)       // %4
+  : "r"(luma),         // %5
+    "rm"(lumacoeff)    // %6
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/row_win.cc b/third_party/libyuv/source/row_win.cc
new file mode 100644
index 0000000..2cfacad
--- /dev/null
+++ b/third_party/libyuv/source/row_win.cc
@@ -0,0 +1,7284 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+#ifdef HAS_ARGBTOYROW_SSSE3
+
+// Constants for ARGB.
+static const vec8 kARGBToY = {
+  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
+};
+
+// JPeg full range.
+static const vec8 kARGBToYJ = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+
+static const vec8 kARGBToU = {
+  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
+};
+
+static const vec8 kARGBToUJ = {
+  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
+};
+
+static const vec8 kARGBToV = {
+  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+};
+
+static const vec8 kARGBToVJ = {
+  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
+};
+
+// vpermd for vphaddw + vpackuswb vpermd.
+static const lvec32 kPermdARGBToY_AVX = {
+  0, 4, 1, 5, 2, 6, 3, 7
+};
+
+// vpshufb for vphaddw + vpackuswb packed to shorts.
+static const lvec8 kShufARGBToUV_AVX = {
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+};
+
+// Constants for BGRA.
+static const vec8 kBGRAToY = {
+  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
+};
+
+static const vec8 kBGRAToU = {
+  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
+};
+
+static const vec8 kBGRAToV = {
+  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
+};
+
+// Constants for ABGR.
+static const vec8 kABGRToY = {
+  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
+};
+
+static const vec8 kABGRToU = {
+  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
+};
+
+static const vec8 kABGRToV = {
+  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
+};
+
+// Constants for RGBA.
+static const vec8 kRGBAToY = {
+  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
+};
+
+static const vec8 kRGBAToU = {
+  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
+};
+
+static const vec8 kRGBAToV = {
+  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
+};
+
+static const uvec8 kAddY16 = {
+  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
+};
+
+static const vec16 kAddYJ64 = {
+  64, 64, 64, 64, 64, 64, 64, 64
+};
+
+static const uvec8 kAddUV128 = {
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
+};
+
+static const uvec16 kAddUVJ128 = {
+  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
+};
+
+// Shuffle table for converting RGB24 to ARGB.
+static const uvec8 kShuffleMaskRGB24ToARGB = {
+  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
+};
+
+// Shuffle table for converting RAW to ARGB.
+static const uvec8 kShuffleMaskRAWToARGB = {
+  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
+};
+
+// Shuffle table for converting ARGB to RGB24.
+static const uvec8 kShuffleMaskARGBToRGB24 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static const uvec8 kShuffleMaskARGBToRAW = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
+static const uvec8 kShuffleMaskARGBToRGB24_0 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static const uvec8 kShuffleMaskARGBToRAW_0 = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u
+};
+
+// Duplicates gray value 3 times and fills in alpha opaque.
+__declspec(naked) __declspec(align(16))
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]        // src_y
+    mov        edx, [esp + 8]        // dst_argb
+    mov        ecx, [esp + 12]       // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
+    pslld      xmm5, 24
+
+    align      4
+  convertloop:
+    movq       xmm0, qword ptr [eax]
+    lea        eax,  [eax + 8]
+    punpcklbw  xmm0, xmm0
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0
+    punpckhwd  xmm1, xmm1
+    por        xmm0, xmm5
+    por        xmm1, xmm5
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb,
+                                  int pix) {
+  __asm {
+    mov        eax, [esp + 4]        // src_y
+    mov        edx, [esp + 8]        // dst_argb
+    mov        ecx, [esp + 12]       // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
+    pslld      xmm5, 24
+
+    align      4
+  convertloop:
+    movq       xmm0, qword ptr [eax]
+    lea        eax,  [eax + 8]
+    punpcklbw  xmm0, xmm0
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0
+    punpckhwd  xmm1, xmm1
+    por        xmm0, xmm5
+    por        xmm1, xmm5
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_rgb24
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pslld     xmm5, 24
+    movdqa    xmm4, kShuffleMaskRGB24ToARGB
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm3, [eax + 32]
+    lea       eax, [eax + 48]
+    movdqa    xmm2, xmm3
+    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    pshufb    xmm2, xmm4
+    por       xmm2, xmm5
+    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    pshufb    xmm0, xmm4
+    movdqa    [edx + 32], xmm2
+    por       xmm0, xmm5
+    pshufb    xmm1, xmm4
+    movdqa    [edx], xmm0
+    por       xmm1, xmm5
+    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    pshufb    xmm3, xmm4
+    movdqa    [edx + 16], xmm1
+    por       xmm3, xmm5
+    sub       ecx, 16
+    movdqa    [edx + 48], xmm3
+    lea       edx, [edx + 64]
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
+                        int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_raw
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pslld     xmm5, 24
+    movdqa    xmm4, kShuffleMaskRAWToARGB
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm3, [eax + 32]
+    lea       eax, [eax + 48]
+    movdqa    xmm2, xmm3
+    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    pshufb    xmm2, xmm4
+    por       xmm2, xmm5
+    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    pshufb    xmm0, xmm4
+    movdqa    [edx + 32], xmm2
+    por       xmm0, xmm5
+    pshufb    xmm1, xmm4
+    movdqa    [edx], xmm0
+    por       xmm1, xmm5
+    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    pshufb    xmm3, xmm4
+    movdqa    [edx + 16], xmm1
+    por       xmm3, xmm5
+    sub       ecx, 16
+    movdqa    [edx + 48], xmm3
+    lea       edx, [edx + 64]
+    jg        convertloop
+    ret
+  }
+}
+
+// pmul method to replicate bits.
+// Math to replicate bits:
+// (v << 8) | (v << 3)
+// v * 256 + v * 8
+// v * (256 + 8)
+// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
+// 20 instructions.
+__declspec(naked) __declspec(align(16))
+void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+                          int pix) {
+  __asm {
+    mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
+    movd      xmm5, eax
+    pshufd    xmm5, xmm5, 0
+    mov       eax, 0x20802080  // multiplier shift by 5 and then repeat 6 bits
+    movd      xmm6, eax
+    pshufd    xmm6, xmm6, 0
+    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    psllw     xmm3, 11
+    pcmpeqb   xmm4, xmm4       // generate mask 0x07e007e0 for Green
+    psllw     xmm4, 10
+    psrlw     xmm4, 5
+    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    psllw     xmm7, 8
+
+    mov       eax, [esp + 4]   // src_rgb565
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of bgr565
+    movdqa    xmm1, xmm0
+    movdqa    xmm2, xmm0
+    pand      xmm1, xmm3    // R in upper 5 bits
+    psllw     xmm2, 11      // B in upper 5 bits
+    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    psllw     xmm1, 8
+    por       xmm1, xmm2    // RB
+    pand      xmm0, xmm4    // G in middle 6 bits
+    pmulhuw   xmm0, xmm6    // << 5 * (256 + 4)
+    por       xmm0, xmm7    // AG
+    movdqa    xmm2, xmm1
+    punpcklbw xmm1, xmm0
+    punpckhbw xmm2, xmm0
+    movdqa    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqa    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+// 24 instructions
+__declspec(naked) __declspec(align(16))
+void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
+    movd      xmm5, eax
+    pshufd    xmm5, xmm5, 0
+    mov       eax, 0x42004200  // multiplier shift by 6 and then repeat 5 bits
+    movd      xmm6, eax
+    pshufd    xmm6, xmm6, 0
+    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    psllw     xmm3, 11
+    movdqa    xmm4, xmm3       // generate mask 0x03e003e0 for Green
+    psrlw     xmm4, 6
+    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    psllw     xmm7, 8
+
+    mov       eax, [esp + 4]   // src_argb1555
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of 1555
+    movdqa    xmm1, xmm0
+    movdqa    xmm2, xmm0
+    psllw     xmm1, 1       // R in upper 5 bits
+    psllw     xmm2, 11      // B in upper 5 bits
+    pand      xmm1, xmm3
+    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    psllw     xmm1, 8
+    por       xmm1, xmm2    // RB
+    movdqa    xmm2, xmm0
+    pand      xmm0, xmm4    // G in middle 5 bits
+    psraw     xmm2, 8       // A
+    pmulhuw   xmm0, xmm6    // << 6 * (256 + 8)
+    pand      xmm2, xmm7
+    por       xmm0, xmm2    // AG
+    movdqa    xmm2, xmm1
+    punpcklbw xmm1, xmm0
+    punpckhbw xmm2, xmm0
+    movdqa    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqa    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+// 18 instructions.
+__declspec(naked) __declspec(align(16))
+void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax, 0x0f0f0f0f  // generate mask 0x0f0f0f0f
+    movd      xmm4, eax
+    pshufd    xmm4, xmm4, 0
+    movdqa    xmm5, xmm4       // 0xf0f0f0f0 for high nibbles
+    pslld     xmm5, 4
+    mov       eax, [esp + 4]   // src_argb4444
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of bgra4444
+    movdqa    xmm2, xmm0
+    pand      xmm0, xmm4    // mask low nibbles
+    pand      xmm2, xmm5    // mask high nibbles
+    movdqa    xmm1, xmm0
+    movdqa    xmm3, xmm2
+    psllw     xmm1, 4
+    psrlw     xmm3, 4
+    por       xmm0, xmm1
+    por       xmm2, xmm3
+    movdqa    xmm1, xmm0
+    punpcklbw xmm0, xmm2
+    punpckhbw xmm1, xmm2
+    movdqa    [eax * 2 + edx], xmm0  // store 4 pixels of ARGB
+    movdqa    [eax * 2 + edx + 16], xmm1  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    movdqa    xmm6, kShuffleMaskARGBToRGB24
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm2, [eax + 32]
+    movdqu    xmm3, [eax + 48]
+    lea       eax, [eax + 64]
+    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm1, xmm6
+    pshufb    xmm2, xmm6
+    pshufb    xmm3, xmm6
+    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
+    psrldq    xmm1, 4      // 8 bytes from 1
+    pslldq    xmm4, 12     // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
+    por       xmm0, xmm4   // 4 bytes from 1 for 0
+    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqu    [edx], xmm0  // store 0
+    por       xmm1, xmm5   // 8 bytes from 2 for 1
+    psrldq    xmm2, 8      // 4 bytes from 2
+    pslldq    xmm3, 4      // 12 bytes from 3 for 2
+    por       xmm2, xmm3   // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1   // store 1
+    movdqu    [edx + 32], xmm2   // store 2
+    lea       edx, [edx + 48]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    movdqa    xmm6, kShuffleMaskARGBToRAW
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm2, [eax + 32]
+    movdqu    xmm3, [eax + 48]
+    lea       eax, [eax + 64]
+    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm1, xmm6
+    pshufb    xmm2, xmm6
+    pshufb    xmm3, xmm6
+    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
+    psrldq    xmm1, 4      // 8 bytes from 1
+    pslldq    xmm4, 12     // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
+    por       xmm0, xmm4   // 4 bytes from 1 for 0
+    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqu    [edx], xmm0  // store 0
+    por       xmm1, xmm5   // 8 bytes from 2 for 1
+    psrldq    xmm2, 8      // 4 bytes from 2
+    pslldq    xmm3, 4      // 12 bytes from 3 for 2
+    por       xmm2, xmm3   // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1   // store 1
+    movdqu    [edx + 32], xmm2   // store 2
+    lea       edx, [edx + 48]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm3, xmm3       // generate mask 0x0000001f
+    psrld     xmm3, 27
+    pcmpeqb   xmm4, xmm4       // generate mask 0x000007e0
+    psrld     xmm4, 26
+    pslld     xmm4, 5
+    pcmpeqb   xmm5, xmm5       // generate mask 0xfffff800
+    pslld     xmm5, 11
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    pslld     xmm0, 8       // R
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 5       // G
+    psrad     xmm0, 16      // R
+    pand      xmm1, xmm3    // B
+    pand      xmm2, xmm4    // G
+    pand      xmm0, xmm5    // R
+    por       xmm1, xmm2    // BG
+    por       xmm0, xmm1    // BGR
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of RGB565
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+// TODO(fbarchard): Improve sign extension/packing.
+__declspec(naked) __declspec(align(16))
+void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm4, xmm4       // generate mask 0x0000001f
+    psrld     xmm4, 27
+    movdqa    xmm5, xmm4       // generate mask 0x000003e0
+    pslld     xmm5, 5
+    movdqa    xmm6, xmm4       // generate mask 0x00007c00
+    pslld     xmm6, 10
+    pcmpeqb   xmm7, xmm7       // generate mask 0xffff8000
+    pslld     xmm7, 15
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    movdqa    xmm3, xmm0    // R
+    psrad     xmm0, 16      // A
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 6       // G
+    psrld     xmm3, 9       // R
+    pand      xmm0, xmm7    // A
+    pand      xmm1, xmm4    // B
+    pand      xmm2, xmm5    // G
+    pand      xmm3, xmm6    // R
+    por       xmm0, xmm1    // BA
+    por       xmm2, xmm3    // GR
+    por       xmm0, xmm2    // BGRA
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB1555
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm4, xmm4       // generate mask 0xf000f000
+    psllw     xmm4, 12
+    movdqa    xmm3, xmm4       // generate mask 0x00f000f0
+    psrlw     xmm3, 8
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0
+    pand      xmm0, xmm3    // low nibble
+    pand      xmm1, xmm4    // high nibble
+    psrl      xmm0, 4
+    psrl      xmm1, 8
+    por       xmm0, xmm1
+    packuswb  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB4444
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+__declspec(naked) __declspec(align(16))
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kARGBToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+__declspec(naked) __declspec(align(16))
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    paddw      xmm0, xmm5  // Add .5 for rounding.
+    paddw      xmm2, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked) __declspec(align(32))
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vbroadcastf128 ymm4, kARGBToY
+    vbroadcastf128 ymm5, kAddY16
+    vmovdqa    ymm6, kPermdARGBToY_AVX
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    vpaddb     ymm0, ymm0, ymm5
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYROW_AVX2
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked) __declspec(align(32))
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vbroadcastf128 ymm4, kARGBToYJ
+    vbroadcastf128 ymm5, kAddYJ64
+    vmovdqa    ymm6, kPermdARGBToY_AVX
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpaddw     ymm0, ymm0, ymm5  // Add .5 for rounding.
+    vpaddw     ymm2, ymm2, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYJROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kARGBToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    paddw      xmm0, xmm5
+    paddw      xmm2, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kBGRAToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kBGRAToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kABGRToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kABGRToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kRGBAToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kRGBAToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToUJ
+    movdqa     xmm6, kARGBToVJ
+    movdqa     xmm5, kAddUVJ128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    paddw      xmm0, xmm5            // +.5 rounding -> unsigned
+    paddw      xmm1, xmm5
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+__declspec(naked) __declspec(align(32))
+void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    vbroadcastf128 ymm5, kAddUV128
+    vbroadcastf128 ymm6, kARGBToV
+    vbroadcastf128 ymm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 32x2 argb pixels to 16x1 */
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    vpavgb     ymm2, ymm2, [eax + esi + 64]
+    vpavgb     ymm3, ymm3, [eax + esi + 96]
+    lea        eax,  [eax + 128]
+    vshufps    ymm4, ymm0, ymm1, 0x88
+    vshufps    ymm0, ymm0, ymm1, 0xdd
+    vpavgb     ymm0, ymm0, ymm4  // mutated by vshufps
+    vshufps    ymm4, ymm2, ymm3, 0x88
+    vshufps    ymm2, ymm2, ymm3, 0xdd
+    vpavgb     ymm2, ymm2, ymm4  // mutated by vshufps
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 32 different pixels, its 16 pixels of U and 16 of V
+    vpmaddubsw ymm1, ymm0, ymm7  // U
+    vpmaddubsw ymm3, ymm2, ymm7
+    vpmaddubsw ymm0, ymm0, ymm6  // V
+    vpmaddubsw ymm2, ymm2, ymm6
+    vphaddw    ymm1, ymm1, ymm3  // mutates
+    vphaddw    ymm0, ymm0, ymm2
+    vpsraw     ymm1, ymm1, 8
+    vpsraw     ymm0, ymm0, 8
+    vpacksswb  ymm0, ymm1, ymm0  // mutates
+    vpermq     ymm0, ymm0, 0xd8  // For vpacksswb
+    vpshufb    ymm0, ymm0, kShufARGBToUV_AVX  // For vshufps + vphaddw
+    vpaddb     ymm0, ymm0, ymm5  // -> unsigned
+
+    // step 3 - store 16 U and 16 V values
+    sub         ecx, 32
+    vextractf128 [edx], ymm0, 0 // U
+    vextractf128 [edx + edi], ymm0, 1 // V
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOUVROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToUJ
+    movdqa     xmm6, kARGBToVJ
+    movdqa     xmm5, kAddUVJ128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    paddw      xmm0, xmm5            // +.5 rounding -> unsigned
+    paddw      xmm1, xmm5
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* convert to U and V */
+    movdqa     xmm0, [eax]          // U
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm7
+    pmaddubsw  xmm1, xmm7
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm3, xmm7
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx,  16
+    movdqa     [edx], xmm0
+
+    movdqa     xmm0, [eax]          // V
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm6
+    pmaddubsw  xmm1, xmm6
+    pmaddubsw  xmm2, xmm6
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    lea        eax,  [eax + 64]
+    movdqa     [edx + edi], xmm0
+    lea        edx,  [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb0,
+                                    uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* convert to U and V */
+    movdqu     xmm0, [eax]          // U
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm7
+    pmaddubsw  xmm1, xmm7
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm3, xmm7
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx,  16
+    movdqu     [edx], xmm0
+
+    movdqu     xmm0, [eax]          // V
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm6
+    pmaddubsw  xmm1, xmm6
+    pmaddubsw  xmm2, xmm6
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    lea        eax,  [eax + 64]
+    movdqu     [edx + edi], xmm0
+    lea        edx,  [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0,
+                                    uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kBGRAToU
+    movdqa     xmm6, kBGRAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kBGRAToU
+    movdqa     xmm6, kBGRAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kABGRToU
+    movdqa     xmm6, kABGRToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kABGRToU
+    movdqa     xmm6, kABGRToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kRGBAToU
+    movdqa     xmm6, kRGBAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kRGBAToU
+    movdqa     xmm6, kRGBAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBTOYROW_SSSE3
+
+#define YG 74 /* (int8)(1.164 * 64 + 0.5) */
+
+#define UB 127 /* min(63,(int8)(2.018 * 64)) */
+#define UG -25 /* (int8)(-0.391 * 64 - 0.5) */
+#define UR 0
+
+#define VB 0
+#define VG -52 /* (int8)(-0.813 * 64 - 0.5) */
+#define VR 102 /* (int8)(1.596 * 64 + 0.5) */
+
+// Bias
+#define BB UB * 128 + VB * 128
+#define BG UG * 128 + VG * 128
+#define BR UR * 128 + VR * 128
+
+#ifdef HAS_I422TOARGBROW_AVX2
+
+static const lvec8 kUVToB_AVX = {
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB,
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB
+};
+static const lvec8 kUVToR_AVX = {
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR,
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR
+};
+static const lvec8 kUVToG_AVX = {
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG
+};
+static const lvec16 kYToRgb_AVX = {
+  YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG
+};
+static const lvec16 kYSub16_AVX = {
+  16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16
+};
+static const lvec16 kUVBiasB_AVX = {
+  BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB
+};
+static const lvec16 kUVBiasG_AVX = {
+  BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG
+};
+static const lvec16 kUVBiasR_AVX = {
+  BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR
+};
+
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToARGBRow_AVX2(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpxor      ymm4, ymm4, ymm4
+
+    align      4
+ convertloop:
+    vmovq      xmm0, qword ptr [esi]          //  U
+    vmovq      xmm1, qword ptr [esi + edi]    //  V
+    lea        esi,  [esi + 8]
+    vpunpcklbw ymm0, ymm0, ymm1               // UV
+    vpermq     ymm0, ymm0, 0xd8
+    vpunpcklwd ymm0, ymm0, ymm0              // UVUV
+    vpmaddubsw ymm2, ymm0, kUVToB_AVX        // scale B UV
+    vpmaddubsw ymm1, ymm0, kUVToG_AVX        // scale G UV
+    vpmaddubsw ymm0, ymm0, kUVToR_AVX        // scale R UV
+    vpsubw     ymm2, ymm2, kUVBiasB_AVX      // unbias back to signed
+    vpsubw     ymm1, ymm1, kUVBiasG_AVX
+    vpsubw     ymm0, ymm0, kUVBiasR_AVX
+
+    // Step 2: Find Y contribution to 16 R,G,B values
+    vmovdqu    xmm3, [eax]                  // NOLINT
+    lea        eax, [eax + 16]
+    vpermq     ymm3, ymm3, 0xd8
+    vpunpcklbw ymm3, ymm3, ymm4
+    vpsubsw    ymm3, ymm3, kYSub16_AVX
+    vpmullw    ymm3, ymm3, kYToRgb_AVX
+    vpaddsw    ymm2, ymm2, ymm3           // B += Y
+    vpaddsw    ymm1, ymm1, ymm3           // G += Y
+    vpaddsw    ymm0, ymm0, ymm3           // R += Y
+    vpsraw     ymm2, ymm2, 6
+    vpsraw     ymm1, ymm1, 6
+    vpsraw     ymm0, ymm0, 6
+    vpackuswb  ymm2, ymm2, ymm2           // B
+    vpackuswb  ymm1, ymm1, ymm1           // G
+    vpackuswb  ymm0, ymm0, ymm0           // R
+
+    // Step 3: Weave into ARGB
+    vpunpcklbw ymm2, ymm2, ymm1           // BG
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklbw ymm0, ymm0, ymm5           // RA
+    vpermq     ymm0, ymm0, 0xd8
+    vpunpcklwd ymm1, ymm2, ymm0           // BGRA first 8 pixels
+    vpunpckhwd ymm2, ymm2, ymm0           // BGRA next 8 pixels
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+    vzeroupper
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_I422TOARGBROW_AVX2
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+
+static const vec8 kUVToB = {
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB
+};
+
+static const vec8 kUVToR = {
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR
+};
+
+static const vec8 kUVToG = {
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG
+};
+
+static const vec8 kVUToB = {
+  VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB,
+};
+
+static const vec8 kVUToR = {
+  VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR,
+};
+
+static const vec8 kVUToG = {
+  VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
+};
+
+static const vec16 kYToRgb = { YG, YG, YG, YG, YG, YG, YG, YG };
+static const vec16 kYSub16 = { 16, 16, 16, 16, 16, 16, 16, 16 };
+static const vec16 kUVBiasB = { BB, BB, BB, BB, BB, BB, BB, BB };
+static const vec16 kUVBiasG = { BG, BG, BG, BG, BG, BG, BG, BG };
+static const vec16 kUVBiasR = { BR, BR, BR, BR, BR, BR, BR, BR };
+
+// TODO(fbarchard): Read that does half size on Y and treats 420 as 444.
+
+// Read 8 UV from 444.
+#define READYUV444 __asm {                                                     \
+    __asm movq       xmm0, qword ptr [esi] /* U */                /* NOLINT */ \
+    __asm movq       xmm1, qword ptr [esi + edi] /* V */          /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+  }
+
+// Read 4 UV from 422, upsample to 8 UV.
+#define READYUV422 __asm {                                                     \
+    __asm movd       xmm0, [esi]          /* U */                              \
+    __asm movd       xmm1, [esi + edi]    /* V */                              \
+    __asm lea        esi,  [esi + 4]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Read 2 UV from 411, upsample to 8 UV.
+#define READYUV411 __asm {                                                     \
+    __asm movzx      ebx, word ptr [esi]        /* U */           /* NOLINT */ \
+    __asm movd       xmm0, ebx                                                 \
+    __asm movzx      ebx, word ptr [esi + edi]  /* V */           /* NOLINT */ \
+    __asm movd       xmm1, ebx                                                 \
+    __asm lea        esi,  [esi + 2]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+    __asm punpckldq  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Read 4 UV from NV12, upsample to 8 UV.
+#define READNV12 __asm {                                                       \
+    __asm movq       xmm0, qword ptr [esi] /* UV */               /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Convert 8 pixels: 8 UV and 8 Y.
+#define YUVTORGB __asm {                                                       \
+    /* Step 1: Find 4 UV contributions to 8 R,G,B values */                    \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm movdqa     xmm2, xmm0                                                \
+    __asm pmaddubsw  xmm0, kUVToB        /* scale B UV */                      \
+    __asm pmaddubsw  xmm1, kUVToG        /* scale G UV */                      \
+    __asm pmaddubsw  xmm2, kUVToR        /* scale R UV */                      \
+    __asm psubw      xmm0, kUVBiasB      /* unbias back to signed */           \
+    __asm psubw      xmm1, kUVBiasG                                            \
+    __asm psubw      xmm2, kUVBiasR                                            \
+    /* Step 2: Find Y contribution to 8 R,G,B values */                        \
+    __asm movq       xmm3, qword ptr [eax]                        /* NOLINT */ \
+    __asm lea        eax, [eax + 8]                                            \
+    __asm punpcklbw  xmm3, xmm4                                                \
+    __asm psubsw     xmm3, kYSub16                                             \
+    __asm pmullw     xmm3, kYToRgb                                             \
+    __asm paddsw     xmm0, xmm3           /* B += Y */                         \
+    __asm paddsw     xmm1, xmm3           /* G += Y */                         \
+    __asm paddsw     xmm2, xmm3           /* R += Y */                         \
+    __asm psraw      xmm0, 6                                                   \
+    __asm psraw      xmm1, 6                                                   \
+    __asm psraw      xmm2, 6                                                   \
+    __asm packuswb   xmm0, xmm0           /* B */                              \
+    __asm packuswb   xmm1, xmm1           /* G */                              \
+    __asm packuswb   xmm2, xmm2           /* R */                              \
+  }
+
+// Convert 8 pixels: 8 VU and 8 Y.
+#define YVUTORGB __asm {                                                       \
+    /* Step 1: Find 4 UV contributions to 8 R,G,B values */                    \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm movdqa     xmm2, xmm0                                                \
+    __asm pmaddubsw  xmm0, kVUToB        /* scale B UV */                      \
+    __asm pmaddubsw  xmm1, kVUToG        /* scale G UV */                      \
+    __asm pmaddubsw  xmm2, kVUToR        /* scale R UV */                      \
+    __asm psubw      xmm0, kUVBiasB      /* unbias back to signed */           \
+    __asm psubw      xmm1, kUVBiasG                                            \
+    __asm psubw      xmm2, kUVBiasR                                            \
+    /* Step 2: Find Y contribution to 8 R,G,B values */                        \
+    __asm movq       xmm3, qword ptr [eax]                        /* NOLINT */ \
+    __asm lea        eax, [eax + 8]                                            \
+    __asm punpcklbw  xmm3, xmm4                                                \
+    __asm psubsw     xmm3, kYSub16                                             \
+    __asm pmullw     xmm3, kYToRgb                                             \
+    __asm paddsw     xmm0, xmm3           /* B += Y */                         \
+    __asm paddsw     xmm1, xmm3           /* G += Y */                         \
+    __asm paddsw     xmm2, xmm3           /* R += Y */                         \
+    __asm psraw      xmm0, 6                                                   \
+    __asm psraw      xmm1, 6                                                   \
+    __asm psraw      xmm2, 6                                                   \
+    __asm packuswb   xmm0, xmm0           /* B */                              \
+    __asm packuswb   xmm1, xmm1           /* G */                              \
+    __asm packuswb   xmm2, xmm2           /* R */                              \
+  }
+
+// 8 pixels, dest aligned 16.
+// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I444ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV444
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToRGB24Row_SSSE3(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* dst_rgb24,
+                          int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgb24
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+    movdqa     xmm5, kShuffleMaskARGBToRGB24_0
+    movdqa     xmm6, kShuffleMaskARGBToRGB24
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RRGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm2           // RR
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
+    pshufb     xmm0, xmm5           // Pack into first 8 and last 4 bytes.
+    pshufb     xmm1, xmm6           // Pack into first 12 bytes.
+    palignr    xmm1, xmm0, 12       // last 4 bytes of xmm0 + 12 from xmm1
+    movq       qword ptr [edx], xmm0  // First 8 bytes
+    movdqu     [edx + 8], xmm1      // Last 16 bytes. = 24 bytes, 8 RGB pixels.
+    lea        edx,  [edx + 24]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToRAWRow_SSSE3(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_raw,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // raw
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+    movdqa     xmm5, kShuffleMaskARGBToRAW_0
+    movdqa     xmm6, kShuffleMaskARGBToRAW
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RRGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm2           // RR
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
+    pshufb     xmm0, xmm5           // Pack into first 8 and last 4 bytes.
+    pshufb     xmm1, xmm6           // Pack into first 12 bytes.
+    palignr    xmm1, xmm0, 12       // last 4 bytes of xmm0 + 12 from xmm1
+    movq       qword ptr [edx], xmm0  // First 8 bytes
+    movdqu     [edx + 8], xmm1      // Last 16 bytes. = 24 bytes, 8 RGB pixels.
+    lea        edx,  [edx + 24]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest unaligned.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToRGB565Row_SSSE3(const uint8* y_buf,
+                           const uint8* u_buf,
+                           const uint8* v_buf,
+                           uint8* rgb565_buf,
+                           int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgb565
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+    pcmpeqb    xmm5, xmm5       // generate mask 0x0000001f
+    psrld      xmm5, 27
+    pcmpeqb    xmm6, xmm6       // generate mask 0x000007e0
+    psrld      xmm6, 26
+    pslld      xmm6, 5
+    pcmpeqb    xmm7, xmm7       // generate mask 0xfffff800
+    pslld      xmm7, 11
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RRGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm2           // RR
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
+
+    // Step 3b: RRGB -> RGB565
+    movdqa     xmm3, xmm0    // B  first 4 pixels of argb
+    movdqa     xmm2, xmm0    // G
+    pslld      xmm0, 8       // R
+    psrld      xmm3, 3       // B
+    psrld      xmm2, 5       // G
+    psrad      xmm0, 16      // R
+    pand       xmm3, xmm5    // B
+    pand       xmm2, xmm6    // G
+    pand       xmm0, xmm7    // R
+    por        xmm3, xmm2    // BG
+    por        xmm0, xmm3    // BGR
+    movdqa     xmm3, xmm1    // B  next 4 pixels of argb
+    movdqa     xmm2, xmm1    // G
+    pslld      xmm1, 8       // R
+    psrld      xmm3, 3       // B
+    psrld      xmm2, 5       // G
+    psrad      xmm1, 16      // R
+    pand       xmm3, xmm5    // B
+    pand       xmm2, xmm6    // G
+    pand       xmm1, xmm7    // R
+    por        xmm3, xmm2    // BG
+    por        xmm1, xmm3    // BGR
+    packssdw   xmm0, xmm1
+    sub        ecx, 8
+    movdqu     [edx], xmm0   // store 8 pixels of RGB565
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// Similar to I420 but duplicate UV once more.
+__declspec(naked) __declspec(align(16))
+void I411ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]   // Y
+    mov        esi, [esp + 12 + 8]   // U
+    mov        edi, [esp + 12 + 12]  // V
+    mov        edx, [esp + 12 + 16]  // argb
+    mov        ecx, [esp + 12 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV411  // modifies EBX
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV12ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* uv_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV21ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* uv_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // VU
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YVUTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, unaligned.
+// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I444ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV444
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, unaligned.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, unaligned.
+// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// Similar to I420 but duplicate UV once more.
+__declspec(naked) __declspec(align(16))
+void I411ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]   // Y
+    mov        esi, [esp + 12 + 8]   // U
+    mov        edi, [esp + 12 + 12]  // V
+    mov        edx, [esp + 12 + 16]  // argb
+    mov        ecx, [esp + 12 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV411  // modifies EBX
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV12ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* uv_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV21ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* uv_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // VU
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YVUTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToBGRARow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_bgra,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // bgra
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into BGRA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm0           // GB
+    punpcklbw  xmm5, xmm2           // AR
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // BGRA first 4 pixels
+    punpckhwd  xmm0, xmm1           // BGRA next 4 pixels
+    movdqa     [edx], xmm5
+    movdqa     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToBGRARow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_bgra,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // bgra
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into BGRA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm0           // GB
+    punpcklbw  xmm5, xmm2           // AR
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // BGRA first 4 pixels
+    punpckhwd  xmm0, xmm1           // BGRA next 4 pixels
+    movdqu     [edx], xmm5
+    movdqu     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToABGRRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_abgr,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // abgr
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm2, xmm1           // RG
+    punpcklbw  xmm0, xmm5           // BA
+    movdqa     xmm1, xmm2
+    punpcklwd  xmm2, xmm0           // RGBA first 4 pixels
+    punpckhwd  xmm1, xmm0           // RGBA next 4 pixels
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToABGRRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_abgr,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // abgr
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm2, xmm1           // RG
+    punpcklbw  xmm0, xmm5           // BA
+    movdqa     xmm1, xmm2
+    punpcklwd  xmm2, xmm0           // RGBA first 4 pixels
+    punpckhwd  xmm1, xmm0           // RGBA next 4 pixels
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToRGBARow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_rgba,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgba
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RGBA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm2           // GR
+    punpcklbw  xmm5, xmm0           // AB
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // RGBA first 4 pixels
+    punpckhwd  xmm0, xmm1           // RGBA next 4 pixels
+    movdqa     [edx], xmm5
+    movdqa     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToRGBARow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_rgba,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgba
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RGBA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm2           // GR
+    punpcklbw  xmm5, xmm0           // AB
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // RGBA first 4 pixels
+    punpckhwd  xmm0, xmm1           // RGBA next 4 pixels
+    movdqu     [edx], xmm5
+    movdqu     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+#ifdef HAS_YTOARGBROW_SSE2
+__declspec(naked) __declspec(align(16))
+void YToARGBRow_SSE2(const uint8* y_buf,
+                     uint8* rgb_buf,
+                     int width) {
+  __asm {
+    pxor       xmm5, xmm5
+    pcmpeqb    xmm4, xmm4           // generate mask 0xff000000
+    pslld      xmm4, 24
+    mov        eax, 0x00100010
+    movd       xmm3, eax
+    pshufd     xmm3, xmm3, 0
+    mov        eax, 0x004a004a       // 74
+    movd       xmm2, eax
+    pshufd     xmm2, xmm2,0
+    mov        eax, [esp + 4]       // Y
+    mov        edx, [esp + 8]       // rgb
+    mov        ecx, [esp + 12]      // width
+
+    align      4
+ convertloop:
+    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+    movq       xmm0, qword ptr [eax]
+    lea        eax, [eax + 8]
+    punpcklbw  xmm0, xmm5           // 0.Y
+    psubusw    xmm0, xmm3
+    pmullw     xmm0, xmm2
+    psrlw      xmm0, 6
+    packuswb   xmm0, xmm0           // G
+
+    // Step 2: Weave into ARGB
+    punpcklbw  xmm0, xmm0           // GG
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm1           // BGRA next 4 pixels
+    por        xmm0, xmm4
+    por        xmm1, xmm4
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_YTOARGBROW_SSE2
+
+#ifdef HAS_MIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static const uvec8 kShuffleMirror = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked) __declspec(align(16))
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    movdqa    xmm5, kShuffleMirror
+    lea       eax, [eax - 16]
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax + ecx]
+    pshufb    xmm0, xmm5
+    sub       ecx, 16
+    movdqa    [edx], xmm0
+    lea       edx, [edx + 16]
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_SSSE3
+
+#ifdef HAS_MIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec8 kShuffleMirror_AVX2 = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u,
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked) __declspec(align(16))
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    vmovdqa   ymm5, kShuffleMirror_AVX2
+    lea       eax, [eax - 32]
+
+    align      4
+ convertloop:
+    vmovdqu   ymm0, [eax + ecx]
+    vpshufb   ymm0, ymm0, ymm5
+    vpermq    ymm0, ymm0, 0x4e  // swap high and low halfs
+    sub       ecx, 32
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSE2
+// SSE2 version has movdqu so it can be used on unaligned buffers when SSSE3
+// version can not.
+__declspec(naked) __declspec(align(16))
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 16]
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax + ecx]
+    movdqa    xmm1, xmm0        // swap bytes
+    psllw     xmm0, 8
+    psrlw     xmm1, 8
+    por       xmm0, xmm1
+    pshuflw   xmm0, xmm0, 0x1b  // swap words
+    pshufhw   xmm0, xmm0, 0x1b
+    pshufd    xmm0, xmm0, 0x4e  // swap qwords
+    sub       ecx, 16
+    movdqu    [edx], xmm0
+    lea       edx, [edx + 16]
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_SSE2
+
+#ifdef HAS_MIRRORROW_UV_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static const uvec8 kShuffleMirrorUV = {
+  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
+};
+
+__declspec(naked) __declspec(align(16))
+void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
+                       int width) {
+  __asm {
+    push      edi
+    mov       eax, [esp + 4 + 4]   // src
+    mov       edx, [esp + 4 + 8]   // dst_u
+    mov       edi, [esp + 4 + 12]  // dst_v
+    mov       ecx, [esp + 4 + 16]  // width
+    movdqa    xmm1, kShuffleMirrorUV
+    lea       eax, [eax + ecx * 2 - 16]
+    sub       edi, edx
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]
+    lea       eax, [eax - 16]
+    pshufb    xmm0, xmm1
+    sub       ecx, 8
+    movlpd    qword ptr [edx], xmm0
+    movhpd    qword ptr [edx + edi], xmm0
+    lea       edx, [edx + 8]
+    jg        convertloop
+
+    pop       edi
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_UV_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static const uvec8 kARGBShuffleMirror = {
+  12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u, 4u, 5u, 6u, 7u, 0u, 1u, 2u, 3u
+};
+
+__declspec(naked) __declspec(align(16))
+void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 16 + ecx * 4]  // last 4 pixels.
+    movdqa    xmm5, kARGBShuffleMirror
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]
+    lea       eax, [eax - 16]
+    pshufb    xmm0, xmm5
+    sub       ecx, 4
+    movdqa    [edx], xmm0
+    lea       edx, [edx + 16]
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec32 kARGBShuffleMirror_AVX2 = {
+  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked) __declspec(align(16))
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 32]
+    vmovdqa   ymm5, kARGBShuffleMirror_AVX2
+
+    align      4
+ convertloop:
+    vpermd    ymm0, ymm5, [eax + ecx * 4]  // permute dword order
+    sub       ecx, 8
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_AVX2
+
+#ifdef HAS_SPLITUVROW_SSE2
+__declspec(naked) __declspec(align(16))
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    movdqa     xmm3, xmm1
+    pand       xmm0, xmm5   // even bytes
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm2, 8      // odd bytes
+    psrlw      xmm3, 8
+    packuswb   xmm2, xmm3
+    movdqa     [edx], xmm0
+    movdqa     [edx + edi], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    movdqa     xmm3, xmm1
+    pand       xmm0, xmm5   // even bytes
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm2, 8      // odd bytes
+    psrlw      xmm3, 8
+    packuswb   xmm2, xmm3
+    movdqu     [edx], xmm0
+    movdqu     [edx + edi], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  // HAS_SPLITUVROW_SSE2
+
+#ifdef HAS_SPLITUVROW_AVX2
+__declspec(naked) __declspec(align(16))
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm2, ymm0, 8      // odd bytes
+    vpsrlw     ymm3, ymm1, 8
+    vpand      ymm0, ymm0, ymm5   // even bytes
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1
+    vpackuswb  ymm2, ymm2, ymm3
+    vpermq     ymm0, ymm0, 0xd8
+    vpermq     ymm2, ymm2, 0xd8
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + edi], ymm2
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SPLITUVROW_AVX2
+
+#ifdef HAS_MERGEUVROW_SSE2
+__declspec(naked) __declspec(align(16))
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]      // read 16 U's
+    movdqa     xmm1, [eax + edx]  // and 16 V's
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, xmm0
+    punpcklbw  xmm0, xmm1       // first 8 UV pairs
+    punpckhbw  xmm2, xmm1       // next 8 UV pairs
+    movdqa     [edi], xmm0
+    movdqa     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
+                               uint8* dst_uv, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]      // read 16 U's
+    movdqu     xmm1, [eax + edx]  // and 16 V's
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, xmm0
+    punpcklbw  xmm0, xmm1       // first 8 UV pairs
+    punpckhbw  xmm2, xmm1       // next 8 UV pairs
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  //  HAS_MERGEUVROW_SSE2
+
+#ifdef HAS_MERGEUVROW_AVX2
+__declspec(naked) __declspec(align(16))
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]           // read 32 U's
+    vmovdqu    ymm1, [eax + edx]     // and 32 V's
+    lea        eax,  [eax + 32]
+    vpunpcklbw ymm2, ymm0, ymm1      // low 16 UV pairs. mutated qqword 0,2
+    vpunpckhbw ymm0, ymm0, ymm1      // high 16 UV pairs. mutated qqword 1,3
+    vperm2i128 ymm1, ymm2, ymm0, 0x20  // low 128 of ymm2 and low 128 of ymm0
+    vperm2i128 ymm2, ymm2, ymm0, 0x31  // high 128 of ymm2 and high 128 of ymm0
+    vmovdqu    [edi], ymm1
+    vmovdqu    [edi + 32], ymm2
+    lea        edi, [edi + 64]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_MERGEUVROW_AVX2
+
+#ifdef HAS_COPYROW_SSE2
+// CopyRow copys 'count' bytes using a 16 byte load/store, 32 bytes at time.
+__declspec(naked) __declspec(align(16))
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_COPYROW_SSE2
+
+// Unaligned Multiple of 1.
+__declspec(naked) __declspec(align(16))
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, esi
+    mov        edx, edi
+    mov        esi, [esp + 4]   // src
+    mov        edi, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    rep movsb
+    mov        edi, edx
+    mov        esi, eax
+    ret
+  }
+}
+
+#ifdef HAS_COPYROW_X86
+__declspec(naked) __declspec(align(16))
+void CopyRow_X86(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, esi
+    mov        edx, edi
+    mov        esi, [esp + 4]   // src
+    mov        edi, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    shr        ecx, 2
+    rep movsd
+    mov        edi, edx
+    mov        esi, eax
+    ret
+  }
+}
+#endif  // HAS_COPYROW_X86
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    pslld      xmm0, 24
+    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    psrld      xmm1, 8
+
+    align      4
+  convertloop:
+    movdqa     xmm2, [eax]
+    movdqa     xmm3, [eax + 16]
+    lea        eax, [eax + 32]
+    movdqa     xmm4, [edx]
+    movdqa     xmm5, [edx + 16]
+    pand       xmm2, xmm0
+    pand       xmm3, xmm0
+    pand       xmm4, xmm1
+    pand       xmm5, xmm1
+    por        xmm2, xmm4
+    por        xmm3, xmm5
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm3
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYALPHAROW_AVX2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+
+    align      4
+  convertloop:
+    vmovdqu    ymm1, [eax]
+    vmovdqu    ymm2, [eax + 32]
+    lea        eax, [eax + 64]
+    vpblendvb  ymm1, ymm1, [edx], ymm0
+    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_AVX2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    pslld      xmm0, 24
+    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    psrld      xmm1, 8
+
+    align      4
+  convertloop:
+    movq       xmm2, qword ptr [eax]  // 8 Y's
+    lea        eax, [eax + 8]
+    punpcklbw  xmm2, xmm2
+    punpckhwd  xmm3, xmm2
+    punpcklwd  xmm2, xmm2
+    movdqa     xmm4, [edx]
+    movdqa     xmm5, [edx + 16]
+    pand       xmm2, xmm0
+    pand       xmm3, xmm0
+    pand       xmm4, xmm1
+    pand       xmm5, xmm1
+    por        xmm2, xmm4
+    por        xmm3, xmm5
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm3
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+
+    align      4
+  convertloop:
+    vpmovzxbd  ymm1, qword ptr [eax]
+    vpmovzxbd  ymm2, qword ptr [eax + 8]
+    lea        eax, [eax + 16]
+    vpslld     ymm1, ymm1, 24
+    vpslld     ymm2, ymm2, 24
+    vpblendvb  ymm1, ymm1, [edx], ymm0
+    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
+
+#ifdef HAS_SETROW_X86
+// SetRow8 writes 'count' bytes using a 32 bit value repeated.
+__declspec(naked) __declspec(align(16))
+void SetRow_X86(uint8* dst, uint32 v32, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        eax, [esp + 8]   // v32
+    mov        ecx, [esp + 12]  // count
+    shr        ecx, 2
+    rep stosd
+    mov        edi, edx
+    ret
+  }
+}
+
+// SetRow32 writes 'count' words using a 32 bit value repeated.
+__declspec(naked) __declspec(align(16))
+void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
+                   int dst_stride, int height) {
+  __asm {
+    push       esi
+    push       edi
+    push       ebp
+    mov        edi, [esp + 12 + 4]   // dst
+    mov        eax, [esp + 12 + 8]   // v32
+    mov        ebp, [esp + 12 + 12]  // width
+    mov        edx, [esp + 12 + 16]  // dst_stride
+    mov        esi, [esp + 12 + 20]  // height
+    lea        ecx, [ebp * 4]
+    sub        edx, ecx             // stride - width * 4
+
+    align      4
+  convertloop:
+    mov        ecx, ebp
+    rep stosd
+    add        edi, edx
+    sub        esi, 1
+    jg         convertloop
+
+    pop        ebp
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SETROW_X86
+
+#ifdef HAS_YUY2TOYROW_AVX2
+__declspec(naked) __declspec(align(16))
+void YUY2ToYRow_AVX2(const uint8* src_yuy2,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // even bytes are Y
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToYRow_AVX2(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // odd bytes are Y
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+    ret
+    vzeroupper
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_YUY2TOYROW_AVX2
+
+#ifdef HAS_YUY2TOYROW_SSE2
+__declspec(naked) __declspec(align(16))
+void YUY2ToYRow_SSE2(const uint8* src_yuy2,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // even bytes are Y
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
+                               uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // even bytes are Y
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToYRow_SSE2(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8    // odd bytes are Y
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
+                               uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8    // odd bytes are Y
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  // HAS_YUY2TOYROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSE2
+// Blend 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm7, xmm7       // generate constant 1
+    psrlw      xmm7, 15
+    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    psrlw      xmm6, 8
+    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    psllw      xmm5, 8
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+
+    sub        ecx, 1
+    je         convertloop1     // only 1 pixel?
+    jl         convertloop1b
+
+    // 1 pixel loop until destination pointer is aligned.
+  alignloop1:
+    test       edx, 15          // aligned?
+    je         alignloop1b
+    movd       xmm3, [eax]
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        alignloop1
+
+  alignloop1b:
+    add        ecx, 1 - 4
+    jl         convertloop4b
+
+    // 4 pixel loop.
+  convertloop4:
+    movdqu     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqu     xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqu     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertloop4
+
+  convertloop4b:
+    add        ecx, 4 - 1
+    jl         convertloop1b
+
+    // 1 pixel loop.
+  convertloop1:
+    movd       xmm3, [eax]      // src argb
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        convertloop1
+
+  convertloop1b:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBBLENDROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSSE3
+// Shuffle table for isolating alpha.
+static const uvec8 kShuffleAlpha = {
+  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
+  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
+};
+// Same as SSE2, but replaces:
+//    psrlw      xmm3, 8          // alpha
+//    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+//    pshuflw    xmm3, xmm3, 0F5h
+// with..
+//    pshufb     xmm3, kShuffleAlpha // alpha
+// Blend 8 pixels at a time.
+
+__declspec(naked) __declspec(align(16))
+void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
+                        uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm7, xmm7       // generate constant 0x0001
+    psrlw      xmm7, 15
+    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    psrlw      xmm6, 8
+    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    psllw      xmm5, 8
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+
+    sub        ecx, 1
+    je         convertloop1     // only 1 pixel?
+    jl         convertloop1b
+
+    // 1 pixel loop until destination pointer is aligned.
+  alignloop1:
+    test       edx, 15          // aligned?
+    je         alignloop1b
+    movd       xmm3, [eax]
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        alignloop1
+
+  alignloop1b:
+    add        ecx, 1 - 4
+    jl         convertloop4b
+
+    test       eax, 15          // unaligned?
+    jne        convertuloop4
+    test       esi, 15          // unaligned?
+    jne        convertuloop4
+
+    // 4 pixel loop.
+  convertloop4:
+    movdqa     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqa     xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqa     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertloop4
+    jmp        convertloop4b
+
+    // 4 pixel unaligned loop.
+  convertuloop4:
+    movdqu     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqu     xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqu     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertuloop4
+
+  convertloop4b:
+    add        ecx, 4 - 1
+    jl         convertloop1b
+
+    // 1 pixel loop.
+  convertloop1:
+    movd       xmm3, [eax]      // src argb
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        convertloop1
+
+  convertloop1b:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBBLENDROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+// Attenuate 4 pixels at a time.
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+    pcmpeqb    xmm5, xmm5       // generate mask 0x00ffffff
+    psrld      xmm5, 8
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]      // read 4 pixels
+    punpcklbw  xmm0, xmm0       // first 2
+    pshufhw    xmm2, xmm0, 0FFh // 8 alpha words
+    pshuflw    xmm2, xmm2, 0FFh
+    pmulhuw    xmm0, xmm2       // rgb * a
+    movdqa     xmm1, [eax]      // read 4 pixels
+    punpckhbw  xmm1, xmm1       // next 2 pixels
+    pshufhw    xmm2, xmm1, 0FFh // 8 alpha words
+    pshuflw    xmm2, xmm2, 0FFh
+    pmulhuw    xmm1, xmm2       // rgb * a
+    movdqa     xmm2, [eax]      // alphas
+    lea        eax, [eax + 16]
+    psrlw      xmm0, 8
+    pand       xmm2, xmm4
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    pand       xmm0, xmm5       // keep original alphas
+    por        xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha0 = {
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+};
+static const uvec8 kShuffleAlpha1 = {
+  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+};
+__declspec(naked) __declspec(align(16))
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm3, xmm3       // generate mask 0xff000000
+    pslld      xmm3, 24
+    movdqa     xmm4, kShuffleAlpha0
+    movdqa     xmm5, kShuffleAlpha1
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    pshufb     xmm0, xmm4       // isolate first 2 alphas
+    movdqu     xmm1, [eax]      // read 4 pixels
+    punpcklbw  xmm1, xmm1       // first 2 pixel rgbs
+    pmulhuw    xmm0, xmm1       // rgb * a
+    movdqu     xmm1, [eax]      // read 4 pixels
+    pshufb     xmm1, xmm5       // isolate next 2 alphas
+    movdqu     xmm2, [eax]      // read 4 pixels
+    punpckhbw  xmm2, xmm2       // next 2 pixel rgbs
+    pmulhuw    xmm1, xmm2       // rgb * a
+    movdqu     xmm2, [eax]      // mask original alpha
+    lea        eax, [eax + 16]
+    pand       xmm2, xmm3
+    psrlw      xmm0, 8
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    por        xmm0, xmm2       // copy original alpha
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const ulvec8 kShuffleAlpha_AVX2 = {
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u,
+  14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u,
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u,
+  14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u,
+};
+__declspec(naked) __declspec(align(16))
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vmovdqa    ymm4, kShuffleAlpha_AVX2
+    vpcmpeqb   ymm5, ymm5, ymm5 // generate mask 0xff000000
+    vpslld     ymm5, ymm5, 24
+
+    align      4
+ convertloop:
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpshufb    ymm2, ymm0, ymm4  // low 4 alphas
+    vpshufb    ymm3, ymm1, ymm4  // high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * a
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * a
+    vpand      ymm6, ymm6, ymm5  // isolate alpha
+    vpsrlw     ymm0, ymm0, 8
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    vpor       ymm0, ymm0, ymm6  // copy original alpha
+    sub        ecx, 8
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+// Unattenuate 4 pixels at a time.
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb0
+    mov        edx, [esp + 8 + 8]   // dst_argb
+    mov        ecx, [esp + 8 + 12]  // width
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    movzx      esi, byte ptr [eax + 3]  // first alpha
+    movzx      edi, byte ptr [eax + 7]  // second alpha
+    punpcklbw  xmm0, xmm0       // first 2
+    movd       xmm2, dword ptr fixed_invtbl8[esi * 4]
+    movd       xmm3, dword ptr fixed_invtbl8[edi * 4]
+    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words.  1, a, a, a
+    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    movlhps    xmm2, xmm3
+    pmulhuw    xmm0, xmm2       // rgb * a
+
+    movdqu     xmm1, [eax]      // read 4 pixels
+    movzx      esi, byte ptr [eax + 11]  // third alpha
+    movzx      edi, byte ptr [eax + 15]  // forth alpha
+    punpckhbw  xmm1, xmm1       // next 2
+    movd       xmm2, dword ptr fixed_invtbl8[esi * 4]
+    movd       xmm3, dword ptr fixed_invtbl8[edi * 4]
+    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words
+    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    movlhps    xmm2, xmm3
+    pmulhuw    xmm1, xmm2       // rgb * a
+    lea        eax, [eax + 16]
+
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBUNATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const ulvec8 kUnattenShuffleAlpha_AVX2 = {
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15,
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15,
+};
+// TODO(fbarchard): Enable USE_GATHER for future hardware if faster.
+// USE_GATHER is not on by default, due to being a slow instruction.
+#ifdef USE_GATHER
+__declspec(naked) __declspec(align(16))
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vmovdqa    ymm4, kUnattenShuffleAlpha_AVX2
+
+    align      4
+ convertloop:
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0xffffffff for gather.
+    vpsrld     ymm2, ymm6, 24    // alpha in low 8 bits.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpgatherdd ymm3, [ymm2 * 4 + fixed_invtbl8], ymm5  // ymm5 cleared.  1, a
+    vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
+    vpunpckhwd ymm3, ymm3, ymm3  // high 4 inverted alphas. mutated.
+    vpshufb    ymm2, ymm2, ymm4  // replicate low 4 alphas. 1, a, a, a
+    vpshufb    ymm3, ymm3, ymm4  // replicate high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    sub        ecx, 8
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#else  // USE_GATHER
+__declspec(naked) __declspec(align(16))
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vmovdqa    ymm5, kUnattenShuffleAlpha_AVX2
+
+    push       esi
+    push       edi
+
+    align      4
+ convertloop:
+    // replace VPGATHER
+    movzx      esi, byte ptr [eax + 3]                 // alpha0
+    movzx      edi, byte ptr [eax + 7]                 // alpha1
+    vmovd      xmm0, dword ptr fixed_invtbl8[esi * 4]  // [1,a0]
+    vmovd      xmm1, dword ptr fixed_invtbl8[edi * 4]  // [1,a1]
+    movzx      esi, byte ptr [eax + 11]                // alpha2
+    movzx      edi, byte ptr [eax + 15]                // alpha3
+    vpunpckldq xmm6, xmm0, xmm1                        // [1,a1,1,a0]
+    vmovd      xmm2, dword ptr fixed_invtbl8[esi * 4]  // [1,a2]
+    vmovd      xmm3, dword ptr fixed_invtbl8[edi * 4]  // [1,a3]
+    movzx      esi, byte ptr [eax + 19]                // alpha4
+    movzx      edi, byte ptr [eax + 23]                // alpha5
+    vpunpckldq xmm7, xmm2, xmm3                        // [1,a3,1,a2]
+    vmovd      xmm0, dword ptr fixed_invtbl8[esi * 4]  // [1,a4]
+    vmovd      xmm1, dword ptr fixed_invtbl8[edi * 4]  // [1,a5]
+    movzx      esi, byte ptr [eax + 27]                // alpha6
+    movzx      edi, byte ptr [eax + 31]                // alpha7
+    vpunpckldq xmm0, xmm0, xmm1                        // [1,a5,1,a4]
+    vmovd      xmm2, dword ptr fixed_invtbl8[esi * 4]  // [1,a6]
+    vmovd      xmm3, dword ptr fixed_invtbl8[edi * 4]  // [1,a7]
+    vpunpckldq xmm2, xmm2, xmm3                        // [1,a7,1,a6]
+    vpunpcklqdq xmm3, xmm6, xmm7                       // [1,a3,1,a2,1,a1,1,a0]
+    vpunpcklqdq xmm0, xmm0, xmm2                       // [1,a7,1,a6,1,a5,1,a4]
+    vinserti128 ymm3, ymm3, xmm0, 1 // [1,a7,1,a6,1,a5,1,a4,1,a3,1,a2,1,a1,1,a0]
+    // end of VPGATHER
+
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
+    vpunpckhwd ymm3, ymm3, ymm3  // high 4 inverted alphas. mutated.
+    vpshufb    ymm2, ymm2, ymm5  // replicate low 4 alphas. 1, a, a, a
+    vpshufb    ymm3, ymm3, ymm5  // replicate high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    sub        ecx, 8
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // USE_GATHER
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBGRAYROW_SSSE3
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels.
+__declspec(naked) __declspec(align(16))
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]  /* width */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // G
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    phaddw     xmm0, xmm1
+    paddw      xmm0, xmm5  // Add .5 for rounding.
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0   // 8 G bytes
+    movdqa     xmm2, [eax]  // A
+    movdqa     xmm3, [eax + 16]
+    lea        eax, [eax + 32]
+    psrld      xmm2, 24
+    psrld      xmm3, 24
+    packuswb   xmm2, xmm3
+    packuswb   xmm2, xmm2   // 8 A bytes
+    movdqa     xmm3, xmm0   // Weave into GG, GA, then GGGA
+    punpcklbw  xmm0, xmm0   // 8 GG words
+    punpcklbw  xmm3, xmm2   // 8 GA words
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm3   // GGGA first 4
+    punpckhwd  xmm1, xmm3   // GGGA next 4
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBGRAYROW_SSSE3
+
+#ifdef HAS_ARGBSEPIAROW_SSSE3
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+// Constant for ARGB color to sepia tone.
+static const vec8 kARGBToSepiaB = {
+  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
+};
+
+static const vec8 kARGBToSepiaG = {
+  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
+};
+
+static const vec8 kARGBToSepiaR = {
+  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
+};
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+__declspec(naked) __declspec(align(16))
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* dst_argb */
+    mov        ecx, [esp + 8]   /* width */
+    movdqa     xmm2, kARGBToSepiaB
+    movdqa     xmm3, kARGBToSepiaG
+    movdqa     xmm4, kARGBToSepiaR
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // B
+    movdqa     xmm6, [eax + 16]
+    pmaddubsw  xmm0, xmm2
+    pmaddubsw  xmm6, xmm2
+    phaddw     xmm0, xmm6
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0   // 8 B values
+    movdqa     xmm5, [eax]  // G
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm5, xmm3
+    pmaddubsw  xmm1, xmm3
+    phaddw     xmm5, xmm1
+    psrlw      xmm5, 7
+    packuswb   xmm5, xmm5   // 8 G values
+    punpcklbw  xmm0, xmm5   // 8 BG values
+    movdqa     xmm5, [eax]  // R
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm5, xmm4
+    pmaddubsw  xmm1, xmm4
+    phaddw     xmm5, xmm1
+    psrlw      xmm5, 7
+    packuswb   xmm5, xmm5   // 8 R values
+    movdqa     xmm6, [eax]  // A
+    movdqa     xmm1, [eax + 16]
+    psrld      xmm6, 24
+    psrld      xmm1, 24
+    packuswb   xmm6, xmm1
+    packuswb   xmm6, xmm6   // 8 A values
+    punpcklbw  xmm5, xmm6   // 8 RA values
+    movdqa     xmm1, xmm0   // Weave BG, RA together
+    punpcklwd  xmm0, xmm5   // BGRA first 4
+    punpckhwd  xmm1, xmm5   // BGRA next 4
+    sub        ecx, 8
+    movdqa     [eax], xmm0
+    movdqa     [eax + 16], xmm1
+    lea        eax, [eax + 32]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBSEPIAROW_SSSE3
+
+#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// Same as Sepia except matrix is provided.
+// TODO(fbarchard): packuswbs only use half of the reg. To make RGBA, combine R
+// and B into a high and low, then G/A, unpackl/hbw and then unpckl/hwd.
+__declspec(naked) __declspec(align(16))
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]  /* matrix_argb */
+    movdqu     xmm5, [ecx]
+    pshufd     xmm2, xmm5, 0x00
+    pshufd     xmm3, xmm5, 0x55
+    pshufd     xmm4, xmm5, 0xaa
+    pshufd     xmm5, xmm5, 0xff
+    mov        ecx, [esp + 16]  /* width */
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // B
+    movdqa     xmm7, [eax + 16]
+    pmaddubsw  xmm0, xmm2
+    pmaddubsw  xmm7, xmm2
+    movdqa     xmm6, [eax]  // G
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm6, xmm3
+    pmaddubsw  xmm1, xmm3
+    phaddsw    xmm0, xmm7   // B
+    phaddsw    xmm6, xmm1   // G
+    psraw      xmm0, 6      // B
+    psraw      xmm6, 6      // G
+    packuswb   xmm0, xmm0   // 8 B values
+    packuswb   xmm6, xmm6   // 8 G values
+    punpcklbw  xmm0, xmm6   // 8 BG values
+    movdqa     xmm1, [eax]  // R
+    movdqa     xmm7, [eax + 16]
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm7, xmm4
+    phaddsw    xmm1, xmm7   // R
+    movdqa     xmm6, [eax]  // A
+    movdqa     xmm7, [eax + 16]
+    pmaddubsw  xmm6, xmm5
+    pmaddubsw  xmm7, xmm5
+    phaddsw    xmm6, xmm7   // A
+    psraw      xmm1, 6      // R
+    psraw      xmm6, 6      // A
+    packuswb   xmm1, xmm1   // 8 R values
+    packuswb   xmm6, xmm6   // 8 A values
+    punpcklbw  xmm1, xmm6   // 8 RA values
+    movdqa     xmm6, xmm0   // Weave BG, RA together
+    punpcklwd  xmm0, xmm1   // BGRA first 4
+    punpckhwd  xmm6, xmm1   // BGRA next 4
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm6
+    lea        eax, [eax + 32]
+    lea        edx, [edx + 32]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
+
+#ifdef HAS_ARGBQUANTIZEROW_SSE2
+// Quantize 4 ARGB pixels (16 bytes).
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  __asm {
+    mov        eax, [esp + 4]    /* dst_argb */
+    movd       xmm2, [esp + 8]   /* scale */
+    movd       xmm3, [esp + 12]  /* interval_size */
+    movd       xmm4, [esp + 16]  /* interval_offset */
+    mov        ecx, [esp + 20]   /* width */
+    pshuflw    xmm2, xmm2, 040h
+    pshufd     xmm2, xmm2, 044h
+    pshuflw    xmm3, xmm3, 040h
+    pshufd     xmm3, xmm3, 044h
+    pshuflw    xmm4, xmm4, 040h
+    pshufd     xmm4, xmm4, 044h
+    pxor       xmm5, xmm5  // constant 0
+    pcmpeqb    xmm6, xmm6  // generate mask 0xff000000
+    pslld      xmm6, 24
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // read 4 pixels
+    punpcklbw  xmm0, xmm5   // first 2 pixels
+    pmulhuw    xmm0, xmm2   // pixel * scale >> 16
+    movdqa     xmm1, [eax]  // read 4 pixels
+    punpckhbw  xmm1, xmm5   // next 2 pixels
+    pmulhuw    xmm1, xmm2
+    pmullw     xmm0, xmm3   // * interval_size
+    movdqa     xmm7, [eax]  // read 4 pixels
+    pmullw     xmm1, xmm3
+    pand       xmm7, xmm6   // mask alpha
+    paddw      xmm0, xmm4   // + interval_size / 2
+    paddw      xmm1, xmm4
+    packuswb   xmm0, xmm1
+    por        xmm0, xmm7
+    sub        ecx, 4
+    movdqa     [eax], xmm0
+    lea        eax, [eax + 16]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBQUANTIZEROW_SSE2
+
+#ifdef HAS_ARGBSHADEROW_SSE2
+// Shade 4 pixels at a time by specified value.
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    movd       xmm2, [esp + 16]  // value
+    punpcklbw  xmm2, xmm2
+    punpcklqdq xmm2, xmm2
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]      // read 4 pixels
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm0       // first 2
+    punpckhbw  xmm1, xmm1       // next 2
+    pmulhuw    xmm0, xmm2       // argb * value
+    pmulhuw    xmm1, xmm2       // argb * value
+    psrlw      xmm0, 8
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBSHADEROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+// Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pxor       xmm5, xmm5  // constant 0
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    movdqu     xmm2, [esi]        // read 4 pixels from src_argb1
+    movdqu     xmm1, xmm0
+    movdqu     xmm3, xmm2
+    punpcklbw  xmm0, xmm0         // first 2
+    punpckhbw  xmm1, xmm1         // next 2
+    punpcklbw  xmm2, xmm5         // first 2
+    punpckhbw  xmm3, xmm5         // next 2
+    pmulhuw    xmm0, xmm2         // src_argb0 * src_argb1 first 2
+    pmulhuw    xmm1, xmm3         // src_argb0 * src_argb1 next 2
+    lea        eax, [eax + 16]
+    lea        esi, [esi + 16]
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBMULTIPLYROW_SSE2
+
+#ifdef HAS_ARGBADDROW_SSE2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+// TODO(fbarchard): Port this to posix, neon and other math functions.
+__declspec(naked) __declspec(align(16))
+void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    sub        ecx, 4
+    jl         convertloop49
+
+    align      4
+ convertloop4:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    lea        eax, [eax + 16]
+    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    lea        esi, [esi + 16]
+    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertloop4
+
+ convertloop49:
+    add        ecx, 4 - 1
+    jl         convertloop19
+
+ convertloop1:
+    movd       xmm0, [eax]        // read 1 pixels from src_argb0
+    lea        eax, [eax + 4]
+    movd       xmm1, [esi]        // read 1 pixels from src_argb1
+    lea        esi, [esi + 4]
+    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        convertloop1
+
+ convertloop19:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBADDROW_SSE2
+
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels together, 4 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    lea        eax, [eax + 16]
+    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    lea        esi, [esi + 16]
+    psubusb    xmm0, xmm1         // src_argb0 - src_argb1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpxor      ymm5, ymm5, ymm5     // constant 0
+
+    align      4
+ convertloop:
+    vmovdqu    ymm1, [eax]        // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vmovdqu    ymm3, [esi]        // read 8 pixels from src_argb1
+    lea        esi, [esi + 32]
+    vpunpcklbw ymm0, ymm1, ymm1   // low 4
+    vpunpckhbw ymm1, ymm1, ymm1   // high 4
+    vpunpcklbw ymm2, ymm3, ymm5   // low 4
+    vpunpckhbw ymm3, ymm3, ymm5   // high 4
+    vpmulhuw   ymm0, ymm0, ymm2   // src_argb0 * src_argb1 low 4
+    vpmulhuw   ymm1, ymm1, ymm3   // src_argb0 * src_argb1 high 4
+    vpackuswb  ymm0, ymm0, ymm1
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBMULTIPLYROW_AVX2
+
+#ifdef HAS_ARGBADDROW_AVX2
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vpaddusb   ymm0, ymm0, [esi]        // add 8 pixels from src_argb1
+    lea        esi, [esi + 32]
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBADDROW_AVX2
+
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+// Subtract 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vpsubusb   ymm0, ymm0, [esi]        // src_argb0 - src_argb1
+    lea        esi, [esi + 32]
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_AVX2
+
+#ifdef HAS_SOBELXROW_SSE2
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+__declspec(naked) __declspec(align(16))
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_y0
+    mov        esi, [esp + 8 + 8]   // src_y1
+    mov        edi, [esp + 8 + 12]  // src_y2
+    mov        edx, [esp + 8 + 16]  // dst_sobelx
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        esi, eax
+    sub        edi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+    align      4
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    movq       xmm2, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + edi]      // read 8 pixels from src_y2[0]
+    movq       xmm3, qword ptr [eax + edi + 2]  // read 8 pixels from src_y2[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    psubw      xmm1, xmm0
+    pmaxsw     xmm0, xmm1
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXROW_SSE2
+
+#ifdef HAS_SOBELYROW_SSE2
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+__declspec(naked) __declspec(align(16))
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_y0
+    mov        esi, [esp + 4 + 8]   // src_y1
+    mov        edx, [esp + 4 + 12]  // dst_sobely
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+    align      4
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + 1]        // read 8 pixels from src_y0[1]
+    movq       xmm2, qword ptr [eax + esi + 1]  // read 8 pixels from src_y1[1]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    movq       xmm3, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    psubw      xmm1, xmm0
+    pmaxsw     xmm0, xmm1
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELYROW_SSE2
+
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+__declspec(naked) __declspec(align(16))
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+    pslld      xmm5, 24             // 0xff000000
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm2, xmm0             // GG
+    punpcklbw  xmm2, xmm0             // First 8
+    punpckhbw  xmm0, xmm0             // Next 8
+    movdqa     xmm1, xmm2             // GGGG
+    punpcklwd  xmm1, xmm2             // First 4
+    punpckhwd  xmm2, xmm2             // Next 4
+    por        xmm1, xmm5             // GGGA
+    por        xmm2, xmm5
+    movdqa     xmm3, xmm0             // GGGG
+    punpcklwd  xmm3, xmm0             // Next 4
+    punpckhwd  xmm0, xmm0             // Last 4
+    por        xmm3, xmm5             // GGGA
+    por        xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm1
+    movdqa     [edx + 16], xmm2
+    movdqa     [edx + 32], xmm3
+    movdqa     [edx + 48], xmm0
+    lea        edx, [edx + 64]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELROW_SSE2
+
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into a plane.
+__declspec(naked) __declspec(align(16))
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELTOPLANEROW_SSE2
+
+#ifdef HAS_SOBELXYROW_SSE2
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+__declspec(naked) __declspec(align(16))
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    movdqa     xmm2, xmm0
+    paddusb    xmm2, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm3, xmm0             // XA
+    punpcklbw  xmm3, xmm5
+    punpckhbw  xmm0, xmm5
+    movdqa     xmm4, xmm1             // YS
+    punpcklbw  xmm4, xmm2
+    punpckhbw  xmm1, xmm2
+    movdqa     xmm6, xmm4             // YSXA
+    punpcklwd  xmm6, xmm3             // First 4
+    punpckhwd  xmm4, xmm3             // Next 4
+    movdqa     xmm7, xmm1             // YSXA
+    punpcklwd  xmm7, xmm0             // Next 4
+    punpckhwd  xmm1, xmm0             // Last 4
+    sub        ecx, 16
+    movdqa     [edx], xmm6
+    movdqa     [edx + 16], xmm4
+    movdqa     [edx + 32], xmm7
+    movdqa     [edx + 48], xmm1
+    lea        edx, [edx + 64]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXYROW_SSE2
+
+#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+// Consider float CumulativeSum.
+// Consider calling CumulativeSum one row at time as needed.
+// Consider circular CumulativeSum buffer of radius * 2 + 1 height.
+// Convert cumulative sum for an area to an average for 1 pixel.
+// topleft is pointer to top left of CumulativeSum buffer for area.
+// botleft is pointer to bottom left of CumulativeSum buffer.
+// width is offset from left to right of area in CumulativeSum buffer measured
+//   in number of ints.
+// area is the number of pixels in the area being averaged.
+// dst points to pixel to store result to.
+// count is number of averaged pixels to produce.
+// Does 4 pixels at a time, requires CumulativeSum pointers to be 16 byte
+// aligned.
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst,
+                                    int count) {
+  __asm {
+    mov        eax, topleft  // eax topleft
+    mov        esi, botleft  // esi botleft
+    mov        edx, width
+    movd       xmm5, area
+    mov        edi, dst
+    mov        ecx, count
+    cvtdq2ps   xmm5, xmm5
+    rcpss      xmm4, xmm5  // 1.0f / area
+    pshufd     xmm4, xmm4, 0
+    sub        ecx, 4
+    jl         l4b
+
+    cmp        area, 128  // 128 pixels will not overflow 15 bits.
+    ja         l4
+
+    pshufd     xmm5, xmm5, 0        // area
+    pcmpeqb    xmm6, xmm6           // constant of 65536.0 - 1 = 65535.0
+    psrld      xmm6, 16
+    cvtdq2ps   xmm6, xmm6
+    addps      xmm5, xmm6           // (65536.0 + area - 1)
+    mulps      xmm5, xmm4           // (65536.0 + area - 1) * 1 / area
+    cvtps2dq   xmm5, xmm5           // 0.16 fixed point
+    packssdw   xmm5, xmm5           // 16 bit shorts
+
+    // 4 pixel loop small blocks.
+    align      4
+  s4:
+    // top left
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+
+    // - top right
+    psubd      xmm0, [eax + edx * 4]
+    psubd      xmm1, [eax + edx * 4 + 16]
+    psubd      xmm2, [eax + edx * 4 + 32]
+    psubd      xmm3, [eax + edx * 4 + 48]
+    lea        eax, [eax + 64]
+
+    // - bottom left
+    psubd      xmm0, [esi]
+    psubd      xmm1, [esi + 16]
+    psubd      xmm2, [esi + 32]
+    psubd      xmm3, [esi + 48]
+
+    // + bottom right
+    paddd      xmm0, [esi + edx * 4]
+    paddd      xmm1, [esi + edx * 4 + 16]
+    paddd      xmm2, [esi + edx * 4 + 32]
+    paddd      xmm3, [esi + edx * 4 + 48]
+    lea        esi, [esi + 64]
+
+    packssdw   xmm0, xmm1  // pack 4 pixels into 2 registers
+    packssdw   xmm2, xmm3
+
+    pmulhuw    xmm0, xmm5
+    pmulhuw    xmm2, xmm5
+
+    packuswb   xmm0, xmm2
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jge        s4
+
+    jmp        l4b
+
+    // 4 pixel loop
+    align      4
+  l4:
+    // top left
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+
+    // - top right
+    psubd      xmm0, [eax + edx * 4]
+    psubd      xmm1, [eax + edx * 4 + 16]
+    psubd      xmm2, [eax + edx * 4 + 32]
+    psubd      xmm3, [eax + edx * 4 + 48]
+    lea        eax, [eax + 64]
+
+    // - bottom left
+    psubd      xmm0, [esi]
+    psubd      xmm1, [esi + 16]
+    psubd      xmm2, [esi + 32]
+    psubd      xmm3, [esi + 48]
+
+    // + bottom right
+    paddd      xmm0, [esi + edx * 4]
+    paddd      xmm1, [esi + edx * 4 + 16]
+    paddd      xmm2, [esi + edx * 4 + 32]
+    paddd      xmm3, [esi + edx * 4 + 48]
+    lea        esi, [esi + 64]
+
+    cvtdq2ps   xmm0, xmm0   // Average = Sum * 1 / Area
+    cvtdq2ps   xmm1, xmm1
+    mulps      xmm0, xmm4
+    mulps      xmm1, xmm4
+    cvtdq2ps   xmm2, xmm2
+    cvtdq2ps   xmm3, xmm3
+    mulps      xmm2, xmm4
+    mulps      xmm3, xmm4
+    cvtps2dq   xmm0, xmm0
+    cvtps2dq   xmm1, xmm1
+    cvtps2dq   xmm2, xmm2
+    cvtps2dq   xmm3, xmm3
+    packssdw   xmm0, xmm1
+    packssdw   xmm2, xmm3
+    packuswb   xmm0, xmm2
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+    align      4
+  l1:
+    movdqa     xmm0, [eax]
+    psubd      xmm0, [eax + edx * 4]
+    lea        eax, [eax + 16]
+    psubd      xmm0, [esi]
+    paddd      xmm0, [esi + edx * 4]
+    lea        esi, [esi + 16]
+    cvtdq2ps   xmm0, xmm0
+    mulps      xmm0, xmm4
+    cvtps2dq   xmm0, xmm0
+    packssdw   xmm0, xmm0
+    packuswb   xmm0, xmm0
+    movd       dword ptr [edi], xmm0
+    lea        edi, [edi + 4]
+    sub        ecx, 1
+    jge        l1
+  l1b:
+  }
+}
+#endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+
+#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
+// Creates a table of cumulative sums where each value is a sum of all values
+// above and to the left of the value.
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width) {
+  __asm {
+    mov        eax, row
+    mov        edx, cumsum
+    mov        esi, previous_cumsum
+    mov        ecx, width
+    pxor       xmm0, xmm0
+    pxor       xmm1, xmm1
+
+    sub        ecx, 4
+    jl         l4b
+    test       edx, 15
+    jne        l4b
+
+    // 4 pixel loop
+    align      4
+  l4:
+    movdqu     xmm2, [eax]  // 4 argb pixels 16 bytes.
+    lea        eax, [eax + 16]
+    movdqa     xmm4, xmm2
+
+    punpcklbw  xmm2, xmm1
+    movdqa     xmm3, xmm2
+    punpcklwd  xmm2, xmm1
+    punpckhwd  xmm3, xmm1
+
+    punpckhbw  xmm4, xmm1
+    movdqa     xmm5, xmm4
+    punpcklwd  xmm4, xmm1
+    punpckhwd  xmm5, xmm1
+
+    paddd      xmm0, xmm2
+    movdqa     xmm2, [esi]  // previous row above.
+    paddd      xmm2, xmm0
+
+    paddd      xmm0, xmm3
+    movdqa     xmm3, [esi + 16]
+    paddd      xmm3, xmm0
+
+    paddd      xmm0, xmm4
+    movdqa     xmm4, [esi + 32]
+    paddd      xmm4, xmm0
+
+    paddd      xmm0, xmm5
+    movdqa     xmm5, [esi + 48]
+    lea        esi, [esi + 64]
+    paddd      xmm5, xmm0
+
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm3
+    movdqa     [edx + 32], xmm4
+    movdqa     [edx + 48], xmm5
+
+    lea        edx, [edx + 64]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+    align      4
+  l1:
+    movd       xmm2, dword ptr [eax]  // 1 argb pixel 4 bytes.
+    lea        eax, [eax + 4]
+    punpcklbw  xmm2, xmm1
+    punpcklwd  xmm2, xmm1
+    paddd      xmm0, xmm2
+    movdqu     xmm2, [esi]
+    lea        esi, [esi + 16]
+    paddd      xmm2, xmm0
+    movdqu     [edx], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 1
+    jge        l1
+
+ l1b:
+  }
+}
+#endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
+
+#ifdef HAS_ARGBAFFINEROW_SSE2
+// Copy ARGB pixels from source image with slope to a row of destination.
+__declspec(naked) __declspec(align(16))
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 12]  // src_argb
+    mov        esi, [esp + 16]  // stride
+    mov        edx, [esp + 20]  // dst_argb
+    mov        ecx, [esp + 24]  // pointer to uv_dudv
+    movq       xmm2, qword ptr [ecx]  // uv
+    movq       xmm7, qword ptr [ecx + 8]  // dudv
+    mov        ecx, [esp + 28]  // width
+    shl        esi, 16          // 4, stride
+    add        esi, 4
+    movd       xmm5, esi
+    sub        ecx, 4
+    jl         l4b
+
+    // setup for 4 pixel loop
+    pshufd     xmm7, xmm7, 0x44  // dup dudv
+    pshufd     xmm5, xmm5, 0  // dup 4, stride
+    movdqa     xmm0, xmm2    // x0, y0, x1, y1
+    addps      xmm0, xmm7
+    movlhps    xmm2, xmm0
+    movdqa     xmm4, xmm7
+    addps      xmm4, xmm4    // dudv *= 2
+    movdqa     xmm3, xmm2    // x2, y2, x3, y3
+    addps      xmm3, xmm4
+    addps      xmm4, xmm4    // dudv *= 4
+
+    // 4 pixel loop
+    align      4
+  l4:
+    cvttps2dq  xmm0, xmm2    // x, y float to int first 2
+    cvttps2dq  xmm1, xmm3    // x, y float to int next 2
+    packssdw   xmm0, xmm1    // x, y as 8 shorts
+    pmaddwd    xmm0, xmm5    // offsets = x * 4 + y * stride.
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       edi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       xmm1, [eax + esi]  // read pixel 0
+    movd       xmm6, [eax + edi]  // read pixel 1
+    punpckldq  xmm1, xmm6     // combine pixel 0 and 1
+    addps      xmm2, xmm4    // x, y += dx, dy first 2
+    movq       qword ptr [edx], xmm1
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       edi, xmm0
+    movd       xmm6, [eax + esi]  // read pixel 2
+    movd       xmm0, [eax + edi]  // read pixel 3
+    punpckldq  xmm6, xmm0     // combine pixel 2 and 3
+    addps      xmm3, xmm4    // x, y += dx, dy next 2
+    sub        ecx, 4
+    movq       qword ptr 8[edx], xmm6
+    lea        edx, [edx + 16]
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+    align      4
+  l1:
+    cvttps2dq  xmm0, xmm2    // x, y float to int
+    packssdw   xmm0, xmm0    // x, y as shorts
+    pmaddwd    xmm0, xmm5    // offset = x * 4 + y * stride
+    addps      xmm2, xmm7    // x, y += dx, dy
+    movd       esi, xmm0
+    movd       xmm0, [eax + esi]  // copy a pixel
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        l1
+  l1b:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBAFFINEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    sub        edi, esi
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    vmovd      xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    vmovd      xmm5, eax  // low fraction 128..1
+    vpunpcklbw xmm5, xmm5, xmm0
+    vpunpcklwd xmm5, xmm5, xmm5
+    vpxor      ymm0, ymm0, ymm0
+    vpermd     ymm5, ymm0, ymm5
+
+    align      4
+  xloop:
+    vmovdqu    ymm0, [esi]
+    vmovdqu    ymm2, [esi + edx]
+    vpunpckhbw ymm1, ymm0, ymm2  // mutates
+    vpunpcklbw ymm0, ymm0, ymm2  // mutates
+    vpmaddubsw ymm0, ymm0, ymm5
+    vpmaddubsw ymm1, ymm1, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm1, ymm1, 7
+    vpackuswb  ymm0, ymm0, ymm1  // unmutates
+    sub        ecx, 32
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    vmovdqu    ymm0, [esi]
+    vpavgb     ymm0, ymm0, [esi + edx]
+    vpavgb     ymm0, ymm0, [esi + edx]
+    sub        ecx, 32
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    vmovdqu    ymm0, [esi]
+    vpavgb     ymm0, ymm0, [esi + edx]
+    sub        ecx, 32
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    vmovdqu    ymm0, [esi + edx]
+    vpavgb     ymm0, ymm0, [esi]
+    vpavgb     ymm0, ymm0, [esi]
+    sub        ecx, 32
+    vmovdqu     [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    rep movsb
+
+  xloop99:
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_AVX2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    movd       xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    movd       xmm5, eax  // low fraction 128..1
+    punpcklbw  xmm5, xmm0
+    punpcklwd  xmm5, xmm5
+    pshufd     xmm5, xmm5, 0
+
+    align      4
+  xloop:
+    movdqa     xmm0, [esi]
+    movdqa     xmm2, [esi + edx]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm2
+    punpckhbw  xmm1, xmm2
+    pmaddubsw  xmm0, xmm5
+    pmaddubsw  xmm1, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm1, 7
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqa     xmm1, [esi]
+    movdqa     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqa     xmm0, [esi]
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 256.  Blend 100 / 0.
+    cmp        eax, 64
+    je         xloop75   // 64 / 256 is 0.25.  Blend 75 / 25.
+    cmp        eax, 128
+    je         xloop50   // 128 / 256 is 0.50.  Blend 50 / 50.
+    cmp        eax, 192
+    je         xloop25   // 192 / 256 is 0.75.  Blend 25 / 75.
+
+    movd       xmm5, eax            // xmm5 = y fraction
+    punpcklbw  xmm5, xmm5
+    psrlw      xmm5, 1
+    punpcklwd  xmm5, xmm5
+    punpckldq  xmm5, xmm5
+    punpcklqdq xmm5, xmm5
+    pxor       xmm4, xmm4
+
+    align      4
+  xloop:
+    movdqa     xmm0, [esi]  // row0
+    movdqa     xmm2, [esi + edx]  // row1
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    psubw      xmm2, xmm0  // row1 - row0
+    psubw      xmm3, xmm1
+    paddw      xmm2, xmm2  // 9 bits * 15 bits = 8.16
+    paddw      xmm3, xmm3
+    pmulhw     xmm2, xmm5  // scale diff
+    pmulhw     xmm3, xmm5
+    paddw      xmm0, xmm2  // sum rows
+    paddw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqa     xmm1, [esi]
+    movdqa     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqa     xmm0, [esi]
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride, int dst_width,
+                                    int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    movd       xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    movd       xmm5, eax  // low fraction 128..1
+    punpcklbw  xmm5, xmm0
+    punpcklwd  xmm5, xmm5
+    pshufd     xmm5, xmm5, 0
+
+    align      4
+  xloop:
+    movdqu     xmm0, [esi]
+    movdqu     xmm2, [esi + edx]
+    movdqu     xmm1, xmm0
+    punpcklbw  xmm0, xmm2
+    punpckhbw  xmm1, xmm2
+    pmaddubsw  xmm0, xmm5
+    pmaddubsw  xmm1, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm1, 7
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqu     xmm1, [esi]
+    movdqu     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqu     xmm0, [esi]
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride, int dst_width,
+                                   int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 256.  Blend 100 / 0.
+    cmp        eax, 64
+    je         xloop75   // 64 / 256 is 0.25.  Blend 75 / 25.
+    cmp        eax, 128
+    je         xloop50   // 128 / 256 is 0.50.  Blend 50 / 50.
+    cmp        eax, 192
+    je         xloop25   // 192 / 256 is 0.75.  Blend 25 / 75.
+
+    movd       xmm5, eax            // xmm5 = y fraction
+    punpcklbw  xmm5, xmm5
+    psrlw      xmm5, 1
+    punpcklwd  xmm5, xmm5
+    punpckldq  xmm5, xmm5
+    punpcklqdq xmm5, xmm5
+    pxor       xmm4, xmm4
+
+    align      4
+  xloop:
+    movdqu     xmm0, [esi]  // row0
+    movdqu     xmm2, [esi + edx]  // row1
+    movdqu     xmm1, xmm0
+    movdqu     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    psubw      xmm2, xmm0  // row1 - row0
+    psubw      xmm3, xmm1
+    paddw      xmm2, xmm2  // 9 bits * 15 bits = 8.16
+    paddw      xmm3, xmm3
+    pmulhw     xmm2, xmm5  // scale diff
+    pmulhw     xmm3, xmm5
+    paddw      xmm0, xmm2  // sum rows
+    paddw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqu     xmm1, [esi]
+    movdqu     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqu     xmm0, [esi]
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+__declspec(naked) __declspec(align(16))
+void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // src_uv_stride
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    sub        edi, eax
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    pavgb      xmm0, [eax + edx]
+    sub        ecx, 16
+    movdqa     [eax + edi], xmm0
+    lea        eax,  [eax + 16]
+    jg         convertloop
+    pop        edi
+    ret
+  }
+}
+
+#ifdef HAS_HALFROW_AVX2
+__declspec(naked) __declspec(align(16))
+void HalfRow_AVX2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // src_uv_stride
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    sub        edi, eax
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vpavgb     ymm0, ymm0, [eax + edx]
+    sub        ecx, 32
+    vmovdqu    [eax + edi], ymm0
+    lea        eax,  [eax + 32]
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_HALFROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                          uint32 selector, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_bayer
+    movd       xmm5, [esp + 12]  // selector
+    mov        ecx, [esp + 16]   // pix
+    pshufd     xmm5, xmm5, 0
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    punpckldq  xmm0, xmm1
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+    ret
+  }
+}
+
+// Specialized ARGB to Bayer that just isolates G channel.
+__declspec(naked) __declspec(align(16))
+void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 selector, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_bayer
+                                 // selector
+    mov        ecx, [esp + 16]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x000000ff
+    psrld      xmm5, 24
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    psrld      xmm0, 8  // Move green to bottom.
+    psrld      xmm1, 8
+    pand       xmm0, xmm5
+    pand       xmm1, xmm5
+    packssdw   xmm0, xmm1
+    packuswb   xmm0, xmm1
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+    ret
+  }
+}
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_argb
+    mov        ecx, [esp + 12]   // shuffler
+    movdqa     xmm5, [ecx]
+    mov        ecx, [esp + 16]   // pix
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                    const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_argb
+    mov        ecx, [esp + 12]   // shuffler
+    movdqa     xmm5, [ecx]
+    mov        ecx, [esp + 16]   // pix
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    sub        ecx, 8
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]     // src_argb
+    mov        edx, [esp + 8]     // dst_argb
+    mov        ecx, [esp + 12]    // shuffler
+    vbroadcastf128 ymm5, [ecx]    // same shuffle in high as low.
+    mov        ecx, [esp + 16]    // pix
+
+    align      4
+  wloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax, [eax + 64]
+    vpshufb    ymm0, ymm0, ymm5
+    vpshufb    ymm1, ymm1, ymm5
+    sub        ecx, 16
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx, [edx + 64]
+    jg         wloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBSHUFFLEROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  __asm {
+    push       ebx
+    push       esi
+    mov        eax, [esp + 8 + 4]    // src_argb
+    mov        edx, [esp + 8 + 8]    // dst_argb
+    mov        esi, [esp + 8 + 12]   // shuffler
+    mov        ecx, [esp + 8 + 16]   // pix
+    pxor       xmm5, xmm5
+
+    mov        ebx, [esi]   // shuffler
+    cmp        ebx, 0x03000102
+    je         shuf_3012
+    cmp        ebx, 0x00010203
+    je         shuf_0123
+    cmp        ebx, 0x00030201
+    je         shuf_0321
+    cmp        ebx, 0x02010003
+    je         shuf_2103
+
+  // TODO(fbarchard): Use one source pointer and 3 offsets.
+  shuf_any1:
+    movzx      ebx, byte ptr [esi]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx], bl
+    movzx      ebx, byte ptr [esi + 1]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 1], bl
+    movzx      ebx, byte ptr [esi + 2]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 2], bl
+    movzx      ebx, byte ptr [esi + 3]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 3], bl
+    lea        eax, [eax + 4]
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jg         shuf_any1
+    jmp        shuf99
+
+    align      4
+  shuf_0123:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 01Bh   // 1B = 00011011 = 0x0123 = BGRAToARGB
+    pshuflw    xmm0, xmm0, 01Bh
+    pshufhw    xmm1, xmm1, 01Bh
+    pshuflw    xmm1, xmm1, 01Bh
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_0123
+    jmp        shuf99
+
+    align      4
+  shuf_0321:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 039h   // 39 = 00111001 = 0x0321 = RGBAToARGB
+    pshuflw    xmm0, xmm0, 039h
+    pshufhw    xmm1, xmm1, 039h
+    pshuflw    xmm1, xmm1, 039h
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_0321
+    jmp        shuf99
+
+    align      4
+  shuf_2103:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 093h   // 93 = 10010011 = 0x2103 = ARGBToRGBA
+    pshuflw    xmm0, xmm0, 093h
+    pshufhw    xmm1, xmm1, 093h
+    pshuflw    xmm1, xmm1, 093h
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_2103
+    jmp        shuf99
+
+    align      4
+  shuf_3012:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 0C6h   // C6 = 11000110 = 0x3012 = ABGRToARGB
+    pshuflw    xmm0, xmm0, 0C6h
+    pshufhw    xmm1, xmm1, 0C6h
+    pshuflw    xmm1, xmm1, 0C6h
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_3012
+
+  shuf99:
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// YUY2 - Macro-pixel = 2 image pixels
+// Y0U0Y1V0....Y2U2Y3V2...Y4U4Y5V4....
+
+// UYVY - Macro-pixel = 2 image pixels
+// U0Y0V0Y1
+
+__declspec(naked) __declspec(align(16))
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_y
+    mov        esi, [esp + 8 + 8]    // src_u
+    mov        edx, [esp + 8 + 12]   // src_v
+    mov        edi, [esp + 8 + 16]   // dst_frame
+    mov        ecx, [esp + 8 + 20]   // width
+    sub        edx, esi
+
+    align      4
+  convertloop:
+    movq       xmm2, qword ptr [esi] // U
+    movq       xmm3, qword ptr [esi + edx] // V
+    lea        esi, [esi + 8]
+    punpcklbw  xmm2, xmm3 // UV
+    movdqu     xmm0, [eax] // Y
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm2 // YUYV
+    punpckhbw  xmm1, xmm2
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm1
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_y
+    mov        esi, [esp + 8 + 8]    // src_u
+    mov        edx, [esp + 8 + 12]   // src_v
+    mov        edi, [esp + 8 + 16]   // dst_frame
+    mov        ecx, [esp + 8 + 20]   // width
+    sub        edx, esi
+
+    align      4
+  convertloop:
+    movq       xmm2, qword ptr [esi] // U
+    movq       xmm3, qword ptr [esi + edx] // V
+    lea        esi, [esi + 8]
+    punpcklbw  xmm2, xmm3 // UV
+    movdqu     xmm0, [eax] // Y
+    movdqa     xmm1, xmm2
+    lea        eax, [eax + 16]
+    punpcklbw  xmm1, xmm0 // UYVY
+    punpckhbw  xmm2, xmm0
+    movdqu     [edi], xmm1
+    movdqu     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+__declspec(naked) __declspec(align(16))
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* src_argb */
+    mov        edx, [esp + 4 + 8]   /* dst_argb */
+    mov        esi, [esp + 4 + 12]  /* poly */
+    mov        ecx, [esp + 4 + 16]  /* width */
+    pxor       xmm3, xmm3  // 0 constant for zero extending bytes to ints.
+
+    // 2 pixel loop.
+    align      4
+ convertloop:
+//    pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
+//    pmovzxbd  xmm4, dword ptr [eax + 4]  // BGRA pixel
+    movq       xmm0, qword ptr [eax]  // BGRABGRA
+    lea        eax, [eax + 8]
+    punpcklbw  xmm0, xmm3
+    movdqa     xmm4, xmm0
+    punpcklwd  xmm0, xmm3  // pixel 0
+    punpckhwd  xmm4, xmm3  // pixel 1
+    cvtdq2ps   xmm0, xmm0  // 4 floats
+    cvtdq2ps   xmm4, xmm4
+    movdqa     xmm1, xmm0  // X
+    movdqa     xmm5, xmm4
+    mulps      xmm0, [esi + 16]  // C1 * X
+    mulps      xmm4, [esi + 16]
+    addps      xmm0, [esi]  // result = C0 + C1 * X
+    addps      xmm4, [esi]
+    movdqa     xmm2, xmm1
+    movdqa     xmm6, xmm5
+    mulps      xmm2, xmm1  // X * X
+    mulps      xmm6, xmm5
+    mulps      xmm1, xmm2  // X * X * X
+    mulps      xmm5, xmm6
+    mulps      xmm2, [esi + 32]  // C2 * X * X
+    mulps      xmm6, [esi + 32]
+    mulps      xmm1, [esi + 48]  // C3 * X * X * X
+    mulps      xmm5, [esi + 48]
+    addps      xmm0, xmm2  // result += C2 * X * X
+    addps      xmm4, xmm6
+    addps      xmm0, xmm1  // result += C3 * X * X * X
+    addps      xmm4, xmm5
+    cvttps2dq  xmm0, xmm0
+    cvttps2dq  xmm4, xmm4
+    packuswb   xmm0, xmm4
+    packuswb   xmm0, xmm0
+    sub        ecx, 2
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         convertloop
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+__declspec(naked) __declspec(align(16))
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]   /* poly */
+    vbroadcastf128 ymm4, [ecx]       // C0
+    vbroadcastf128 ymm5, [ecx + 16]  // C1
+    vbroadcastf128 ymm6, [ecx + 32]  // C2
+    vbroadcastf128 ymm7, [ecx + 48]  // C3
+    mov        ecx, [esp + 16]  /* width */
+
+    // 2 pixel loop.
+    align      4
+ convertloop:
+    vpmovzxbd   ymm0, qword ptr [eax]  // 2 BGRA pixels
+    lea         eax, [eax + 8]
+    vcvtdq2ps   ymm0, ymm0        // X 8 floats
+    vmulps      ymm2, ymm0, ymm0  // X * X
+    vmulps      ymm3, ymm0, ymm7  // C3 * X
+    vfmadd132ps ymm0, ymm4, ymm5  // result = C0 + C1 * X
+    vfmadd231ps ymm0, ymm2, ymm6  // result += C2 * X * X
+    vfmadd231ps ymm0, ymm2, ymm3  // result += C3 * X * X * X
+    vcvttps2dq  ymm0, ymm0
+    vpackusdw   ymm0, ymm0, ymm0  // b0g0r0a0_00000000_b0g0r0a0_00000000
+    vpermq      ymm0, ymm0, 0xd8  // b0g0r0a0_b0g0r0a0_00000000_00000000
+    vpackuswb   xmm0, xmm0, xmm0  // bgrabgra_00000000_00000000_00000000
+    sub         ecx, 2
+    vmovq       qword ptr [edx], xmm0
+    lea         edx, [edx + 8]
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
+#ifdef HAS_ARGBCOLORTABLEROW_X86
+// Tranform ARGB pixels with color table.
+__declspec(naked) __declspec(align(16))
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+                           int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* dst_argb */
+    mov        esi, [esp + 4 + 8]   /* table_argb */
+    mov        ecx, [esp + 4 + 12]  /* width */
+
+    // 1 pixel loop.
+    align      4
+  convertloop:
+    movzx      edx, byte ptr [eax]
+    lea        eax, [eax + 4]
+    movzx      edx, byte ptr [esi + edx * 4]
+    mov        byte ptr [eax - 4], dl
+    movzx      edx, byte ptr [eax - 4 + 1]
+    movzx      edx, byte ptr [esi + edx * 4 + 1]
+    mov        byte ptr [eax - 4 + 1], dl
+    movzx      edx, byte ptr [eax - 4 + 2]
+    movzx      edx, byte ptr [esi + edx * 4 + 2]
+    mov        byte ptr [eax - 4 + 2], dl
+    movzx      edx, byte ptr [eax - 4 + 3]
+    movzx      edx, byte ptr [esi + edx * 4 + 3]
+    mov        byte ptr [eax - 4 + 3], dl
+    dec        ecx
+    jg         convertloop
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBCOLORTABLEROW_X86
+
+#ifdef HAS_RGBCOLORTABLEROW_X86
+// Tranform RGB pixels with color table.
+__declspec(naked) __declspec(align(16))
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* dst_argb */
+    mov        esi, [esp + 4 + 8]   /* table_argb */
+    mov        ecx, [esp + 4 + 12]  /* width */
+
+    // 1 pixel loop.
+    align      4
+  convertloop:
+    movzx      edx, byte ptr [eax]
+    lea        eax, [eax + 4]
+    movzx      edx, byte ptr [esi + edx * 4]
+    mov        byte ptr [eax - 4], dl
+    movzx      edx, byte ptr [eax - 4 + 1]
+    movzx      edx, byte ptr [esi + edx * 4 + 1]
+    mov        byte ptr [eax - 4 + 1], dl
+    movzx      edx, byte ptr [eax - 4 + 2]
+    movzx      edx, byte ptr [esi + edx * 4 + 2]
+    mov        byte ptr [eax - 4 + 2], dl
+    dec        ecx
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_RGBCOLORTABLEROW_X86
+
+#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// Tranform RGB pixels with luma table.
+__declspec(naked) __declspec(align(16))
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   /* src_argb */
+    mov        edi, [esp + 8 + 8]   /* dst_argb */
+    mov        ecx, [esp + 8 + 12]  /* width */
+    movd       xmm2, dword ptr [esp + 8 + 16]  // luma table
+    movd       xmm3, dword ptr [esp + 8 + 20]  // lumacoeff
+    pshufd     xmm2, xmm2, 0
+    pshufd     xmm3, xmm3, 0
+    pcmpeqb    xmm4, xmm4        // generate mask 0xff00ff00
+    psllw      xmm4, 8
+    pxor       xmm5, xmm5
+
+    // 4 pixel loop.
+    align      4
+  convertloop:
+    movdqu     xmm0, qword ptr [eax]      // generate luma ptr
+    pmaddubsw  xmm0, xmm3
+    phaddw     xmm0, xmm0
+    pand       xmm0, xmm4  // mask out low bits
+    punpcklwd  xmm0, xmm5
+    paddd      xmm0, xmm2  // add table base
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi], dl
+    movzx      edx, byte ptr [eax + 1]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 1], dl
+    movzx      edx, byte ptr [eax + 2]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 2], dl
+    movzx      edx, byte ptr [eax + 3]  // copy alpha.
+    mov        byte ptr [edi + 3], dl
+
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax + 4]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 4], dl
+    movzx      edx, byte ptr [eax + 5]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 5], dl
+    movzx      edx, byte ptr [eax + 6]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 6], dl
+    movzx      edx, byte ptr [eax + 7]  // copy alpha.
+    mov        byte ptr [edi + 7], dl
+
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax + 8]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 8], dl
+    movzx      edx, byte ptr [eax + 9]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 9], dl
+    movzx      edx, byte ptr [eax + 10]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 10], dl
+    movzx      edx, byte ptr [eax + 11]  // copy alpha.
+    mov        byte ptr [edi + 11], dl
+
+    movd       esi, xmm0
+
+    movzx      edx, byte ptr [eax + 12]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 12], dl
+    movzx      edx, byte ptr [eax + 13]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 13], dl
+    movzx      edx, byte ptr [eax + 14]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 14], dl
+    movzx      edx, byte ptr [eax + 15]  // copy alpha.
+    mov        byte ptr [edi + 15], dl
+
+    sub        ecx, 4
+    lea        eax, [eax + 16]
+    lea        edi, [edi + 16]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/scale.c b/third_party/libyuv/source/scale.c
deleted file mode 100644
index 1809300..0000000
--- a/third_party/libyuv/source/scale.c
+++ /dev/null
@@ -1,3884 +0,0 @@
-/*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "third_party/libyuv/include/libyuv/scale.h"
-
-#include <assert.h>
-#include <string.h>
-
-#include "third_party/libyuv/include/libyuv/cpu_id.h"
-#include "third_party/libyuv/source/row.h"
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-/*
- * Note: Defining YUV_DISABLE_ASM allows to use c version.
- */
-//#define YUV_DISABLE_ASM
-
-#if defined(_MSC_VER)
-#define ALIGN16(var) __declspec(align(16)) var
-#else
-#define ALIGN16(var) var __attribute__((aligned(16)))
-#endif
-
-// Note: A Neon reference manual
-// http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0204j/CJAJIIGG.html
-// Note: Some SSE2 reference manuals
-// cpuvol1.pdf agner_instruction_tables.pdf 253666.pdf 253667.pdf
-
-// Set the following flag to true to revert to only
-// using the reference implementation ScalePlaneBox(), and
-// NOT the optimized versions. Useful for debugging and
-// when comparing the quality of the resulting YUV planes
-// as produced by the optimized and non-optimized versions.
-
-static int use_reference_impl_ = 0;
-
-void SetUseReferenceImpl(int use) {
-  use_reference_impl_ = use;
-}
-
-// ScaleRowDown2Int also used by planar functions
-
-/**
- * NEON downscalers with interpolation.
- *
- * Provided by Fritz Koenig
- *
- */
-
-#if defined(__ARM_NEON__) && !defined(YUV_DISABLE_ASM)
-#define HAS_SCALEROWDOWN2_NEON
-void ScaleRowDown2_NEON(const uint8* src_ptr, int  src_stride,
-                        uint8* dst, int dst_width) {
-  asm volatile (
-    "1:                                        \n"
-    "vld2.u8    {q0,q1}, [%0]!                 \n"  // load even pixels into q0, odd into q1
-    "vst1.u8    {q0}, [%1]!                    \n"  // store even pixels
-    "subs       %2, %2, #16                    \n"  // 16 processed per loop
-    "bhi        1b                             \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst),              // %1
-      "+r"(dst_width)         // %2
-    :
-    : "q0", "q1"              // Clobber List
-  );
-}
-
-void ScaleRowDown2Int_NEON(const uint8* src_ptr, int src_stride,
-                           uint8* dst, int dst_width) {
-  asm volatile (
-    "add        %1, %0                         \n"  // change the stride to row 2 pointer
-    "1:                                        \n"
-    "vld1.u8    {q0,q1}, [%0]!                 \n"  // load row 1 and post increment
-    "vld1.u8    {q2,q3}, [%1]!                 \n"  // load row 2 and post increment
-    "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
-    "vpaddl.u8  q1, q1                         \n"
-    "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent, add row 1 to row 2
-    "vpadal.u8  q1, q3                         \n"
-    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
-    "vrshrn.u16 d1, q1, #2                     \n"
-    "vst1.u8    {q0}, [%2]!                    \n"
-    "subs       %3, %3, #16                    \n"  // 16 processed per loop
-    "bhi        1b                             \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(src_stride),       // %1
-      "+r"(dst),              // %2
-      "+r"(dst_width)         // %3
-    :
-    : "q0", "q1", "q2", "q3"     // Clobber List
-   );
-}
-
-#define HAS_SCALEROWDOWN4_NEON
-static void ScaleRowDown4_NEON(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "1:                                        \n"
-    "vld2.u8    {d0, d1}, [%0]!                \n"
-    "vtrn.u8    d1, d0                         \n"
-    "vshrn.u16  d0, q0, #8                     \n"
-    "vst1.u32   {d0[1]}, [%1]!                 \n"
-
-    "subs       %2, #4                         \n"
-    "bhi        1b                             \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width)         // %2
-    :
-    : "q0", "q1", "memory", "cc"
-  );
-}
-
-static void ScaleRowDown4Int_NEON(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "add        r4, %0, %3                     \n"
-    "add        r5, r4, %3                     \n"
-    "add        %3, r5, %3                     \n"
-    "1:                                        \n"
-    "vld1.u8    {q0}, [%0]!                    \n"   // load up 16x4 block of input data
-    "vld1.u8    {q1}, [r4]!                    \n"
-    "vld1.u8    {q2}, [r5]!                    \n"
-    "vld1.u8    {q3}, [%3]!                    \n"
-
-    "vpaddl.u8  q0, q0                         \n"
-    "vpadal.u8  q0, q1                         \n"
-    "vpadal.u8  q0, q2                         \n"
-    "vpadal.u8  q0, q3                         \n"
-
-    "vpaddl.u16 q0, q0                         \n"
-
-    "vrshrn.u32 d0, q0, #4                     \n"   // divide by 16 w/rounding
-
-    "vmovn.u16  d0, q0                         \n"
-    "vst1.u32   {d0[0]}, [%1]!                 \n"
-
-    "subs       %2, #4                         \n"
-    "bhi        1b                             \n"
-
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width)         // %2
-    : "r"(src_stride)         // %3
-    : "r4", "r5", "q0", "q1", "q2", "q3", "memory", "cc"
-  );
-}
-
-#define HAS_SCALEROWDOWN34_NEON
-// Down scale from 4 to 3 pixels.  Use the neon multilane read/write
-//  to load up the every 4th pixel into a 4 different registers.
-// Point samples 32 pixels to 24 pixels.
-static void ScaleRowDown34_NEON(const uint8* src_ptr, int src_stride,
-                                uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "1:                                        \n"
-    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n" // src line 0
-    "vmov         d2, d3                       \n" // order needs to be d0, d1, d2
-    "vst3.u8      {d0, d1, d2}, [%1]!          \n"
-    "subs         %2, #24                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width)         // %2
-    :
-    : "d0", "d1", "d2", "d3", "memory", "cc"
-  );
-}
-
-static void ScaleRowDown34_0_Int_NEON(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vmov.u8      d24, #3                      \n"
-    "add          %3, %0                       \n"
-    "1:                                        \n"
-    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n" // src line 0
-    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n" // src line 1
-
-    // filter src line 0 with src line 1
-    // expand chars to shorts to allow for room
-    // when adding lines together
-    "vmovl.u8     q8, d4                       \n"
-    "vmovl.u8     q9, d5                       \n"
-    "vmovl.u8     q10, d6                      \n"
-    "vmovl.u8     q11, d7                      \n"
-
-    // 3 * line_0 + line_1
-    "vmlal.u8     q8, d0, d24                  \n"
-    "vmlal.u8     q9, d1, d24                  \n"
-    "vmlal.u8     q10, d2, d24                 \n"
-    "vmlal.u8     q11, d3, d24                 \n"
-
-    // (3 * line_0 + line_1) >> 2
-    "vqrshrn.u16  d0, q8, #2                   \n"
-    "vqrshrn.u16  d1, q9, #2                   \n"
-    "vqrshrn.u16  d2, q10, #2                  \n"
-    "vqrshrn.u16  d3, q11, #2                  \n"
-
-    // a0 = (src[0] * 3 + s[1] * 1) >> 2
-    "vmovl.u8     q8, d1                       \n"
-    "vmlal.u8     q8, d0, d24                  \n"
-    "vqrshrn.u16  d0, q8, #2                   \n"
-
-    // a1 = (src[1] * 1 + s[2] * 1) >> 1
-    "vrhadd.u8    d1, d1, d2                   \n"
-
-    // a2 = (src[2] * 1 + s[3] * 3) >> 2
-    "vmovl.u8     q8, d2                       \n"
-    "vmlal.u8     q8, d3, d24                  \n"
-    "vqrshrn.u16  d2, q8, #2                   \n"
-
-    "vst3.u8      {d0, d1, d2}, [%1]!          \n"
-
-    "subs         %2, #24                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width),        // %2
-      "+r"(src_stride)        // %3
-    :
-    : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc"
-  );
-}
-
-static void ScaleRowDown34_1_Int_NEON(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vmov.u8      d24, #3                      \n"
-    "add          %3, %0                       \n"
-    "1:                                        \n"
-    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n" // src line 0
-    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n" // src line 1
-
-    // average src line 0 with src line 1
-    "vrhadd.u8    q0, q0, q2                   \n"
-    "vrhadd.u8    q1, q1, q3                   \n"
-
-    // a0 = (src[0] * 3 + s[1] * 1) >> 2
-    "vmovl.u8     q3, d1                       \n"
-    "vmlal.u8     q3, d0, d24                  \n"
-    "vqrshrn.u16  d0, q3, #2                   \n"
-
-    // a1 = (src[1] * 1 + s[2] * 1) >> 1
-    "vrhadd.u8    d1, d1, d2                   \n"
-
-    // a2 = (src[2] * 1 + s[3] * 3) >> 2
-    "vmovl.u8     q3, d2                       \n"
-    "vmlal.u8     q3, d3, d24                  \n"
-    "vqrshrn.u16  d2, q3, #2                   \n"
-
-    "vst3.u8      {d0, d1, d2}, [%1]!          \n"
-
-    "subs         %2, #24                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width),        // %2
-      "+r"(src_stride)        // %3
-    :
-    : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
-  );
-}
-
-#define HAS_SCALEROWDOWN38_NEON
-const uint8 shuf38[16] __attribute__ ((aligned(16))) =
-  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
-const uint8 shuf38_2[16] __attribute__ ((aligned(16))) =
-  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
-const unsigned short mult38_div6[8] __attribute__ ((aligned(16))) =
-  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
-    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
-const unsigned short mult38_div9[8] __attribute__ ((aligned(16))) =
-  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
-    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
-
-// 32 -> 12
-static void ScaleRowDown38_NEON(const uint8* src_ptr, int src_stride,
-                                uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vld1.u8      {q3}, [%3]                   \n"
-    "1:                                        \n"
-    "vld1.u8      {d0, d1, d2, d3}, [%0]!      \n"
-    "vtbl.u8      d4, {d0, d1, d2, d3}, d6     \n"
-    "vtbl.u8      d5, {d0, d1, d2, d3}, d7     \n"
-    "vst1.u8      {d4}, [%1]!                  \n"
-    "vst1.u32     {d5[0]}, [%1]!               \n"
-    "subs         %2, #12                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width)         // %2
-    : "r"(shuf38)             // %3
-    : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
-  );
-}
-
-// 32x3 -> 12x1
-static void ScaleRowDown38_3_Int_NEON(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vld1.u16     {q13}, [%4]                  \n"
-    "vld1.u8      {q14}, [%5]                  \n"
-    "vld1.u8      {q15}, [%6]                  \n"
-    "add          r4, %0, %3, lsl #1           \n"
-    "add          %3, %0                       \n"
-    "1:                                        \n"
-
-    // d0 = 00 40 01 41 02 42 03 43
-    // d1 = 10 50 11 51 12 52 13 53
-    // d2 = 20 60 21 61 22 62 23 63
-    // d3 = 30 70 31 71 32 72 33 73
-    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n"
-    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n"
-    "vld4.u8      {d16, d17, d18, d19}, [r4]!  \n"
-
-    // Shuffle the input data around to get align the data
-    //  so adjacent data can be added.  0,1 - 2,3 - 4,5 - 6,7
-    // d0 = 00 10 01 11 02 12 03 13
-    // d1 = 40 50 41 51 42 52 43 53
-    "vtrn.u8      d0, d1                       \n"
-    "vtrn.u8      d4, d5                       \n"
-    "vtrn.u8      d16, d17                     \n"
-
-    // d2 = 20 30 21 31 22 32 23 33
-    // d3 = 60 70 61 71 62 72 63 73
-    "vtrn.u8      d2, d3                       \n"
-    "vtrn.u8      d6, d7                       \n"
-    "vtrn.u8      d18, d19                     \n"
-
-    // d0 = 00+10 01+11 02+12 03+13
-    // d2 = 40+50 41+51 42+52 43+53
-    "vpaddl.u8    q0, q0                       \n"
-    "vpaddl.u8    q2, q2                       \n"
-    "vpaddl.u8    q8, q8                       \n"
-
-    // d3 = 60+70 61+71 62+72 63+73
-    "vpaddl.u8    d3, d3                       \n"
-    "vpaddl.u8    d7, d7                       \n"
-    "vpaddl.u8    d19, d19                     \n"
-
-    // combine source lines
-    "vadd.u16     q0, q2                       \n"
-    "vadd.u16     q0, q8                       \n"
-    "vadd.u16     d4, d3, d7                   \n"
-    "vadd.u16     d4, d19                      \n"
-
-    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
-    //             + s[6 + st * 1] + s[7 + st * 1]
-    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
-    "vqrdmulh.s16 q2, q13                      \n"
-    "vmovn.u16    d4, q2                       \n"
-
-    // Shuffle 2,3 reg around so that 2 can be added to the
-    //  0,1 reg and 3 can be added to the 4,5 reg.  This
-    //  requires expanding from u8 to u16 as the 0,1 and 4,5
-    //  registers are already expanded.  Then do transposes
-    //  to get aligned.
-    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-    "vmovl.u8     q1, d2                       \n"
-    "vmovl.u8     q3, d6                       \n"
-    "vmovl.u8     q9, d18                      \n"
-
-    // combine source lines
-    "vadd.u16     q1, q3                       \n"
-    "vadd.u16     q1, q9                       \n"
-
-    // d4 = xx 20 xx 30 xx 22 xx 32
-    // d5 = xx 21 xx 31 xx 23 xx 33
-    "vtrn.u32     d2, d3                       \n"
-
-    // d4 = xx 20 xx 21 xx 22 xx 23
-    // d5 = xx 30 xx 31 xx 32 xx 33
-    "vtrn.u16     d2, d3                       \n"
-
-    // 0+1+2, 3+4+5
-    "vadd.u16     q0, q1                       \n"
-
-    // Need to divide, but can't downshift as the the value
-    //  isn't a power of 2.  So multiply by 65536 / n
-    //  and take the upper 16 bits.
-    "vqrdmulh.s16 q0, q15                      \n"
-
-    // Align for table lookup, vtbl requires registers to
-    //  be adjacent
-    "vmov.u8      d2, d4                       \n"
-
-    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
-    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
-
-    "vst1.u8      {d3}, [%1]!                  \n"
-    "vst1.u32     {d4[0]}, [%1]!               \n"
-    "subs         %2, #12                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width),        // %2
-      "+r"(src_stride)        // %3
-    : "r"(mult38_div6),       // %4
-      "r"(shuf38_2),          // %5
-      "r"(mult38_div9)        // %6
-    : "r4", "q0", "q1", "q2", "q3", "q8", "q9",
-      "q13", "q14", "q15", "memory", "cc"
-  );
-}
-
-// 32x2 -> 12x1
-static void ScaleRowDown38_2_Int_NEON(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vld1.u16     {q13}, [%4]                  \n"
-    "vld1.u8      {q14}, [%5]                  \n"
-    "add          %3, %0                       \n"
-    "1:                                        \n"
-
-    // d0 = 00 40 01 41 02 42 03 43
-    // d1 = 10 50 11 51 12 52 13 53
-    // d2 = 20 60 21 61 22 62 23 63
-    // d3 = 30 70 31 71 32 72 33 73
-    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n"
-    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n"
-
-    // Shuffle the input data around to get align the data
-    //  so adjacent data can be added.  0,1 - 2,3 - 4,5 - 6,7
-    // d0 = 00 10 01 11 02 12 03 13
-    // d1 = 40 50 41 51 42 52 43 53
-    "vtrn.u8      d0, d1                       \n"
-    "vtrn.u8      d4, d5                       \n"
-
-    // d2 = 20 30 21 31 22 32 23 33
-    // d3 = 60 70 61 71 62 72 63 73
-    "vtrn.u8      d2, d3                       \n"
-    "vtrn.u8      d6, d7                       \n"
-
-    // d0 = 00+10 01+11 02+12 03+13
-    // d2 = 40+50 41+51 42+52 43+53
-    "vpaddl.u8    q0, q0                       \n"
-    "vpaddl.u8    q2, q2                       \n"
-
-    // d3 = 60+70 61+71 62+72 63+73
-    "vpaddl.u8    d3, d3                       \n"
-    "vpaddl.u8    d7, d7                       \n"
-
-    // combine source lines
-    "vadd.u16     q0, q2                       \n"
-    "vadd.u16     d4, d3, d7                   \n"
-
-    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
-    "vqrshrn.u16  d4, q2, #2                   \n"
-
-    // Shuffle 2,3 reg around so that 2 can be added to the
-    //  0,1 reg and 3 can be added to the 4,5 reg.  This
-    //  requires expanding from u8 to u16 as the 0,1 and 4,5
-    //  registers are already expanded.  Then do transposes
-    //  to get aligned.
-    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-    "vmovl.u8     q1, d2                       \n"
-    "vmovl.u8     q3, d6                       \n"
-
-    // combine source lines
-    "vadd.u16     q1, q3                       \n"
-
-    // d4 = xx 20 xx 30 xx 22 xx 32
-    // d5 = xx 21 xx 31 xx 23 xx 33
-    "vtrn.u32     d2, d3                       \n"
-
-    // d4 = xx 20 xx 21 xx 22 xx 23
-    // d5 = xx 30 xx 31 xx 32 xx 33
-    "vtrn.u16     d2, d3                       \n"
-
-    // 0+1+2, 3+4+5
-    "vadd.u16     q0, q1                       \n"
-
-    // Need to divide, but can't downshift as the the value
-    //  isn't a power of 2.  So multiply by 65536 / n
-    //  and take the upper 16 bits.
-    "vqrdmulh.s16 q0, q13                      \n"
-
-    // Align for table lookup, vtbl requires registers to
-    //  be adjacent
-    "vmov.u8      d2, d4                       \n"
-
-    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
-    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
-
-    "vst1.u8      {d3}, [%1]!                  \n"
-    "vst1.u32     {d4[0]}, [%1]!               \n"
-    "subs         %2, #12                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width),        // %2
-      "+r"(src_stride)        // %3
-    : "r"(mult38_div6),       // %4
-      "r"(shuf38_2)           // %5
-    : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc"
-  );
-}
-
-/**
- * SSE2 downscalers with interpolation.
- *
- * Provided by Frank Barchard (fbarchard@google.com)
- *
- */
-
-// Constants for SSE2 code
-#elif (defined(_M_IX86) || defined(__i386__) || defined(__x86_64__)) && \
-    !defined(YUV_DISABLE_ASM)
-#if defined(_MSC_VER)
-#define TALIGN16(t, var) __declspec(align(16)) t _ ## var
-#elif (defined(__APPLE__) || defined(__MINGW32__) || defined(__CYGWIN__)) && defined(__i386__)
-#define TALIGN16(t, var) t var __attribute__((aligned(16)))
-#else
-#define TALIGN16(t, var) t _ ## var __attribute__((aligned(16)))
-#endif
-
-#if (defined(__APPLE__) || defined(__MINGW32__) || defined(__CYGWIN__)) && \
-    defined(__i386__)
-#define DECLARE_FUNCTION(name)                                                 \
-    ".text                                     \n"                             \
-    ".globl _" #name "                         \n"                             \
-"_" #name ":                                   \n"
-#else
-#define DECLARE_FUNCTION(name)                                                 \
-    ".text                                     \n"                             \
-    ".global " #name "                         \n"                             \
-#name ":                                       \n"
-#endif
-
-
-// Offsets for source bytes 0 to 9
-//extern "C"
-TALIGN16(const uint8, shuf0[16]) =
-  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
-
-// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
-//extern "C"
-TALIGN16(const uint8, shuf1[16]) =
-  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
-
-// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
-//extern "C"
-TALIGN16(const uint8, shuf2[16]) =
-  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
-
-// Offsets for source bytes 0 to 10
-//extern "C"
-TALIGN16(const uint8, shuf01[16]) =
-  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
-
-// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
-//extern "C"
-TALIGN16(const uint8, shuf11[16]) =
-  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
-
-// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
-//extern "C"
-TALIGN16(const uint8, shuf21[16]) =
-  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
-
-// Coefficients for source bytes 0 to 10
-//extern "C"
-TALIGN16(const uint8, madd01[16]) =
-  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
-
-// Coefficients for source bytes 10 to 21
-//extern "C"
-TALIGN16(const uint8, madd11[16]) =
-  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
-
-// Coefficients for source bytes 21 to 31
-//extern "C"
-TALIGN16(const uint8, madd21[16]) =
-  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
-
-// Coefficients for source bytes 21 to 31
-//extern "C"
-TALIGN16(const int16, round34[8]) =
-  { 2, 2, 2, 2, 2, 2, 2, 2 };
-
-//extern "C"
-TALIGN16(const uint8, shuf38a[16]) =
-  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
-
-//extern "C"
-TALIGN16(const uint8, shuf38b[16]) =
-  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
-
-// Arrange words 0,3,6 into 0,1,2
-//extern "C"
-TALIGN16(const uint8, shufac0[16]) =
-  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
-
-// Arrange words 0,3,6 into 3,4,5
-//extern "C"
-TALIGN16(const uint8, shufac3[16]) =
-  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
-
-// Scaling values for boxes of 3x3 and 2x3
-//extern "C"
-TALIGN16(const uint16, scaleac3[8]) =
-  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
-
-// Arrange first value for pixels 0,1,2,3,4,5
-//extern "C"
-TALIGN16(const uint8, shufab0[16]) =
-  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
-
-// Arrange second value for pixels 0,1,2,3,4,5
-//extern "C"
-TALIGN16(const uint8, shufab1[16]) =
-  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
-
-// Arrange third value for pixels 0,1,2,3,4,5
-//extern "C"
-TALIGN16(const uint8, shufab2[16]) =
-  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
-
-// Scaling values for boxes of 3x2 and 2x2
-//extern "C"
-TALIGN16(const uint16, scaleab2[8]) =
-  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
-#endif
-
-#if defined(_M_IX86) && !defined(YUV_DISABLE_ASM) && defined(_MSC_VER)
-
-#define HAS_SCALEROWDOWN2_SSE2
-// Reads 32 pixels, throws half away and writes 16 pixels.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
-__declspec(naked)
-static void ScaleRowDown2_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  __asm {
-    mov        eax, [esp + 4]        // src_ptr
-                                     // src_stride ignored
-    mov        edx, [esp + 12]       // dst_ptr
-    mov        ecx, [esp + 16]       // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-
-  wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax,  [eax + 32]
-    pand       xmm0, xmm5
-    pand       xmm1, xmm5
-    packuswb   xmm0, xmm1
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    sub        ecx, 16
-    ja         wloop
-
-    ret
-  }
-}
-// Blends 32x2 rectangle to 16x1.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
-__declspec(naked)
-void ScaleRowDown2Int_SSE2(const uint8* src_ptr, int src_stride,
-                           uint8* dst_ptr, int dst_width) {
-  __asm {
-    push       esi
-    mov        eax, [esp + 4 + 4]    // src_ptr
-    mov        esi, [esp + 4 + 8]    // src_stride
-    mov        edx, [esp + 4 + 12]   // dst_ptr
-    mov        ecx, [esp + 4 + 16]   // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-
-  wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + esi]
-    movdqa     xmm3, [eax + esi + 16]
-    lea        eax,  [eax + 32]
-    pavgb      xmm0, xmm2            // average rows
-    pavgb      xmm1, xmm3
-
-    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
-    psrlw      xmm0, 8
-    movdqa     xmm3, xmm1
-    psrlw      xmm1, 8
-    pand       xmm2, xmm5
-    pand       xmm3, xmm5
-    pavgw      xmm0, xmm2
-    pavgw      xmm1, xmm3
-    packuswb   xmm0, xmm1
-
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    sub        ecx, 16
-    ja         wloop
-
-    pop        esi
-    ret
-  }
-}
-
-#define HAS_SCALEROWDOWN4_SSE2
-// Point samples 32 pixels to 8 pixels.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleRowDown4_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-                                     // src_stride ignored
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x000000ff
-    psrld      xmm5, 24
-
-  wloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + 16]
-    lea        esi,  [esi + 32]
-    pand       xmm0, xmm5
-    pand       xmm1, xmm5
-    packuswb   xmm0, xmm1
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi], xmm0
-    lea        edi, [edi + 8]
-    sub        ecx, 8
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-// Blends 32x4 rectangle to 8x1.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleRowDown4Int_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        ebx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    pcmpeqb    xmm7, xmm7            // generate mask 0x00ff00ff
-    psrlw      xmm7, 8
-    lea        edx, [ebx + ebx * 2]  // src_stride * 3
-
-  wloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + 16]
-    movdqa     xmm2, [esi + ebx]
-    movdqa     xmm3, [esi + ebx + 16]
-    pavgb      xmm0, xmm2            // average rows
-    pavgb      xmm1, xmm3
-    movdqa     xmm2, [esi + ebx * 2]
-    movdqa     xmm3, [esi + ebx * 2 + 16]
-    movdqa     xmm4, [esi + edx]
-    movdqa     xmm5, [esi + edx + 16]
-    lea        esi, [esi + 32]
-    pavgb      xmm2, xmm4
-    pavgb      xmm3, xmm5
-    pavgb      xmm0, xmm2
-    pavgb      xmm1, xmm3
-
-    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
-    psrlw      xmm0, 8
-    movdqa     xmm3, xmm1
-    psrlw      xmm1, 8
-    pand       xmm2, xmm7
-    pand       xmm3, xmm7
-    pavgw      xmm0, xmm2
-    pavgw      xmm1, xmm3
-    packuswb   xmm0, xmm1
-
-    movdqa     xmm2, xmm0            // average columns (16 to 8 pixels)
-    psrlw      xmm0, 8
-    pand       xmm2, xmm7
-    pavgw      xmm0, xmm2
-    packuswb   xmm0, xmm0
-
-    movq       qword ptr [edi], xmm0
-    lea        edi, [edi + 8]
-    sub        ecx, 8
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-#define HAS_SCALEROWDOWN8_SSE2
-// Point samples 32 pixels to 4 pixels.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 4 byte aligned.
-__declspec(naked)
-static void ScaleRowDown8_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-                                     // src_stride ignored
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask isolating 1 src 8 bytes
-    psrlq      xmm5, 56
-
-  wloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + 16]
-    lea        esi,  [esi + 32]
-    pand       xmm0, xmm5
-    pand       xmm1, xmm5
-    packuswb   xmm0, xmm1  // 32->16
-    packuswb   xmm0, xmm0  // 16->8
-    packuswb   xmm0, xmm0  // 8->4
-    movd       dword ptr [edi], xmm0
-    lea        edi, [edi + 4]
-    sub        ecx, 4
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-// Blends 32x8 rectangle to 4x1.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 4 byte aligned.
-__declspec(naked)
-static void ScaleRowDown8Int_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        ebx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    lea        edx, [ebx + ebx * 2]  // src_stride * 3
-    pxor       xmm7, xmm7
-
-  wloop:
-    movdqa     xmm0, [esi]           // average 8 rows to 1
-    movdqa     xmm1, [esi + 16]
-    movdqa     xmm2, [esi + ebx]
-    movdqa     xmm3, [esi + ebx + 16]
-    pavgb      xmm0, xmm2
-    pavgb      xmm1, xmm3
-    movdqa     xmm2, [esi + ebx * 2]
-    movdqa     xmm3, [esi + ebx * 2 + 16]
-    movdqa     xmm4, [esi + edx]
-    movdqa     xmm5, [esi + edx + 16]
-    lea        ebp, [esi + ebx * 4]
-    lea        esi, [esi + 32]
-    pavgb      xmm2, xmm4
-    pavgb      xmm3, xmm5
-    pavgb      xmm0, xmm2
-    pavgb      xmm1, xmm3
-
-    movdqa     xmm2, [ebp]
-    movdqa     xmm3, [ebp + 16]
-    movdqa     xmm4, [ebp + ebx]
-    movdqa     xmm5, [ebp + ebx + 16]
-    pavgb      xmm2, xmm4
-    pavgb      xmm3, xmm5
-    movdqa     xmm4, [ebp + ebx * 2]
-    movdqa     xmm5, [ebp + ebx * 2 + 16]
-    movdqa     xmm6, [ebp + edx]
-    pavgb      xmm4, xmm6
-    movdqa     xmm6, [ebp + edx + 16]
-    pavgb      xmm5, xmm6
-    pavgb      xmm2, xmm4
-    pavgb      xmm3, xmm5
-    pavgb      xmm0, xmm2
-    pavgb      xmm1, xmm3
-
-    psadbw     xmm0, xmm7            // average 32 pixels to 4
-    psadbw     xmm1, xmm7
-    pshufd     xmm0, xmm0, 0xd8      // x1x0 -> xx01
-    pshufd     xmm1, xmm1, 0x8d      // x3x2 -> 32xx
-    por        xmm0, xmm1            //      -> 3201
-    psrlw      xmm0, 3
-    packuswb   xmm0, xmm0
-    packuswb   xmm0, xmm0
-    movd       dword ptr [edi], xmm0
-
-    lea        edi, [edi + 4]
-    sub        ecx, 4
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-#define HAS_SCALEROWDOWN34_SSSE3
-// Point samples 32 pixels to 24 pixels.
-// Produces three 8 byte values.  For each 8 bytes, 16 bytes are read.
-// Then shuffled to do the scaling.
-
-// Note that movdqa+palign may be better than movdqu.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleRowDown34_SSSE3(const uint8* src_ptr, int src_stride,
-                                 uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-                                     // src_stride ignored
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm3, _shuf0
-    movdqa     xmm4, _shuf1
-    movdqa     xmm5, _shuf2
-
-  wloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + 16]
-    lea        esi,  [esi + 32]
-    movdqa     xmm2, xmm1
-    palignr    xmm1, xmm0, 8
-    pshufb     xmm0, xmm3
-    pshufb     xmm1, xmm4
-    pshufb     xmm2, xmm5
-    movq       qword ptr [edi], xmm0
-    movq       qword ptr [edi + 8], xmm1
-    movq       qword ptr [edi + 16], xmm2
-    lea        edi, [edi + 24]
-    sub        ecx, 24
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-// Blends 32x2 rectangle to 24x1
-// Produces three 8 byte values.  For each 8 bytes, 16 bytes are read.
-// Then shuffled to do the scaling.
-
-// Register usage:
-// xmm0 src_row 0
-// xmm1 src_row 1
-// xmm2 shuf 0
-// xmm3 shuf 1
-// xmm4 shuf 2
-// xmm5 madd 0
-// xmm6 madd 1
-// xmm7 round34
-
-// Note that movdqa+palign may be better than movdqu.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleRowDown34_1_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        ebx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm2, _shuf01
-    movdqa     xmm3, _shuf11
-    movdqa     xmm4, _shuf21
-    movdqa     xmm5, _madd01
-    movdqa     xmm6, _madd11
-    movdqa     xmm7, _round34
-
-  wloop:
-    movdqa     xmm0, [esi]           // pixels 0..7
-    movdqa     xmm1, [esi+ebx]
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm2
-    pmaddubsw  xmm0, xmm5
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi], xmm0
-    movdqu     xmm0, [esi+8]         // pixels 8..15
-    movdqu     xmm1, [esi+ebx+8]
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm3
-    pmaddubsw  xmm0, xmm6
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi+8], xmm0
-    movdqa     xmm0, [esi+16]        // pixels 16..23
-    movdqa     xmm1, [esi+ebx+16]
-    lea        esi, [esi+32]
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm4
-    movdqa     xmm1, _madd21
-    pmaddubsw  xmm0, xmm1
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi+16], xmm0
-    lea        edi, [edi+24]
-    sub        ecx, 24
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-// Note that movdqa+palign may be better than movdqu.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleRowDown34_0_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        ebx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm2, _shuf01
-    movdqa     xmm3, _shuf11
-    movdqa     xmm4, _shuf21
-    movdqa     xmm5, _madd01
-    movdqa     xmm6, _madd11
-    movdqa     xmm7, _round34
-
-  wloop:
-    movdqa     xmm0, [esi]           // pixels 0..7
-    movdqa     xmm1, [esi+ebx]
-    pavgb      xmm1, xmm0
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm2
-    pmaddubsw  xmm0, xmm5
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi], xmm0
-    movdqu     xmm0, [esi+8]         // pixels 8..15
-    movdqu     xmm1, [esi+ebx+8]
-    pavgb      xmm1, xmm0
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm3
-    pmaddubsw  xmm0, xmm6
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi+8], xmm0
-    movdqa     xmm0, [esi+16]        // pixels 16..23
-    movdqa     xmm1, [esi+ebx+16]
-    lea        esi, [esi+32]
-    pavgb      xmm1, xmm0
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm4
-    movdqa     xmm1, _madd21
-    pmaddubsw  xmm0, xmm1
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi+16], xmm0
-    lea        edi, [edi+24]
-    sub        ecx, 24
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-#define HAS_SCALEROWDOWN38_SSSE3
-// 3/8 point sampler
-
-// Scale 32 pixels to 12
-__declspec(naked)
-static void ScaleRowDown38_SSSE3(const uint8* src_ptr, int src_stride,
-                                 uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        edx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm4, _shuf38a
-    movdqa     xmm5, _shuf38b
-
-  xloop:
-    movdqa     xmm0, [esi]           // 16 pixels -> 0,1,2,3,4,5
-    movdqa     xmm1, [esi + 16]      // 16 pixels -> 6,7,8,9,10,11
-    lea        esi, [esi + 32]
-    pshufb     xmm0, xmm4
-    pshufb     xmm1, xmm5
-    paddusb    xmm0, xmm1
-
-    movq       qword ptr [edi], xmm0 // write 12 pixels
-    movhlps    xmm1, xmm0
-    movd       [edi + 8], xmm1
-    lea        edi, [edi + 12]
-    sub        ecx, 12
-    ja         xloop
-
-    popad
-    ret
-  }
-}
-
-// Scale 16x3 pixels to 6x1 with interpolation
-__declspec(naked)
-static void ScaleRowDown38_3_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        edx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm4, _shufac0
-    movdqa     xmm5, _shufac3
-    movdqa     xmm6, _scaleac3
-    pxor       xmm7, xmm7
-
-  xloop:
-    movdqa     xmm0, [esi]           // sum up 3 rows into xmm0/1
-    movdqa     xmm2, [esi + edx]
-    movhlps    xmm1, xmm0
-    movhlps    xmm3, xmm2
-    punpcklbw  xmm0, xmm7
-    punpcklbw  xmm1, xmm7
-    punpcklbw  xmm2, xmm7
-    punpcklbw  xmm3, xmm7
-    paddusw    xmm0, xmm2
-    paddusw    xmm1, xmm3
-    movdqa     xmm2, [esi + edx * 2]
-    lea        esi, [esi + 16]
-    movhlps    xmm3, xmm2
-    punpcklbw  xmm2, xmm7
-    punpcklbw  xmm3, xmm7
-    paddusw    xmm0, xmm2
-    paddusw    xmm1, xmm3
-
-    movdqa     xmm2, xmm0            // 8 pixels -> 0,1,2 of xmm2
-    psrldq     xmm0, 2
-    paddusw    xmm2, xmm0
-    psrldq     xmm0, 2
-    paddusw    xmm2, xmm0
-    pshufb     xmm2, xmm4
-
-    movdqa     xmm3, xmm1            // 8 pixels -> 3,4,5 of xmm2
-    psrldq     xmm1, 2
-    paddusw    xmm3, xmm1
-    psrldq     xmm1, 2
-    paddusw    xmm3, xmm1
-    pshufb     xmm3, xmm5
-    paddusw    xmm2, xmm3
-
-    pmulhuw    xmm2, xmm6            // divide by 9,9,6, 9,9,6
-    packuswb   xmm2, xmm2
-
-    movd       [edi], xmm2           // write 6 pixels
-    pextrw     eax, xmm2, 2
-    mov        [edi + 4], ax
-    lea        edi, [edi + 6]
-    sub        ecx, 6
-    ja         xloop
-
-    popad
-    ret
-  }
-}
-
-// Scale 16x2 pixels to 6x1 with interpolation
-__declspec(naked)
-static void ScaleRowDown38_2_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        edx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm4, _shufab0
-    movdqa     xmm5, _shufab1
-    movdqa     xmm6, _shufab2
-    movdqa     xmm7, _scaleab2
-
-  xloop:
-    movdqa     xmm2, [esi]           // average 2 rows into xmm2
-    pavgb      xmm2, [esi + edx]
-    lea        esi, [esi + 16]
-
-    movdqa     xmm0, xmm2            // 16 pixels -> 0,1,2,3,4,5 of xmm0
-    pshufb     xmm0, xmm4
-    movdqa     xmm1, xmm2
-    pshufb     xmm1, xmm5
-    paddusw    xmm0, xmm1
-    pshufb     xmm2, xmm6
-    paddusw    xmm0, xmm2
-
-    pmulhuw    xmm0, xmm7            // divide by 3,3,2, 3,3,2
-    packuswb   xmm0, xmm0
-
-    movd       [edi], xmm0           // write 6 pixels
-    pextrw     eax, xmm0, 2
-    mov        [edi + 4], ax
-    lea        edi, [edi + 6]
-    sub        ecx, 6
-    ja         xloop
-
-    popad
-    ret
-  }
-}
-
-#define HAS_SCALEADDROWS_SSE2
-
-// Reads 8xN bytes and produces 16 shorts at a time.
-__declspec(naked)
-static void ScaleAddRows_SSE2(const uint8* src_ptr, int src_stride,
-                              uint16* dst_ptr, int src_width,
-                              int src_height) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        edx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    mov        ebx, [esp + 32 + 20]  // height
-    pxor       xmm5, xmm5
-    dec        ebx
-
-  xloop:
-    // first row
-    movdqa     xmm2, [esi]
-    lea        eax, [esi + edx]
-    movhlps    xmm3, xmm2
-    mov        ebp, ebx
-    punpcklbw  xmm2, xmm5
-    punpcklbw  xmm3, xmm5
-
-    // sum remaining rows
-  yloop:
-    movdqa     xmm0, [eax]       // read 16 pixels
-    lea        eax, [eax + edx]  // advance to next row
-    movhlps    xmm1, xmm0
-    punpcklbw  xmm0, xmm5
-    punpcklbw  xmm1, xmm5
-    paddusw    xmm2, xmm0        // sum 16 words
-    paddusw    xmm3, xmm1
-    sub        ebp, 1
-    ja         yloop
-
-    movdqa     [edi], xmm2
-    movdqa     [edi + 16], xmm3
-    lea        edi, [edi + 32]
-    lea        esi, [esi + 16]
-
-    sub        ecx, 16
-    ja         xloop
-
-    popad
-    ret
-  }
-}
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSE2 version.
-#define HAS_SCALEFILTERROWS_SSE2
-__declspec(naked)
-static void ScaleFilterRows_SSE2(uint8* dst_ptr, const uint8* src_ptr,
-                                 int src_stride, int dst_width,
-                                 int source_y_fraction) {
-  __asm {
-    push       esi
-    push       edi
-    mov        edi, [esp + 8 + 4]   // dst_ptr
-    mov        esi, [esp + 8 + 8]   // src_ptr
-    mov        edx, [esp + 8 + 12]  // src_stride
-    mov        ecx, [esp + 8 + 16]  // dst_width
-    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
-    cmp        eax, 0
-    je         xloop1
-    cmp        eax, 128
-    je         xloop2
-
-    movd       xmm6, eax            // xmm6 = y fraction
-    punpcklwd  xmm6, xmm6
-    pshufd     xmm6, xmm6, 0
-    neg        eax                  // xmm5 = 256 - y fraction
-    add        eax, 256
-    movd       xmm5, eax
-    punpcklwd  xmm5, xmm5
-    pshufd     xmm5, xmm5, 0
-    pxor       xmm7, xmm7
-
-  xloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm2, [esi + edx]
-    lea        esi, [esi + 16]
-    movdqa     xmm1, xmm0
-    movdqa     xmm3, xmm2
-    punpcklbw  xmm0, xmm7
-    punpcklbw  xmm2, xmm7
-    punpckhbw  xmm1, xmm7
-    punpckhbw  xmm3, xmm7
-    pmullw     xmm0, xmm5           // scale row 0
-    pmullw     xmm1, xmm5
-    pmullw     xmm2, xmm6           // scale row 1
-    pmullw     xmm3, xmm6
-    paddusw    xmm0, xmm2           // sum rows
-    paddusw    xmm1, xmm3
-    psrlw      xmm0, 8
-    psrlw      xmm1, 8
-    packuswb   xmm0, xmm1
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-
-  xloop1:
-    movdqa     xmm0, [esi]
-    lea        esi, [esi + 16]
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop1
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-
-  xloop2:
-    movdqa     xmm0, [esi]
-    movdqa     xmm2, [esi + edx]
-    lea        esi, [esi + 16]
-    pavgb      xmm0, xmm2
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop2
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSSE3 version.
-#define HAS_SCALEFILTERROWS_SSSE3
-__declspec(naked)
-static void ScaleFilterRows_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                                  int src_stride, int dst_width,
-                                  int source_y_fraction) {
-  __asm {
-    push       esi
-    push       edi
-    mov        edi, [esp + 8 + 4]   // dst_ptr
-    mov        esi, [esp + 8 + 8]   // src_ptr
-    mov        edx, [esp + 8 + 12]  // src_stride
-    mov        ecx, [esp + 8 + 16]  // dst_width
-    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
-    shr        eax, 1
-    cmp        eax, 0
-    je         xloop1
-    cmp        eax, 64
-    je         xloop2
-
-    mov        ah,al
-    neg        al
-    add        al, 128
-    movd       xmm5, eax
-    punpcklwd  xmm5, xmm5
-    pshufd     xmm5, xmm5, 0
-
-  xloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm2, [esi + edx]
-    lea        esi, [esi + 16]
-    movdqa     xmm1, xmm0
-    punpcklbw  xmm0, xmm2
-    punpckhbw  xmm1, xmm2
-    pmaddubsw  xmm0, xmm5
-    pmaddubsw  xmm1, xmm5
-    psrlw      xmm0, 7
-    psrlw      xmm1, 7
-    packuswb   xmm0, xmm1
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-
-  xloop1:
-    movdqa     xmm0, [esi]
-    lea        esi, [esi + 16]
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop1
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-
-  xloop2:
-    movdqa     xmm0, [esi]
-    movdqa     xmm2, [esi + edx]
-    lea        esi, [esi + 16]
-    pavgb      xmm0, xmm2
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop2
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-
-  }
-}
-
-// Note that movdqa+palign may be better than movdqu.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleFilterCols34_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                                    int dst_width) {
-  __asm {
-    mov        edx, [esp + 4]    // dst_ptr
-    mov        eax, [esp + 8]    // src_ptr
-    mov        ecx, [esp + 12]   // dst_width
-    movdqa     xmm1, _round34
-    movdqa     xmm2, _shuf01
-    movdqa     xmm3, _shuf11
-    movdqa     xmm4, _shuf21
-    movdqa     xmm5, _madd01
-    movdqa     xmm6, _madd11
-    movdqa     xmm7, _madd21
-
-  wloop:
-    movdqa     xmm0, [eax]           // pixels 0..7
-    pshufb     xmm0, xmm2
-    pmaddubsw  xmm0, xmm5
-    paddsw     xmm0, xmm1
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edx], xmm0
-    movdqu     xmm0, [eax+8]         // pixels 8..15
-    pshufb     xmm0, xmm3
-    pmaddubsw  xmm0, xmm6
-    paddsw     xmm0, xmm1
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edx+8], xmm0
-    movdqa     xmm0, [eax+16]        // pixels 16..23
-    lea        eax, [eax+32]
-    pshufb     xmm0, xmm4
-    pmaddubsw  xmm0, xmm7
-    paddsw     xmm0, xmm1
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edx+16], xmm0
-    lea        edx, [edx+24]
-    sub        ecx, 24
-    ja         wloop
-    ret
-  }
-}
-
-#elif (defined(__x86_64__) || defined(__i386__)) && !defined(YUV_DISABLE_ASM)
-
-// GCC versions of row functions are verbatim conversions from Visual C.
-// Generated using gcc disassembly on Visual C object file:
-// objdump -D yuvscaler.obj >yuvscaler.txt
-#define HAS_SCALEROWDOWN2_SSE2
-static void ScaleRowDown2_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "pcmpeqb    %%xmm5,%%xmm5                    \n"
-  "psrlw      $0x8,%%xmm5                      \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "pand       %%xmm5,%%xmm0                    \n"
-  "pand       %%xmm5,%%xmm1                    \n"
-  "packuswb   %%xmm1,%%xmm0                    \n"
-  "movdqa     %%xmm0,(%1)                      \n"
-  "lea        0x10(%1),%1                      \n"
-  "sub        $0x10,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-);
-}
-
-static void ScaleRowDown2Int_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "pcmpeqb    %%xmm5,%%xmm5                    \n"
-  "psrlw      $0x8,%%xmm5                      \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "movdqa     (%0,%3,1),%%xmm2                 \n"
-  "movdqa     0x10(%0,%3,1),%%xmm3             \n"
-  "lea        0x20(%0),%0                      \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "movdqa     %%xmm0,%%xmm2                    \n"
-  "psrlw      $0x8,%%xmm0                      \n"
-  "movdqa     %%xmm1,%%xmm3                    \n"
-  "psrlw      $0x8,%%xmm1                      \n"
-  "pand       %%xmm5,%%xmm2                    \n"
-  "pand       %%xmm5,%%xmm3                    \n"
-  "pavgw      %%xmm2,%%xmm0                    \n"
-  "pavgw      %%xmm3,%%xmm1                    \n"
-  "packuswb   %%xmm1,%%xmm0                    \n"
-  "movdqa     %%xmm0,(%1)                      \n"
-  "lea        0x10(%1),%1                      \n"
-  "sub        $0x10,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc"
-);
-}
-
-#define HAS_SCALEROWDOWN4_SSE2
-static void ScaleRowDown4_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "pcmpeqb    %%xmm5,%%xmm5                    \n"
-  "psrld      $0x18,%%xmm5                     \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "pand       %%xmm5,%%xmm0                    \n"
-  "pand       %%xmm5,%%xmm1                    \n"
-  "packuswb   %%xmm1,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "lea        0x8(%1),%1                       \n"
-  "sub        $0x8,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-);
-}
-
-static void ScaleRowDown4Int_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  intptr_t temp = 0;
-  asm volatile (
-  "pcmpeqb    %%xmm7,%%xmm7                    \n"
-  "psrlw      $0x8,%%xmm7                      \n"
-  "lea        (%4,%4,2),%3                     \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "movdqa     (%0,%4,1),%%xmm2                 \n"
-  "movdqa     0x10(%0,%4,1),%%xmm3             \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "movdqa     (%0,%4,2),%%xmm2                 \n"
-  "movdqa     0x10(%0,%4,2),%%xmm3             \n"
-  "movdqa     (%0,%3,1),%%xmm4                 \n"
-  "movdqa     0x10(%0,%3,1),%%xmm5             \n"
-  "lea        0x20(%0),%0                      \n"
-  "pavgb      %%xmm4,%%xmm2                    \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm5,%%xmm3                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "movdqa     %%xmm0,%%xmm2                    \n"
-  "psrlw      $0x8,%%xmm0                      \n"
-  "movdqa     %%xmm1,%%xmm3                    \n"
-  "psrlw      $0x8,%%xmm1                      \n"
-  "pand       %%xmm7,%%xmm2                    \n"
-  "pand       %%xmm7,%%xmm3                    \n"
-  "pavgw      %%xmm2,%%xmm0                    \n"
-  "pavgw      %%xmm3,%%xmm1                    \n"
-  "packuswb   %%xmm1,%%xmm0                    \n"
-  "movdqa     %%xmm0,%%xmm2                    \n"
-  "psrlw      $0x8,%%xmm0                      \n"
-  "pand       %%xmm7,%%xmm2                    \n"
-  "pavgw      %%xmm2,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "lea        0x8(%1),%1                       \n"
-  "sub        $0x8,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width),   // %2
-    "+r"(temp)         // %3
-  : "r"((intptr_t)(src_stride))    // %4
-  : "memory", "cc"
-#if defined(__x86_64__)
-    , "xmm6", "xmm7"
-#endif
-);
-}
-
-#define HAS_SCALEROWDOWN8_SSE2
-static void ScaleRowDown8_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "pcmpeqb    %%xmm5,%%xmm5                    \n"
-  "psrlq      $0x38,%%xmm5                     \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "pand       %%xmm5,%%xmm0                    \n"
-  "pand       %%xmm5,%%xmm1                    \n"
-  "packuswb   %%xmm1,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movd       %%xmm0,(%1)                      \n"
-  "lea        0x4(%1),%1                       \n"
-  "sub        $0x4,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-);
-}
-
-#if defined(__i386__)
-void ScaleRowDown8Int_SSE2(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown8Int_SSE2)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%ebx                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "lea    (%ebx,%ebx,2),%edx                 \n"
-    "pxor   %xmm7,%xmm7                        \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa 0x10(%esi),%xmm1                   \n"
-    "movdqa (%esi,%ebx,1),%xmm2                \n"
-    "movdqa 0x10(%esi,%ebx,1),%xmm3            \n"
-    "pavgb  %xmm2,%xmm0                        \n"
-    "pavgb  %xmm3,%xmm1                        \n"
-    "movdqa (%esi,%ebx,2),%xmm2                \n"
-    "movdqa 0x10(%esi,%ebx,2),%xmm3            \n"
-    "movdqa (%esi,%edx,1),%xmm4                \n"
-    "movdqa 0x10(%esi,%edx,1),%xmm5            \n"
-    "lea    (%esi,%ebx,4),%ebp                 \n"
-    "lea    0x20(%esi),%esi                    \n"
-    "pavgb  %xmm4,%xmm2                        \n"
-    "pavgb  %xmm5,%xmm3                        \n"
-    "pavgb  %xmm2,%xmm0                        \n"
-    "pavgb  %xmm3,%xmm1                        \n"
-    "movdqa 0x0(%ebp),%xmm2                    \n"
-    "movdqa 0x10(%ebp),%xmm3                   \n"
-    "movdqa 0x0(%ebp,%ebx,1),%xmm4             \n"
-    "movdqa 0x10(%ebp,%ebx,1),%xmm5            \n"
-    "pavgb  %xmm4,%xmm2                        \n"
-    "pavgb  %xmm5,%xmm3                        \n"
-    "movdqa 0x0(%ebp,%ebx,2),%xmm4             \n"
-    "movdqa 0x10(%ebp,%ebx,2),%xmm5            \n"
-    "movdqa 0x0(%ebp,%edx,1),%xmm6             \n"
-    "pavgb  %xmm6,%xmm4                        \n"
-    "movdqa 0x10(%ebp,%edx,1),%xmm6            \n"
-    "pavgb  %xmm6,%xmm5                        \n"
-    "pavgb  %xmm4,%xmm2                        \n"
-    "pavgb  %xmm5,%xmm3                        \n"
-    "pavgb  %xmm2,%xmm0                        \n"
-    "pavgb  %xmm3,%xmm1                        \n"
-    "psadbw %xmm7,%xmm0                        \n"
-    "psadbw %xmm7,%xmm1                        \n"
-    "pshufd $0xd8,%xmm0,%xmm0                  \n"
-    "pshufd $0x8d,%xmm1,%xmm1                  \n"
-    "por    %xmm1,%xmm0                        \n"
-    "psrlw  $0x3,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movd   %xmm0,(%edi)                       \n"
-    "lea    0x4(%edi),%edi                     \n"
-    "sub    $0x4,%ecx                          \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-// fpic is used for magiccam plugin
-#if !defined(__PIC__)
-#define HAS_SCALEROWDOWN34_SSSE3
-void ScaleRowDown34_SSSE3(const uint8* src_ptr, int src_stride,
-                                     uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown34_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shuf0,%xmm3                       \n"
-    "movdqa _shuf1,%xmm4                       \n"
-    "movdqa _shuf2,%xmm5                       \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa 0x10(%esi),%xmm2                   \n"
-    "lea    0x20(%esi),%esi                    \n"
-    "movdqa %xmm2,%xmm1                        \n"
-    "palignr $0x8,%xmm0,%xmm1                  \n"
-    "pshufb %xmm3,%xmm0                        \n"
-    "pshufb %xmm4,%xmm1                        \n"
-    "pshufb %xmm5,%xmm2                        \n"
-    "movq   %xmm0,(%edi)                       \n"
-    "movq   %xmm1,0x8(%edi)                    \n"
-    "movq   %xmm2,0x10(%edi)                   \n"
-    "lea    0x18(%edi),%edi                    \n"
-    "sub    $0x18,%ecx                         \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-void ScaleRowDown34_1_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                           uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown34_1_Int_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%ebp                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shuf01,%xmm2                      \n"
-    "movdqa _shuf11,%xmm3                      \n"
-    "movdqa _shuf21,%xmm4                      \n"
-    "movdqa _madd01,%xmm5                      \n"
-    "movdqa _madd11,%xmm6                      \n"
-    "movdqa _round34,%xmm7                     \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%ebp),%xmm1                  \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm2,%xmm0                        \n"
-    "pmaddubsw %xmm5,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,(%edi)                       \n"
-    "movdqu 0x8(%esi),%xmm0                    \n"
-    "movdqu 0x8(%esi,%ebp),%xmm1               \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm3,%xmm0                        \n"
-    "pmaddubsw %xmm6,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,0x8(%edi)                    \n"
-    "movdqa 0x10(%esi),%xmm0                   \n"
-    "movdqa 0x10(%esi,%ebp),%xmm1              \n"
-    "lea    0x20(%esi),%esi                    \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm4,%xmm0                        \n"
-    "movdqa  _madd21,%xmm1                     \n"
-    "pmaddubsw %xmm1,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,0x10(%edi)                   \n"
-    "lea    0x18(%edi),%edi                    \n"
-    "sub    $0x18,%ecx                         \n"
-    "ja     1b                                 \n"
-
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-void ScaleRowDown34_0_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                           uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown34_0_Int_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%ebp                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shuf01,%xmm2                      \n"
-    "movdqa _shuf11,%xmm3                      \n"
-    "movdqa _shuf21,%xmm4                      \n"
-    "movdqa _madd01,%xmm5                      \n"
-    "movdqa _madd11,%xmm6                      \n"
-    "movdqa _round34,%xmm7                     \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%ebp,1),%xmm1                \n"
-    "pavgb  %xmm0,%xmm1                        \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm2,%xmm0                        \n"
-    "pmaddubsw %xmm5,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,(%edi)                       \n"
-    "movdqu 0x8(%esi),%xmm0                    \n"
-    "movdqu 0x8(%esi,%ebp,1),%xmm1             \n"
-    "pavgb  %xmm0,%xmm1                        \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm3,%xmm0                        \n"
-    "pmaddubsw %xmm6,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,0x8(%edi)                    \n"
-    "movdqa 0x10(%esi),%xmm0                   \n"
-    "movdqa 0x10(%esi,%ebp,1),%xmm1            \n"
-    "lea    0x20(%esi),%esi                    \n"
-    "pavgb  %xmm0,%xmm1                        \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm4,%xmm0                        \n"
-    "movdqa  _madd21,%xmm1                     \n"
-    "pmaddubsw %xmm1,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,0x10(%edi)                   \n"
-    "lea    0x18(%edi),%edi                    \n"
-    "sub    $0x18,%ecx                         \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-#define HAS_SCALEROWDOWN38_SSSE3
-void ScaleRowDown38_SSSE3(const uint8* src_ptr, int src_stride,
-                                     uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown38_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%edx                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shuf38a ,%xmm4                    \n"
-    "movdqa _shuf38b ,%xmm5                    \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa 0x10(%esi),%xmm1                   \n"
-    "lea    0x20(%esi),%esi                    \n"
-    "pshufb %xmm4,%xmm0                        \n"
-    "pshufb %xmm5,%xmm1                        \n"
-    "paddusb %xmm1,%xmm0                       \n"
-    "movq   %xmm0,(%edi)                       \n"
-    "movhlps %xmm0,%xmm1                       \n"
-    "movd   %xmm1,0x8(%edi)                    \n"
-    "lea    0xc(%edi),%edi                     \n"
-    "sub    $0xc,%ecx                          \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-void ScaleRowDown38_3_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                           uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown38_3_Int_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%edx                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shufac0,%xmm4                     \n"
-    "movdqa _shufac3,%xmm5                     \n"
-    "movdqa _scaleac3,%xmm6                    \n"
-    "pxor   %xmm7,%xmm7                        \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%edx,1),%xmm2                \n"
-    "movhlps %xmm0,%xmm1                       \n"
-    "movhlps %xmm2,%xmm3                       \n"
-    "punpcklbw %xmm7,%xmm0                     \n"
-    "punpcklbw %xmm7,%xmm1                     \n"
-    "punpcklbw %xmm7,%xmm2                     \n"
-    "punpcklbw %xmm7,%xmm3                     \n"
-    "paddusw %xmm2,%xmm0                       \n"
-    "paddusw %xmm3,%xmm1                       \n"
-    "movdqa (%esi,%edx,2),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movhlps %xmm2,%xmm3                       \n"
-    "punpcklbw %xmm7,%xmm2                     \n"
-    "punpcklbw %xmm7,%xmm3                     \n"
-    "paddusw %xmm2,%xmm0                       \n"
-    "paddusw %xmm3,%xmm1                       \n"
-    "movdqa %xmm0,%xmm2                        \n"
-    "psrldq $0x2,%xmm0                         \n"
-    "paddusw %xmm0,%xmm2                       \n"
-    "psrldq $0x2,%xmm0                         \n"
-    "paddusw %xmm0,%xmm2                       \n"
-    "pshufb %xmm4,%xmm2                        \n"
-    "movdqa %xmm1,%xmm3                        \n"
-    "psrldq $0x2,%xmm1                         \n"
-    "paddusw %xmm1,%xmm3                       \n"
-    "psrldq $0x2,%xmm1                         \n"
-    "paddusw %xmm1,%xmm3                       \n"
-    "pshufb %xmm5,%xmm3                        \n"
-    "paddusw %xmm3,%xmm2                       \n"
-    "pmulhuw %xmm6,%xmm2                       \n"
-    "packuswb %xmm2,%xmm2                      \n"
-    "movd   %xmm2,(%edi)                       \n"
-    "pextrw $0x2,%xmm2,%eax                    \n"
-    "mov    %ax,0x4(%edi)                      \n"
-    "lea    0x6(%edi),%edi                     \n"
-    "sub    $0x6,%ecx                          \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-void ScaleRowDown38_2_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                           uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown38_2_Int_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%edx                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shufab0,%xmm4                     \n"
-    "movdqa _shufab1,%xmm5                     \n"
-    "movdqa _shufab2,%xmm6                     \n"
-    "movdqa _scaleab2,%xmm7                    \n"
-
-"1:"
-    "movdqa (%esi),%xmm2                       \n"
-    "pavgb  (%esi,%edx,1),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movdqa %xmm2,%xmm0                        \n"
-    "pshufb %xmm4,%xmm0                        \n"
-    "movdqa %xmm2,%xmm1                        \n"
-    "pshufb %xmm5,%xmm1                        \n"
-    "paddusw %xmm1,%xmm0                       \n"
-    "pshufb %xmm6,%xmm2                        \n"
-    "paddusw %xmm2,%xmm0                       \n"
-    "pmulhuw %xmm7,%xmm0                       \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movd   %xmm0,(%edi)                       \n"
-    "pextrw $0x2,%xmm0,%eax                    \n"
-    "mov    %ax,0x4(%edi)                      \n"
-    "lea    0x6(%edi),%edi                     \n"
-    "sub    $0x6,%ecx                          \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-#endif // __PIC__
-
-#define HAS_SCALEADDROWS_SSE2
-void ScaleAddRows_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint16* dst_ptr, int src_width,
-                                  int src_height);
-  asm(
-    DECLARE_FUNCTION(ScaleAddRows_SSE2)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%edx                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "mov    0x34(%esp),%ebx                    \n"
-    "pxor   %xmm5,%xmm5                        \n"
-
-"1:"
-    "movdqa (%esi),%xmm2                       \n"
-    "lea    (%esi,%edx,1),%eax                 \n"
-    "movhlps %xmm2,%xmm3                       \n"
-    "lea    -0x1(%ebx),%ebp                    \n"
-    "punpcklbw %xmm5,%xmm2                     \n"
-    "punpcklbw %xmm5,%xmm3                     \n"
-
-"2:"
-    "movdqa (%eax),%xmm0                       \n"
-    "lea    (%eax,%edx,1),%eax                 \n"
-    "movhlps %xmm0,%xmm1                       \n"
-    "punpcklbw %xmm5,%xmm0                     \n"
-    "punpcklbw %xmm5,%xmm1                     \n"
-    "paddusw %xmm0,%xmm2                       \n"
-    "paddusw %xmm1,%xmm3                       \n"
-    "sub    $0x1,%ebp                          \n"
-    "ja     2b                                 \n"
-
-    "movdqa %xmm2,(%edi)                       \n"
-    "movdqa %xmm3,0x10(%edi)                   \n"
-    "lea    0x20(%edi),%edi                    \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSE2 version
-#define HAS_SCALEFILTERROWS_SSE2
-void ScaleFilterRows_SSE2(uint8* dst_ptr,
-                                     const uint8* src_ptr, int src_stride,
-                                     int dst_width, int source_y_fraction);
-  asm(
-    DECLARE_FUNCTION(ScaleFilterRows_SSE2)
-    "push   %esi                               \n"
-    "push   %edi                               \n"
-    "mov    0xc(%esp),%edi                     \n"
-    "mov    0x10(%esp),%esi                    \n"
-    "mov    0x14(%esp),%edx                    \n"
-    "mov    0x18(%esp),%ecx                    \n"
-    "mov    0x1c(%esp),%eax                    \n"
-    "cmp    $0x0,%eax                          \n"
-    "je     2f                                 \n"
-    "cmp    $0x80,%eax                         \n"
-    "je     3f                                 \n"
-    "movd   %eax,%xmm6                         \n"
-    "punpcklwd %xmm6,%xmm6                     \n"
-    "pshufd $0x0,%xmm6,%xmm6                   \n"
-    "neg    %eax                               \n"
-    "add    $0x100,%eax                        \n"
-    "movd   %eax,%xmm5                         \n"
-    "punpcklwd %xmm5,%xmm5                     \n"
-    "pshufd $0x0,%xmm5,%xmm5                   \n"
-    "pxor   %xmm7,%xmm7                        \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%edx,1),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movdqa %xmm0,%xmm1                        \n"
-    "movdqa %xmm2,%xmm3                        \n"
-    "punpcklbw %xmm7,%xmm0                     \n"
-    "punpcklbw %xmm7,%xmm2                     \n"
-    "punpckhbw %xmm7,%xmm1                     \n"
-    "punpckhbw %xmm7,%xmm3                     \n"
-    "pmullw %xmm5,%xmm0                        \n"
-    "pmullw %xmm5,%xmm1                        \n"
-    "pmullw %xmm6,%xmm2                        \n"
-    "pmullw %xmm6,%xmm3                        \n"
-    "paddusw %xmm2,%xmm0                       \n"
-    "paddusw %xmm3,%xmm1                       \n"
-    "psrlw  $0x8,%xmm0                         \n"
-    "psrlw  $0x8,%xmm1                         \n"
-    "packuswb %xmm1,%xmm0                      \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     1b                                 \n"
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-
-"2:"
-    "movdqa (%esi),%xmm0                       \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     2b                                 \n"
-
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-
-"3:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%edx,1),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "pavgb  %xmm2,%xmm0                        \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     3b                                 \n"
-
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-);
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSSE3 version
-#define HAS_SCALEFILTERROWS_SSSE3
-void ScaleFilterRows_SSSE3(uint8* dst_ptr,
-                                      const uint8* src_ptr, int src_stride,
-                                      int dst_width, int source_y_fraction);
-  asm(
-    DECLARE_FUNCTION(ScaleFilterRows_SSSE3)
-    "push   %esi                               \n"
-    "push   %edi                               \n"
-    "mov    0xc(%esp),%edi                     \n"
-    "mov    0x10(%esp),%esi                    \n"
-    "mov    0x14(%esp),%edx                    \n"
-    "mov    0x18(%esp),%ecx                    \n"
-    "mov    0x1c(%esp),%eax                    \n"
-    "shr    %eax                               \n"
-    "cmp    $0x0,%eax                          \n"
-    "je     2f                                 \n"
-    "cmp    $0x40,%eax                         \n"
-    "je     3f                                 \n"
-    "mov    %al,%ah                            \n"
-    "neg    %al                                \n"
-    "add    $0x80,%al                          \n"
-    "movd   %eax,%xmm5                         \n"
-    "punpcklwd %xmm5,%xmm5                     \n"
-    "pshufd $0x0,%xmm5,%xmm5                   \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%edx,1),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movdqa %xmm0,%xmm1                        \n"
-    "punpcklbw %xmm2,%xmm0                     \n"
-    "punpckhbw %xmm2,%xmm1                     \n"
-    "pmaddubsw %xmm5,%xmm0                     \n"
-    "pmaddubsw %xmm5,%xmm1                     \n"
-    "psrlw  $0x7,%xmm0                         \n"
-    "psrlw  $0x7,%xmm1                         \n"
-    "packuswb %xmm1,%xmm0                      \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     1b                                 \n"
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-
-"2:"
-    "movdqa (%esi),%xmm0                       \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     2b                                 \n"
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-
-"3:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%edx,1),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "pavgb  %xmm2,%xmm0                        \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     3b                                 \n"
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-);
-
-#elif defined(__x86_64__)
-static void ScaleRowDown8Int_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "lea        (%3,%3,2),%%r10                  \n"
-  "pxor       %%xmm7,%%xmm7                    \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "movdqa     (%0,%3,1),%%xmm2                 \n"
-  "movdqa     0x10(%0,%3,1),%%xmm3             \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "movdqa     (%0,%3,2),%%xmm2                 \n"
-  "movdqa     0x10(%0,%3,2),%%xmm3             \n"
-  "movdqa     (%0,%%r10,1),%%xmm4              \n"
-  "movdqa     0x10(%0,%%r10,1),%%xmm5          \n"
-  "lea        (%0,%3,4),%%r11                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "pavgb      %%xmm4,%%xmm2                    \n"
-  "pavgb      %%xmm5,%%xmm3                    \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "movdqa     0x0(%%r11),%%xmm2                \n"
-  "movdqa     0x10(%%r11),%%xmm3               \n"
-  "movdqa     0x0(%%r11,%3,1),%%xmm4           \n"
-  "movdqa     0x10(%%r11,%3,1),%%xmm5          \n"
-  "pavgb      %%xmm4,%%xmm2                    \n"
-  "pavgb      %%xmm5,%%xmm3                    \n"
-  "movdqa     0x0(%%r11,%3,2),%%xmm4           \n"
-  "movdqa     0x10(%%r11,%3,2),%%xmm5          \n"
-  "movdqa     0x0(%%r11,%%r10,1),%%xmm6        \n"
-  "pavgb      %%xmm6,%%xmm4                    \n"
-  "movdqa     0x10(%%r11,%%r10,1),%%xmm6       \n"
-  "pavgb      %%xmm6,%%xmm5                    \n"
-  "pavgb      %%xmm4,%%xmm2                    \n"
-  "pavgb      %%xmm5,%%xmm3                    \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "psadbw     %%xmm7,%%xmm0                    \n"
-  "psadbw     %%xmm7,%%xmm1                    \n"
-  "pshufd     $0xd8,%%xmm0,%%xmm0              \n"
-  "pshufd     $0x8d,%%xmm1,%%xmm1              \n"
-  "por        %%xmm1,%%xmm0                    \n"
-  "psrlw      $0x3,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movd       %%xmm0,(%1)                      \n"
-  "lea        0x4(%1),%1                       \n"
-  "sub        $0x4,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc", "r10", "r11", "xmm6", "xmm7"
-);
-}
-
-#define HAS_SCALEROWDOWN34_SSSE3
-static void ScaleRowDown34_SSSE3(const uint8* src_ptr, int src_stride,
-                                 uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%3),%%xmm3                      \n"
-  "movdqa     (%4),%%xmm4                      \n"
-  "movdqa     (%5),%%xmm5                      \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm2                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "movdqa     %%xmm2,%%xmm1                    \n"
-  "palignr    $0x8,%%xmm0,%%xmm1               \n"
-  "pshufb     %%xmm3,%%xmm0                    \n"
-  "pshufb     %%xmm4,%%xmm1                    \n"
-  "pshufb     %%xmm5,%%xmm2                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "movq       %%xmm1,0x8(%1)                   \n"
-  "movq       %%xmm2,0x10(%1)                  \n"
-  "lea        0x18(%1),%1                      \n"
-  "sub        $0x18,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"(_shuf0),   // %3
-    "r"(_shuf1),   // %4
-    "r"(_shuf2)    // %5
-  : "memory", "cc"
-);
-}
-
-static void ScaleRowDown34_1_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%4),%%xmm2                      \n"  // _shuf01
-  "movdqa     (%5),%%xmm3                      \n"  // _shuf11
-  "movdqa     (%6),%%xmm4                      \n"  // _shuf21
-  "movdqa     (%7),%%xmm5                      \n"  // _madd01
-  "movdqa     (%8),%%xmm6                      \n"  // _madd11
-  "movdqa     (%9),%%xmm7                      \n"  // _round34
-  "movdqa     (%10),%%xmm8                     \n"  // _madd21
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     (%0,%3),%%xmm1                   \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm2,%%xmm0                    \n"
-  "pmaddubsw  %%xmm5,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "movdqu     0x8(%0),%%xmm0                   \n"
-  "movdqu     0x8(%0,%3),%%xmm1                \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm3,%%xmm0                    \n"
-  "pmaddubsw  %%xmm6,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,0x8(%1)                   \n"
-  "movdqa     0x10(%0),%%xmm0                  \n"
-  "movdqa     0x10(%0,%3),%%xmm1               \n"
-  "lea        0x20(%0),%0                      \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm4,%%xmm0                    \n"
-  "pmaddubsw  %%xmm8,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,0x10(%1)                  \n"
-  "lea        0x18(%1),%1                      \n"
-  "sub        $0x18,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride)),  // %3
-    "r"(_shuf01),   // %4
-    "r"(_shuf11),   // %5
-    "r"(_shuf21),   // %6
-    "r"(_madd01),   // %7
-    "r"(_madd11),   // %8
-    "r"(_round34),  // %9
-    "r"(_madd21)    // %10
-  : "memory", "cc", "xmm6", "xmm7", "xmm8"
-);
-}
-
-static void ScaleRowDown34_0_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%4),%%xmm2                      \n"  // _shuf01
-  "movdqa     (%5),%%xmm3                      \n"  // _shuf11
-  "movdqa     (%6),%%xmm4                      \n"  // _shuf21
-  "movdqa     (%7),%%xmm5                      \n"  // _madd01
-  "movdqa     (%8),%%xmm6                      \n"  // _madd11
-  "movdqa     (%9),%%xmm7                      \n"  // _round34
-  "movdqa     (%10),%%xmm8                     \n"  // _madd21
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     (%0,%3,1),%%xmm1                 \n"
-  "pavgb      %%xmm0,%%xmm1                    \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm2,%%xmm0                    \n"
-  "pmaddubsw  %%xmm5,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "movdqu     0x8(%0),%%xmm0                   \n"
-  "movdqu     0x8(%0,%3,1),%%xmm1              \n"
-  "pavgb      %%xmm0,%%xmm1                    \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm3,%%xmm0                    \n"
-  "pmaddubsw  %%xmm6,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,0x8(%1)                   \n"
-  "movdqa     0x10(%0),%%xmm0                  \n"
-  "movdqa     0x10(%0,%3,1),%%xmm1             \n"
-  "lea        0x20(%0),%0                      \n"
-  "pavgb      %%xmm0,%%xmm1                    \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm4,%%xmm0                    \n"
-  "pmaddubsw  %%xmm8,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,0x10(%1)                  \n"
-  "lea        0x18(%1),%1                      \n"
-  "sub        $0x18,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride)),  // %3
-    "r"(_shuf01),   // %4
-    "r"(_shuf11),   // %5
-    "r"(_shuf21),   // %6
-    "r"(_madd01),   // %7
-    "r"(_madd11),   // %8
-    "r"(_round34),  // %9
-    "r"(_madd21)    // %10
-  : "memory", "cc", "xmm6", "xmm7", "xmm8"
-);
-}
-
-#define HAS_SCALEROWDOWN38_SSSE3
-static void ScaleRowDown38_SSSE3(const uint8* src_ptr, int src_stride,
-                                 uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%3),%%xmm4                      \n"
-  "movdqa     (%4),%%xmm5                      \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "pshufb     %%xmm4,%%xmm0                    \n"
-  "pshufb     %%xmm5,%%xmm1                    \n"
-  "paddusb    %%xmm1,%%xmm0                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "movhlps    %%xmm0,%%xmm1                    \n"
-  "movd       %%xmm1,0x8(%1)                   \n"
-  "lea        0xc(%1),%1                       \n"
-  "sub        $0xc,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"(_shuf38a),  // %3
-    "r"(_shuf38b)   // %4
-  : "memory", "cc"
-);
-}
-
-static void ScaleRowDown38_3_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%4),%%xmm4                      \n"
-  "movdqa     (%5),%%xmm5                      \n"
-  "movdqa     (%6),%%xmm6                      \n"
-  "pxor       %%xmm7,%%xmm7                    \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     (%0,%3,1),%%xmm2                 \n"
-  "movhlps    %%xmm0,%%xmm1                    \n"
-  "movhlps    %%xmm2,%%xmm3                    \n"
-  "punpcklbw  %%xmm7,%%xmm0                    \n"
-  "punpcklbw  %%xmm7,%%xmm1                    \n"
-  "punpcklbw  %%xmm7,%%xmm2                    \n"
-  "punpcklbw  %%xmm7,%%xmm3                    \n"
-  "paddusw    %%xmm2,%%xmm0                    \n"
-  "paddusw    %%xmm3,%%xmm1                    \n"
-  "movdqa     (%0,%3,2),%%xmm2                 \n"
-  "lea        0x10(%0),%0                      \n"
-  "movhlps    %%xmm2,%%xmm3                    \n"
-  "punpcklbw  %%xmm7,%%xmm2                    \n"
-  "punpcklbw  %%xmm7,%%xmm3                    \n"
-  "paddusw    %%xmm2,%%xmm0                    \n"
-  "paddusw    %%xmm3,%%xmm1                    \n"
-  "movdqa     %%xmm0,%%xmm2                    \n"
-  "psrldq     $0x2,%%xmm0                      \n"
-  "paddusw    %%xmm0,%%xmm2                    \n"
-  "psrldq     $0x2,%%xmm0                      \n"
-  "paddusw    %%xmm0,%%xmm2                    \n"
-  "pshufb     %%xmm4,%%xmm2                    \n"
-  "movdqa     %%xmm1,%%xmm3                    \n"
-  "psrldq     $0x2,%%xmm1                      \n"
-  "paddusw    %%xmm1,%%xmm3                    \n"
-  "psrldq     $0x2,%%xmm1                      \n"
-  "paddusw    %%xmm1,%%xmm3                    \n"
-  "pshufb     %%xmm5,%%xmm3                    \n"
-  "paddusw    %%xmm3,%%xmm2                    \n"
-  "pmulhuw    %%xmm6,%%xmm2                    \n"
-  "packuswb   %%xmm2,%%xmm2                    \n"
-  "movd       %%xmm2,(%1)                      \n"
-  "pextrw     $0x2,%%xmm2,%%eax                \n"
-  "mov        %%ax,0x4(%1)                     \n"
-  "lea        0x6(%1),%1                       \n"
-  "sub        $0x6,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride)),  // %3
-    "r"(_shufac0),   // %4
-    "r"(_shufac3),   // %5
-    "r"(_scaleac3)   // %6
-  : "memory", "cc", "rax", "xmm6", "xmm7"
-);
-}
-
-static void ScaleRowDown38_2_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%4),%%xmm4                      \n"
-  "movdqa     (%5),%%xmm5                      \n"
-  "movdqa     (%6),%%xmm6                      \n"
-  "movdqa     (%7),%%xmm7                      \n"
-"1:"
-  "movdqa     (%0),%%xmm2                      \n"
-  "pavgb      (%0,%3,1),%%xmm2                 \n"
-  "lea        0x10(%0),%0                      \n"
-  "movdqa     %%xmm2,%%xmm0                    \n"
-  "pshufb     %%xmm4,%%xmm0                    \n"
-  "movdqa     %%xmm2,%%xmm1                    \n"
-  "pshufb     %%xmm5,%%xmm1                    \n"
-  "paddusw    %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm6,%%xmm2                    \n"
-  "paddusw    %%xmm2,%%xmm0                    \n"
-  "pmulhuw    %%xmm7,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movd       %%xmm0,(%1)                      \n"
-  "pextrw     $0x2,%%xmm0,%%eax                \n"
-  "mov        %%ax,0x4(%1)                     \n"
-  "lea        0x6(%1),%1                       \n"
-  "sub        $0x6,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride)),  // %3
-    "r"(_shufab0),   // %4
-    "r"(_shufab1),   // %5
-    "r"(_shufab2),   // %6
-    "r"(_scaleab2)   // %7
-  : "memory", "cc", "rax", "xmm6", "xmm7"
-);
-}
-
-#define HAS_SCALEADDROWS_SSE2
-static void ScaleAddRows_SSE2(const uint8* src_ptr, int src_stride,
-                              uint16* dst_ptr, int src_width,
-                              int src_height) {
-  asm volatile (
-  "pxor       %%xmm5,%%xmm5                    \n"
-"1:"
-  "movdqa     (%0),%%xmm2                      \n"
-  "lea        (%0,%4,1),%%r10                  \n"
-  "movhlps    %%xmm2,%%xmm3                    \n"
-  "lea        -0x1(%3),%%r11                   \n"
-  "punpcklbw  %%xmm5,%%xmm2                    \n"
-  "punpcklbw  %%xmm5,%%xmm3                    \n"
-
-"2:"
-  "movdqa     (%%r10),%%xmm0                   \n"
-  "lea        (%%r10,%4,1),%%r10               \n"
-  "movhlps    %%xmm0,%%xmm1                    \n"
-  "punpcklbw  %%xmm5,%%xmm0                    \n"
-  "punpcklbw  %%xmm5,%%xmm1                    \n"
-  "paddusw    %%xmm0,%%xmm2                    \n"
-  "paddusw    %%xmm1,%%xmm3                    \n"
-  "sub        $0x1,%%r11                       \n"
-  "ja         2b                               \n"
-
-  "movdqa     %%xmm2,(%1)                      \n"
-  "movdqa     %%xmm3,0x10(%1)                  \n"
-  "lea        0x20(%1),%1                      \n"
-  "lea        0x10(%0),%0                      \n"
-  "sub        $0x10,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(src_width),   // %2
-    "+r"(src_height)   // %3
-  : "r"((intptr_t)(src_stride))  // %4
-  : "memory", "cc", "r10", "r11"
-);
-}
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSE2 version
-#define HAS_SCALEFILTERROWS_SSE2
-static void ScaleFilterRows_SSE2(uint8* dst_ptr,
-                                 const uint8* src_ptr, int src_stride,
-                                 int dst_width, int source_y_fraction) {
-  if (source_y_fraction == 0) {
-    asm volatile (
-    "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "lea        0x10(%1),%1                  \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width)    // %2
-      :
-      : "memory", "cc", "rax"
-    );
-    return;
-  } else if (source_y_fraction == 128) {
-    asm volatile (
-    "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "movdqa     (%1,%3,1),%%xmm2             \n"
-      "lea        0x10(%1),%1                  \n"
-      "pavgb      %%xmm2,%%xmm0                \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width)    // %2
-      : "r"((intptr_t)(src_stride))  // %3
-      : "memory", "cc", "rax"
-    );
-    return;
-  } else {
-    asm volatile (
-      "mov        %3,%%eax                     \n"
-      "movd       %%eax,%%xmm6                 \n"
-      "punpcklwd  %%xmm6,%%xmm6                \n"
-      "pshufd     $0x0,%%xmm6,%%xmm6           \n"
-      "neg        %%eax                        \n"
-      "add        $0x100,%%eax                 \n"
-      "movd       %%eax,%%xmm5                 \n"
-      "punpcklwd  %%xmm5,%%xmm5                \n"
-      "pshufd     $0x0,%%xmm5,%%xmm5           \n"
-      "pxor       %%xmm7,%%xmm7                \n"
-    "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "movdqa     (%1,%4,1),%%xmm2             \n"
-      "lea        0x10(%1),%1                  \n"
-      "movdqa     %%xmm0,%%xmm1                \n"
-      "movdqa     %%xmm2,%%xmm3                \n"
-      "punpcklbw  %%xmm7,%%xmm0                \n"
-      "punpcklbw  %%xmm7,%%xmm2                \n"
-      "punpckhbw  %%xmm7,%%xmm1                \n"
-      "punpckhbw  %%xmm7,%%xmm3                \n"
-      "pmullw     %%xmm5,%%xmm0                \n"
-      "pmullw     %%xmm5,%%xmm1                \n"
-      "pmullw     %%xmm6,%%xmm2                \n"
-      "pmullw     %%xmm6,%%xmm3                \n"
-      "paddusw    %%xmm2,%%xmm0                \n"
-      "paddusw    %%xmm3,%%xmm1                \n"
-      "psrlw      $0x8,%%xmm0                  \n"
-      "psrlw      $0x8,%%xmm1                  \n"
-      "packuswb   %%xmm1,%%xmm0                \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width),   // %2
-        "+r"(source_y_fraction)  // %3
-      : "r"((intptr_t)(src_stride))  // %4
-      : "memory", "cc", "rax", "xmm6", "xmm7"
-    );
-  }
-  return;
-}
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSSE3 version
-#define HAS_SCALEFILTERROWS_SSSE3
-static void ScaleFilterRows_SSSE3(uint8* dst_ptr,
-                                  const uint8* src_ptr, int src_stride,
-                                  int dst_width, int source_y_fraction) {
-  source_y_fraction >>= 1;
-  if (source_y_fraction == 0) {
-    asm volatile (
-   "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "lea        0x10(%1),%1                  \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width)    // %2
-      :
-      : "memory", "cc", "rax"
-    );
-    return;
-  } else if (source_y_fraction == 64) {
-    asm volatile (
-    "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "movdqa     (%1,%3,1),%%xmm2             \n"
-      "lea        0x10(%1),%1                  \n"
-      "pavgb      %%xmm2,%%xmm0                \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width)    // %2
-      : "r"((intptr_t)(src_stride))  // %3
-     : "memory", "cc", "rax"
-    );
-    return;
-  } else {
-    asm volatile (
-      "mov        %3,%%eax                     \n"
-      "mov        %%al,%%ah                    \n"
-      "neg        %%al                         \n"
-      "add        $0x80,%%al                   \n"
-      "movd       %%eax,%%xmm5                 \n"
-      "punpcklwd  %%xmm5,%%xmm5                \n"
-      "pshufd     $0x0,%%xmm5,%%xmm5           \n"
-    "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "movdqa     (%1,%4,1),%%xmm2             \n"
-      "lea        0x10(%1),%1                  \n"
-      "movdqa     %%xmm0,%%xmm1                \n"
-      "punpcklbw  %%xmm2,%%xmm0                \n"
-      "punpckhbw  %%xmm2,%%xmm1                \n"
-      "pmaddubsw  %%xmm5,%%xmm0                \n"
-      "pmaddubsw  %%xmm5,%%xmm1                \n"
-      "psrlw      $0x7,%%xmm0                  \n"
-      "psrlw      $0x7,%%xmm1                  \n"
-      "packuswb   %%xmm1,%%xmm0                \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width),   // %2
-        "+r"(source_y_fraction)  // %3
-      : "r"((intptr_t)(src_stride))  // %4
-      : "memory", "cc", "rax"
-    );
-  }
-  return;
-}
-#endif
-#endif
-
-// CPU agnostic row functions
-static void ScaleRowDown2_C(const uint8* src_ptr, int src_stride,
-                            uint8* dst, int dst_width) {
-  int x;
-  for (x = 0; x < dst_width; ++x) {
-    *dst++ = *src_ptr;
-    src_ptr += 2;
-  }
-}
-
-static void ScaleRowDown2Int_C(const uint8* src_ptr, int src_stride,
-                               uint8* dst, int dst_width) {
-  int x;
-  for (x = 0; x < dst_width; ++x) {
-    *dst++ = (src_ptr[0] + src_ptr[1] +
-              src_ptr[src_stride] + src_ptr[src_stride + 1] + 2) >> 2;
-    src_ptr += 2;
-  }
-}
-
-static void ScaleRowDown4_C(const uint8* src_ptr, int src_stride,
-                            uint8* dst, int dst_width) {
-  int x;
-  for (x = 0; x < dst_width; ++x) {
-    *dst++ = *src_ptr;
-    src_ptr += 4;
-  }
-}
-
-static void ScaleRowDown4Int_C(const uint8* src_ptr, int src_stride,
-                               uint8* dst, int dst_width) {
-  int x;
-  for (x = 0; x < dst_width; ++x) {
-    *dst++ = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
-              src_ptr[src_stride + 0] + src_ptr[src_stride + 1] +
-              src_ptr[src_stride + 2] + src_ptr[src_stride + 3] +
-              src_ptr[src_stride * 2 + 0] + src_ptr[src_stride * 2 + 1] +
-              src_ptr[src_stride * 2 + 2] + src_ptr[src_stride * 2 + 3] +
-              src_ptr[src_stride * 3 + 0] + src_ptr[src_stride * 3 + 1] +
-              src_ptr[src_stride * 3 + 2] + src_ptr[src_stride * 3 + 3] +
-              8) >> 4;
-    src_ptr += 4;
-  }
-}
-
-// 640 output pixels is enough to allow 5120 input pixels with 1/8 scale down.
-// Keeping the total buffer under 4096 bytes avoids a stackcheck, saving 4% cpu.
-// The following 2 lines cause error on Windows.
-//static const int kMaxOutputWidth = 640;
-//static const int kMaxRow12 = 1280;         //kMaxOutputWidth * 2;
-#define kMaxOutputWidth   640
-#define kMaxRow12         1280
-
-static void ScaleRowDown8_C(const uint8* src_ptr, int src_stride,
-                            uint8* dst, int dst_width) {
-  int x;
-  for (x = 0; x < dst_width; ++x) {
-    *dst++ = *src_ptr;
-    src_ptr += 8;
-  }
-}
-
-// Note calling code checks width is less than max and if not
-// uses ScaleRowDown8_C instead.
-static void ScaleRowDown8Int_C(const uint8* src_ptr, int src_stride,
-                               uint8* dst, int dst_width) {
-  ALIGN16(uint8 src_row[kMaxRow12 * 2]);
-  assert(dst_width <= kMaxOutputWidth);
-  ScaleRowDown4Int_C(src_ptr, src_stride, src_row, dst_width * 2);
-  ScaleRowDown4Int_C(src_ptr + src_stride * 4, src_stride,
-                     src_row + kMaxOutputWidth,
-                     dst_width * 2);
-  ScaleRowDown2Int_C(src_row, kMaxOutputWidth, dst, dst_width);
-}
-
-static void ScaleRowDown34_C(const uint8* src_ptr, int src_stride,
-                             uint8* dst, int dst_width) {
-  uint8* dend;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  dend = dst + dst_width;
-  do {
-    dst[0] = src_ptr[0];
-    dst[1] = src_ptr[1];
-    dst[2] = src_ptr[3];
-    dst += 3;
-    src_ptr += 4;
-  } while (dst < dend);
-}
-
-// Filter rows 0 and 1 together, 3 : 1
-static void ScaleRowDown34_0_Int_C(const uint8* src_ptr, int src_stride,
-                                   uint8* d, int dst_width) {
-  uint8* dend;
-  const uint8* s;
-  const uint8* t;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  dend = d + dst_width;
-  s = src_ptr;
-  t = src_ptr + src_stride;
-  do {
-    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
-    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
-    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
-    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
-    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
-    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
-    d[0] = (a0 * 3 + b0 + 2) >> 2;
-    d[1] = (a1 * 3 + b1 + 2) >> 2;
-    d[2] = (a2 * 3 + b2 + 2) >> 2;
-    d += 3;
-    s += 4;
-    t += 4;
-  } while (d < dend);
-}
-
-// Filter rows 1 and 2 together, 1 : 1
-static void ScaleRowDown34_1_Int_C(const uint8* src_ptr, int src_stride,
-                                   uint8* d, int dst_width) {
-  uint8* dend;
-  const uint8* s;
-  const uint8* t;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  dend = d + dst_width;
-  s = src_ptr;
-  t = src_ptr + src_stride;
-  do {
-    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
-    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
-    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
-    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
-    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
-    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
-    d[0] = (a0 + b0 + 1) >> 1;
-    d[1] = (a1 + b1 + 1) >> 1;
-    d[2] = (a2 + b2 + 1) >> 1;
-    d += 3;
-    s += 4;
-    t += 4;
-  } while (d < dend);
-}
-
-#if defined(HAS_SCALEFILTERROWS_SSE2)
-// Filter row to 3/4
-static void ScaleFilterCols34_C(uint8* dst_ptr, const uint8* src_ptr,
-                                int dst_width) {
-  uint8* dend;
-  const uint8* s;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  dend = dst_ptr + dst_width;
-  s = src_ptr;
-  do {
-    dst_ptr[0] = (s[0] * 3 + s[1] * 1 + 2) >> 2;
-    dst_ptr[1] = (s[1] * 1 + s[2] * 1 + 1) >> 1;
-    dst_ptr[2] = (s[2] * 1 + s[3] * 3 + 2) >> 2;
-    dst_ptr += 3;
-    s += 4;
-  } while (dst_ptr < dend);
-}
-#endif
-
-static void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
-                              int dst_width, int dx) {
-  int x = 0;
-  int j;
-  for (j = 0; j < dst_width; ++j) {
-    int xi = x >> 16;
-    int xf1 = x & 0xffff;
-    int xf0 = 65536 - xf1;
-
-    *dst_ptr++ = (src_ptr[xi] * xf0 + src_ptr[xi + 1] * xf1) >> 16;
-    x += dx;
-  }
-}
-
-//Not work on Windows
-//static const int kMaxInputWidth = 2560;
-#define kMaxInputWidth    2560
-#if defined(HAS_SCALEFILTERROWS_SSE2)
-#define HAS_SCALEROWDOWN34_SSE2
-// Filter rows 0 and 1 together, 3 : 1
-static void ScaleRowDown34_0_Int_SSE2(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  ALIGN16(uint8 row[kMaxInputWidth]);
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  ScaleFilterRows_SSE2(row, src_ptr, src_stride, dst_width * 4 / 3, 256 / 4);
-  ScaleFilterCols34_C(dst_ptr, row, dst_width);
-}
-
-// Filter rows 1 and 2 together, 1 : 1
-static void ScaleRowDown34_1_Int_SSE2(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  ALIGN16(uint8 row[kMaxInputWidth]);
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  ScaleFilterRows_SSE2(row, src_ptr, src_stride, dst_width * 4 / 3, 256 / 2);
-  ScaleFilterCols34_C(dst_ptr, row, dst_width);
-}
-#endif
-
-static void ScaleRowDown38_C(const uint8* src_ptr, int src_stride,
-                             uint8* dst, int dst_width) {
-  int x;
-  assert(dst_width % 3 == 0);
-  for (x = 0; x < dst_width; x += 3) {
-    dst[0] = src_ptr[0];
-    dst[1] = src_ptr[3];
-    dst[2] = src_ptr[6];
-    dst += 3;
-    src_ptr += 8;
-  }
-}
-
-// 8x3 -> 3x1
-static void ScaleRowDown38_3_Int_C(const uint8* src_ptr, int src_stride,
-                                   uint8* dst_ptr, int dst_width) {
-  int i;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  for (i = 0; i < dst_width; i+=3) {
-    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
-        src_ptr[src_stride + 0] + src_ptr[src_stride + 1] +
-        src_ptr[src_stride + 2] + src_ptr[src_stride * 2 + 0] +
-        src_ptr[src_stride * 2 + 1] + src_ptr[src_stride * 2 + 2]) *
-        (65536 / 9) >> 16;
-    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
-        src_ptr[src_stride + 3] + src_ptr[src_stride + 4] +
-        src_ptr[src_stride + 5] + src_ptr[src_stride * 2 + 3] +
-        src_ptr[src_stride * 2 + 4] + src_ptr[src_stride * 2 + 5]) *
-        (65536 / 9) >> 16;
-    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
-        src_ptr[src_stride + 6] + src_ptr[src_stride + 7] +
-        src_ptr[src_stride * 2 + 6] + src_ptr[src_stride * 2 + 7]) *
-        (65536 / 6) >> 16;
-    src_ptr += 8;
-    dst_ptr += 3;
-  }
-}
-
-// 8x2 -> 3x1
-static void ScaleRowDown38_2_Int_C(const uint8* src_ptr, int src_stride,
-                                   uint8* dst_ptr, int dst_width) {
-  int i;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  for (i = 0; i < dst_width; i+=3) {
-    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
-        src_ptr[src_stride + 0] + src_ptr[src_stride + 1] +
-        src_ptr[src_stride + 2]) * (65536 / 6) >> 16;
-    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
-        src_ptr[src_stride + 3] + src_ptr[src_stride + 4] +
-        src_ptr[src_stride + 5]) * (65536 / 6) >> 16;
-    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
-        src_ptr[src_stride + 6] + src_ptr[src_stride + 7]) *
-        (65536 / 4) >> 16;
-    src_ptr += 8;
-    dst_ptr += 3;
-  }
-}
-
-// C version 8x2 -> 8x1
-static void ScaleFilterRows_C(uint8* dst_ptr,
-                              const uint8* src_ptr, int src_stride,
-                              int dst_width, int source_y_fraction) {
-  int y1_fraction;
-  int y0_fraction;
-  const uint8* src_ptr1;
-  uint8* end;
-  assert(dst_width > 0);
-  y1_fraction = source_y_fraction;
-  y0_fraction = 256 - y1_fraction;
-  src_ptr1 = src_ptr + src_stride;
-  end = dst_ptr + dst_width;
-  do {
-    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
-    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
-    dst_ptr[2] = (src_ptr[2] * y0_fraction + src_ptr1[2] * y1_fraction) >> 8;
-    dst_ptr[3] = (src_ptr[3] * y0_fraction + src_ptr1[3] * y1_fraction) >> 8;
-    dst_ptr[4] = (src_ptr[4] * y0_fraction + src_ptr1[4] * y1_fraction) >> 8;
-    dst_ptr[5] = (src_ptr[5] * y0_fraction + src_ptr1[5] * y1_fraction) >> 8;
-    dst_ptr[6] = (src_ptr[6] * y0_fraction + src_ptr1[6] * y1_fraction) >> 8;
-    dst_ptr[7] = (src_ptr[7] * y0_fraction + src_ptr1[7] * y1_fraction) >> 8;
-    src_ptr += 8;
-    src_ptr1 += 8;
-    dst_ptr += 8;
-  } while (dst_ptr < end);
-  dst_ptr[0] = dst_ptr[-1];
-}
-
-void ScaleAddRows_C(const uint8* src_ptr, int src_stride,
-                    uint16* dst_ptr, int src_width, int src_height) {
-  int x,y;
-  assert(src_width > 0);
-  assert(src_height > 0);
-  for (x = 0; x < src_width; ++x) {
-    const uint8* s = src_ptr + x;
-    int sum = 0;
-    for (y = 0; y < src_height; ++y) {
-      sum += s[0];
-      s += src_stride;
-    }
-    dst_ptr[x] = sum;
-  }
-}
-
-/**
- * Scale plane, 1/2
- *
- * This is an optimized version for scaling down a plane to 1/2 of
- * its original size.
- *
- */
-static void ScalePlaneDown2(int src_width, int src_height,
-                            int dst_width, int dst_height,
-                            int src_stride, int dst_stride,
-                            const uint8* src_ptr, uint8* dst_ptr,
-                            FilterModeEnum filtering) {
-  void (*ScaleRowDown2)(const uint8* src_ptr, int src_stride,
-                        uint8* dst_ptr, int dst_width);
-  assert(IS_ALIGNED(src_width, 2));
-  assert(IS_ALIGNED(src_height, 2));
-
-#if defined(HAS_SCALEROWDOWN2_NEON)
-  if (TestCpuFlag(kCpuHasNEON) &&
-      IS_ALIGNED(dst_width, 16)) {
-    ScaleRowDown2 = filtering ? ScaleRowDown2Int_NEON : ScaleRowDown2_NEON;
-  } else
-#endif
-#if defined(HAS_SCALEROWDOWN2_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(dst_width, 16) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
-    ScaleRowDown2 = filtering ? ScaleRowDown2Int_SSE2 : ScaleRowDown2_SSE2;
-  } else
-#endif
-  {
-    ScaleRowDown2 = filtering ? ScaleRowDown2Int_C : ScaleRowDown2_C;
-  }
-
-  {
-    int y;
-    for (y = 0; y < dst_height; ++y) {
-      ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
-      src_ptr += (src_stride << 1);
-      dst_ptr += dst_stride;
-    }
-  }
-}
-
-/**
- * Scale plane, 1/4
- *
- * This is an optimized version for scaling down a plane to 1/4 of
- * its original size.
- */
-static void ScalePlaneDown4(int src_width, int src_height,
-                            int dst_width, int dst_height,
-                            int src_stride, int dst_stride,
-                            const uint8* src_ptr, uint8* dst_ptr,
-                            FilterModeEnum filtering) {
-  void (*ScaleRowDown4)(const uint8* src_ptr, int src_stride,
-                        uint8* dst_ptr, int dst_width);
-  assert(IS_ALIGNED(src_width, 4));
-  assert(IS_ALIGNED(src_height, 4));
-
-#if defined(HAS_SCALEROWDOWN4_NEON)
-  if (TestCpuFlag(kCpuHasNEON) &&
-      IS_ALIGNED(dst_width, 4)) {
-    ScaleRowDown4 = filtering ? ScaleRowDown4Int_NEON : ScaleRowDown4_NEON;
-  } else
-#endif
-#if defined(HAS_SCALEROWDOWN4_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(dst_width, 8) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_ptr, 8) && IS_ALIGNED(dst_stride, 8)) {
-    ScaleRowDown4 = filtering ? ScaleRowDown4Int_SSE2 : ScaleRowDown4_SSE2;
-  } else
-#endif
-  {
-    ScaleRowDown4 = filtering ? ScaleRowDown4Int_C : ScaleRowDown4_C;
-  }
-
-  {
-    int y;
-    for (y = 0; y < dst_height; ++y) {
-      ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
-      src_ptr += (src_stride << 2);
-      dst_ptr += dst_stride;
-    }
-  }
-}
-
-/**
- * Scale plane, 1/8
- *
- * This is an optimized version for scaling down a plane to 1/8
- * of its original size.
- *
- */
-static void ScalePlaneDown8(int src_width, int src_height,
-                            int dst_width, int dst_height,
-                            int src_stride, int dst_stride,
-                            const uint8* src_ptr, uint8* dst_ptr,
-                            FilterModeEnum filtering) {
-  void (*ScaleRowDown8)(const uint8* src_ptr, int src_stride,
-                        uint8* dst_ptr, int dst_width);
-  assert(IS_ALIGNED(src_width, 8));
-  assert(IS_ALIGNED(src_height, 8));
-
-#if defined(HAS_SCALEROWDOWN8_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(dst_width, 4) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
-    ScaleRowDown8 = filtering ? ScaleRowDown8Int_SSE2 : ScaleRowDown8_SSE2;
-  } else
-#endif
-  {
-    ScaleRowDown8 = filtering && (dst_width <= kMaxOutputWidth) ?
-        ScaleRowDown8Int_C : ScaleRowDown8_C;
-  }
-
-  {
-    int y;
-    for (y = 0; y < dst_height; ++y) {
-      ScaleRowDown8(src_ptr, src_stride, dst_ptr, dst_width);
-      src_ptr += (src_stride << 3);
-      dst_ptr += dst_stride;
-    }
-  }
-}
-
-/**
- * Scale plane down, 3/4
- *
- * Provided by Frank Barchard (fbarchard@google.com)
- *
- */
-static void ScalePlaneDown34(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint8* src_ptr, uint8* dst_ptr,
-                             FilterModeEnum filtering) {
-  void (*ScaleRowDown34_0)(const uint8* src_ptr, int src_stride,
-                           uint8* dst_ptr, int dst_width);
-  void (*ScaleRowDown34_1)(const uint8* src_ptr, int src_stride,
-                           uint8* dst_ptr, int dst_width);
-  assert(dst_width % 3 == 0);
-#if defined(HAS_SCALEROWDOWN34_NEON)
-  if (TestCpuFlag(kCpuHasNEON) &&
-      (dst_width % 24 == 0)) {
-    if (!filtering) {
-      ScaleRowDown34_0 = ScaleRowDown34_NEON;
-      ScaleRowDown34_1 = ScaleRowDown34_NEON;
-    } else {
-      ScaleRowDown34_0 = ScaleRowDown34_0_Int_NEON;
-      ScaleRowDown34_1 = ScaleRowDown34_1_Int_NEON;
-    }
-  } else
-#endif
-
-#if defined(HAS_SCALEROWDOWN34_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) &&
-      (dst_width % 24 == 0) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_ptr, 8) && IS_ALIGNED(dst_stride, 8)) {
-    if (!filtering) {
-      ScaleRowDown34_0 = ScaleRowDown34_SSSE3;
-      ScaleRowDown34_1 = ScaleRowDown34_SSSE3;
-    } else {
-      ScaleRowDown34_0 = ScaleRowDown34_0_Int_SSSE3;
-      ScaleRowDown34_1 = ScaleRowDown34_1_Int_SSSE3;
-    }
-  } else
-#endif
-#if defined(HAS_SCALEROWDOWN34_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      (dst_width % 24 == 0) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_stride, 8) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(dst_ptr, 8) &&
-      filtering) {
-    ScaleRowDown34_0 = ScaleRowDown34_0_Int_SSE2;
-    ScaleRowDown34_1 = ScaleRowDown34_1_Int_SSE2;
-  } else
-#endif
-  {
-    if (!filtering) {
-      ScaleRowDown34_0 = ScaleRowDown34_C;
-      ScaleRowDown34_1 = ScaleRowDown34_C;
-    } else {
-      ScaleRowDown34_0 = ScaleRowDown34_0_Int_C;
-      ScaleRowDown34_1 = ScaleRowDown34_1_Int_C;
-    }
-  }
-  {
-  int src_row = 0;
-    int y;
-    for (y = 0; y < dst_height; ++y) {
-    switch (src_row) {
-      case 0:
-        ScaleRowDown34_0(src_ptr, src_stride, dst_ptr, dst_width);
-        break;
-
-      case 1:
-        ScaleRowDown34_1(src_ptr, src_stride, dst_ptr, dst_width);
-        break;
-
-      case 2:
-        ScaleRowDown34_0(src_ptr + src_stride, -src_stride,
-                         dst_ptr, dst_width);
-        break;
-    }
-    ++src_row;
-    src_ptr += src_stride;
-    dst_ptr += dst_stride;
-    if (src_row >= 3) {
-      src_ptr += src_stride;
-      src_row = 0;
-    }
-  }
-}
-}
-
-/**
- * Scale plane, 3/8
- *
- * This is an optimized version for scaling down a plane to 3/8
- * of its original size.
- *
- * Reduces 16x3 to 6x1
- */
-static void ScalePlaneDown38(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint8* src_ptr, uint8* dst_ptr,
-                             FilterModeEnum filtering) {
-  void (*ScaleRowDown38_3)(const uint8* src_ptr, int src_stride,
-                           uint8* dst_ptr, int dst_width);
-  void (*ScaleRowDown38_2)(const uint8* src_ptr, int src_stride,
-                           uint8* dst_ptr, int dst_width);
-  assert(dst_width % 3 == 0);
-#if defined(HAS_SCALEROWDOWN38_NEON)
-  if (TestCpuFlag(kCpuHasNEON) &&
-      (dst_width % 12 == 0)) {
-    if (!filtering) {
-      ScaleRowDown38_3 = ScaleRowDown38_NEON;
-      ScaleRowDown38_2 = ScaleRowDown38_NEON;
-    } else {
-      ScaleRowDown38_3 = ScaleRowDown38_3_Int_NEON;
-      ScaleRowDown38_2 = ScaleRowDown38_2_Int_NEON;
-    }
-  } else
-#endif
-
-#if defined(HAS_SCALEROWDOWN38_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) &&
-      (dst_width % 24 == 0) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_stride, 8) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(dst_ptr, 8)) {
-    if (!filtering) {
-      ScaleRowDown38_3 = ScaleRowDown38_SSSE3;
-      ScaleRowDown38_2 = ScaleRowDown38_SSSE3;
-    } else {
-      ScaleRowDown38_3 = ScaleRowDown38_3_Int_SSSE3;
-      ScaleRowDown38_2 = ScaleRowDown38_2_Int_SSSE3;
-    }
-  } else
-#endif
-  {
-    if (!filtering) {
-      ScaleRowDown38_3 = ScaleRowDown38_C;
-      ScaleRowDown38_2 = ScaleRowDown38_C;
-    } else {
-      ScaleRowDown38_3 = ScaleRowDown38_3_Int_C;
-      ScaleRowDown38_2 = ScaleRowDown38_2_Int_C;
-    }
-  }
-  {
-  int src_row = 0;
-    int y;
-    for (y = 0; y < dst_height; ++y) {
-    switch (src_row) {
-      case 0:
-      case 1:
-        ScaleRowDown38_3(src_ptr, src_stride, dst_ptr, dst_width);
-        src_ptr += src_stride * 3;
-        ++src_row;
-        break;
-
-      case 2:
-        ScaleRowDown38_2(src_ptr, src_stride, dst_ptr, dst_width);
-        src_ptr += src_stride * 2;
-        src_row = 0;
-        break;
-    }
-    dst_ptr += dst_stride;
-  }
-}
-}
-
-__inline static uint32 SumBox(int iboxwidth, int iboxheight,
-                            int src_stride, const uint8* src_ptr) {
-  int x, y;
-  uint32 sum;
-  assert(iboxwidth > 0);
-  assert(iboxheight > 0);
-  sum = 0u;
-  for (y = 0; y < iboxheight; ++y) {
-    for (x = 0; x < iboxwidth; ++x) {
-      sum += src_ptr[x];
-    }
-    src_ptr += src_stride;
-  }
-  return sum;
-}
-
-static void ScalePlaneBoxRow(int dst_width, int boxheight,
-                             int dx, int src_stride,
-                             const uint8* src_ptr, uint8* dst_ptr) {
-  int x = 0;
-  int i;
-  for (i = 0; i < dst_width; ++i) {
-    int ix = x >> 16;
-    int boxwidth;
-    x += dx;
-    boxwidth = (x >> 16) - ix;
-    *dst_ptr++ = SumBox(boxwidth, boxheight, src_stride, src_ptr + ix) /
-        (boxwidth * boxheight);
-  }
-}
-
-__inline static uint32 SumPixels(int iboxwidth, const uint16* src_ptr) {
-  uint32 sum;
-  int x;
-  assert(iboxwidth > 0);
-  sum = 0u;
-  for (x = 0; x < iboxwidth; ++x) {
-    sum += src_ptr[x];
-  }
-  return sum;
-}
-
-static void ScaleAddCols2_C(int dst_width, int boxheight, int dx,
-                            const uint16* src_ptr, uint8* dst_ptr) {
-  int scaletbl[2];
-  int minboxwidth = (dx >> 16);
-  scaletbl[0] = 65536 / (minboxwidth * boxheight);
-  scaletbl[1] = 65536 / ((minboxwidth + 1) * boxheight);
-  {
-  int *scaleptr = scaletbl - minboxwidth;
-  int x = 0;
-    int i;
-    for (i = 0; i < dst_width; ++i) {
-    int ix = x >> 16;
-      int boxwidth;
-    x += dx;
-      boxwidth = (x >> 16) - ix;
-    *dst_ptr++ = SumPixels(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16;
-    }
-  }
-}
-
-static void ScaleAddCols1_C(int dst_width, int boxheight, int dx,
-                            const uint16* src_ptr, uint8* dst_ptr) {
-  int boxwidth = (dx >> 16);
-  int scaleval = 65536 / (boxwidth * boxheight);
-  int x = 0;
-  int i;
-  for (i = 0; i < dst_width; ++i) {
-    *dst_ptr++ = SumPixels(boxwidth, src_ptr + x) * scaleval >> 16;
-    x += boxwidth;
-  }
-}
-
-/**
- * Scale plane down to any dimensions, with interpolation.
- * (boxfilter).
- *
- * Same method as SimpleScale, which is fixed point, outputting
- * one pixel of destination using fixed point (16.16) to step
- * through source, sampling a box of pixel with simple
- * averaging.
- */
-static void ScalePlaneBox(int src_width, int src_height,
-                          int dst_width, int dst_height,
-                          int src_stride, int dst_stride,
-                          const uint8* src_ptr, uint8* dst_ptr) {
-  int dx, dy;
-  assert(dst_width > 0);
-  assert(dst_height > 0);
-  dy = (src_height << 16) / dst_height;
-  dx = (src_width << 16) / dst_width;
-  if (!IS_ALIGNED(src_width, 16) || (src_width > kMaxInputWidth) ||
-      dst_height * 2 > src_height) {
-    uint8* dst = dst_ptr;
-    int dy = (src_height << 16) / dst_height;
-    int dx = (src_width << 16) / dst_width;
-    int y = 0;
-    int j;
-    for (j = 0; j < dst_height; ++j) {
-      int iy = y >> 16;
-      const uint8* const src = src_ptr + iy * src_stride;
-      int boxheight;
-      y += dy;
-      if (y > (src_height << 16)) {
-        y = (src_height << 16);
-      }
-      boxheight = (y >> 16) - iy;
-      ScalePlaneBoxRow(dst_width, boxheight,
-                       dx, src_stride,
-                       src, dst);
-
-      dst += dst_stride;
-    }
-  } else {
-    ALIGN16(uint16 row[kMaxInputWidth]);
-    void (*ScaleAddRows)(const uint8* src_ptr, int src_stride,
-                         uint16* dst_ptr, int src_width, int src_height);
-    void (*ScaleAddCols)(int dst_width, int boxheight, int dx,
-                         const uint16* src_ptr, uint8* dst_ptr);
-#if defined(HAS_SCALEADDROWS_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) &&
-        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(src_ptr, 16) &&
-        IS_ALIGNED(src_width, 16)) {
-      ScaleAddRows = ScaleAddRows_SSE2;
-    } else
-#endif
-    {
-      ScaleAddRows = ScaleAddRows_C;
-    }
-    if (dx & 0xffff) {
-      ScaleAddCols = ScaleAddCols2_C;
-    } else {
-      ScaleAddCols = ScaleAddCols1_C;
-    }
-
-    {
-    int y = 0;
-      int j;
-      for (j = 0; j < dst_height; ++j) {
-      int iy = y >> 16;
-      const uint8* const src = src_ptr + iy * src_stride;
-        int boxheight;
-      y += dy;
-      if (y > (src_height << 16)) {
-        y = (src_height << 16);
-      }
-        boxheight = (y >> 16) - iy;
-      ScaleAddRows(src, src_stride, row, src_width, boxheight);
-      ScaleAddCols(dst_width, boxheight, dx, row, dst_ptr);
-      dst_ptr += dst_stride;
-      }
-    }
-  }
-}
-
-/**
- * Scale plane to/from any dimensions, with interpolation.
- */
-static void ScalePlaneBilinearSimple(int src_width, int src_height,
-                                     int dst_width, int dst_height,
-                                     int src_stride, int dst_stride,
-                                     const uint8* src_ptr, uint8* dst_ptr) {
-  int i, j;
-  uint8* dst = dst_ptr;
-  int dx = (src_width << 16) / dst_width;
-  int dy = (src_height << 16) / dst_height;
-  int maxx = ((src_width - 1) << 16) - 1;
-  int maxy = ((src_height - 1) << 16) - 1;
-  int y = (dst_height < src_height) ? 32768 :
-      (src_height << 16) / dst_height - 32768;
-  for (i = 0; i < dst_height; ++i) {
-    int cy = (y < 0) ? 0 : y;
-    int yi = cy >> 16;
-    int yf = cy & 0xffff;
-    const uint8* const src = src_ptr + yi * src_stride;
-    int x = (dst_width < src_width) ? 32768 :
-        (src_width << 16) / dst_width - 32768;
-    for (j = 0; j < dst_width; ++j) {
-      int cx = (x < 0) ? 0 : x;
-      int xi = cx >> 16;
-      int xf = cx & 0xffff;
-      int r0 = (src[xi] * (65536 - xf) + src[xi + 1] * xf) >> 16;
-      int r1 = (src[xi + src_stride] * (65536 - xf) +
-          src[xi + src_stride + 1] * xf) >> 16;
-      *dst++ = (r0 * (65536 - yf) + r1 * yf) >> 16;
-      x += dx;
-      if (x > maxx)
-        x = maxx;
-    }
-    dst += dst_stride - dst_width;
-    y += dy;
-    if (y > maxy)
-      y = maxy;
-  }
-}
-
-/**
- * Scale plane to/from any dimensions, with bilinear
- * interpolation.
- */
-static void ScalePlaneBilinear(int src_width, int src_height,
-                               int dst_width, int dst_height,
-                               int src_stride, int dst_stride,
-                               const uint8* src_ptr, uint8* dst_ptr) {
-  int dy;
-  int dx;
-  assert(dst_width > 0);
-  assert(dst_height > 0);
-  dy = (src_height << 16) / dst_height;
-  dx = (src_width << 16) / dst_width;
-  if (!IS_ALIGNED(src_width, 8) || (src_width > kMaxInputWidth)) {
-    ScalePlaneBilinearSimple(src_width, src_height, dst_width, dst_height,
-                             src_stride, dst_stride, src_ptr, dst_ptr);
-
-  } else {
-    ALIGN16(uint8 row[kMaxInputWidth + 1]);
-    void (*ScaleFilterRows)(uint8* dst_ptr, const uint8* src_ptr,
-                            int src_stride,
-                            int dst_width, int source_y_fraction);
-    void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
-                            int dst_width, int dx);
-#if defined(HAS_SCALEFILTERROWS_SSSE3)
-    if (TestCpuFlag(kCpuHasSSSE3) &&
-        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(src_ptr, 16) &&
-        IS_ALIGNED(src_width, 16)) {
-      ScaleFilterRows = ScaleFilterRows_SSSE3;
-    } else
-#endif
-#if defined(HAS_SCALEFILTERROWS_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) &&
-        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(src_ptr, 16) &&
-        IS_ALIGNED(src_width, 16)) {
-      ScaleFilterRows = ScaleFilterRows_SSE2;
-    } else
-#endif
-    {
-      ScaleFilterRows = ScaleFilterRows_C;
-    }
-    ScaleFilterCols = ScaleFilterCols_C;
-
-    {
-    int y = 0;
-    int maxy = ((src_height - 1) << 16) - 1; // max is filter of last 2 rows.
-      int j;
-      for (j = 0; j < dst_height; ++j) {
-      int iy = y >> 16;
-      int fy = (y >> 8) & 255;
-      const uint8* const src = src_ptr + iy * src_stride;
-      ScaleFilterRows(row, src, src_stride, src_width, fy);
-      ScaleFilterCols(dst_ptr, row, dst_width, dx);
-      dst_ptr += dst_stride;
-      y += dy;
-      if (y > maxy) {
-        y = maxy;
-      }
-    }
-  }
-}
-}
-
-/**
- * Scale plane to/from any dimensions, without interpolation.
- * Fixed point math is used for performance: The upper 16 bits
- * of x and dx is the integer part of the source position and
- * the lower 16 bits are the fixed decimal part.
- */
-static void ScalePlaneSimple(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint8* src_ptr, uint8* dst_ptr) {
-  uint8* dst = dst_ptr;
-  int dx = (src_width << 16) / dst_width;
-  int y;
-  for (y = 0; y < dst_height; ++y) {
-    const uint8* const src = src_ptr + (y * src_height / dst_height) *
-        src_stride;
-    // TODO(fbarchard): Round X coordinate by setting x=0x8000.
-    int x = 0;
-    int i;
-    for (i = 0; i < dst_width; ++i) {
-      *dst++ = src[x >> 16];
-      x += dx;
-    }
-    dst += dst_stride - dst_width;
-  }
-}
-
-/**
- * Scale plane to/from any dimensions.
- */
-static void ScalePlaneAnySize(int src_width, int src_height,
-                              int dst_width, int dst_height,
-                              int src_stride, int dst_stride,
-                              const uint8* src_ptr, uint8* dst_ptr,
-                              FilterModeEnum filtering) {
-  if (!filtering) {
-    ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
-                     src_stride, dst_stride, src_ptr, dst_ptr);
-  } else {
-    // fall back to non-optimized version
-    ScalePlaneBilinear(src_width, src_height, dst_width, dst_height,
-                       src_stride, dst_stride, src_ptr, dst_ptr);
-  }
-}
-
-/**
- * Scale plane down, any size
- *
- * This is an optimized version for scaling down a plane to any size.
- * The current implementation is ~10 times faster compared to the
- * reference implementation for e.g. XGA->LowResPAL
- *
- */
-static void ScalePlaneDown(int src_width, int src_height,
-                           int dst_width, int dst_height,
-                           int src_stride, int dst_stride,
-                           const uint8* src_ptr, uint8* dst_ptr,
-                           FilterModeEnum filtering) {
-  if (!filtering) {
-    ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
-                     src_stride, dst_stride, src_ptr, dst_ptr);
-  } else if (filtering == kFilterBilinear || src_height * 2 > dst_height) {
-    // between 1/2x and 1x use bilinear
-    ScalePlaneBilinear(src_width, src_height, dst_width, dst_height,
-                       src_stride, dst_stride, src_ptr, dst_ptr);
-  } else {
-    ScalePlaneBox(src_width, src_height, dst_width, dst_height,
-                  src_stride, dst_stride, src_ptr, dst_ptr);
-  }
-}
-
-/**
- * Copy plane, no scaling
- *
- * This simply copies the given plane without scaling.
- * The current implementation is ~115 times faster
- * compared to the reference implementation.
- *
- */
-static void CopyPlane(int src_width, int src_height,
-                      int dst_width, int dst_height,
-                      int src_stride, int dst_stride,
-                      const uint8* src_ptr, uint8* dst_ptr) {
-  if (src_stride == src_width && dst_stride == dst_width) {
-    // All contiguous, so can use REALLY fast path.
-    memcpy(dst_ptr, src_ptr, src_width * src_height);
-  } else {
-    // Not all contiguous; must copy scanlines individually
-    const uint8* src = src_ptr;
-    uint8* dst = dst_ptr;
-    int i;
-    for (i = 0; i < src_height; ++i) {
-      memcpy(dst, src, src_width);
-      dst += dst_stride;
-      src += src_stride;
-    }
-  }
-}
-
-static void ScalePlane(const uint8* src, int src_stride,
-                       int src_width, int src_height,
-                       uint8* dst, int dst_stride,
-                       int dst_width, int dst_height,
-                       FilterModeEnum filtering, int use_ref) {
-  // Use specialized scales to improve performance for common resolutions.
-  // For example, all the 1/2 scalings will use ScalePlaneDown2()
-  if (dst_width == src_width && dst_height == src_height) {
-    // Straight copy.
-    CopyPlane(src_width, src_height, dst_width, dst_height, src_stride,
-              dst_stride, src, dst);
-  } else if (dst_width <= src_width && dst_height <= src_height) {
-    // Scale down.
-    if (use_ref) {
-      // For testing, allow the optimized versions to be disabled.
-      ScalePlaneDown(src_width, src_height, dst_width, dst_height,
-                     src_stride, dst_stride, src, dst, filtering);
-    } else if (4 * dst_width == 3 * src_width &&
-               4 * dst_height == 3 * src_height) {
-      // optimized, 3/4
-      ScalePlaneDown34(src_width, src_height, dst_width, dst_height,
-                       src_stride, dst_stride, src, dst, filtering);
-    } else if (2 * dst_width == src_width && 2 * dst_height == src_height) {
-      // optimized, 1/2
-      ScalePlaneDown2(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst, filtering);
-    // 3/8 rounded up for odd sized chroma height.
-    } else if (8 * dst_width == 3 * src_width &&
-               dst_height == ((src_height * 3 + 7) / 8)) {
-      // optimized, 3/8
-      ScalePlaneDown38(src_width, src_height, dst_width, dst_height,
-                       src_stride, dst_stride, src, dst, filtering);
-    } else if (4 * dst_width == src_width && 4 * dst_height == src_height) {
-      // optimized, 1/4
-      ScalePlaneDown4(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst, filtering);
-    } else if (8 * dst_width == src_width && 8 * dst_height == src_height) {
-      // optimized, 1/8
-      ScalePlaneDown8(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst, filtering);
-    } else {
-      // Arbitrary downsample
-      ScalePlaneDown(src_width, src_height, dst_width, dst_height,
-                     src_stride, dst_stride, src, dst, filtering);
-    }
-  } else {
-    // Arbitrary scale up and/or down.
-    ScalePlaneAnySize(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst, filtering);
-  }
-}
-
-/**
- * Scale a plane.
- *
- * This function in turn calls a scaling function
- * suitable for handling the desired resolutions.
- *
- */
-
-int I420Scale(const uint8* src_y, int src_stride_y,
-              const uint8* src_u, int src_stride_u,
-              const uint8* src_v, int src_stride_v,
-              int src_width, int src_height,
-              uint8* dst_y, int dst_stride_y,
-              uint8* dst_u, int dst_stride_u,
-              uint8* dst_v, int dst_stride_v,
-              int dst_width, int dst_height,
-              FilterModeEnum filtering) {
-  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
-      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (src_height < 0) {
-    int halfheight;
-    src_height = -src_height;
-    halfheight = (src_height + 1) >> 1;
-    src_y = src_y + (src_height - 1) * src_stride_y;
-    src_u = src_u + (halfheight - 1) * src_stride_u;
-    src_v = src_v + (halfheight - 1) * src_stride_v;
-    src_stride_y = -src_stride_y;
-    src_stride_u = -src_stride_u;
-    src_stride_v = -src_stride_v;
-  }
-  {
-  int src_halfwidth = (src_width + 1) >> 1;
-  int src_halfheight = (src_height + 1) >> 1;
-  int dst_halfwidth = (dst_width + 1) >> 1;
-  int dst_halfheight = (dst_height + 1) >> 1;
-
-    ScalePlane(src_y, src_stride_y, src_width, src_height,
-               dst_y, dst_stride_y, dst_width, dst_height,
-               filtering, use_reference_impl_);
-  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight,
-             dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
-             filtering, use_reference_impl_);
-  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight,
-             dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
-             filtering, use_reference_impl_);
-  }
-  return 0;
-}
-
-// Deprecated api
-int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
-          int src_stride_y, int src_stride_u, int src_stride_v,
-          int src_width, int src_height,
-          uint8* dst_y, uint8* dst_u, uint8* dst_v,
-          int dst_stride_y, int dst_stride_u, int dst_stride_v,
-          int dst_width, int dst_height,
-          int interpolate) {
-  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
-      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (src_height < 0) {
-    int halfheight;
-    src_height = -src_height;
-    halfheight = (src_height + 1) >> 1;
-    src_y = src_y + (src_height - 1) * src_stride_y;
-    src_u = src_u + (halfheight - 1) * src_stride_u;
-    src_v = src_v + (halfheight - 1) * src_stride_v;
-    src_stride_y = -src_stride_y;
-    src_stride_u = -src_stride_u;
-    src_stride_v = -src_stride_v;
-  }
-  {
-  int src_halfwidth = (src_width + 1) >> 1;
-  int src_halfheight = (src_height + 1) >> 1;
-  int dst_halfwidth = (dst_width + 1) >> 1;
-  int dst_halfheight = (dst_height + 1) >> 1;
-  FilterModeEnum filtering = interpolate ? kFilterBox : kFilterNone;
-
-  ScalePlane(src_y, src_stride_y, src_width, src_height,
-             dst_y, dst_stride_y, dst_width, dst_height,
-             filtering, use_reference_impl_);
-  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight,
-             dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
-             filtering, use_reference_impl_);
-  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight,
-             dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
-             filtering, use_reference_impl_);
-  }
-  return 0;
-}
-
-// Deprecated api
-int ScaleOffset(const uint8* src, int src_width, int src_height,
-                uint8* dst, int dst_width, int dst_height, int dst_yoffset,
-          int interpolate) {
-  if (!src || src_width <= 0 || src_height <= 0 ||
-      !dst || dst_width <= 0 || dst_height <= 0 || dst_yoffset < 0 ||
-      dst_yoffset >= dst_height) {
-    return -1;
-  }
-  dst_yoffset = dst_yoffset & ~1;  // chroma requires offset to multiple of 2.
-  {
-  int src_halfwidth = (src_width + 1) >> 1;
-  int src_halfheight = (src_height + 1) >> 1;
-  int dst_halfwidth = (dst_width + 1) >> 1;
-  int dst_halfheight = (dst_height + 1) >> 1;
-  int aheight = dst_height - dst_yoffset * 2;  // actual output height
-  const uint8* const src_y = src;
-  const uint8* const src_u = src + src_width * src_height;
-  const uint8* const src_v = src + src_width * src_height +
-                             src_halfwidth * src_halfheight;
-  uint8* dst_y = dst + dst_yoffset * dst_width;
-  uint8* dst_u = dst + dst_width * dst_height +
-                 (dst_yoffset >> 1) * dst_halfwidth;
-  uint8* dst_v = dst + dst_width * dst_height + dst_halfwidth * dst_halfheight +
-                 (dst_yoffset >> 1) * dst_halfwidth;
-  return Scale(src_y, src_u, src_v, src_width, src_halfwidth, src_halfwidth,
-               src_width, src_height, dst_y, dst_u, dst_v, dst_width,
-               dst_halfwidth, dst_halfwidth, dst_width, aheight, interpolate);
-  }
-}
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
diff --git a/third_party/libyuv/source/scale.cc b/third_party/libyuv/source/scale.cc
new file mode 100644
index 0000000..31cedf1
--- /dev/null
+++ b/third_party/libyuv/source/scale.cc
@@ -0,0 +1,1716 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "third_party/libyuv/include/libyuv/cpu_id.h"
+#include "third_party/libyuv/include/libyuv/planar_functions.h"  // CopyPlane
+#include "third_party/libyuv/include/libyuv/row.h"
+#include "third_party/libyuv/include/libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Remove this macro if OVERREAD is safe.
+#define AVOID_OVERREAD 1
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+
+// Scale plane, 1/2
+// This is an optimized version for scaling down a plane to 1/2 of
+// its original size.
+
+static void ScalePlaneDown2(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) =
+    filtering == kFilterNone ? ScaleRowDown2_C :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_C :
+        ScaleRowDown2Box_C);
+  int row_stride = src_stride << 1;
+  if (!filtering) {
+    src_ptr += src_stride;  // Point to odd rows.
+    src_stride = 0;
+  }
+
+#if defined(HAS_SCALEROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering ? ScaleRowDown2Box_NEON : ScaleRowDown2_NEON;
+  }
+#elif defined(HAS_SCALEROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Unaligned_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Unaligned_SSE2 :
+        ScaleRowDown2Box_Unaligned_SSE2);
+    if (IS_ALIGNED(src_ptr, 16) &&
+        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(row_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 :
+          ScaleRowDown2Box_SSE2);
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN2_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
+      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown2 = filtering ?
+        ScaleRowDown2Box_MIPS_DSPR2 : ScaleRowDown2_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  // TODO(fbarchard): Loop through source height to allow odd height.
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void ScalePlaneDown2_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown2)(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst_ptr, int dst_width) =
+    filtering == kFilterNone ? ScaleRowDown2_16_C :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_16_C :
+        ScaleRowDown2Box_16_C);
+  int row_stride = src_stride << 1;
+  if (!filtering) {
+    src_ptr += src_stride;  // Point to odd rows.
+    src_stride = 0;
+  }
+
+#if defined(HAS_SCALEROWDOWN2_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering ? ScaleRowDown2Box_16_NEON :
+        ScaleRowDown2_16_NEON;
+  }
+#elif defined(HAS_SCALEROWDOWN2_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering == kFilterNone ?
+        ScaleRowDown2_Unaligned_16_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Unaligned_16_SSE2 :
+        ScaleRowDown2Box_Unaligned_16_SSE2);
+    if (IS_ALIGNED(src_ptr, 16) &&
+        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(row_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_16_SSE2 :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_16_SSE2 :
+          ScaleRowDown2Box_16_SSE2);
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN2_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
+      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown2 = filtering ?
+        ScaleRowDown2Box_16_MIPS_DSPR2 : ScaleRowDown2_16_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  // TODO(fbarchard): Loop through source height to allow odd height.
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+// Scale plane, 1/4
+// This is an optimized version for scaling down a plane to 1/4 of
+// its original size.
+
+static void ScalePlaneDown4(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown4)(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) =
+      filtering ? ScaleRowDown4Box_C : ScaleRowDown4_C;
+  int row_stride = src_stride << 2;
+  if (!filtering) {
+    src_ptr += src_stride * 2;  // Point to row 2.
+    src_stride = 0;
+  }
+#if defined(HAS_SCALEROWDOWN4_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_NEON : ScaleRowDown4_NEON;
+  }
+#elif defined(HAS_SCALEROWDOWN4_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(dst_width, 8) && IS_ALIGNED(row_stride, 16) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_SSE2 : ScaleRowDown4_SSE2;
+  }
+#elif defined(HAS_SCALEROWDOWN4_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_MIPS_DSPR2 : ScaleRowDown4_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void ScalePlaneDown4_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown4)(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst_ptr, int dst_width) =
+      filtering ? ScaleRowDown4Box_16_C : ScaleRowDown4_16_C;
+  int row_stride = src_stride << 2;
+  if (!filtering) {
+    src_ptr += src_stride * 2;  // Point to row 2.
+    src_stride = 0;
+  }
+#if defined(HAS_SCALEROWDOWN4_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_NEON :
+        ScaleRowDown4_16_NEON;
+  }
+#elif defined(HAS_SCALEROWDOWN4_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(dst_width, 8) && IS_ALIGNED(row_stride, 16) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_SSE2 :
+        ScaleRowDown4_16_SSE2;
+  }
+#elif defined(HAS_SCALEROWDOWN4_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_16_MIPS_DSPR2 : ScaleRowDown4_16_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+// Scale plane down, 3/4
+
+static void ScalePlaneDown34(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr,
+                             enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown34_0)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown34_0 = ScaleRowDown34_C;
+    ScaleRowDown34_1 = ScaleRowDown34_C;
+  } else {
+    ScaleRowDown34_0 = ScaleRowDown34_0_Box_C;
+    ScaleRowDown34_1 = ScaleRowDown34_1_Box_C;
+  }
+#if defined(HAS_SCALEROWDOWN34_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_NEON;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_SSSE3;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_MIPS_DSPR2;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
+                     dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static void ScalePlaneDown34_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr,
+                                enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown34_0)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown34_0 = ScaleRowDown34_16_C;
+    ScaleRowDown34_1 = ScaleRowDown34_16_C;
+  } else {
+    ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_C;
+    ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_C;
+  }
+#if defined(HAS_SCALEROWDOWN34_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_16_NEON;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_16_SSSE3;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_16_MIPS_DSPR2;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
+                     dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+
+// Scale plane, 3/8
+// This is an optimized version for scaling down a plane to 3/8
+// of its original size.
+//
+// Uses box filter arranges like this
+// aaabbbcc -> abc
+// aaabbbcc    def
+// aaabbbcc    ghi
+// dddeeeff
+// dddeeeff
+// dddeeeff
+// ggghhhii
+// ggghhhii
+// Boxes are 3x3, 2x3, 3x2 and 2x2
+
+static void ScalePlaneDown38(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr,
+                             enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown38_3)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown38_3 = ScaleRowDown38_C;
+    ScaleRowDown38_2 = ScaleRowDown38_C;
+  } else {
+    ScaleRowDown38_3 = ScaleRowDown38_3_Box_C;
+    ScaleRowDown38_2 = ScaleRowDown38_2_Box_C;
+  }
+#if defined(HAS_SCALEROWDOWN38_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 12 == 0)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_NEON;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_NEON;
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN38_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_SSSE3;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_SSSE3;
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN38_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_MIPS_DSPR2;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static void ScalePlaneDown38_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr,
+                                enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown38_3)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown38_3 = ScaleRowDown38_16_C;
+    ScaleRowDown38_2 = ScaleRowDown38_16_C;
+  } else {
+    ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_C;
+    ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_C;
+  }
+#if defined(HAS_SCALEROWDOWN38_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 12 == 0)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_16_NEON;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_NEON;
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN38_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_16_SSSE3;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_SSSE3;
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN38_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_16_MIPS_DSPR2;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static __inline uint32 SumBox(int iboxwidth, int iboxheight,
+                              ptrdiff_t src_stride, const uint8* src_ptr) {
+  uint32 sum = 0u;
+  int y;
+  assert(iboxwidth > 0);
+  assert(iboxheight > 0);
+  for (y = 0; y < iboxheight; ++y) {
+    int x;
+    for (x = 0; x < iboxwidth; ++x) {
+      sum += src_ptr[x];
+    }
+    src_ptr += src_stride;
+  }
+  return sum;
+}
+
+static __inline uint32 SumBox_16(int iboxwidth, int iboxheight,
+                                 ptrdiff_t src_stride, const uint16* src_ptr) {
+  uint32 sum = 0u;
+  int y;
+  assert(iboxwidth > 0);
+  assert(iboxheight > 0);
+  for (y = 0; y < iboxheight; ++y) {
+    int x;
+    for (x = 0; x < iboxwidth; ++x) {
+      sum += src_ptr[x];
+    }
+    src_ptr += src_stride;
+  }
+  return sum;
+}
+
+static void ScalePlaneBoxRow_C(int dst_width, int boxheight,
+                               int x, int dx, ptrdiff_t src_stride,
+                               const uint8* src_ptr, uint8* dst_ptr) {
+  int i;
+  int boxwidth;
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = (x >> 16) - ix;
+    *dst_ptr++ = SumBox(boxwidth, boxheight, src_stride, src_ptr + ix) /
+        (boxwidth * boxheight);
+  }
+}
+
+static void ScalePlaneBoxRow_16_C(int dst_width, int boxheight,
+                                  int x, int dx, ptrdiff_t src_stride,
+                                  const uint16* src_ptr, uint16* dst_ptr) {
+  int i;
+  int boxwidth;
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = (x >> 16) - ix;
+    *dst_ptr++ = SumBox_16(boxwidth, boxheight, src_stride, src_ptr + ix) /
+        (boxwidth * boxheight);
+  }
+}
+
+static __inline uint32 SumPixels(int iboxwidth, const uint16* src_ptr) {
+  uint32 sum = 0u;
+  int x;
+  assert(iboxwidth > 0);
+  for (x = 0; x < iboxwidth; ++x) {
+    sum += src_ptr[x];
+  }
+  return sum;
+}
+
+static __inline uint32 SumPixels_16(int iboxwidth, const uint32* src_ptr) {
+  uint32 sum = 0u;
+  int x;
+  assert(iboxwidth > 0);
+  for (x = 0; x < iboxwidth; ++x) {
+    sum += src_ptr[x];
+  }
+  return sum;
+}
+
+static void ScaleAddCols2_C(int dst_width, int boxheight, int x, int dx,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int i;
+  int scaletbl[2];
+  int minboxwidth = (dx >> 16);
+  int* scaleptr = scaletbl - minboxwidth;
+  int boxwidth;
+  scaletbl[0] = 65536 / (minboxwidth * boxheight);
+  scaletbl[1] = 65536 / ((minboxwidth + 1) * boxheight);
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = (x >> 16) - ix;
+    *dst_ptr++ = SumPixels(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16;
+  }
+}
+
+static void ScaleAddCols2_16_C(int dst_width, int boxheight, int x, int dx,
+                               const uint32* src_ptr, uint16* dst_ptr) {
+  int i;
+  int scaletbl[2];
+  int minboxwidth = (dx >> 16);
+  int* scaleptr = scaletbl - minboxwidth;
+  int boxwidth;
+  scaletbl[0] = 65536 / (minboxwidth * boxheight);
+  scaletbl[1] = 65536 / ((minboxwidth + 1) * boxheight);
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = (x >> 16) - ix;
+    *dst_ptr++ = SumPixels_16(boxwidth, src_ptr + ix) *
+        scaleptr[boxwidth] >> 16;
+  }
+}
+
+static void ScaleAddCols1_C(int dst_width, int boxheight, int x, int dx,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int boxwidth = (dx >> 16);
+  int scaleval = 65536 / (boxwidth * boxheight);
+  int i;
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = SumPixels(boxwidth, src_ptr + x) * scaleval >> 16;
+    x += boxwidth;
+  }
+}
+
+static void ScaleAddCols1_16_C(int dst_width, int boxheight, int x, int dx,
+                               const uint32* src_ptr, uint16* dst_ptr) {
+  int boxwidth = (dx >> 16);
+  int scaleval = 65536 / (boxwidth * boxheight);
+  int i;
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = SumPixels_16(boxwidth, src_ptr + x) * scaleval >> 16;
+    x += boxwidth;
+  }
+}
+
+// Scale plane down to any dimensions, with interpolation.
+// (boxfilter).
+//
+// Same method as SimpleScale, which is fixed point, outputting
+// one pixel of destination using fixed point (16.16) to step
+// through source, sampling a box of pixel with simple
+// averaging.
+static void ScalePlaneBox(int src_width, int src_height,
+                          int dst_width, int dst_height,
+                          int src_stride, int dst_stride,
+                          const uint8* src_ptr, uint8* dst_ptr) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height << 16);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  // TODO(fbarchard): Remove this and make AddRows handle boxheight 1.
+  if (!IS_ALIGNED(src_width, 16) || dst_height * 2 > src_height) {
+    uint8* dst = dst_ptr;
+    int j;
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint8* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+      boxheight = (y >> 16) - iy;
+      ScalePlaneBoxRow_C(dst_width, boxheight,
+                         x, dx, src_stride,
+                         src, dst);
+      dst += dst_stride;
+    }
+    return;
+  }
+  {
+    // Allocate a row buffer of uint16.
+    align_buffer_64(row16, src_width * 2);
+    void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
+        const uint16* src_ptr, uint8* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_C: ScaleAddCols1_C;
+    void (*ScaleAddRows)(const uint8* src_ptr, ptrdiff_t src_stride,
+        uint16* dst_ptr, int src_width, int src_height) = ScaleAddRows_C;
+
+#if defined(HAS_SCALEADDROWS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) &&
+#ifdef AVOID_OVERREAD
+        IS_ALIGNED(src_width, 16) &&
+#endif
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+      ScaleAddRows = ScaleAddRows_SSE2;
+    }
+#endif
+
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint8* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > (src_height << 16)) {
+        y = (src_height << 16);
+      }
+      boxheight = (y >> 16) - iy;
+      ScaleAddRows(src, src_stride, (uint16*)(row16),
+                 src_width, boxheight);
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint16*)(row16),
+                 dst_ptr);
+      dst_ptr += dst_stride;
+    }
+    free_aligned_buffer_64(row16);
+  }
+}
+
+static void ScalePlaneBox_16(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint16* src_ptr, uint16* dst_ptr) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height << 16);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  // TODO(fbarchard): Remove this and make AddRows handle boxheight 1.
+  if (!IS_ALIGNED(src_width, 16) || dst_height * 2 > src_height) {
+    uint16* dst = dst_ptr;
+    int j;
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint16* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+      boxheight = (y >> 16) - iy;
+      ScalePlaneBoxRow_16_C(dst_width, boxheight,
+                            x, dx, src_stride,
+                            src, dst);
+      dst += dst_stride;
+    }
+    return;
+  }
+  {
+    // Allocate a row buffer of uint32.
+    align_buffer_64(row32, src_width * 4);
+    void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
+        const uint32* src_ptr, uint16* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_16_C: ScaleAddCols1_16_C;
+    void (*ScaleAddRows)(const uint16* src_ptr, ptrdiff_t src_stride,
+        uint32* dst_ptr, int src_width, int src_height) = ScaleAddRows_16_C;
+
+#if defined(HAS_SCALEADDROWS_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) &&
+#ifdef AVOID_OVERREAD
+        IS_ALIGNED(src_width, 16) &&
+#endif
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+      ScaleAddRows = ScaleAddRows_16_SSE2;
+    }
+#endif
+
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint16* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > (src_height << 16)) {
+        y = (src_height << 16);
+      }
+      boxheight = (y >> 16) - iy;
+      ScaleAddRows(src, src_stride, (uint32*)(row32),
+                 src_width, boxheight);
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint32*)(row32),
+                 dst_ptr);
+      dst_ptr += dst_stride;
+    }
+    free_aligned_buffer_64(row32);
+  }
+}
+
+// Scale plane down with bilinear interpolation.
+void ScalePlaneBilinearDown(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row buffer.
+  align_buffer_64(row, src_width);
+
+  const int max_y = (src_height - 1) << 16;
+  int j;
+  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleFilterCols64_C : ScaleFilterCols_C;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && src_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && src_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(src_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+
+#if defined(HAS_SCALEFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_SSSE3;
+  }
+#endif
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    int yi = y >> 16;
+    const uint8* src = src_ptr + yi * src_stride;
+    if (filtering == kFilterLinear) {
+      ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow(row, src, src_stride, src_width, yf);
+      ScaleFilterCols(dst_ptr, row, dst_width, x, dx);
+    }
+    dst_ptr += dst_stride;
+    y += dy;
+    if (y > max_y) {
+      y = max_y;
+    }
+  }
+  free_aligned_buffer_64(row);
+}
+
+void ScalePlaneBilinearDown_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row buffer.
+  align_buffer_64(row, src_width * 2);
+
+  const int max_y = (src_height - 1) << 16;
+  int j;
+  void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleFilterCols64_16_C : ScaleFilterCols_16_C;
+  void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSE2;
+      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSSE3;
+      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && src_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && src_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(src_width, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+
+#if defined(HAS_SCALEFILTERCOLS_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_16_SSSE3;
+  }
+#endif
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    int yi = y >> 16;
+    const uint16* src = src_ptr + yi * src_stride;
+    if (filtering == kFilterLinear) {
+      ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow((uint16*)row, src, src_stride, src_width, yf);
+      ScaleFilterCols(dst_ptr, (uint16*)row, dst_width, x, dx);
+    }
+    dst_ptr += dst_stride;
+    y += dy;
+    if (y > max_y) {
+      y = max_y;
+    }
+  }
+  free_aligned_buffer_64(row);
+}
+
+// Scale up down with bilinear interpolation.
+void ScalePlaneBilinearUp(int src_width, int src_height,
+                          int dst_width, int dst_height,
+                          int src_stride, int dst_stride,
+                          const uint8* src_ptr, uint8* dst_ptr,
+                          enum FilterMode filtering) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height - 1) << 16;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
+       int dst_width, int x, int dx) =
+       filtering ? ScaleFilterCols_C : ScaleCols_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (filtering && src_width >= 32768) {
+    ScaleFilterCols = ScaleFilterCols64_C;
+  }
+#if defined(HAS_SCALEFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_SSSE3;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleFilterCols = ScaleColsUp2_C;
+#if defined(HAS_SCALECOLS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleFilterCols = ScaleColsUp2_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+  {
+    int yi = y >> 16;
+    const uint8* src = src_ptr + yi * src_stride;
+
+    // Allocate 2 row buffers.
+    const int kRowSize = (dst_width + 15) & ~15;
+    align_buffer_64(row, kRowSize * 2);
+
+    uint8* rowptr = row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_ptr + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf);
+      }
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+void ScalePlaneBilinearUp_16(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint16* src_ptr, uint16* dst_ptr,
+                             enum FilterMode filtering) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height - 1) << 16;
+  void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr,
+       int dst_width, int x, int dx) =
+       filtering ? ScaleFilterCols_16_C : ScaleCols_16_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSE2;
+      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSSE3;
+      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (filtering && src_width >= 32768) {
+    ScaleFilterCols = ScaleFilterCols64_16_C;
+  }
+#if defined(HAS_SCALEFILTERCOLS_16_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_16_SSSE3;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleFilterCols = ScaleColsUp2_16_C;
+#if defined(HAS_SCALECOLS_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleFilterCols = ScaleColsUp2_16_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+  {
+    int yi = y >> 16;
+    const uint16* src = src_ptr + yi * src_stride;
+
+    // Allocate 2 row buffers.
+    const int kRowSize = (dst_width + 15) & ~15;
+    align_buffer_64(row, kRowSize * 4);
+
+    uint16* rowptr = (uint16*)row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_ptr + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf);
+      }
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+// Scale Plane to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScalePlaneSimple(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr) {
+  int i;
+  void (*ScaleCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) = ScaleCols_C;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleCols = ScaleColsUp2_C;
+#if defined(HAS_SCALECOLS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleCols = ScaleColsUp2_SSE2;
+    }
+#endif
+  }
+
+  for (i = 0; i < dst_height; ++i) {
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride,
+              dst_width, x, dx);
+    dst_ptr += dst_stride;
+    y += dy;
+  }
+}
+
+static void ScalePlaneSimple_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr) {
+  int i;
+  void (*ScaleCols)(uint16* dst_ptr, const uint16* src_ptr,
+      int dst_width, int x, int dx) = ScaleCols_16_C;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleCols = ScaleColsUp2_16_C;
+#if defined(HAS_SCALECOLS_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleCols = ScaleColsUp2_16_SSE2;
+    }
+#endif
+  }
+
+  for (i = 0; i < dst_height; ++i) {
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride,
+              dst_width, x, dx);
+    dst_ptr += dst_stride;
+    y += dy;
+  }
+}
+
+// Scale a plane.
+// This function dispatches to a specialized scaler based on scale factor.
+
+LIBYUV_API
+void ScalePlane(const uint8* src, int src_stride,
+                int src_width, int src_height,
+                uint8* dst, int dst_stride,
+                int dst_width, int dst_height,
+                enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  // Use specialized scales to improve performance for common resolutions.
+  // For example, all the 1/2 scalings will use ScalePlaneDown2()
+  if (dst_width == src_width && dst_height == src_height) {
+    // Straight copy.
+    CopyPlane(src, src_stride, dst, dst_stride, dst_width, dst_height);
+    return;
+  }
+  if (dst_width == src_width) {
+    int dy = FixedDiv(src_height, dst_height);
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical(src_height,
+                       dst_width, dst_height,
+                       src_stride, dst_stride, src, dst,
+                       0, 0, dy, 1, filtering);
+    return;
+  }
+  if (dst_width <= Abs(src_width) && dst_height <= src_height) {
+    // Scale down.
+    if (4 * dst_width == 3 * src_width &&
+        4 * dst_height == 3 * src_height) {
+      // optimized, 3/4
+      ScalePlaneDown34(src_width, src_height, dst_width, dst_height,
+                       src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (2 * dst_width == src_width && 2 * dst_height == src_height) {
+      // optimized, 1/2
+      ScalePlaneDown2(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    // 3/8 rounded up for odd sized chroma height.
+    if (8 * dst_width == 3 * src_width &&
+        dst_height == ((src_height * 3 + 7) / 8)) {
+      // optimized, 3/8
+      ScalePlaneDown38(src_width, src_height, dst_width, dst_height,
+                       src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (4 * dst_width == src_width && 4 * dst_height == src_height &&
+               filtering != kFilterBilinear) {
+      // optimized, 1/4
+      ScalePlaneDown4(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+  }
+  if (filtering == kFilterBox && dst_height * 2 < src_height) {
+    ScalePlaneBox(src_width, src_height, dst_width, dst_height,
+                  src_stride, dst_stride, src, dst);
+    return;
+  }
+  if (filtering && dst_height > src_height) {
+    ScalePlaneBilinearUp(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  if (filtering) {
+    ScalePlaneBilinearDown(src_width, src_height, dst_width, dst_height,
+                           src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
+                   src_stride, dst_stride, src, dst);
+}
+
+LIBYUV_API
+void ScalePlane_16(const uint16* src, int src_stride,
+                  int src_width, int src_height,
+                  uint16* dst, int dst_stride,
+                  int dst_width, int dst_height,
+                  enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  // Use specialized scales to improve performance for common resolutions.
+  // For example, all the 1/2 scalings will use ScalePlaneDown2()
+  if (dst_width == src_width && dst_height == src_height) {
+    // Straight copy.
+    CopyPlane_16(src, src_stride, dst, dst_stride, dst_width, dst_height);
+    return;
+  }
+  if (dst_width == src_width) {
+    int dy = FixedDiv(src_height, dst_height);
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical_16(src_height,
+                          dst_width, dst_height,
+                          src_stride, dst_stride, src, dst,
+                          0, 0, dy, 1, filtering);
+    return;
+  }
+  if (dst_width <= Abs(src_width) && dst_height <= src_height) {
+    // Scale down.
+    if (4 * dst_width == 3 * src_width &&
+        4 * dst_height == 3 * src_height) {
+      // optimized, 3/4
+      ScalePlaneDown34_16(src_width, src_height, dst_width, dst_height,
+                          src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (2 * dst_width == src_width && 2 * dst_height == src_height) {
+      // optimized, 1/2
+      ScalePlaneDown2_16(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    // 3/8 rounded up for odd sized chroma height.
+    if (8 * dst_width == 3 * src_width &&
+        dst_height == ((src_height * 3 + 7) / 8)) {
+      // optimized, 3/8
+      ScalePlaneDown38_16(src_width, src_height, dst_width, dst_height,
+                          src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (4 * dst_width == src_width && 4 * dst_height == src_height &&
+               filtering != kFilterBilinear) {
+      // optimized, 1/4
+      ScalePlaneDown4_16(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+  }
+  if (filtering == kFilterBox && dst_height * 2 < src_height) {
+    ScalePlaneBox_16(src_width, src_height, dst_width, dst_height,
+                     src_stride, dst_stride, src, dst);
+    return;
+  }
+  if (filtering && dst_height > src_height) {
+    ScalePlaneBilinearUp_16(src_width, src_height, dst_width, dst_height,
+                            src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  if (filtering) {
+    ScalePlaneBilinearDown_16(src_width, src_height, dst_width, dst_height,
+                              src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  ScalePlaneSimple_16(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst);
+}
+
+// Scale an I420 image.
+// This function in turn calls a scaling function for each plane.
+
+LIBYUV_API
+int I420Scale(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              int src_width, int src_height,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int dst_width, int dst_height,
+              enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane(src_y, src_stride_y, src_width, src_height,
+             dst_y, dst_stride_y, dst_width, dst_height,
+             filtering);
+  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight,
+             dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
+             filtering);
+  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight,
+             dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
+             filtering);
+  return 0;
+}
+
+LIBYUV_API
+int I420Scale_16(const uint16* src_y, int src_stride_y,
+                 const uint16* src_u, int src_stride_u,
+                 const uint16* src_v, int src_stride_v,
+                 int src_width, int src_height,
+                 uint16* dst_y, int dst_stride_y,
+                 uint16* dst_u, int dst_stride_u,
+                 uint16* dst_v, int dst_stride_v,
+                 int dst_width, int dst_height,
+                 enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_16(src_y, src_stride_y, src_width, src_height,
+                dst_y, dst_stride_y, dst_width, dst_height,
+                filtering);
+  ScalePlane_16(src_u, src_stride_u, src_halfwidth, src_halfheight,
+                dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
+                filtering);
+  ScalePlane_16(src_v, src_stride_v, src_halfwidth, src_halfheight,
+                dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
+                filtering);
+  return 0;
+}
+
+// Deprecated api
+LIBYUV_API
+int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
+          int src_stride_y, int src_stride_u, int src_stride_v,
+          int src_width, int src_height,
+          uint8* dst_y, uint8* dst_u, uint8* dst_v,
+          int dst_stride_y, int dst_stride_u, int dst_stride_v,
+          int dst_width, int dst_height,
+          LIBYUV_BOOL interpolate) {
+  return I420Scale(src_y, src_stride_y,
+                   src_u, src_stride_u,
+                   src_v, src_stride_v,
+                   src_width, src_height,
+                   dst_y, dst_stride_y,
+                   dst_u, dst_stride_u,
+                   dst_v, dst_stride_v,
+                   dst_width, dst_height,
+                   interpolate ? kFilterBox : kFilterNone);
+}
+
+// Deprecated api
+LIBYUV_API
+int ScaleOffset(const uint8* src, int src_width, int src_height,
+                uint8* dst, int dst_width, int dst_height, int dst_yoffset,
+                LIBYUV_BOOL interpolate) {
+  // Chroma requires offset to multiple of 2.
+  int dst_yoffset_even = dst_yoffset & ~1;
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  int aheight = dst_height - dst_yoffset_even * 2;  // actual output height
+  const uint8* src_y = src;
+  const uint8* src_u = src + src_width * src_height;
+  const uint8* src_v = src + src_width * src_height +
+                             src_halfwidth * src_halfheight;
+  uint8* dst_y = dst + dst_yoffset_even * dst_width;
+  uint8* dst_u = dst + dst_width * dst_height +
+                 (dst_yoffset_even >> 1) * dst_halfwidth;
+  uint8* dst_v = dst + dst_width * dst_height + dst_halfwidth * dst_halfheight +
+                 (dst_yoffset_even >> 1) * dst_halfwidth;
+  if (!src || src_width <= 0 || src_height <= 0 ||
+      !dst || dst_width <= 0 || dst_height <= 0 || dst_yoffset_even < 0 ||
+      dst_yoffset_even >= dst_height) {
+    return -1;
+  }
+  return I420Scale(src_y, src_width,
+                   src_u, src_halfwidth,
+                   src_v, src_halfwidth,
+                   src_width, src_height,
+                   dst_y, dst_width,
+                   dst_u, dst_halfwidth,
+                   dst_v, dst_halfwidth,
+                   dst_width, aheight,
+                   interpolate ? kFilterBox : kFilterNone);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/scale_common.cc b/third_party/libyuv/source/scale_common.cc
new file mode 100644
index 0000000..595ad66
--- /dev/null
+++ b/third_party/libyuv/source/scale_common.cc
@@ -0,0 +1,1165 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "third_party/libyuv/include/libyuv/cpu_id.h"
+#include "third_party/libyuv/include/libyuv/planar_functions.h"  // CopyARGB
+#include "third_party/libyuv/include/libyuv/row.h"
+#include "third_party/libyuv/include/libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// CPU agnostic row functions
+void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[1];
+    dst[1] = src_ptr[3];
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[1];
+  }
+}
+
+void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[1];
+    dst[1] = src_ptr[3];
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[1];
+  }
+}
+
+void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+    dst[1] = (s[2] + s[3] + 1) >> 1;
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+  }
+}
+
+void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                              uint16* dst, int dst_width) {
+  const uint16* s = src_ptr;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+    dst[1] = (s[2] + s[3] + 1) >> 1;
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+  }
+}
+
+void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+  }
+}
+
+void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+  }
+}
+
+void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[2];
+    dst[1] = src_ptr[6];
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[2];
+  }
+}
+
+void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[2];
+    dst[1] = src_ptr[6];
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[2];
+  }
+}
+
+void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+    dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
+             src_ptr[stride + 4] + src_ptr[stride + 5] +
+             src_ptr[stride + 6] + src_ptr[stride + 7] +
+             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
+             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
+             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
+             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
+             8) >> 4;
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+  }
+}
+
+void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+    dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
+             src_ptr[stride + 4] + src_ptr[stride + 5] +
+             src_ptr[stride + 6] + src_ptr[stride + 7] +
+             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
+             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
+             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
+             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
+             8) >> 4;
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+  }
+}
+
+void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width) {
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[1];
+    dst[2] = src_ptr[3];
+    dst += 3;
+    src_ptr += 4;
+  }
+}
+
+void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width) {
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[1];
+    dst[2] = src_ptr[3];
+    dst += 3;
+    src_ptr += 4;
+  }
+}
+
+// Filter rows 0 and 1 together, 3 : 1
+void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 * 3 + b0 + 2) >> 2;
+    d[1] = (a1 * 3 + b1 + 2) >> 2;
+    d[2] = (a2 * 3 + b2 + 2) >> 2;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 * 3 + b0 + 2) >> 2;
+    d[1] = (a1 * 3 + b1 + 2) >> 2;
+    d[2] = (a2 * 3 + b2 + 2) >> 2;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+// Filter rows 1 and 2 together, 1 : 1
+void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 + b0 + 1) >> 1;
+    d[1] = (a1 + b1 + 1) >> 1;
+    d[2] = (a2 + b2 + 1) >> 1;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 + b0 + 1) >> 1;
+    d[1] = (a1 + b1 + 1) >> 1;
+    d[2] = (a2 + b2 + 1) >> 1;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                 int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[0] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr[1] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[x >> 16];
+  }
+}
+
+void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                    int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[0] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr[1] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
+                    int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[1] = dst_ptr[0] = src_ptr[0];
+    src_ptr += 1;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[0];
+  }
+}
+
+void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[1] = dst_ptr[0] = src_ptr[0];
+    src_ptr += 1;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[0];
+  }
+}
+
+// (1-f)a + fb can be replaced with a + f(b-a)
+#define BLENDER(a, b, f) (uint8)((int)(a) + \
+    ((int)(f) * ((int)(b) - (int)(a)) >> 16))
+
+void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+
+void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
+                         int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+#undef BLENDER
+
+#define BLENDER(a, b, f) (uint16)((int)(a) + \
+    ((int)(f) * ((int)(b) - (int)(a)) >> 16))
+
+void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+
+void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+#undef BLENDER
+
+void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width) {
+  int x;
+  assert(dst_width % 3 == 0);
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[3];
+    dst[2] = src_ptr[6];
+    dst += 3;
+    src_ptr += 8;
+  }
+}
+
+void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width) {
+  int x;
+  assert(dst_width % 3 == 0);
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[3];
+    dst[2] = src_ptr[6];
+    dst += 3;
+    src_ptr += 8;
+  }
+}
+
+// 8x3 -> 3x1
+void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+        (65536 / 9) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+        (65536 / 9) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7] +
+        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+        (65536 / 6) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+        (65536 / 9) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+        (65536 / 9) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7] +
+        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+        (65536 / 6) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+// 8x2 -> 3x1
+void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2]) * (65536 / 6) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5]) * (65536 / 6) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7]) *
+        (65536 / 4) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2]) * (65536 / 6) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5]) * (65536 / 6) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7]) *
+        (65536 / 4) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleAddRows_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height) {
+  int x;
+  assert(src_width > 0);
+  assert(src_height > 0);
+  for (x = 0; x < src_width; ++x) {
+    const uint8* s = src_ptr + x;
+    unsigned int sum = 0u;
+    int y;
+    for (y = 0; y < src_height; ++y) {
+      sum += s[0];
+      s += src_stride;
+    }
+    // TODO(fbarchard): Consider limitting height to 256 to avoid overflow.
+    dst_ptr[x] = sum < 65535u ? sum : 65535u;
+  }
+}
+
+void ScaleAddRows_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                       uint32* dst_ptr, int src_width, int src_height) {
+  int x;
+  assert(src_width > 0);
+  assert(src_height > 0);
+  for (x = 0; x < src_width; ++x) {
+    const uint16* s = src_ptr + x;
+    unsigned int sum = 0u;
+    int y;
+    for (y = 0; y < src_height; ++y) {
+      sum += s[0];
+      s += src_stride;
+    }
+    // No risk of overflow here now
+    dst_ptr[x] = sum;
+  }
+}
+
+void ScaleARGBRowDown2_C(const uint8* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8* dst_argb, int dst_width) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[1];
+    dst[1] = src[3];
+    src += 4;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[1];
+  }
+}
+
+void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] + 1) >> 1;
+    dst_argb[1] = (src_argb[1] + src_argb[5] + 1) >> 1;
+    dst_argb[2] = (src_argb[2] + src_argb[6] + 1) >> 1;
+    dst_argb[3] = (src_argb[3] + src_argb[7] + 1) >> 1;
+    src_argb += 8;
+    dst_argb += 4;
+  }
+}
+
+void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] +
+                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] +
+                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] +
+                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] +
+                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    src_argb += 8;
+    dst_argb += 4;
+  }
+}
+
+void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            int src_stepx,
+                            uint8* dst_argb, int dst_width) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[0];
+    dst[1] = src[src_stepx];
+    src += src_stepx * 2;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] +
+                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] +
+                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] +
+                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] +
+                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    src_argb += src_stepx * 4;
+    dst_argb += 4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
+                     int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[1] = dst[0] = src[0];
+    src += 1;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+// Mimics SSSE3 blender
+#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b) * f) >> 7
+#define BLENDERC(a, b, f, s) (uint32)( \
+    BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
+#define BLENDER(a, b, f) \
+    BLENDERC(a, b, f, 24) | BLENDERC(a, b, f, 16) | \
+    BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0)
+
+void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+
+void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
+                             int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+#undef BLENDER1
+#undef BLENDERC
+#undef BLENDER
+
+// Scale plane vertically with bilinear interpolation.
+void ScalePlaneVertical(int src_height,
+                        int dst_width, int dst_height,
+                        int src_stride, int dst_stride,
+                        const uint8* src_argb, uint8* dst_argb,
+                        int x, int y, int dy,
+                        int bpp, enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher bpp.
+  int dst_width_bytes = dst_width * bpp;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(bpp >= 1 && bpp <= 4);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * bpp;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width_bytes >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width_bytes >= 4 &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width_bytes, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow(dst_argb, src_argb + yi * src_stride,
+                   src_stride, dst_width_bytes, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+void ScalePlaneVertical_16(int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint16* src_argb, uint16* dst_argb,
+                           int x, int y, int dy,
+                           int wpp, enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher wpp.
+  int dst_width_words = dst_width * wpp;
+  void (*InterpolateRow)(uint16* dst_argb, const uint16* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(wpp >= 1 && wpp <= 2);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * wpp;
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSE2;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width_bytes >= 32) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width_bytes >= 4 &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width_bytes, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow(dst_argb, src_argb + yi * src_stride,
+                   src_stride, dst_width_words, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
+// Simplify the filtering based on scale factors.
+enum FilterMode ScaleFilterReduce(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  enum FilterMode filtering) {
+  if (src_width < 0) {
+    src_width = -src_width;
+  }
+  if (src_height < 0) {
+    src_height = -src_height;
+  }
+  if (filtering == kFilterBox) {
+    // If scaling both axis to 0.5 or larger, switch from Box to Bilinear.
+    if (dst_width * 2 >= src_width && dst_height * 2 >= src_height) {
+      filtering = kFilterBilinear;
+    }
+    // If scaling to larger, switch from Box to Bilinear.
+    if (dst_width >= src_width || dst_height >= src_height) {
+      filtering = kFilterBilinear;
+    }
+  }
+  if (filtering == kFilterBilinear) {
+    if (src_height == 1) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to Linear.
+    if (dst_height == src_height || dst_height * 3 == src_height) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Remove 1 pixel wide filter restriction, which is to
+    // avoid reading 2 pixels horizontally that causes memory exception.
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+  }
+  if (filtering == kFilterLinear) {
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to None.
+    if (dst_width == src_width || dst_width * 3 == src_width) {
+      filtering = kFilterNone;
+    }
+  }
+  return filtering;
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_C(int num, int div) {
+  return (int)(((int64)(num) << 16) / div);
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv1_C(int num, int div) {
+  return (int)((((int64)(num) << 16) - 0x00010001) /
+                          (div - 1));
+}
+
+#define CENTERSTART(dx, s) (dx < 0) ? -((-dx >> 1) + s) : ((dx >> 1) + s)
+
+// Compute slope values for stepping.
+void ScaleSlope(int src_width, int src_height,
+                int dst_width, int dst_height,
+                enum FilterMode filtering,
+                int* x, int* y, int* dx, int* dy) {
+  assert(x != NULL);
+  assert(y != NULL);
+  assert(dx != NULL);
+  assert(dy != NULL);
+  assert(src_width != 0);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  // Check for 1 pixel and avoid FixedDiv overflow.
+  if (dst_width == 1 && src_width >= 32768) {
+    dst_width = src_width;
+  }
+  if (dst_height == 1 && src_height >= 32768) {
+    dst_height = src_height;
+  }
+  if (filtering == kFilterBox) {
+    // Scale step for point sampling duplicates all pixels equally.
+    *dx = FixedDiv(Abs(src_width), dst_width);
+    *dy = FixedDiv(src_height, dst_height);
+    *x = 0;
+    *y = 0;
+  } else if (filtering == kFilterBilinear) {
+    // Scale step for bilinear sampling renders last pixel once for upsample.
+    if (dst_width <= Abs(src_width)) {
+      *dx = FixedDiv(Abs(src_width), dst_width);
+      *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_width > 1) {
+      *dx = FixedDiv1(Abs(src_width), dst_width);
+      *x = 0;
+    }
+    if (dst_height <= src_height) {
+      *dy = FixedDiv(src_height,  dst_height);
+      *y = CENTERSTART(*dy, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_height > 1) {
+      *dy = FixedDiv1(src_height, dst_height);
+      *y = 0;
+    }
+  } else if (filtering == kFilterLinear) {
+    // Scale step for bilinear sampling renders last pixel once for upsample.
+    if (dst_width <= Abs(src_width)) {
+      *dx = FixedDiv(Abs(src_width), dst_width);
+      *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_width > 1) {
+      *dx = FixedDiv1(Abs(src_width), dst_width);
+      *x = 0;
+    }
+    *dy = FixedDiv(src_height, dst_height);
+    *y = *dy >> 1;
+  } else {
+    // Scale step for point sampling duplicates all pixels equally.
+    *dx = FixedDiv(Abs(src_width), dst_width);
+    *dy = FixedDiv(src_height, dst_height);
+    *x = CENTERSTART(*dx, 0);
+    *y = CENTERSTART(*dy, 0);
+  }
+  // Negative src_width means horizontally mirror.
+  if (src_width < 0) {
+    *x += (dst_width - 1) * *dx;
+    *dx = -*dx;
+    // src_width = -src_width;   // Caller must do this.
+  }
+}
+#undef CENTERSTART
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/scale_mips.cc b/third_party/libyuv/source/scale_mips.cc
new file mode 100644
index 0000000..5722dea
--- /dev/null
+++ b/third_party/libyuv/source/scale_mips.cc
@@ -0,0 +1,653 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/basic_types.h"
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC MIPS DSPR2
+#if !defined(LIBYUV_DISABLE_MIPS) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width) {
+  __asm__ __volatile__(
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+
+    "srl            $t9, %[dst_width], 4           \n"  // iterations -> by 16
+    "beqz           $t9, 2f                        \n"
+    " nop                                          \n"
+
+    ".p2align       2                              \n"
+  "1:                                              \n"
+    "lw             $t0, 0(%[src_ptr])             \n"  // |3|2|1|0|
+    "lw             $t1, 4(%[src_ptr])             \n"  // |7|6|5|4|
+    "lw             $t2, 8(%[src_ptr])             \n"  // |11|10|9|8|
+    "lw             $t3, 12(%[src_ptr])            \n"  // |15|14|13|12|
+    "lw             $t4, 16(%[src_ptr])            \n"  // |19|18|17|16|
+    "lw             $t5, 20(%[src_ptr])            \n"  // |23|22|21|20|
+    "lw             $t6, 24(%[src_ptr])            \n"  // |27|26|25|24|
+    "lw             $t7, 28(%[src_ptr])            \n"  // |31|30|29|28|
+    // TODO(fbarchard): Use odd pixels instead of even.
+    "precr.qb.ph    $t8, $t1, $t0                  \n"  // |6|4|2|0|
+    "precr.qb.ph    $t0, $t3, $t2                  \n"  // |14|12|10|8|
+    "precr.qb.ph    $t1, $t5, $t4                  \n"  // |22|20|18|16|
+    "precr.qb.ph    $t2, $t7, $t6                  \n"  // |30|28|26|24|
+    "addiu          %[src_ptr], %[src_ptr], 32     \n"
+    "addiu          $t9, $t9, -1                   \n"
+    "sw             $t8, 0(%[dst])                 \n"
+    "sw             $t0, 4(%[dst])                 \n"
+    "sw             $t1, 8(%[dst])                 \n"
+    "sw             $t2, 12(%[dst])                \n"
+    "bgtz           $t9, 1b                        \n"
+    " addiu         %[dst], %[dst], 16             \n"
+
+  "2:                                              \n"
+    "andi           $t9, %[dst_width], 0xf         \n"  // residue
+    "beqz           $t9, 3f                        \n"
+    " nop                                          \n"
+
+  "21:                                             \n"
+    "lbu            $t0, 0(%[src_ptr])             \n"
+    "addiu          %[src_ptr], %[src_ptr], 2      \n"
+    "addiu          $t9, $t9, -1                   \n"
+    "sb             $t0, 0(%[dst])                 \n"
+    "bgtz           $t9, 21b                       \n"
+    " addiu         %[dst], %[dst], 1              \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+  : [src_ptr] "+r" (src_ptr),
+    [dst] "+r" (dst)
+  : [dst_width] "r" (dst_width)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width) {
+  const uint8* t = src_ptr + src_stride;
+
+  __asm__ __volatile__ (
+    ".set push                                    \n"
+    ".set noreorder                               \n"
+
+    "srl            $t9, %[dst_width], 3          \n"  // iterations -> step 8
+    "bltz           $t9, 2f                       \n"
+    " nop                                         \n"
+
+    ".p2align       2                             \n"
+  "1:                                             \n"
+    "lw             $t0, 0(%[src_ptr])            \n"  // |3|2|1|0|
+    "lw             $t1, 4(%[src_ptr])            \n"  // |7|6|5|4|
+    "lw             $t2, 8(%[src_ptr])            \n"  // |11|10|9|8|
+    "lw             $t3, 12(%[src_ptr])           \n"  // |15|14|13|12|
+    "lw             $t4, 0(%[t])                  \n"  // |19|18|17|16|
+    "lw             $t5, 4(%[t])                  \n"  // |23|22|21|20|
+    "lw             $t6, 8(%[t])                  \n"  // |27|26|25|24|
+    "lw             $t7, 12(%[t])                 \n"  // |31|30|29|28|
+    "addiu          $t9, $t9, -1                  \n"
+    "srl            $t8, $t0, 16                  \n"  // |X|X|3|2|
+    "ins            $t0, $t4, 16, 16              \n"  // |17|16|1|0|
+    "ins            $t4, $t8, 0, 16               \n"  // |19|18|3|2|
+    "raddu.w.qb     $t0, $t0                      \n"  // |17+16+1+0|
+    "raddu.w.qb     $t4, $t4                      \n"  // |19+18+3+2|
+    "shra_r.w       $t0, $t0, 2                   \n"  // |t0+2|>>2
+    "shra_r.w       $t4, $t4, 2                   \n"  // |t4+2|>>2
+    "srl            $t8, $t1, 16                  \n"  // |X|X|7|6|
+    "ins            $t1, $t5, 16, 16              \n"  // |21|20|5|4|
+    "ins            $t5, $t8, 0, 16               \n"  // |22|23|7|6|
+    "raddu.w.qb     $t1, $t1                      \n"  // |21+20+5+4|
+    "raddu.w.qb     $t5, $t5                      \n"  // |23+22+7+6|
+    "shra_r.w       $t1, $t1, 2                   \n"  // |t1+2|>>2
+    "shra_r.w       $t5, $t5, 2                   \n"  // |t5+2|>>2
+    "srl            $t8, $t2, 16                  \n"  // |X|X|11|10|
+    "ins            $t2, $t6, 16, 16              \n"  // |25|24|9|8|
+    "ins            $t6, $t8, 0, 16               \n"  // |27|26|11|10|
+    "raddu.w.qb     $t2, $t2                      \n"  // |25+24+9+8|
+    "raddu.w.qb     $t6, $t6                      \n"  // |27+26+11+10|
+    "shra_r.w       $t2, $t2, 2                   \n"  // |t2+2|>>2
+    "shra_r.w       $t6, $t6, 2                   \n"  // |t5+2|>>2
+    "srl            $t8, $t3, 16                  \n"  // |X|X|15|14|
+    "ins            $t3, $t7, 16, 16              \n"  // |29|28|13|12|
+    "ins            $t7, $t8, 0, 16               \n"  // |31|30|15|14|
+    "raddu.w.qb     $t3, $t3                      \n"  // |29+28+13+12|
+    "raddu.w.qb     $t7, $t7                      \n"  // |31+30+15+14|
+    "shra_r.w       $t3, $t3, 2                   \n"  // |t3+2|>>2
+    "shra_r.w       $t7, $t7, 2                   \n"  // |t7+2|>>2
+    "addiu          %[src_ptr], %[src_ptr], 16    \n"
+    "addiu          %[t], %[t], 16                \n"
+    "sb             $t0, 0(%[dst])                \n"
+    "sb             $t4, 1(%[dst])                \n"
+    "sb             $t1, 2(%[dst])                \n"
+    "sb             $t5, 3(%[dst])                \n"
+    "sb             $t2, 4(%[dst])                \n"
+    "sb             $t6, 5(%[dst])                \n"
+    "sb             $t3, 6(%[dst])                \n"
+    "sb             $t7, 7(%[dst])                \n"
+    "bgtz           $t9, 1b                       \n"
+    " addiu         %[dst], %[dst], 8             \n"
+
+  "2:                                             \n"
+    "andi           $t9, %[dst_width], 0x7        \n"  // x = residue
+    "beqz           $t9, 3f                       \n"
+    " nop                                         \n"
+
+    "21:                                          \n"
+    "lwr            $t1, 0(%[src_ptr])            \n"
+    "lwl            $t1, 3(%[src_ptr])            \n"
+    "lwr            $t2, 0(%[t])                  \n"
+    "lwl            $t2, 3(%[t])                  \n"
+    "srl            $t8, $t1, 16                  \n"
+    "ins            $t1, $t2, 16, 16              \n"
+    "ins            $t2, $t8, 0, 16               \n"
+    "raddu.w.qb     $t1, $t1                      \n"
+    "raddu.w.qb     $t2, $t2                      \n"
+    "shra_r.w       $t1, $t1, 2                   \n"
+    "shra_r.w       $t2, $t2, 2                   \n"
+    "sb             $t1, 0(%[dst])                \n"
+    "sb             $t2, 1(%[dst])                \n"
+    "addiu          %[src_ptr], %[src_ptr], 4     \n"
+    "addiu          $t9, $t9, -2                  \n"
+    "addiu          %[t], %[t], 4                 \n"
+    "bgtz           $t9, 21b                      \n"
+    " addiu         %[dst], %[dst], 2             \n"
+
+  "3:                                             \n"
+    ".set pop                                     \n"
+
+  : [src_ptr] "+r" (src_ptr),
+    [dst] "+r" (dst), [t] "+r" (t)
+  : [dst_width] "r" (dst_width)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                    \n"
+      ".set noreorder                               \n"
+
+      "srl            $t9, %[dst_width], 3          \n"
+      "beqz           $t9, 2f                       \n"
+      " nop                                         \n"
+
+      ".p2align       2                             \n"
+     "1:                                            \n"
+      "lw             $t1, 0(%[src_ptr])            \n"  // |3|2|1|0|
+      "lw             $t2, 4(%[src_ptr])            \n"  // |7|6|5|4|
+      "lw             $t3, 8(%[src_ptr])            \n"  // |11|10|9|8|
+      "lw             $t4, 12(%[src_ptr])           \n"  // |15|14|13|12|
+      "lw             $t5, 16(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw             $t6, 20(%[src_ptr])           \n"  // |23|22|21|20|
+      "lw             $t7, 24(%[src_ptr])           \n"  // |27|26|25|24|
+      "lw             $t8, 28(%[src_ptr])           \n"  // |31|30|29|28|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |6|4|2|0|
+      "precr.qb.ph    $t2, $t4, $t3                 \n"  // |14|12|10|8|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |22|20|18|16|
+      "precr.qb.ph    $t6, $t8, $t7                 \n"  // |30|28|26|24|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |12|8|4|0|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |28|24|20|16|
+      "addiu          %[src_ptr], %[src_ptr], 32    \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sw             $t1, 0(%[dst])                \n"
+      "sw             $t5, 4(%[dst])                \n"
+      "bgtz           $t9, 1b                       \n"
+      " addiu         %[dst], %[dst], 8             \n"
+
+    "2:                                             \n"
+      "andi           $t9, %[dst_width], 7          \n"  // residue
+      "beqz           $t9, 3f                       \n"
+      " nop                                         \n"
+
+    "21:                                            \n"
+      "lbu            $t1, 0(%[src_ptr])            \n"
+      "addiu          %[src_ptr], %[src_ptr], 4     \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sb             $t1, 0(%[dst])                \n"
+      "bgtz           $t9, 21b                      \n"
+      " addiu         %[dst], %[dst], 1             \n"
+
+    "3:                                             \n"
+      ".set pop                                     \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst)
+      : [dst_width] "r" (dst_width)
+      : "t1", "t2", "t3", "t4", "t5",
+        "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  const uint8* s2 = s1 + stride;
+  const uint8* s3 = s2 + stride;
+
+  __asm__ __volatile__ (
+      ".set push                                  \n"
+      ".set noreorder                             \n"
+
+      "srl           $t9, %[dst_width], 1         \n"
+      "andi          $t8, %[dst_width], 1         \n"
+
+      ".p2align      2                            \n"
+     "1:                                          \n"
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "lw            $t4, 4(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw            $t5, 4(%[s1])                \n"  // |23|22|21|20|
+      "lw            $t6, 4(%[s2])                \n"  // |27|26|25|24|
+      "lw            $t7, 4(%[s3])                \n"  // |31|30|29|28|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "raddu.w.qb    $t4, $t4                     \n"  // |19 + 18 + 17 + 16|
+      "raddu.w.qb    $t5, $t5                     \n"  // |23 + 22 + 21 + 20|
+      "raddu.w.qb    $t6, $t6                     \n"  // |27 + 26 + 25 + 24|
+      "raddu.w.qb    $t7, $t7                     \n"  // |31 + 30 + 29 + 28|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "add           $t4, $t4, $t5                \n"
+      "add           $t6, $t6, $t7                \n"
+      "add           $t4, $t4, $t6                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "shra_r.w      $t4, $t4, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+      "sb            $t4, 1(%[dst])               \n"
+      "addiu         %[src_ptr], %[src_ptr], 8    \n"
+      "addiu         %[s1], %[s1], 8              \n"
+      "addiu         %[s2], %[s2], 8              \n"
+      "addiu         %[s3], %[s3], 8              \n"
+      "addiu         $t9, $t9, -1                 \n"
+      "bgtz          $t9, 1b                      \n"
+      " addiu        %[dst], %[dst], 2            \n"
+      "beqz          $t8, 2f                      \n"
+      " nop                                       \n"
+
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+
+      "2:                                         \n"
+      ".set pop                                   \n"
+
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [s1] "+r" (s1),
+        [s2] "+r" (s2),
+        [s3] "+r" (s3)
+      : [dst_width] "r" (dst_width)
+      : "t0", "t1", "t2", "t3", "t4", "t5",
+        "t6","t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                          \n"
+      ".set noreorder                                     \n"
+      ".p2align        2                                  \n"
+    "1:                                                   \n"
+      "lw              $t1, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw              $t2, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw              $t3, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw              $t4, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw              $t5, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw              $t6, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw              $t7, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw              $t8, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "precrq.qb.ph    $t0, $t2, $t4                      \n"  // |7|5|15|13|
+      "precrq.qb.ph    $t9, $t6, $t8                      \n"  // |23|21|31|30|
+      "addiu           %[dst_width], %[dst_width], -24    \n"
+      "ins             $t1, $t1, 8, 16                    \n"  // |3|1|0|X|
+      "ins             $t4, $t0, 8, 16                    \n"  // |X|15|13|12|
+      "ins             $t5, $t5, 8, 16                    \n"  // |19|17|16|X|
+      "ins             $t8, $t9, 8, 16                    \n"  // |X|31|29|28|
+      "addiu           %[src_ptr], %[src_ptr], 32         \n"
+      "packrl.ph       $t0, $t3, $t0                      \n"  // |9|8|7|5|
+      "packrl.ph       $t9, $t7, $t9                      \n"  // |25|24|23|21|
+      "prepend         $t1, $t2, 8                        \n"  // |4|3|1|0|
+      "prepend         $t3, $t4, 24                       \n"  // |15|13|12|11|
+      "prepend         $t5, $t6, 8                        \n"  // |20|19|17|16|
+      "prepend         $t7, $t8, 24                       \n"  // |31|29|28|27|
+      "sw              $t1, 0(%[dst])                     \n"
+      "sw              $t0, 4(%[dst])                     \n"
+      "sw              $t3, 8(%[dst])                     \n"
+      "sw              $t5, 12(%[dst])                    \n"
+      "sw              $t9, 16(%[dst])                    \n"
+      "sw              $t7, 20(%[dst])                    \n"
+      "bnez            %[dst_width], 1b                   \n"
+      " addiu          %[dst], %[dst], 24                 \n"
+      ".set pop                                           \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4", "t5",
+        "t6","t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "repl.ph           $t3, 3                          \n"  // 0x00030003
+
+     ".p2align           2                               \n"
+    "1:                                                  \n"
+      "lw                $t0, 0(%[src_ptr])              \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])  \n"  // |T3|T2|T1|T0|
+      "rotr              $t2, $t0, 8                     \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                     \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t4, $t2, $t3                   \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t3                   \n"  // |T0*3|T3*3|
+      "andi              $t0, $t2, 0xFFFF                \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                        \n"
+      "raddu.w.qb        $t1, $t1                        \n"
+      "shra_r.w          $t0, $t0, 1                     \n"
+      "shra_r.w          $t1, $t1, 1                     \n"
+      "preceu.ph.qbr     $t2, $t2                        \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                        \n"  // |0|T2|0|T1|
+      "rotr              $t2, $t2, 16                    \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                    \n"  // |0|T1|0|T2|
+      "addu.ph           $t2, $t2, $t4                   \n"
+      "addu.ph           $t6, $t6, $t5                   \n"
+      "sll               $t5, $t0, 1                     \n"
+      "add               $t0, $t5, $t0                   \n"
+      "shra_r.ph         $t2, $t2, 2                     \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "shll.ph           $t4, $t2, 1                     \n"
+      "addq.ph           $t4, $t4, $t2                   \n"
+      "addu              $t0, $t0, $t1                   \n"
+      "addiu             %[src_ptr], %[src_ptr], 4       \n"
+      "shra_r.w          $t0, $t0, 2                     \n"
+      "addu.ph           $t6, $t6, $t4                   \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "srl               $t1, $t6, 16                    \n"
+      "addiu             %[dst_width], %[dst_width], -3  \n"
+      "sb                $t1, 0(%[d])                    \n"
+      "sb                $t0, 1(%[d])                    \n"
+      "sb                $t6, 2(%[d])                    \n"
+      "bgtz              %[dst_width], 1b                \n"
+      " addiu            %[d], %[d], 3                   \n"
+    "3:                                                  \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [src_stride] "+r" (src_stride),
+        [d] "+r" (d),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3",
+        "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                           \n"
+      ".set noreorder                                      \n"
+      "repl.ph           $t2, 3                            \n"  // 0x00030003
+
+      ".p2align          2                                 \n"
+    "1:                                                    \n"
+      "lw                $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])    \n"  // |T3|T2|T1|T0|
+      "rotr              $t4, $t0, 8                       \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                       \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t3, $t4, $t2                     \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t2                     \n"  // |T0*3|T3*3|
+      "andi              $t0, $t4, 0xFFFF                  \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                  \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                          \n"
+      "raddu.w.qb        $t1, $t1                          \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "shra_r.w          $t1, $t1, 1                       \n"
+      "preceu.ph.qbr     $t4, $t4                          \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                          \n"  // |0|T2|0|T1|
+      "rotr              $t4, $t4, 16                      \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                      \n"  // |0|T1|0|T2|
+      "addu.ph           $t4, $t4, $t3                     \n"
+      "addu.ph           $t6, $t6, $t5                     \n"
+      "shra_r.ph         $t6, $t6, 2                       \n"
+      "shra_r.ph         $t4, $t4, 2                       \n"
+      "addu.ph           $t6, $t6, $t4                     \n"
+      "addiu             %[src_ptr], %[src_ptr], 4         \n"
+      "shra_r.ph         $t6, $t6, 1                       \n"
+      "addu              $t0, $t0, $t1                     \n"
+      "addiu             %[dst_width], %[dst_width], -3    \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "srl               $t1, $t6, 16                      \n"
+      "sb                $t1, 0(%[d])                      \n"
+      "sb                $t0, 1(%[d])                      \n"
+      "sb                $t6, 2(%[d])                      \n"
+      "bgtz              %[dst_width], 1b                  \n"
+      " addiu            %[d], %[d], 3                     \n"
+    "3:                                                    \n"
+      ".set pop                                            \n"
+      : [src_ptr] "+r" (src_ptr),
+        [src_stride] "+r" (src_stride),
+        [d] "+r" (d),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3",
+        "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                     \n"
+      ".set noreorder                                \n"
+
+      ".p2align   2                                  \n"
+    "1:                                              \n"
+      "lw         $t0, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw         $t1, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw         $t2, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw         $t3, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw         $t4, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw         $t5, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw         $t6, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw         $t7, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "wsbh       $t0, $t0                           \n"  // |2|3|0|1|
+      "wsbh       $t6, $t6                           \n"  // |26|27|24|25|
+      "srl        $t0, $t0, 8                        \n"  // |X|2|3|0|
+      "srl        $t3, $t3, 16                       \n"  // |X|X|15|14|
+      "srl        $t5, $t5, 16                       \n"  // |X|X|23|22|
+      "srl        $t7, $t7, 16                       \n"  // |X|X|31|30|
+      "ins        $t1, $t2, 24, 8                    \n"  // |8|6|5|4|
+      "ins        $t6, $t5, 0, 8                     \n"  // |26|27|24|22|
+      "ins        $t1, $t0, 0, 16                    \n"  // |8|6|3|0|
+      "ins        $t6, $t7, 24, 8                    \n"  // |30|27|24|22|
+      "prepend    $t2, $t3, 24                       \n"  // |X|15|14|11|
+      "ins        $t4, $t4, 16, 8                    \n"  // |19|16|17|X|
+      "ins        $t4, $t2, 0, 16                    \n"  // |19|16|14|11|
+      "addiu      %[src_ptr], %[src_ptr], 32         \n"
+      "addiu      %[dst_width], %[dst_width], -12    \n"
+      "addiu      $t8,%[dst_width], -12              \n"
+      "sw         $t1, 0(%[dst])                     \n"
+      "sw         $t4, 4(%[dst])                     \n"
+      "sw         $t6, 8(%[dst])                     \n"
+      "bgez       $t8, 1b                            \n"
+      " addiu     %[dst], %[dst], 12                 \n"
+      ".set pop                                      \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4",
+        "t5", "t6", "t7", "t8"
+  );
+}
+
+void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* t = src_ptr + stride;
+  const int c = 0x2AAA;
+
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      ".p2align        2                                 \n"
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[t])                      \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[t])                      \n"  // |T7|T6|T5|T4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t4, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "packrl.ph       $t5, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t4, $t4                          \n"  // S7+S6+T7+T6
+      "raddu.w.qb      $t5, $t5                          \n"  // T5+T4+S5+S4
+      "precrq.qb.ph    $t6, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t6, $t6, $t6                     \n"  // |S3|T3|S3|T3|
+      "srl             $t4, $t4, 2                       \n"  // t4 / 4
+      "srl             $t6, $t6, 16                      \n"  // |0|0|S3|T3|
+      "raddu.w.qb      $t6, $t6                          \n"  // 0+0+S3+T3
+      "addu            $t6, $t5, $t6                     \n"
+      "mul             $t6, $t6, %[c]                    \n"  // t6 * 0x2AAA
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "raddu.w.qb      $t0, $t0                          \n"  // S2+S1+S0+0
+      "raddu.w.qb      $t2, $t2                          \n"  // T2+T1+T0+0
+      "addu            $t0, $t0, $t2                     \n"
+      "mul             $t0, $t0, %[c]                    \n"  // t0 * 0x2AAA
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[t], %[t], 8                     \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t4, -1(%[dst_ptr])               \n"
+      "sb              $t6, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst_ptr] "+r" (dst_ptr),
+        [t] "+r" (t),
+        [dst_width] "+r" (dst_width)
+      : [c] "r" (c)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  stride += stride;
+  const uint8* s2 = src_ptr + stride;
+  const int c1 = 0x1C71;
+  const int c2 = 0x2AAA;
+
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      ".p2align        2                                 \n"
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[s1])                     \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[s1])                     \n"  // |T7|T6|T5|T4|
+      "lw              $t4, 0(%[s2])                     \n"  // |R3|R2|R1|R0|
+      "lw              $t5, 4(%[s2])                     \n"  // |R7|R6|R5|R4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t6, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "raddu.w.qb      $t6, $t6                          \n"  // S7+S6+T7+T6
+      "packrl.ph       $t7, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t7, $t7                          \n"  // T5+T4+S5+S4
+      "sll             $t8, $t5, 16                      \n"  // |R5|R4|0|0|
+      "raddu.w.qb      $t8, $t8                          \n"  // R5+R4
+      "addu            $t7, $t7, $t8                     \n"
+      "srl             $t8, $t5, 16                      \n"  // |0|0|R7|R6|
+      "raddu.w.qb      $t8, $t8                          \n"  // R7 + R6
+      "addu            $t6, $t6, $t8                     \n"
+      "mul             $t6, $t6, %[c2]                   \n"  // t6 * 0x2AAA
+      "precrq.qb.ph    $t8, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t8, $t8, $t4                     \n"  // |S3|T3|R3|R1|
+      "srl             $t8, $t8, 8                       \n"  // |0|S3|T3|R3|
+      "raddu.w.qb      $t8, $t8                          \n"  // S3 + T3 + R3
+      "addu            $t7, $t7, $t8                     \n"
+      "mul             $t7, $t7, %[c1]                   \n"  // t7 * 0x1C71
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "sll             $t4, $t4, 8                       \n"  // |R2|R1|R0|0|
+      "raddu.w.qb      $t0, $t0                          \n"
+      "raddu.w.qb      $t2, $t2                          \n"
+      "raddu.w.qb      $t4, $t4                          \n"
+      "addu            $t0, $t0, $t2                     \n"
+      "addu            $t0, $t0, $t4                     \n"
+      "mul             $t0, $t0, %[c1]                   \n"  // t0 * 0x1C71
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[s1], %[s1], 8                   \n"
+      "addiu           %[s2], %[s2], 8                   \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t7, $t7, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t6, -1(%[dst_ptr])               \n"
+      "sb              $t7, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst_ptr] "+r" (dst_ptr),
+        [s1] "+r" (s1),
+        [s2] "+r" (s2),
+        [dst_width] "+r" (dst_width)
+      : [c1] "r" (c1), [c2] "r" (c2)
+      : "t0", "t1", "t2", "t3", "t4",
+        "t5", "t6", "t7", "t8"
+  );
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/third_party/libyuv/source/scale_neon.cc b/third_party/libyuv/source/scale_neon.cc
new file mode 100644
index 0000000..704cfd2
--- /dev/null
+++ b/third_party/libyuv/source/scale_neon.cc
@@ -0,0 +1,684 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+
+// NEON downscalers with interpolation.
+// Provided by Fritz Koenig
+
+// Read 32x1 throw away even pixels, and write 16x1.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    "vld2.8     {q0, q1}, [%0]!                \n"
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1"              // Clobber List
+  );
+}
+
+// Read 32x2 average down and write 16x1.
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 1 and post inc
+    "vld1.8     {q2, q3}, [%1]!                \n"  // load row 2 and post inc
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
+    "vpaddl.u8  q1, q1                         \n"
+    "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent + row1
+    "vpadal.u8  q1, q3                         \n"
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "q0", "q1", "q2", "q3"     // Clobber List
+  );
+}
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n" // src line 0
+    "subs       %2, %2, #8                     \n" // 8 processed per loop
+    "vst1.8     {d2}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1", "memory", "cc"
+  );
+}
+
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "add        r4, %0, %3                     \n"
+    "add        r5, r4, %3                     \n"
+    "add        %3, r5, %3                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"   // load up 16x4
+    "vld1.8     {q1}, [r4]!                    \n"
+    "vld1.8     {q2}, [r5]!                    \n"
+    "vld1.8     {q3}, [%3]!                    \n"
+    "subs       %2, %2, #4                     \n"
+    "vpaddl.u8  q0, q0                         \n"
+    "vpadal.u8  q0, q1                         \n"
+    "vpadal.u8  q0, q2                         \n"
+    "vpadal.u8  q0, q3                         \n"
+    "vpaddl.u16 q0, q0                         \n"
+    "vrshrn.u32 d0, q0, #4                     \n"   // divide by 16 w/rounding
+    "vmovn.u16  d0, q0                         \n"
+    "vst1.32    {d0[0]}, [%1]!                 \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  : "r"(src_stride)         // %3
+  : "r4", "r5", "q0", "q1", "q2", "q3", "memory", "cc"
+  );
+}
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+// to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "subs       %2, %2, #24                  \n"
+    "vmov       d2, d3                       \n" // order d0, d1, d2
+    "vst3.8     {d0, d1, d2}, [%1]!          \n"
+    "bgt        1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "d0", "d1", "d2", "d3", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vmov.u8    d24, #3                        \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
+    "subs         %2, %2, #24                  \n"
+
+    // filter src line 0 with src line 1
+    // expand chars to shorts to allow for room
+    // when adding lines together
+    "vmovl.u8     q8, d4                       \n"
+    "vmovl.u8     q9, d5                       \n"
+    "vmovl.u8     q10, d6                      \n"
+    "vmovl.u8     q11, d7                      \n"
+
+    // 3 * line_0 + line_1
+    "vmlal.u8     q8, d0, d24                  \n"
+    "vmlal.u8     q9, d1, d24                  \n"
+    "vmlal.u8     q10, d2, d24                 \n"
+    "vmlal.u8     q11, d3, d24                 \n"
+
+    // (3 * line_0 + line_1) >> 2
+    "vqrshrn.u16  d0, q8, #2                   \n"
+    "vqrshrn.u16  d1, q9, #2                   \n"
+    "vqrshrn.u16  d2, q10, #2                  \n"
+    "vqrshrn.u16  d3, q11, #2                  \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "vmovl.u8     q8, d1                       \n"
+    "vmlal.u8     q8, d0, d24                  \n"
+    "vqrshrn.u16  d0, q8, #2                   \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "vrhadd.u8    d1, d1, d2                   \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "vmovl.u8     q8, d2                       \n"
+    "vmlal.u8     q8, d3, d24                  \n"
+    "vqrshrn.u16  d2, q8, #2                   \n"
+
+    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vmov.u8    d24, #3                        \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
+    "subs         %2, %2, #24                  \n"
+    // average src line 0 with src line 1
+    "vrhadd.u8    q0, q0, q2                   \n"
+    "vrhadd.u8    q1, q1, q3                   \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "vmovl.u8     q3, d1                       \n"
+    "vmlal.u8     q3, d0, d24                  \n"
+    "vqrshrn.u16  d0, q3, #2                   \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "vrhadd.u8    d1, d1, d2                   \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "vmovl.u8     q3, d2                       \n"
+    "vmlal.u8     q3, d3, d24                  \n"
+    "vqrshrn.u16  d2, q3, #2                   \n"
+
+    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
+  );
+}
+
+#define HAS_SCALEROWDOWN38_NEON
+static uvec8 kShuf38 =
+  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
+static uvec8 kShuf38_2 =
+  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
+static vec16 kMult38_Div6 =
+  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
+    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
+static vec16 kMult38_Div9 =
+  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
+    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vld1.8     {q3}, [%3]                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"
+    "subs       %2, %2, #12                    \n"
+    "vtbl.u8    d4, {d0, d1, d2, d3}, d6       \n"
+    "vtbl.u8    d5, {d0, d1, d2, d3}, d7       \n"
+    "vst1.8     {d4}, [%1]!                    \n"
+    "vst1.32    {d5[0]}, [%1]!                 \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  : "r"(&kShuf38)           // %3
+  : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
+  );
+}
+
+// 32x3 -> 12x1
+void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vld1.16    {q13}, [%4]                    \n"
+    "vld1.8     {q14}, [%5]                    \n"
+    "vld1.8     {q15}, [%6]                    \n"
+    "add        r4, %0, %3, lsl #1             \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+
+    // d0 = 00 40 01 41 02 42 03 43
+    // d1 = 10 50 11 51 12 52 13 53
+    // d2 = 20 60 21 61 22 62 23 63
+    // d3 = 30 70 31 71 32 72 33 73
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    "vld4.8       {d16, d17, d18, d19}, [r4]!  \n"
+    "subs         %2, %2, #12                  \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // d0 = 00 10 01 11 02 12 03 13
+    // d1 = 40 50 41 51 42 52 43 53
+    "vtrn.u8      d0, d1                       \n"
+    "vtrn.u8      d4, d5                       \n"
+    "vtrn.u8      d16, d17                     \n"
+
+    // d2 = 20 30 21 31 22 32 23 33
+    // d3 = 60 70 61 71 62 72 63 73
+    "vtrn.u8      d2, d3                       \n"
+    "vtrn.u8      d6, d7                       \n"
+    "vtrn.u8      d18, d19                     \n"
+
+    // d0 = 00+10 01+11 02+12 03+13
+    // d2 = 40+50 41+51 42+52 43+53
+    "vpaddl.u8    q0, q0                       \n"
+    "vpaddl.u8    q2, q2                       \n"
+    "vpaddl.u8    q8, q8                       \n"
+
+    // d3 = 60+70 61+71 62+72 63+73
+    "vpaddl.u8    d3, d3                       \n"
+    "vpaddl.u8    d7, d7                       \n"
+    "vpaddl.u8    d19, d19                     \n"
+
+    // combine source lines
+    "vadd.u16     q0, q2                       \n"
+    "vadd.u16     q0, q8                       \n"
+    "vadd.u16     d4, d3, d7                   \n"
+    "vadd.u16     d4, d19                      \n"
+
+    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
+    //             + s[6 + st * 1] + s[7 + st * 1]
+    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
+    "vqrdmulh.s16 q2, q2, q13                  \n"
+    "vmovn.u16    d4, q2                       \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "vmovl.u8     q1, d2                       \n"
+    "vmovl.u8     q3, d6                       \n"
+    "vmovl.u8     q9, d18                      \n"
+
+    // combine source lines
+    "vadd.u16     q1, q3                       \n"
+    "vadd.u16     q1, q9                       \n"
+
+    // d4 = xx 20 xx 30 xx 22 xx 32
+    // d5 = xx 21 xx 31 xx 23 xx 33
+    "vtrn.u32     d2, d3                       \n"
+
+    // d4 = xx 20 xx 21 xx 22 xx 23
+    // d5 = xx 30 xx 31 xx 32 xx 33
+    "vtrn.u16     d2, d3                       \n"
+
+    // 0+1+2, 3+4+5
+    "vadd.u16     q0, q1                       \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "vqrdmulh.s16 q0, q0, q15                  \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "vmov.u8      d2, d4                       \n"
+
+    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+    "vst1.8       {d3}, [%1]!                  \n"
+    "vst1.32      {d4[0]}, [%1]!               \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  : "r"(&kMult38_Div6),     // %4
+    "r"(&kShuf38_2),        // %5
+    "r"(&kMult38_Div9)      // %6
+  : "r4", "q0", "q1", "q2", "q3", "q8", "q9",
+    "q13", "q14", "q15", "memory", "cc"
+  );
+}
+
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vld1.16    {q13}, [%4]                    \n"
+    "vld1.8     {q14}, [%5]                    \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+
+    // d0 = 00 40 01 41 02 42 03 43
+    // d1 = 10 50 11 51 12 52 13 53
+    // d2 = 20 60 21 61 22 62 23 63
+    // d3 = 30 70 31 71 32 72 33 73
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    "subs         %2, %2, #12                  \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // d0 = 00 10 01 11 02 12 03 13
+    // d1 = 40 50 41 51 42 52 43 53
+    "vtrn.u8      d0, d1                       \n"
+    "vtrn.u8      d4, d5                       \n"
+
+    // d2 = 20 30 21 31 22 32 23 33
+    // d3 = 60 70 61 71 62 72 63 73
+    "vtrn.u8      d2, d3                       \n"
+    "vtrn.u8      d6, d7                       \n"
+
+    // d0 = 00+10 01+11 02+12 03+13
+    // d2 = 40+50 41+51 42+52 43+53
+    "vpaddl.u8    q0, q0                       \n"
+    "vpaddl.u8    q2, q2                       \n"
+
+    // d3 = 60+70 61+71 62+72 63+73
+    "vpaddl.u8    d3, d3                       \n"
+    "vpaddl.u8    d7, d7                       \n"
+
+    // combine source lines
+    "vadd.u16     q0, q2                       \n"
+    "vadd.u16     d4, d3, d7                   \n"
+
+    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
+    "vqrshrn.u16  d4, q2, #2                   \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "vmovl.u8     q1, d2                       \n"
+    "vmovl.u8     q3, d6                       \n"
+
+    // combine source lines
+    "vadd.u16     q1, q3                       \n"
+
+    // d4 = xx 20 xx 30 xx 22 xx 32
+    // d5 = xx 21 xx 31 xx 23 xx 33
+    "vtrn.u32     d2, d3                       \n"
+
+    // d4 = xx 20 xx 21 xx 22 xx 23
+    // d5 = xx 30 xx 31 xx 32 xx 33
+    "vtrn.u16     d2, d3                       \n"
+
+    // 0+1+2, 3+4+5
+    "vadd.u16     q0, q1                       \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "vqrdmulh.s16 q0, q0, q13                  \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "vmov.u8      d2, d4                       \n"
+
+    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+    "vst1.8       {d3}, [%1]!                  \n"
+    "vst1.32      {d4[0]}, [%1]!               \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),       // %0
+    "+r"(dst_ptr),       // %1
+    "+r"(dst_width),     // %2
+    "+r"(src_stride)     // %3
+  : "r"(&kMult38_Div6),  // %4
+    "r"(&kShuf38_2)      // %5
+  : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc"
+  );
+}
+
+// 16x2 -> 16x1
+void ScaleFilterRows_NEON(uint8* dst_ptr,
+                          const uint8* src_ptr, ptrdiff_t src_stride,
+                          int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp          %4, #0                       \n"
+    "beq          100f                         \n"
+    "add          %2, %1                       \n"
+    "cmp          %4, #64                      \n"
+    "beq          75f                          \n"
+    "cmp          %4, #128                     \n"
+    "beq          50f                          \n"
+    "cmp          %4, #192                     \n"
+    "beq          25f                          \n"
+
+    "vdup.8       d5, %4                       \n"
+    "rsb          %4, #256                     \n"
+    "vdup.8       d4, %4                       \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    "vld1.8       {q0}, [%1]!                  \n"
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vmull.u8     q13, d0, d4                  \n"
+    "vmull.u8     q14, d1, d4                  \n"
+    "vmlal.u8     q13, d2, d5                  \n"
+    "vmlal.u8     q14, d3, d5                  \n"
+    "vrshrn.u16   d0, q13, #8                  \n"
+    "vrshrn.u16   d1, q14, #8                  \n"
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          1b                           \n"
+    "b            99f                          \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    "vld1.8       {q0}, [%1]!                  \n"
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          25b                          \n"
+    "b            99f                          \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    "vld1.8       {q0}, [%1]!                  \n"
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          50b                          \n"
+    "b            99f                          \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    "vld1.8       {q1}, [%1]!                  \n"
+    "vld1.8       {q0}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          75b                          \n"
+    "b            99f                          \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    "vld1.8       {q0}, [%1]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          100b                         \n"
+
+  "99:                                         \n"
+    "vst1.8       {d1[7]}, [%0]                \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc"
+  );
+}
+
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    "vld2.32    {q0, q1}, [%0]!                \n"
+    "vld2.32    {q2, q3}, [%0]!                \n"
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    "vst1.8     {q3}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %1, %0                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
+    "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB pixels.
+    "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
+    "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    "vrshrn.u16 d2, q2, #2                     \n"
+    "vrshrn.u16 d3, q3, #2                     \n"
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %3, lsl #2                \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.32    {d0[0]}, [%0], r12             \n"
+    "vld1.32    {d0[1]}, [%0], r12             \n"
+    "vld1.32    {d1[0]}, [%0], r12             \n"
+    "vld1.32    {d1[1]}, [%0], r12             \n"
+    "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
+    "vst1.8     {q0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(dst_width)    // %2
+  : "r"(src_stepx)     // %3
+  : "memory", "cc", "r12", "q0"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %4, lsl #2                \n"
+    "add        %1, %1, %0                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0}, [%0], r12                \n"  // Read 4 2x2 blocks -> 2x1
+    "vld1.8     {d1}, [%1], r12                \n"
+    "vld1.8     {d2}, [%0], r12                \n"
+    "vld1.8     {d3}, [%1], r12                \n"
+    "vld1.8     {d4}, [%0], r12                \n"
+    "vld1.8     {d5}, [%1], r12                \n"
+    "vld1.8     {d6}, [%0], r12                \n"
+    "vld1.8     {d7}, [%1], r12                \n"
+    "vaddl.u8   q0, d0, d1                     \n"
+    "vaddl.u8   q1, d2, d3                     \n"
+    "vaddl.u8   q2, d4, d5                     \n"
+    "vaddl.u8   q3, d6, d7                     \n"
+    "vswp.8     d1, d2                         \n"  // ab_cd -> ac_bd
+    "vswp.8     d5, d6                         \n"  // ef_gh -> eg_fh
+    "vadd.u16   q0, q0, q1                     \n"  // (a+b)_(c+d)
+    "vadd.u16   q2, q2, q3                     \n"  // (e+f)_(g+h)
+    "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
+    "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
+    "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(src_stride),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(dst_width)    // %3
+  : "r"(src_stepx)     // %4
+  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"
+  );
+}
+
+#endif  // __ARM_NEON__
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/scale_posix.cc b/third_party/libyuv/source/scale_posix.cc
new file mode 100644
index 0000000..18b0810
--- /dev/null
+++ b/third_party/libyuv/source/scale_posix.cc
@@ -0,0 +1,1315 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+// Offsets for source bytes 0 to 9
+static uvec8 kShuf0 =
+  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
+static uvec8 kShuf1 =
+  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf2 =
+  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 0 to 10
+static uvec8 kShuf01 =
+  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+
+// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
+static uvec8 kShuf11 =
+  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf21 =
+  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+
+// Coefficients for source bytes 0 to 10
+static uvec8 kMadd01 =
+  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+
+// Coefficients for source bytes 10 to 21
+static uvec8 kMadd11 =
+  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+
+// Coefficients for source bytes 21 to 31
+static uvec8 kMadd21 =
+  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+
+// Coefficients for source bytes 21 to 31
+static vec16 kRound34 =
+  { 2, 2, 2, 2, 2, 2, 2, 2 };
+
+static uvec8 kShuf38a =
+  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+static uvec8 kShuf38b =
+  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 0,1,2
+static uvec8 kShufAc =
+  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 3,4,5
+static uvec8 kShufAc3 =
+  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x3 and 2x3
+static uvec16 kScaleAc33 =
+  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+
+// Arrange first value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb0 =
+  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+
+// Arrange second value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb1 =
+  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+
+// Arrange third value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb2 =
+  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x2 and 2x2
+static uvec16 kScaleAb2 =
+  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+
+// GCC versions of row functions are verbatim conversions from Visual C.
+// Generated using gcc disassembly on Visual C object file:
+// objdump -D yuvscaler.obj >yuvscaler.txt
+
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10, 0) ",%%xmm1  \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm2)           //  movdqa  (%0,%3,1),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,3,1,xmm3)           //  movdqa  0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x18,%%xmm5                    \n"
+    "pslld     $0x10,%%xmm5                    \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  intptr_t stridex3 = 0;
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0x8,%%xmm7                     \n"
+    "lea       " MEMLEA4(0x00,4,4,2) ",%3      \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    MEMOPREG(movdqa,0x00,0,4,2,xmm2)           //  movdqa  (%0,%4,2),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,4,2,xmm3)           //  movdqa  0x10(%0,%4,2),%%xmm3
+    MEMOPREG(movdqa,0x00,0,3,1,xmm4)           //  movdqa  (%0,%3,1),%%xmm4
+    MEMOPREG(movdqa,0x10,0,3,1,xmm5)           //  movdqa  0x10(%0,%3,1),%%xmm5
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm4,%%xmm2                   \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm5,%%xmm3                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "pand      %%xmm7,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(dst_width),   // %2
+    "+r"(stridex3)     // %3
+  : "r"((intptr_t)(src_stride))    // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm7"
+#endif
+  );
+}
+
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm3                       \n"
+    "movdqa    %1,%%xmm4                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kShuf0),  // %0
+    "m"(kShuf1),  // %1
+    "m"(kShuf2)   // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm2   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "palignr   $0x8,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm3,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "movq      %%xmm1," MEMACCESS2(0x8,1) "    \n"
+    "movq      %%xmm2," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"  // kShuf01
+    "movdqa    %1,%%xmm3                       \n"  // kShuf11
+    "movdqa    %2,%%xmm4                       \n"  // kShuf21
+  :
+  : "m"(kShuf01),  // %0
+    "m"(kShuf11),  // %1
+    "m"(kShuf21)   // %2
+  );
+  asm volatile (
+    "movdqa    %0,%%xmm5                       \n"  // kMadd01
+    "movdqa    %1,%%xmm0                       \n"  // kMadd11
+    "movdqa    %2,%%xmm1                       \n"  // kRound34
+  :
+  : "m"(kMadd01),  // %0
+    "m"(kMadd11),  // %1
+    "m"(kRound34)  // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm7)           //  movdqa  (%0,%3),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
+    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm0,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,3,1,xmm7)           //  movdqa  0x10(%0,%3),%%xmm7
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm4,%%xmm6                   \n"
+    "pmaddubsw %4,%%xmm6                       \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  : "r"((intptr_t)(src_stride)),  // %3
+    "m"(kMadd21)     // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"  // kShuf01
+    "movdqa    %1,%%xmm3                       \n"  // kShuf11
+    "movdqa    %2,%%xmm4                       \n"  // kShuf21
+  :
+  : "m"(kShuf01),  // %0
+    "m"(kShuf11),  // %1
+    "m"(kShuf21)   // %2
+  );
+  asm volatile (
+    "movdqa    %0,%%xmm5                       \n"  // kMadd01
+    "movdqa    %1,%%xmm0                       \n"  // kMadd11
+    "movdqa    %2,%%xmm1                       \n"  // kRound34
+  :
+  : "m"(kMadd01),  // %0
+    "m"(kMadd11),  // %1
+    "m"(kRound34)  // %2
+  );
+
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm7)           //  movdqa  (%0,%3,1),%%xmm7
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
+    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3,1),%%xmm7
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm0,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    MEMOPREG(movdqa,0x10,0,3,1,xmm7)           //  movdqa  0x10(%0,%3,1),%%xmm7
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm4,%%xmm6                   \n"
+    "pmaddubsw %4,%%xmm6                       \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+    : "+r"(src_ptr),   // %0
+      "+r"(dst_ptr),   // %1
+      "+r"(dst_width)  // %2
+    : "r"((intptr_t)(src_stride)),  // %3
+      "m"(kMadd21)     // %4
+    : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "movhlps   %%xmm0,%%xmm1                   \n"
+    "movd      %%xmm1," MEMACCESS2(0x8,1) "    \n"
+    "lea       " MEMLEA(0xc,1) ",%1            \n"
+    "sub       $0xc,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  : "m"(kShuf38a),   // %3
+    "m"(kShuf38b)    // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+      , "xmm0", "xmm1", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm4                       \n"
+    "movdqa    %3,%%xmm5                       \n"
+  :
+  : "m"(kShufAb0),   // %0
+    "m"(kShufAb1),   // %1
+    "m"(kShufAb2),   // %2
+    "m"(kScaleAb2)   // %3
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(pavgb,0x00,0,3,1,xmm0)            //  pavgb   (%0,%3,1),%%xmm0
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pshufb    %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "paddusw   %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "paddusw   %%xmm0,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "sub       $0x6,%2                         \n"
+    "movd      %%xmm1," MEMACCESS(1) "         \n"
+    "psrlq     $0x10,%%xmm1                    \n"
+    "movd      %%xmm1," MEMACCESS2(0x2,1) "    \n"
+    "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(dst_width)    // %2
+  : "r"((intptr_t)(src_stride))  // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm4                       \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+  :
+  : "m"(kShufAc),    // %0
+    "m"(kShufAc3),   // %1
+    "m"(kScaleAc33)  // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm6)           //  movdqa  (%0,%3,1),%%xmm6
+    "movhlps   %%xmm0,%%xmm1                   \n"
+    "movhlps   %%xmm6,%%xmm7                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm6                   \n"
+    "punpcklbw %%xmm5,%%xmm7                   \n"
+    "paddusw   %%xmm6,%%xmm0                   \n"
+    "paddusw   %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqa,0x00,0,3,2,xmm6)           //  movdqa  (%0,%3,2),%%xmm6
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movhlps   %%xmm6,%%xmm7                   \n"
+    "punpcklbw %%xmm5,%%xmm6                   \n"
+    "punpcklbw %%xmm5,%%xmm7                   \n"
+    "paddusw   %%xmm6,%%xmm0                   \n"
+    "paddusw   %%xmm7,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "psrldq    $0x2,%%xmm0                     \n"
+    "paddusw   %%xmm0,%%xmm6                   \n"
+    "psrldq    $0x2,%%xmm0                     \n"
+    "paddusw   %%xmm0,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "movdqa    %%xmm1,%%xmm7                   \n"
+    "psrldq    $0x2,%%xmm1                     \n"
+    "paddusw   %%xmm1,%%xmm7                   \n"
+    "psrldq    $0x2,%%xmm1                     \n"
+    "paddusw   %%xmm1,%%xmm7                   \n"
+    "pshufb    %%xmm3,%%xmm7                   \n"
+    "paddusw   %%xmm7,%%xmm6                   \n"
+    "pmulhuw   %%xmm4,%%xmm6                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "sub       $0x6,%2                         \n"
+    "movd      %%xmm6," MEMACCESS(1) "         \n"
+    "psrlq     $0x10,%%xmm6                    \n"
+    "movd      %%xmm6," MEMACCESS2(0x2,1) "    \n"
+    "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width, int src_height) {
+  int tmp_height = 0;
+  intptr_t tmp_src = 0;
+  asm volatile (
+    "pxor      %%xmm4,%%xmm4                   \n"
+    "sub       $0x1,%5                         \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "mov       %0,%3                           \n"
+    "add       %6,%0                           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "mov       %5,%2                           \n"
+    "test      %2,%2                           \n"
+    "je        3f                              \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "add       %6,%0                           \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "paddusw   %%xmm2,%%xmm0                   \n"
+    "paddusw   %%xmm3,%%xmm1                   \n"
+    "sub       $0x1,%2                         \n"
+    "jg        2b                              \n"
+
+    LABELALIGN
+  "3:                                          \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x10,3) ",%0           \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x10,%4                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(tmp_height),  // %2
+    "+r"(tmp_src),     // %3
+    "+r"(src_width),   // %4
+    "+rm"(src_height)  // %5
+  : "rm"((intptr_t)(src_stride))  // %6
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+
+// Bilinear column filtering. SSSE3 version.
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0, temp_pixel = 0;
+  asm volatile (
+    "movd      %6,%%xmm2                       \n"
+    "movd      %7,%%xmm3                       \n"
+    "movl      $0x04040000,%k2                 \n"
+    "movd      %k2,%%xmm5                      \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x9,%%xmm6                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "subl      $0x2,%5                         \n"
+    "jl        29f                             \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "punpckldq %%xmm0,%%xmm2                   \n"
+    "punpckldq %%xmm3,%%xmm3                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
+    "movd      %k2,%%xmm0                      \n"
+    "psrlw     $0x9,%%xmm1                     \n"
+    BUNDLEALIGN
+    MEMOPARG(movzwl,0x00,1,4,1,k2)             //  movzwl  (%1,%4,1),%k2
+    "movd      %k2,%%xmm4                      \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "punpcklwd %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm1                   \n"
+    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0,%k2                      \n"
+    "mov       %w2," MEMACCESS(0) "            \n"
+    "lea       " MEMLEA(0x2,0) ",%0            \n"
+    "sub       $0x2,%5                         \n"
+    "jge       2b                              \n"
+
+    LABELALIGN
+  "29:                                         \n"
+    "addl      $0x1,%5                         \n"
+    "jl        99f                             \n"
+    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
+    "movd      %k2,%%xmm0                      \n"
+    "psrlw     $0x9,%%xmm2                     \n"
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "pxor      %%xmm6,%%xmm2                   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0,%k2                      \n"
+    "mov       %b2," MEMACCESS(0) "            \n"
+  "99:                                         \n"
+  : "+r"(dst_ptr),     // %0
+    "+r"(src_ptr),     // %1
+    "+a"(temp_pixel),  // %2
+    "+r"(x0),          // %3
+    "+r"(x1),          // %4
+    "+rm"(dst_width)   // %5
+  : "rm"(x),           // %6
+    "rm"(dx)           // %7
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "sub       $0x20,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "jg        1b                              \n"
+
+  : "+r"(dst_ptr),     // %0
+    "+r"(src_ptr),     // %1
+    "+r"(dst_width)    // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "shufps    $0xdd,%%xmm1,%%xmm0             \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(dst_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(dst_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x00,0,3,1,xmm2)           //  movdqa   (%0,%3,1),%%xmm2
+    MEMOPREG(movdqa,0x10,0,3,1,xmm3)           //  movdqa   0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: dst_argb 16 byte aligned.
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
+  intptr_t src_stepx_x12 = 0;
+  asm volatile (
+    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
+    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movd      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,2,xmm2)             //  movd      (%0,%1,2),%%xmm2
+    MEMOPREG(movd,0x00,0,4,1,xmm3)             //  movd      (%0,%4,1),%%xmm3
+    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
+    "punpckldq %%xmm3,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm0                  \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),      // %0
+    "+r"(src_stepx_x4),  // %1
+    "+r"(dst_argb),      // %2
+    "+r"(dst_width),     // %3
+    "+r"(src_stepx_x12)  // %4
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+
+// Blends four 2x2 to 4x1.
+// Alignment requirement: dst_argb 16 byte aligned.
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride, int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
+  intptr_t src_stepx_x12 = 0;
+  intptr_t row1 = (intptr_t)(src_stride);
+  asm volatile (
+    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
+    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
+    "lea       " MEMLEA4(0x00,0,5,1) ",%5      \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movhps,0x00,0,1,1,xmm0)           //  movhps    (%0,%1,1),%%xmm0
+    MEMOPREG(movq,0x00,0,1,2,xmm1)             //  movq      (%0,%1,2),%%xmm1
+    BUNDLEALIGN
+    MEMOPREG(movhps,0x00,0,4,1,xmm1)           //  movhps    (%0,%4,1),%%xmm1
+    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
+    "movq      " MEMACCESS(5) ",%%xmm2         \n"
+    BUNDLEALIGN
+    MEMOPREG(movhps,0x00,5,1,1,xmm2)           //  movhps    (%5,%1,1),%%xmm2
+    MEMOPREG(movq,0x00,5,1,2,xmm3)             //  movq      (%5,%1,2),%%xmm3
+    MEMOPREG(movhps,0x00,5,4,1,xmm3)           //  movhps    (%5,%4,1),%%xmm3
+    "lea       " MEMLEA4(0x00,5,1,4) ",%5      \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),       // %0
+    "+r"(src_stepx_x4),   // %1
+    "+r"(dst_argb),       // %2
+    "+rm"(dst_width),     // %3
+    "+r"(src_stepx_x12),  // %4
+    "+r"(row1)            // %5
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0;
+  asm volatile (
+    "movd      %5,%%xmm2                       \n"
+    "movd      %6,%%xmm3                       \n"
+    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    "pshufd    $0x11,%%xmm3,%%xmm0             \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pshufd    $0x5,%%xmm3,%%xmm0              \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pextrw    $0x1,%%xmm2,%k0                 \n"
+    "pextrw    $0x3,%%xmm2,%k1                 \n"
+    "cmp       $0x0,%4                         \n"
+    "jl        99f                             \n"
+    "sub       $0x4,%4                         \n"
+    "jl        49f                             \n"
+
+    LABELALIGN
+  "40:                                         \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
+    "pextrw    $0x5,%%xmm2,%k0                 \n"
+    "pextrw    $0x7,%%xmm2,%k1                 \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm1)             //  movd      (%3,%0,4),%%xmm1
+    MEMOPREG(movd,0x00,3,1,4,xmm4)             //  movd      (%3,%1,4),%%xmm4
+    "pextrw    $0x1,%%xmm2,%k0                 \n"
+    "pextrw    $0x3,%%xmm2,%k1                 \n"
+    "punpckldq %%xmm4,%%xmm1                   \n"
+    "punpcklqdq %%xmm1,%%xmm0                  \n"
+    "sub       $0x4,%4                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "test      $0x2,%4                         \n"
+    "je        29f                             \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
+    "pextrw    $0x5,%%xmm2,%k0                 \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x8,2) ",%2            \n"
+  "29:                                         \n"
+    "test      $0x1,%4                         \n"
+    "je        99f                             \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+  "99:                                         \n"
+  : "+a"(x0),          // %0
+    "+d"(x1),          // %1
+    "+r"(dst_argb),    // %2
+    "+r"(src_argb),    // %3
+    "+r"(dst_width)    // %4
+  : "rm"(x),           // %5
+    "rm"(dx)           // %6
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpckldq %%xmm0,%%xmm0                   \n"
+    "punpckhdq %%xmm1,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "jg        1b                              \n"
+
+  : "+r"(dst_argb),    // %0
+    "+r"(src_argb),    // %1
+    "+r"(dst_width)    // %2
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
+static uvec8 kShuffleColARGB = {
+  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
+  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+};
+
+// Shuffle table for duplicating 2 fractions into 8 bytes each
+static uvec8 kShuffleFractions = {
+  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+};
+
+// Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0;
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm5                       \n"
+  :
+  : "m"(kShuffleColARGB),  // %0
+    "m"(kShuffleFractions)  // %1
+  );
+
+  asm volatile (
+    "movd      %5,%%xmm2                       \n"
+    "movd      %6,%%xmm3                       \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x9,%%xmm6                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "sub       $0x2,%2                         \n"
+    "jl        29f                             \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "punpckldq %%xmm0,%%xmm2                   \n"
+    "punpckldq %%xmm3,%%xmm3                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
+    "psrlw     $0x9,%%xmm1                     \n"
+    BUNDLEALIGN
+    MEMOPREG(movhps,0x00,1,4,4,xmm0)           //  movhps    (%1,%4,4),%%xmm0
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm1                   \n"
+    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(0) "         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x2,%2                         \n"
+    "jge       2b                              \n"
+
+    LABELALIGN
+  "29:                                         \n"
+    "add       $0x1,%2                         \n"
+    "jl        99f                             \n"
+    "psrlw     $0x9,%%xmm2                     \n"
+    BUNDLEALIGN
+    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm2                   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(0) "         \n"
+
+    LABELALIGN
+  "99:                                         \n"
+  : "+r"(dst_argb),    // %0
+    "+r"(src_argb),    // %1
+    "+rm"(dst_width),  // %2
+    "+r"(x0),          // %3
+    "+r"(x1)           // %4
+  : "rm"(x),           // %5
+    "rm"(dx)           // %6
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_X86(int num, int div) {
+  asm volatile (
+    "cdq                                       \n"
+    "shld      $0x10,%%eax,%%edx               \n"
+    "shl       $0x10,%%eax                     \n"
+    "idiv      %1                              \n"
+    "mov       %0, %%eax                       \n"
+    : "+a"(num)  // %0
+    : "c"(div)   // %1
+    : "memory", "cc", "edx"
+  );
+  return num;
+}
+
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
+int FixedDiv1_X86(int num, int div) {
+  asm volatile (
+    "cdq                                       \n"
+    "shld      $0x10,%%eax,%%edx               \n"
+    "shl       $0x10,%%eax                     \n"
+    "sub       $0x10001,%%eax                  \n"
+    "sbb       $0x0,%%edx                      \n"
+    "sub       $0x1,%1                         \n"
+    "idiv      %1                              \n"
+    "mov       %0, %%eax                       \n"
+    : "+a"(num)  // %0
+    : "c"(div)   // %1
+    : "memory", "cc", "edx"
+  );
+  return num;
+}
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/libyuv/source/scale_win.cc b/third_party/libyuv/source/scale_win.cc
new file mode 100644
index 0000000..bd5cca8
--- /dev/null
+++ b/third_party/libyuv/source/scale_win.cc
@@ -0,0 +1,1320 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/libyuv/include/libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+// Offsets for source bytes 0 to 9
+static uvec8 kShuf0 =
+  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
+static uvec8 kShuf1 =
+  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf2 =
+  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 0 to 10
+static uvec8 kShuf01 =
+  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+
+// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
+static uvec8 kShuf11 =
+  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf21 =
+  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+
+// Coefficients for source bytes 0 to 10
+static uvec8 kMadd01 =
+  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+
+// Coefficients for source bytes 10 to 21
+static uvec8 kMadd11 =
+  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+
+// Coefficients for source bytes 21 to 31
+static uvec8 kMadd21 =
+  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+
+// Coefficients for source bytes 21 to 31
+static vec16 kRound34 =
+  { 2, 2, 2, 2, 2, 2, 2, 2 };
+
+static uvec8 kShuf38a =
+  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+static uvec8 kShuf38b =
+  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 0,1,2
+static uvec8 kShufAc =
+  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 3,4,5
+static uvec8 kShufAc3 =
+  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x3 and 2x3
+static uvec16 kScaleAc33 =
+  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+
+// Arrange first value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb0 =
+  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+
+// Arrange second value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb1 =
+  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+
+// Arrange third value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb2 =
+  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x2 and 2x2
+static uvec16 kScaleAb2 =
+  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+
+// Reads 32 pixels, throws half away and writes 16 pixels.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8               // isolate odd pixels.
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x1 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 32 pixels, throws half away and writes 16 pixels.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8               // isolate odd pixels.
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x1 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Point samples 32 pixels to 8 pixels.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff0000
+    psrld      xmm5, 24
+    pslld      xmm5, 16
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm0, 8
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x4 rectangle to 8x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_ptr
+    mov        esi, [esp + 8 + 8]    // src_stride
+    mov        edx, [esp + 8 + 12]   // dst_ptr
+    mov        ecx, [esp + 8 + 16]   // dst_width
+    lea        edi, [esi + esi * 2]  // src_stride * 3
+    pcmpeqb    xmm7, xmm7            // generate mask 0x00ff00ff
+    psrlw      xmm7, 8
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, [eax + esi * 2]
+    movdqa     xmm3, [eax + esi * 2 + 16]
+    movdqa     xmm4, [eax + edi]
+    movdqa     xmm5, [eax + edi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm2, xmm4
+    pavgb      xmm3, xmm5
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm7
+    pand       xmm3, xmm7
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    movdqa     xmm2, xmm0            // average columns (16 to 8 pixels)
+    psrlw      xmm0, 8
+    pand       xmm2, xmm7
+    pavgw      xmm0, xmm2
+    packuswb   xmm0, xmm0
+
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Point samples 32 pixels to 24 pixels.
+// Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
+// Then shuffled to do the scaling.
+
+// Note that movdqa+palign may be better than movdqu.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    movdqa     xmm3, kShuf0
+    movdqa     xmm4, kShuf1
+    movdqa     xmm5, kShuf2
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm1
+    palignr    xmm1, xmm0, 8
+    pshufb     xmm0, xmm3
+    pshufb     xmm1, xmm4
+    pshufb     xmm2, xmm5
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + 8], xmm1
+    movq       qword ptr [edx + 16], xmm2
+    lea        edx, [edx + 24]
+    sub        ecx, 24
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 24x1
+// Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
+// Then shuffled to do the scaling.
+
+// Register usage:
+// xmm0 src_row 0
+// xmm1 src_row 1
+// xmm2 shuf 0
+// xmm3 shuf 1
+// xmm4 shuf 2
+// xmm5 madd 0
+// xmm6 madd 1
+// xmm7 kRound34
+
+// Note that movdqa+palign may be better than movdqu.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShuf01
+    movdqa     xmm3, kShuf11
+    movdqa     xmm4, kShuf21
+    movdqa     xmm5, kMadd01
+    movdqa     xmm6, kMadd11
+    movdqa     xmm7, kRound34
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]           // pixels 0..7
+    movdqa     xmm1, [eax + esi]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm2
+    pmaddubsw  xmm0, xmm5
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm1, [eax + esi + 8]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm3
+    pmaddubsw  xmm0, xmm6
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 8], xmm0
+    movdqa     xmm0, [eax + 16]      // pixels 16..23
+    movdqa     xmm1, [eax + esi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm4
+    movdqa     xmm1, kMadd21
+    pmaddubsw  xmm0, xmm1
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    sub        ecx, 24
+    movq       qword ptr [edx + 16], xmm0
+    lea        edx, [edx + 24]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Note that movdqa+palign may be better than movdqu.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShuf01
+    movdqa     xmm3, kShuf11
+    movdqa     xmm4, kShuf21
+    movdqa     xmm5, kMadd01
+    movdqa     xmm6, kMadd11
+    movdqa     xmm7, kRound34
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]           // pixels 0..7
+    movdqa     xmm1, [eax + esi]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm2
+    pmaddubsw  xmm0, xmm5
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm1, [eax + esi + 8]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm3
+    pmaddubsw  xmm0, xmm6
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 8], xmm0
+    movdqa     xmm0, [eax + 16]      // pixels 16..23
+    movdqa     xmm1, [eax + esi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm4
+    movdqa     xmm1, kMadd21
+    pmaddubsw  xmm0, xmm1
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    sub        ecx, 24
+    movq       qword ptr [edx + 16], xmm0
+    lea        edx, [edx+24]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 3/8 point sampler
+
+// Scale 32 pixels to 12
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    movdqa     xmm4, kShuf38a
+    movdqa     xmm5, kShuf38b
+
+    align      4
+  xloop:
+    movdqa     xmm0, [eax]           // 16 pixels -> 0,1,2,3,4,5
+    movdqa     xmm1, [eax + 16]      // 16 pixels -> 6,7,8,9,10,11
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm4
+    pshufb     xmm1, xmm5
+    paddusb    xmm0, xmm1
+
+    sub        ecx, 12
+    movq       qword ptr [edx], xmm0  // write 12 pixels
+    movhlps    xmm1, xmm0
+    movd       [edx + 8], xmm1
+    lea        edx, [edx + 12]
+    jg         xloop
+
+    ret
+  }
+}
+
+// Scale 16x3 pixels to 6x1 with interpolation
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShufAc
+    movdqa     xmm3, kShufAc3
+    movdqa     xmm4, kScaleAc33
+    pxor       xmm5, xmm5
+
+    align      4
+  xloop:
+    movdqa     xmm0, [eax]           // sum up 3 rows into xmm0/1
+    movdqa     xmm6, [eax + esi]
+    movhlps    xmm1, xmm0
+    movhlps    xmm7, xmm6
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm6, xmm5
+    punpcklbw  xmm7, xmm5
+    paddusw    xmm0, xmm6
+    paddusw    xmm1, xmm7
+    movdqa     xmm6, [eax + esi * 2]
+    lea        eax, [eax + 16]
+    movhlps    xmm7, xmm6
+    punpcklbw  xmm6, xmm5
+    punpcklbw  xmm7, xmm5
+    paddusw    xmm0, xmm6
+    paddusw    xmm1, xmm7
+
+    movdqa     xmm6, xmm0            // 8 pixels -> 0,1,2 of xmm6
+    psrldq     xmm0, 2
+    paddusw    xmm6, xmm0
+    psrldq     xmm0, 2
+    paddusw    xmm6, xmm0
+    pshufb     xmm6, xmm2
+
+    movdqa     xmm7, xmm1            // 8 pixels -> 3,4,5 of xmm6
+    psrldq     xmm1, 2
+    paddusw    xmm7, xmm1
+    psrldq     xmm1, 2
+    paddusw    xmm7, xmm1
+    pshufb     xmm7, xmm3
+    paddusw    xmm6, xmm7
+
+    pmulhuw    xmm6, xmm4            // divide by 9,9,6, 9,9,6
+    packuswb   xmm6, xmm6
+
+    sub        ecx, 6
+    movd       [edx], xmm6           // write 6 pixels
+    psrlq      xmm6, 16
+    movd       [edx + 2], xmm6
+    lea        edx, [edx + 6]
+    jg         xloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Scale 16x2 pixels to 6x1 with interpolation
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShufAb0
+    movdqa     xmm3, kShufAb1
+    movdqa     xmm4, kShufAb2
+    movdqa     xmm5, kScaleAb2
+
+    align      4
+  xloop:
+    movdqa     xmm0, [eax]           // average 2 rows into xmm0
+    pavgb      xmm0, [eax + esi]
+    lea        eax, [eax + 16]
+
+    movdqa     xmm1, xmm0            // 16 pixels -> 0,1,2,3,4,5 of xmm1
+    pshufb     xmm1, xmm2
+    movdqa     xmm6, xmm0
+    pshufb     xmm6, xmm3
+    paddusw    xmm1, xmm6
+    pshufb     xmm0, xmm4
+    paddusw    xmm1, xmm0
+
+    pmulhuw    xmm1, xmm5            // divide by 3,3,2, 3,3,2
+    packuswb   xmm1, xmm1
+
+    sub        ecx, 6
+    movd       [edx], xmm1           // write 6 pixels
+    psrlq      xmm1, 16
+    movd       [edx + 2], xmm1
+    lea        edx, [edx + 6]
+    jg         xloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 16xN bytes and produces 16 shorts at a time.
+// TODO(fbarchard): Make this handle 4xN bytes for any width ARGB.
+__declspec(naked) __declspec(align(16))
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width,
+                       int src_height) {
+  __asm {
+    push       esi
+    push       edi
+    push       ebx
+    push       ebp
+    mov        esi, [esp + 16 + 4]   // src_ptr
+    mov        edx, [esp + 16 + 8]   // src_stride
+    mov        edi, [esp + 16 + 12]  // dst_ptr
+    mov        ecx, [esp + 16 + 16]  // dst_width
+    mov        ebx, [esp + 16 + 20]  // height
+    pxor       xmm4, xmm4
+    dec        ebx
+
+    align      4
+  xloop:
+    // first row
+    movdqa     xmm0, [esi]
+    lea        eax, [esi + edx]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    lea        esi, [esi + 16]
+    mov        ebp, ebx
+    test       ebp, ebp
+    je         ydone
+
+    // sum remaining rows
+    align      4
+  yloop:
+    movdqa     xmm2, [eax]       // read 16 pixels
+    lea        eax, [eax + edx]  // advance to next row
+    movdqa     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    paddusw    xmm0, xmm2        // sum 16 words
+    paddusw    xmm1, xmm3
+    sub        ebp, 1
+    jg         yloop
+
+    align      4
+  ydone:
+    movdqa     [edi], xmm0
+    movdqa     [edi + 16], xmm1
+    lea        edi, [edi + 32]
+
+    sub        ecx, 16
+    jg         xloop
+
+    pop        ebp
+    pop        ebx
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Bilinear column filtering. SSSE3 version.
+// TODO(fbarchard): Port to Neon
+// TODO(fbarchard): Switch the following:
+//    xor        ebx, ebx
+//    mov        bx, word ptr [esi + eax]  // 2 source x0 pixels
+// To
+//    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+// when drmemory bug fixed.
+// https://code.google.com/p/drmemory/issues/detail?id=1396
+
+__declspec(naked) __declspec(align(16))
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        edi, [esp + 12 + 4]    // dst_ptr
+    mov        esi, [esp + 12 + 8]    // src_ptr
+    mov        ecx, [esp + 12 + 12]   // dst_width
+    movd       xmm2, [esp + 12 + 16]  // x
+    movd       xmm3, [esp + 12 + 20]  // dx
+    mov        eax, 0x04040000      // shuffle to line up fractions with pixel.
+    movd       xmm5, eax
+    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    psrlw      xmm6, 9
+    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    sub        ecx, 2
+    jl         xloop29
+
+    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    paddd      xmm0, xmm3
+    punpckldq  xmm2, xmm0           // x0 x1
+    punpckldq  xmm3, xmm3           // dx dx
+    paddd      xmm3, xmm3           // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+
+    // 2 Pixel loop.
+    align      4
+  xloop2:
+    movdqa     xmm1, xmm2           // x0, x1 fractions.
+    paddd      xmm2, xmm3           // x += dx
+    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+    movd       xmm0, ebx
+    psrlw      xmm1, 9              // 7 bit fractions.
+    movzx      ebx, word ptr [esi + edx]  // 2 source x1 pixels
+    movd       xmm4, ebx
+    pshufb     xmm1, xmm5           // 0011
+    punpcklwd  xmm0, xmm4
+    pxor       xmm1, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1           // 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    psrlw      xmm0, 7              // 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // 8 bits, 2 pixels.
+    movd       ebx, xmm0
+    mov        [edi], bx
+    lea        edi, [edi + 2]
+    sub        ecx, 2               // 2 pixels
+    jge        xloop2
+
+    align      4
+ xloop29:
+
+    add        ecx, 2 - 1
+    jl         xloop99
+
+    // 1 pixel remainder
+    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+    movd       xmm0, ebx
+    psrlw      xmm2, 9              // 7 bit fractions.
+    pshufb     xmm2, xmm5           // 0011
+    pxor       xmm2, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2           // 16 bit
+    psrlw      xmm0, 7              // 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // 8 bits
+    movd       ebx, xmm0
+    mov        [edi], bl
+
+    align      4
+ xloop99:
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// Reads 16 pixels, duplicates them and writes 32 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  __asm {
+    mov        edx, [esp + 4]    // dst_ptr
+    mov        eax, [esp + 8]    // src_ptr
+    mov        ecx, [esp + 12]   // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm0
+    punpckhbw  xmm1, xmm1
+    sub        ecx, 32
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Reads 8 pixels, throws half away and writes 4 even pixels (0, 2, 4, 6)
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_argb
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_argb
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    shufps     xmm0, xmm1, 0xdd
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 8x1 rectangle to 4x1.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_argb
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_argb
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 8x2 rectangle to 4x1.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_argb
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_argb
+    mov        ecx, [esp + 4 + 16]   // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  __asm {
+    push       ebx
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_argb
+                                     // src_stride ignored
+    mov        ebx, [esp + 8 + 12]   // src_stepx
+    mov        edx, [esp + 8 + 16]   // dst_argb
+    mov        ecx, [esp + 8 + 20]   // dst_width
+    lea        ebx, [ebx * 4]
+    lea        edi, [ebx + ebx * 2]
+
+    align      4
+  wloop:
+    movd       xmm0, [eax]
+    movd       xmm1, [eax + ebx]
+    punpckldq  xmm0, xmm1
+    movd       xmm2, [eax + ebx * 2]
+    movd       xmm3, [eax + edi]
+    lea        eax,  [eax + ebx * 4]
+    punpckldq  xmm2, xmm3
+    punpcklqdq xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        edi
+    pop        ebx
+    ret
+  }
+}
+
+// Blends four 2x2 to 4x1.
+// Alignment requirement: dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]    // src_argb
+    mov        esi, [esp + 12 + 8]    // src_stride
+    mov        ebx, [esp + 12 + 12]   // src_stepx
+    mov        edx, [esp + 12 + 16]   // dst_argb
+    mov        ecx, [esp + 12 + 20]   // dst_width
+    lea        esi, [eax + esi]       // row1 pointer
+    lea        ebx, [ebx * 4]
+    lea        edi, [ebx + ebx * 2]
+
+    align      4
+  wloop:
+    movq       xmm0, qword ptr [eax]  // row0 4 pairs
+    movhps     xmm0, qword ptr [eax + ebx]
+    movq       xmm1, qword ptr [eax + ebx * 2]
+    movhps     xmm1, qword ptr [eax + edi]
+    lea        eax,  [eax + ebx * 4]
+    movq       xmm2, qword ptr [esi]  // row1 4 pairs
+    movhps     xmm2, qword ptr [esi + ebx]
+    movq       xmm3, qword ptr [esi + ebx * 2]
+    movhps     xmm3, qword ptr [esi + edi]
+    lea        esi,  [esi + ebx * 4]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// Column scaling unfiltered. SSE2 version.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  __asm {
+    push       edi
+    push       esi
+    mov        edi, [esp + 8 + 4]    // dst_argb
+    mov        esi, [esp + 8 + 8]    // src_argb
+    mov        ecx, [esp + 8 + 12]   // dst_width
+    movd       xmm2, [esp + 8 + 16]  // x
+    movd       xmm3, [esp + 8 + 20]  // dx
+
+    pshufd     xmm2, xmm2, 0         // x0 x0 x0 x0
+    pshufd     xmm0, xmm3, 0x11      // dx  0 dx  0
+    paddd      xmm2, xmm0
+    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 2
+    pshufd     xmm0, xmm3, 0x05      // dx * 2, dx * 2, 0, 0
+    paddd      xmm2, xmm0            // x3 x2 x1 x0
+    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 4
+    pshufd     xmm3, xmm3, 0         // dx * 4, dx * 4, dx * 4, dx * 4
+
+    pextrw     eax, xmm2, 1          // get x0 integer.
+    pextrw     edx, xmm2, 3          // get x1 integer.
+
+    cmp        ecx, 0
+    jle        xloop99
+    sub        ecx, 4
+    jl         xloop49
+
+    // 4 Pixel loop.
+    align      4
+ xloop4:
+    movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
+    movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
+    pextrw     eax, xmm2, 5           // get x2 integer.
+    pextrw     edx, xmm2, 7           // get x3 integer.
+    paddd      xmm2, xmm3             // x += dx
+    punpckldq  xmm0, xmm1             // x0 x1
+
+    movd       xmm1, [esi + eax * 4]  // 1 source x2 pixels
+    movd       xmm4, [esi + edx * 4]  // 1 source x3 pixels
+    pextrw     eax, xmm2, 1           // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3           // get x1 integer. next iteration.
+    punpckldq  xmm1, xmm4             // x2 x3
+    punpcklqdq xmm0, xmm1             // x0 x1 x2 x3
+    sub        ecx, 4                 // 4 pixels
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    jge        xloop4
+
+    align      4
+ xloop49:
+    test       ecx, 2
+    je         xloop29
+
+    // 2 Pixels.
+    movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
+    movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
+    pextrw     eax, xmm2, 5           // get x2 integer.
+    punpckldq  xmm0, xmm1             // x0 x1
+
+    movq       qword ptr [edi], xmm0
+    lea        edi, [edi + 8]
+
+ xloop29:
+    test       ecx, 1
+    je         xloop99
+
+    // 1 Pixels.
+    movd       xmm0, [esi + eax * 4]  // 1 source x2 pixels
+    movd       dword ptr [edi], xmm0
+    align      4
+ xloop99:
+
+    pop        esi
+    pop        edi
+    ret
+  }
+}
+
+// Bilinear row filtering combines 2x1 -> 1x1. SSSE3 version.
+// TODO(fbarchard): Port to Neon
+
+// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
+static uvec8 kShuffleColARGB = {
+  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
+  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+};
+
+// Shuffle table for duplicating 2 fractions into 8 bytes each
+static uvec8 kShuffleFractions = {
+  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+};
+
+__declspec(naked) __declspec(align(16))
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]    // dst_argb
+    mov        esi, [esp + 8 + 8]    // src_argb
+    mov        ecx, [esp + 8 + 12]   // dst_width
+    movd       xmm2, [esp + 8 + 16]  // x
+    movd       xmm3, [esp + 8 + 20]  // dx
+    movdqa     xmm4, kShuffleColARGB
+    movdqa     xmm5, kShuffleFractions
+    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    psrlw      xmm6, 9
+    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    sub        ecx, 2
+    jl         xloop29
+
+    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    paddd      xmm0, xmm3
+    punpckldq  xmm2, xmm0           // x0 x1
+    punpckldq  xmm3, xmm3           // dx dx
+    paddd      xmm3, xmm3           // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+
+    // 2 Pixel loop.
+    align      4
+  xloop2:
+    movdqa     xmm1, xmm2           // x0, x1 fractions.
+    paddd      xmm2, xmm3           // x += dx
+    movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
+    psrlw      xmm1, 9              // 7 bit fractions.
+    movhps     xmm0, qword ptr [esi + edx * 4]  // 2 source x1 pixels
+    pshufb     xmm1, xmm5           // 0000000011111111
+    pshufb     xmm0, xmm4           // arrange pixels into pairs
+    pxor       xmm1, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1           // argb_argb 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    psrlw      xmm0, 7              // argb 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // argb_argb 8 bits, 2 pixels.
+    movq       qword ptr [edi], xmm0
+    lea        edi, [edi + 8]
+    sub        ecx, 2               // 2 pixels
+    jge        xloop2
+
+    align      4
+ xloop29:
+
+    add        ecx, 2 - 1
+    jl         xloop99
+
+    // 1 pixel remainder
+    psrlw      xmm2, 9              // 7 bit fractions.
+    movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
+    pshufb     xmm2, xmm5           // 00000000
+    pshufb     xmm0, xmm4           // arrange pixels into pairs
+    pxor       xmm2, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2           // argb 16 bit, 1 pixel.
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0           // argb 8 bits, 1 pixel.
+    movd       [edi], xmm0
+
+    align      4
+ xloop99:
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  __asm {
+    mov        edx, [esp + 4]    // dst_argb
+    mov        eax, [esp + 8]    // src_argb
+    mov        ecx, [esp + 12]   // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm1, xmm0
+    punpckldq  xmm0, xmm0
+    punpckhdq  xmm1, xmm1
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+__declspec(naked) __declspec(align(16))
+int FixedDiv_X86(int num, int div) {
+  __asm {
+    mov        eax, [esp + 4]    // num
+    cdq                          // extend num to 64 bits
+    shld       edx, eax, 16      // 32.16
+    shl        eax, 16
+    idiv       dword ptr [esp + 8]
+    ret
+  }
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+__declspec(naked) __declspec(align(16))
+int FixedDiv1_X86(int num, int div) {
+  __asm {
+    mov        eax, [esp + 4]    // num
+    mov        ecx, [esp + 8]    // denom
+    cdq                          // extend num to 64 bits
+    shld       edx, eax, 16      // 32.16
+    shl        eax, 16
+    sub        eax, 0x00010001
+    sbb        edx, 0
+    sub        ecx, 1
+    idiv       ecx
+    ret
+  }
+}
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/x86inc/x86inc.asm b/third_party/x86inc/x86inc.asm
index 5b22b94..2134676 100644
--- a/third_party/x86inc/x86inc.asm
+++ b/third_party/x86inc/x86inc.asm
@@ -395,6 +395,23 @@
     %assign n_arg_names %0
 %endmacro
 
+%if ARCH_X86_64
+%macro ALLOC_STACK 2  ; stack_size, num_regs
+  %assign %%stack_aligment ((mmsize + 15) & ~15)
+  %assign stack_size_padded %1
+
+  %assign %%reg_num (%2 - 1)
+  %xdefine rsp_tmp r %+ %%reg_num
+  mov  rsp_tmp, rsp
+  sub  rsp, stack_size_padded
+  and  rsp, ~(%%stack_aligment - 1)
+%endmacro
+
+%macro RESTORE_STACK 0  ; reset rsp register
+  mov  rsp, rsp_tmp
+%endmacro
+%endif
+
 %if WIN64 ; Windows x64 ;=================================================
 
 DECLARE_REG 0,  rcx, ecx,  cx,   cl
diff --git a/tools_common.h b/tools_common.h
index 2124882..e033de2 100644
--- a/tools_common.h
+++ b/tools_common.h
@@ -89,7 +89,7 @@
   enum VideoFileType file_type;
   uint32_t width;
   uint32_t height;
-  int use_i420;
+  vpx_img_fmt_t fmt;
   int only_i420;
   uint32_t fourcc;
   struct VpxRational framerate;
diff --git a/vp8/common/arm/dequantize_arm.c b/vp8/common/arm/dequantize_arm.c
index 3e37e08..1f8157f 100644
--- a/vp8/common/arm/dequantize_arm.c
+++ b/vp8/common/arm/dequantize_arm.c
@@ -12,26 +12,9 @@
 #include "vpx_config.h"
 #include "vp8/common/blockd.h"
 
-#if HAVE_NEON_ASM
-extern void vp8_dequantize_b_loop_neon(short *Q, short *DQC, short *DQ);
-#endif
-
 #if HAVE_MEDIA
 extern void vp8_dequantize_b_loop_v6(short *Q, short *DQC, short *DQ);
-#endif
 
-#if HAVE_NEON_ASM
-
-void vp8_dequantize_b_neon(BLOCKD *d, short *DQC)
-{
-    short *DQ  = d->dqcoeff;
-    short *Q   = d->qcoeff;
-
-    vp8_dequantize_b_loop_neon(Q, DQC, DQ);
-}
-#endif
-
-#if HAVE_MEDIA
 void vp8_dequantize_b_v6(BLOCKD *d, short *DQC)
 {
     short *DQ  = d->dqcoeff;
diff --git a/vp8/common/arm/loopfilter_arm.c b/vp8/common/arm/loopfilter_arm.c
index 5d693c6..f37ca63 100644
--- a/vp8/common/arm/loopfilter_arm.c
+++ b/vp8/common/arm/loopfilter_arm.c
@@ -25,20 +25,24 @@
 extern prototype_loopfilter(vp8_mbloop_filter_vertical_edge_armv6);
 #endif
 
-#if HAVE_NEON_ASM
+#if HAVE_NEON_ASM || HAVE_NEON
 typedef void loopfilter_y_neon(unsigned char *src, int pitch,
         unsigned char blimit, unsigned char limit, unsigned char thresh);
 typedef void loopfilter_uv_neon(unsigned char *u, int pitch,
         unsigned char blimit, unsigned char limit, unsigned char thresh,
         unsigned char *v);
+#endif
 
+#if HAVE_NEON_ASM
 extern loopfilter_y_neon vp8_loop_filter_horizontal_edge_y_neon;
 extern loopfilter_y_neon vp8_loop_filter_vertical_edge_y_neon;
-extern loopfilter_y_neon vp8_mbloop_filter_horizontal_edge_y_neon;
-extern loopfilter_y_neon vp8_mbloop_filter_vertical_edge_y_neon;
-
 extern loopfilter_uv_neon vp8_loop_filter_horizontal_edge_uv_neon;
 extern loopfilter_uv_neon vp8_loop_filter_vertical_edge_uv_neon;
+#endif
+
+#if HAVE_NEON
+extern loopfilter_y_neon vp8_mbloop_filter_horizontal_edge_y_neon;
+extern loopfilter_y_neon vp8_mbloop_filter_vertical_edge_y_neon;
 extern loopfilter_uv_neon vp8_mbloop_filter_horizontal_edge_uv_neon;
 extern loopfilter_uv_neon vp8_mbloop_filter_vertical_edge_uv_neon;
 #endif
@@ -118,7 +122,7 @@
 }
 #endif
 
-#if HAVE_NEON_ASM
+#if HAVE_NEON
 /* NEON loopfilter functions */
 /* Horizontal MB filtering */
 void vp8_loop_filter_mbh_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
@@ -146,7 +150,9 @@
     if (u_ptr)
         vp8_mbloop_filter_vertical_edge_uv_neon(u_ptr, uv_stride, mblim, lim, hev_thr, v_ptr);
 }
+#endif
 
+#if HAVE_NEON_ASM
 /* Horizontal B Filtering */
 void vp8_loop_filter_bh_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
                              int y_stride, int uv_stride, loop_filter_info *lfi)
diff --git a/vp8/common/arm/neon/dequantizeb_neon.c b/vp8/common/arm/neon/dequantizeb_neon.c
index 60f69c8..54e709d 100644
--- a/vp8/common/arm/neon/dequantizeb_neon.c
+++ b/vp8/common/arm/neon/dequantizeb_neon.c
@@ -10,18 +10,16 @@
 
 #include <arm_neon.h>
 
-void vp8_dequantize_b_loop_neon(
-        int16_t *Q,
-        int16_t *DQC,
-        int16_t *DQ) {
+#include "vp8/common/blockd.h"
+
+void vp8_dequantize_b_neon(BLOCKD *d, short *DQC) {
     int16x8x2_t qQ, qDQC, qDQ;
 
-    qQ   = vld2q_s16(Q);
+    qQ   = vld2q_s16(d->qcoeff);
     qDQC = vld2q_s16(DQC);
 
     qDQ.val[0] = vmulq_s16(qQ.val[0], qDQC.val[0]);
     qDQ.val[1] = vmulq_s16(qQ.val[1], qDQC.val[1]);
 
-    vst2q_s16(DQ, qDQ);
-    return;
+    vst2q_s16(d->dqcoeff, qDQ);
 }
diff --git a/vp8/common/arm/neon/idct_dequant_0_2x_neon.asm b/vp8/common/arm/neon/idct_dequant_0_2x_neon.asm
new file mode 100644
index 0000000..3a39210
--- /dev/null
+++ b/vp8/common/arm/neon/idct_dequant_0_2x_neon.asm
@@ -0,0 +1,81 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+
+    EXPORT  |idct_dequant_0_2x_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+;void idct_dequant_0_2x_neon(short *q, short dq,
+;                            unsigned char *dst, int stride);
+; r0   *q
+; r1   dq
+; r2   *dst
+; r3   stride
+|idct_dequant_0_2x_neon| PROC
+    push            {r4, r5}
+    vpush           {d8-d15}
+
+    add             r12, r2, #4
+    vld1.32         {d2[0]}, [r2], r3
+    vld1.32         {d8[0]}, [r12], r3
+    vld1.32         {d2[1]}, [r2], r3
+    vld1.32         {d8[1]}, [r12], r3
+    vld1.32         {d4[0]}, [r2], r3
+    vld1.32         {d10[0]}, [r12], r3
+    vld1.32         {d4[1]}, [r2], r3
+    vld1.32         {d10[1]}, [r12], r3
+
+    ldrh            r12, [r0]               ; lo q
+    ldrh            r4, [r0, #32]           ; hi q
+    mov             r5, #0
+    strh            r5, [r0]
+    strh            r5, [r0, #32]
+
+    sxth            r12, r12                ; lo
+    mul             r0, r12, r1
+    add             r0, r0, #4
+    asr             r0, r0, #3
+    vdup.16         q0, r0
+    sxth            r4, r4                  ; hi
+    mul             r0, r4, r1
+    add             r0, r0, #4
+    asr             r0, r0, #3
+    vdup.16         q3, r0
+
+    vaddw.u8        q1, q0, d2              ; lo
+    vaddw.u8        q2, q0, d4
+    vaddw.u8        q4, q3, d8              ; hi
+    vaddw.u8        q5, q3, d10
+
+    sub             r2, r2, r3, lsl #2      ; dst - 4*stride
+    add             r0, r2, #4
+
+    vqmovun.s16     d2, q1                  ; lo
+    vqmovun.s16     d4, q2
+    vqmovun.s16     d8, q4                  ; hi
+    vqmovun.s16     d10, q5
+
+    vst1.32         {d2[0]}, [r2], r3       ; lo
+    vst1.32         {d8[0]}, [r0], r3       ; hi
+    vst1.32         {d2[1]}, [r2], r3
+    vst1.32         {d8[1]}, [r0], r3
+    vst1.32         {d4[0]}, [r2], r3
+    vst1.32         {d10[0]}, [r0], r3
+    vst1.32         {d4[1]}, [r2]
+    vst1.32         {d10[1]}, [r0]
+
+    vpop            {d8-d15}
+    pop             {r4, r5}
+    bx              lr
+
+    ENDP            ; |idct_dequant_0_2x_neon|
+    END
diff --git a/vp8/common/arm/neon/idct_dequant_0_2x_neon.c b/vp8/common/arm/neon/idct_dequant_0_2x_neon.c
deleted file mode 100644
index 967c322..0000000
--- a/vp8/common/arm/neon/idct_dequant_0_2x_neon.c
+++ /dev/null
@@ -1,62 +0,0 @@
-/*
- *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include <arm_neon.h>
-
-void idct_dequant_0_2x_neon(
-        int16_t *q,
-        int16_t dq,
-        unsigned char *dst,
-        int stride) {
-    unsigned char *dst0;
-    int i, a0, a1;
-    int16x8x2_t q2Add;
-    int32x2_t d2s32, d4s32;
-    uint8x8_t d2u8, d4u8;
-    uint16x8_t q1u16, q2u16;
-
-    a0 = ((q[0] * dq) + 4) >> 3;
-    a1 = ((q[16] * dq) + 4) >> 3;
-    q[0] = q[16] = 0;
-    q2Add.val[0] = vdupq_n_s16((int16_t)a0);
-    q2Add.val[1] = vdupq_n_s16((int16_t)a1);
-
-    for (i = 0; i < 2; i++, dst += 4) {
-        dst0 = dst;
-        d2s32 = vld1_lane_s32((const int32_t *)dst0, d2s32, 0);
-        dst0 += stride;
-        d2s32 = vld1_lane_s32((const int32_t *)dst0, d2s32, 1);
-        dst0 += stride;
-        d4s32 = vld1_lane_s32((const int32_t *)dst0, d4s32, 0);
-        dst0 += stride;
-        d4s32 = vld1_lane_s32((const int32_t *)dst0, d4s32, 1);
-
-        q1u16 = vaddw_u8(vreinterpretq_u16_s16(q2Add.val[i]),
-                         vreinterpret_u8_s32(d2s32));
-        q2u16 = vaddw_u8(vreinterpretq_u16_s16(q2Add.val[i]),
-                         vreinterpret_u8_s32(d4s32));
-
-        d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q1u16));
-        d4u8 = vqmovun_s16(vreinterpretq_s16_u16(q2u16));
-
-        d2s32 = vreinterpret_s32_u8(d2u8);
-        d4s32 = vreinterpret_s32_u8(d4u8);
-
-        dst0 = dst;
-        vst1_lane_s32((int32_t *)dst0, d2s32, 0);
-        dst0 += stride;
-        vst1_lane_s32((int32_t *)dst0, d2s32, 1);
-        dst0 += stride;
-        vst1_lane_s32((int32_t *)dst0, d4s32, 0);
-        dst0 += stride;
-        vst1_lane_s32((int32_t *)dst0, d4s32, 1);
-    }
-    return;
-}
diff --git a/vp8/common/arm/neon/idct_dequant_full_2x_neon.asm b/vp8/common/arm/neon/idct_dequant_full_2x_neon.asm
new file mode 100644
index 0000000..8da0fa0
--- /dev/null
+++ b/vp8/common/arm/neon/idct_dequant_full_2x_neon.asm
@@ -0,0 +1,199 @@
+;
+;  Copyright (c) 2010 The Webm project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |idct_dequant_full_2x_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+;void idct_dequant_full_2x_neon(short *q, short *dq,
+;                               unsigned char *dst, int stride);
+; r0    *q,
+; r1    *dq,
+; r2    *dst
+; r3    stride
+|idct_dequant_full_2x_neon| PROC
+    vpush           {d8-d15}
+
+    vld1.16         {q0, q1}, [r1]          ; dq (same l/r)
+    vld1.16         {q2, q3}, [r0]          ; l q
+    add             r0, r0, #32
+    vld1.16         {q4, q5}, [r0]          ; r q
+    add             r12, r2, #4
+
+    ; interleave the predictors
+    vld1.32         {d28[0]}, [r2],  r3     ; l pre
+    vld1.32         {d28[1]}, [r12], r3     ; r pre
+    vld1.32         {d29[0]}, [r2],  r3
+    vld1.32         {d29[1]}, [r12], r3
+    vld1.32         {d30[0]}, [r2],  r3
+    vld1.32         {d30[1]}, [r12], r3
+    vld1.32         {d31[0]}, [r2],  r3
+    vld1.32         {d31[1]}, [r12]
+
+    adr             r1, cospi8sqrt2minus1   ; pointer to the first constant
+
+    ; dequant: q[i] = q[i] * dq[i]
+    vmul.i16        q2, q2, q0
+    vmul.i16        q3, q3, q1
+    vmul.i16        q4, q4, q0
+    vmul.i16        q5, q5, q1
+
+    vld1.16         {d0}, [r1]
+
+    ; q2: l0r0  q3: l8r8
+    ; q4: l4r4  q5: l12r12
+    vswp            d5, d8
+    vswp            d7, d10
+
+    ; _CONSTANTS_ * 4,12 >> 16
+    ; q6:  4 * sinpi : c1/temp1
+    ; q7: 12 * sinpi : d1/temp2
+    ; q8:  4 * cospi
+    ; q9: 12 * cospi
+    vqdmulh.s16     q6, q4, d0[2]           ; sinpi8sqrt2
+    vqdmulh.s16     q7, q5, d0[2]
+    vqdmulh.s16     q8, q4, d0[0]           ; cospi8sqrt2minus1
+    vqdmulh.s16     q9, q5, d0[0]
+
+    vqadd.s16       q10, q2, q3             ; a1 = 0 + 8
+    vqsub.s16       q11, q2, q3             ; b1 = 0 - 8
+
+    ; vqdmulh only accepts signed values. this was a problem because
+    ; our constant had the high bit set, and was treated as a negative value.
+    ; vqdmulh also doubles the value before it shifts by 16. we need to
+    ; compensate for this. in the case of sinpi8sqrt2, the lowest bit is 0,
+    ; so we can shift the constant without losing precision. this avoids
+    ; shift again afterward, but also avoids the sign issue. win win!
+    ; for cospi8sqrt2minus1 the lowest bit is 1, so we lose precision if we
+    ; pre-shift it
+    vshr.s16        q8, q8, #1
+    vshr.s16        q9, q9, #1
+
+    ; q4:  4 +  4 * cospi : d1/temp1
+    ; q5: 12 + 12 * cospi : c1/temp2
+    vqadd.s16       q4, q4, q8
+    vqadd.s16       q5, q5, q9
+
+    ; c1 = temp1 - temp2
+    ; d1 = temp1 + temp2
+    vqsub.s16       q2, q6, q5
+    vqadd.s16       q3, q4, q7
+
+    ; [0]: a1+d1
+    ; [1]: b1+c1
+    ; [2]: b1-c1
+    ; [3]: a1-d1
+    vqadd.s16       q4, q10, q3
+    vqadd.s16       q5, q11, q2
+    vqsub.s16       q6, q11, q2
+    vqsub.s16       q7, q10, q3
+
+    ; rotate
+    vtrn.32         q4, q6
+    vtrn.32         q5, q7
+    vtrn.16         q4, q5
+    vtrn.16         q6, q7
+    ; idct loop 2
+    ; q4: l 0, 4, 8,12 r 0, 4, 8,12
+    ; q5: l 1, 5, 9,13 r 1, 5, 9,13
+    ; q6: l 2, 6,10,14 r 2, 6,10,14
+    ; q7: l 3, 7,11,15 r 3, 7,11,15
+
+    ; q8:  1 * sinpi : c1/temp1
+    ; q9:  3 * sinpi : d1/temp2
+    ; q10: 1 * cospi
+    ; q11: 3 * cospi
+    vqdmulh.s16     q8, q5, d0[2]           ; sinpi8sqrt2
+    vqdmulh.s16     q9, q7, d0[2]
+    vqdmulh.s16     q10, q5, d0[0]          ; cospi8sqrt2minus1
+    vqdmulh.s16     q11, q7, d0[0]
+
+    vqadd.s16       q2, q4, q6             ; a1 = 0 + 2
+    vqsub.s16       q3, q4, q6             ; b1 = 0 - 2
+
+    ; see note on shifting above
+    vshr.s16        q10, q10, #1
+    vshr.s16        q11, q11, #1
+
+    ; q10: 1 + 1 * cospi : d1/temp1
+    ; q11: 3 + 3 * cospi : c1/temp2
+    vqadd.s16       q10, q5, q10
+    vqadd.s16       q11, q7, q11
+
+    ; q8: c1 = temp1 - temp2
+    ; q9: d1 = temp1 + temp2
+    vqsub.s16       q8, q8, q11
+    vqadd.s16       q9, q10, q9
+
+    ; a1+d1
+    ; b1+c1
+    ; b1-c1
+    ; a1-d1
+    vqadd.s16       q4, q2, q9
+    vqadd.s16       q5, q3, q8
+    vqsub.s16       q6, q3, q8
+    vqsub.s16       q7, q2, q9
+
+    ; +4 >> 3 (rounding)
+    vrshr.s16       q4, q4, #3              ; lo
+    vrshr.s16       q5, q5, #3
+    vrshr.s16       q6, q6, #3              ; hi
+    vrshr.s16       q7, q7, #3
+
+    vtrn.32         q4, q6
+    vtrn.32         q5, q7
+    vtrn.16         q4, q5
+    vtrn.16         q6, q7
+
+    ; adding pre
+    ; input is still packed. pre was read interleaved
+    vaddw.u8        q4, q4, d28
+    vaddw.u8        q5, q5, d29
+    vaddw.u8        q6, q6, d30
+    vaddw.u8        q7, q7, d31
+
+    vmov.i16        q14, #0
+    vmov            q15, q14
+    vst1.16         {q14, q15}, [r0]        ; write over high input
+    sub             r0, r0, #32
+    vst1.16         {q14, q15}, [r0]        ; write over low input
+
+    sub             r2, r2, r3, lsl #2      ; dst - 4*stride
+    add             r1, r2, #4              ; hi
+
+    ;saturate and narrow
+    vqmovun.s16     d0, q4                  ; lo
+    vqmovun.s16     d1, q5
+    vqmovun.s16     d2, q6                  ; hi
+    vqmovun.s16     d3, q7
+
+    vst1.32         {d0[0]}, [r2], r3       ; lo
+    vst1.32         {d0[1]}, [r1], r3       ; hi
+    vst1.32         {d1[0]}, [r2], r3
+    vst1.32         {d1[1]}, [r1], r3
+    vst1.32         {d2[0]}, [r2], r3
+    vst1.32         {d2[1]}, [r1], r3
+    vst1.32         {d3[0]}, [r2]
+    vst1.32         {d3[1]}, [r1]
+
+    vpop            {d8-d15}
+    bx             lr
+
+    ENDP           ; |idct_dequant_full_2x_neon|
+
+; Constant Pool
+cospi8sqrt2minus1 DCD 0x4e7b
+; because the lowest bit in 0x8a8c is 0, we can pre-shift this
+sinpi8sqrt2       DCD 0x4546
+
+    END
diff --git a/vp8/common/arm/neon/idct_dequant_full_2x_neon.c b/vp8/common/arm/neon/idct_dequant_full_2x_neon.c
deleted file mode 100644
index a60ed46..0000000
--- a/vp8/common/arm/neon/idct_dequant_full_2x_neon.c
+++ /dev/null
@@ -1,185 +0,0 @@
-/*
- *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include <arm_neon.h>
-
-static const int16_t cospi8sqrt2minus1 = 20091;
-static const int16_t sinpi8sqrt2       = 17734;
-// because the lowest bit in 0x8a8c is 0, we can pre-shift this
-
-void idct_dequant_full_2x_neon(
-        int16_t *q,
-        int16_t *dq,
-        unsigned char *dst,
-        int stride) {
-    unsigned char *dst0, *dst1;
-    int32x2_t d28, d29, d30, d31;
-    int16x8_t q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11;
-    int16x8_t qEmpty = vdupq_n_s16(0);
-    int32x4x2_t q2tmp0, q2tmp1;
-    int16x8x2_t q2tmp2, q2tmp3;
-    int16x4_t dLow0, dLow1, dHigh0, dHigh1;
-
-    d28 = d29 = d30 = d31 = vdup_n_s32(0);
-
-    // load dq
-    q0 = vld1q_s16(dq);
-    dq += 8;
-    q1 = vld1q_s16(dq);
-
-    // load q
-    q2 = vld1q_s16(q);
-    vst1q_s16(q, qEmpty);
-    q += 8;
-    q3 = vld1q_s16(q);
-    vst1q_s16(q, qEmpty);
-    q += 8;
-    q4 = vld1q_s16(q);
-    vst1q_s16(q, qEmpty);
-    q += 8;
-    q5 = vld1q_s16(q);
-    vst1q_s16(q, qEmpty);
-
-    // load src from dst
-    dst0 = dst;
-    dst1 = dst + 4;
-    d28 = vld1_lane_s32((const int32_t *)dst0, d28, 0);
-    dst0 += stride;
-    d28 = vld1_lane_s32((const int32_t *)dst1, d28, 1);
-    dst1 += stride;
-    d29 = vld1_lane_s32((const int32_t *)dst0, d29, 0);
-    dst0 += stride;
-    d29 = vld1_lane_s32((const int32_t *)dst1, d29, 1);
-    dst1 += stride;
-
-    d30 = vld1_lane_s32((const int32_t *)dst0, d30, 0);
-    dst0 += stride;
-    d30 = vld1_lane_s32((const int32_t *)dst1, d30, 1);
-    dst1 += stride;
-    d31 = vld1_lane_s32((const int32_t *)dst0, d31, 0);
-    d31 = vld1_lane_s32((const int32_t *)dst1, d31, 1);
-
-    q2 = vmulq_s16(q2, q0);
-    q3 = vmulq_s16(q3, q1);
-    q4 = vmulq_s16(q4, q0);
-    q5 = vmulq_s16(q5, q1);
-
-    // vswp
-    dLow0 = vget_low_s16(q2);
-    dHigh0 = vget_high_s16(q2);
-    dLow1 = vget_low_s16(q4);
-    dHigh1 = vget_high_s16(q4);
-    q2 = vcombine_s16(dLow0, dLow1);
-    q4 = vcombine_s16(dHigh0, dHigh1);
-
-    dLow0 = vget_low_s16(q3);
-    dHigh0 = vget_high_s16(q3);
-    dLow1 = vget_low_s16(q5);
-    dHigh1 = vget_high_s16(q5);
-    q3 = vcombine_s16(dLow0, dLow1);
-    q5 = vcombine_s16(dHigh0, dHigh1);
-
-    q6 = vqdmulhq_n_s16(q4, sinpi8sqrt2);
-    q7 = vqdmulhq_n_s16(q5, sinpi8sqrt2);
-    q8 = vqdmulhq_n_s16(q4, cospi8sqrt2minus1);
-    q9 = vqdmulhq_n_s16(q5, cospi8sqrt2minus1);
-
-    q10 = vqaddq_s16(q2, q3);
-    q11 = vqsubq_s16(q2, q3);
-
-    q8 = vshrq_n_s16(q8, 1);
-    q9 = vshrq_n_s16(q9, 1);
-
-    q4 = vqaddq_s16(q4, q8);
-    q5 = vqaddq_s16(q5, q9);
-
-    q2 = vqsubq_s16(q6, q5);
-    q3 = vqaddq_s16(q7, q4);
-
-    q4 = vqaddq_s16(q10, q3);
-    q5 = vqaddq_s16(q11, q2);
-    q6 = vqsubq_s16(q11, q2);
-    q7 = vqsubq_s16(q10, q3);
-
-    q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6));
-    q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7));
-    q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]),
-                       vreinterpretq_s16_s32(q2tmp1.val[0]));
-    q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]),
-                       vreinterpretq_s16_s32(q2tmp1.val[1]));
-
-    // loop 2
-    q8  = vqdmulhq_n_s16(q2tmp2.val[1], sinpi8sqrt2);
-    q9  = vqdmulhq_n_s16(q2tmp3.val[1], sinpi8sqrt2);
-    q10 = vqdmulhq_n_s16(q2tmp2.val[1], cospi8sqrt2minus1);
-    q11 = vqdmulhq_n_s16(q2tmp3.val[1], cospi8sqrt2minus1);
-
-    q2 = vqaddq_s16(q2tmp2.val[0], q2tmp3.val[0]);
-    q3 = vqsubq_s16(q2tmp2.val[0], q2tmp3.val[0]);
-
-    q10 = vshrq_n_s16(q10, 1);
-    q11 = vshrq_n_s16(q11, 1);
-
-    q10 = vqaddq_s16(q2tmp2.val[1], q10);
-    q11 = vqaddq_s16(q2tmp3.val[1], q11);
-
-    q8 = vqsubq_s16(q8, q11);
-    q9 = vqaddq_s16(q9, q10);
-
-    q4 = vqaddq_s16(q2, q9);
-    q5 = vqaddq_s16(q3, q8);
-    q6 = vqsubq_s16(q3, q8);
-    q7 = vqsubq_s16(q2, q9);
-
-    q4 = vrshrq_n_s16(q4, 3);
-    q5 = vrshrq_n_s16(q5, 3);
-    q6 = vrshrq_n_s16(q6, 3);
-    q7 = vrshrq_n_s16(q7, 3);
-
-    q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6));
-    q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7));
-    q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]),
-                       vreinterpretq_s16_s32(q2tmp1.val[0]));
-    q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]),
-                       vreinterpretq_s16_s32(q2tmp1.val[1]));
-
-    q4 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[0]),
-                                          vreinterpret_u8_s32(d28)));
-    q5 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[1]),
-                                          vreinterpret_u8_s32(d29)));
-    q6 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[0]),
-                                          vreinterpret_u8_s32(d30)));
-    q7 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[1]),
-                                          vreinterpret_u8_s32(d31)));
-
-    d28 = vreinterpret_s32_u8(vqmovun_s16(q4));
-    d29 = vreinterpret_s32_u8(vqmovun_s16(q5));
-    d30 = vreinterpret_s32_u8(vqmovun_s16(q6));
-    d31 = vreinterpret_s32_u8(vqmovun_s16(q7));
-
-    dst0 = dst;
-    dst1 = dst + 4;
-    vst1_lane_s32((int32_t *)dst0, d28, 0);
-    dst0 += stride;
-    vst1_lane_s32((int32_t *)dst1, d28, 1);
-    dst1 += stride;
-    vst1_lane_s32((int32_t *)dst0, d29, 0);
-    dst0 += stride;
-    vst1_lane_s32((int32_t *)dst1, d29, 1);
-    dst1 += stride;
-
-    vst1_lane_s32((int32_t *)dst0, d30, 0);
-    dst0 += stride;
-    vst1_lane_s32((int32_t *)dst1, d30, 1);
-    dst1 += stride;
-    vst1_lane_s32((int32_t *)dst0, d31, 0);
-    vst1_lane_s32((int32_t *)dst1, d31, 1);
-    return;
-}
diff --git a/vp8/common/arm/reconintra_arm.c b/vp8/common/arm/reconintra_arm.c
index 765fc3a..e55a33c 100644
--- a/vp8/common/arm/reconintra_arm.c
+++ b/vp8/common/arm/reconintra_arm.c
@@ -14,7 +14,7 @@
 #include "vp8/common/blockd.h"
 #include "vpx_mem/vpx_mem.h"
 
-#if HAVE_NEON_ARM
+#if HAVE_NEON_ASM
 extern void vp8_build_intra_predictors_mby_neon_func(
     unsigned char *y_buffer,
     unsigned char *ypred_ptr,
diff --git a/vp8/common/rtcd_defs.pl b/vp8/common/rtcd_defs.pl
index dcf5b8e..cbfd76a 100644
--- a/vp8/common/rtcd_defs.pl
+++ b/vp8/common/rtcd_defs.pl
@@ -29,9 +29,8 @@
 # Dequant
 #
 add_proto qw/void vp8_dequantize_b/, "struct blockd*, short *dqc";
-specialize qw/vp8_dequantize_b mmx media neon_asm/;
+specialize qw/vp8_dequantize_b mmx media neon/;
 $vp8_dequantize_b_media=vp8_dequantize_b_v6;
-$vp8_dequantize_b_neon_asm=vp8_dequantize_b_neon;
 
 add_proto qw/void vp8_dequant_idct_add/, "short *input, short *dq, unsigned char *output, int stride";
 specialize qw/vp8_dequant_idct_add mmx media neon dspr2/;
@@ -54,9 +53,8 @@
 # Loopfilter
 #
 add_proto qw/void vp8_loop_filter_mbv/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
-specialize qw/vp8_loop_filter_mbv mmx sse2 media neon_asm dspr2/;
+specialize qw/vp8_loop_filter_mbv mmx sse2 media neon dspr2/;
 $vp8_loop_filter_mbv_media=vp8_loop_filter_mbv_armv6;
-$vp8_loop_filter_mbv_neon_asm=vp8_loop_filter_mbv_neon;
 $vp8_loop_filter_mbv_dspr2=vp8_loop_filter_mbv_dspr2;
 
 add_proto qw/void vp8_loop_filter_bv/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
@@ -66,9 +64,8 @@
 $vp8_loop_filter_bv_dspr2=vp8_loop_filter_bv_dspr2;
 
 add_proto qw/void vp8_loop_filter_mbh/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
-specialize qw/vp8_loop_filter_mbh mmx sse2 media neon_asm dspr2/;
+specialize qw/vp8_loop_filter_mbh mmx sse2 media neon dspr2/;
 $vp8_loop_filter_mbh_media=vp8_loop_filter_mbh_armv6;
-$vp8_loop_filter_mbh_neon_asm=vp8_loop_filter_mbh_neon;
 $vp8_loop_filter_mbh_dspr2=vp8_loop_filter_mbh_dspr2;
 
 add_proto qw/void vp8_loop_filter_bh/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
@@ -87,12 +84,12 @@
 $vp8_loop_filter_simple_mbv_neon_asm=vp8_loop_filter_mbvs_neon;
 
 add_proto qw/void vp8_loop_filter_simple_mbh/, "unsigned char *y, int ystride, const unsigned char *blimit";
-specialize qw/vp8_loop_filter_simple_mbh mmx sse2 media neon_asm/;
+specialize qw/vp8_loop_filter_simple_mbh mmx sse2 media neon/;
 $vp8_loop_filter_simple_mbh_c=vp8_loop_filter_simple_horizontal_edge_c;
 $vp8_loop_filter_simple_mbh_mmx=vp8_loop_filter_simple_horizontal_edge_mmx;
 $vp8_loop_filter_simple_mbh_sse2=vp8_loop_filter_simple_horizontal_edge_sse2;
 $vp8_loop_filter_simple_mbh_media=vp8_loop_filter_simple_horizontal_edge_armv6;
-$vp8_loop_filter_simple_mbh_neon_asm=vp8_loop_filter_mbhs_neon;
+$vp8_loop_filter_simple_mbh_neon=vp8_loop_filter_mbhs_neon;
 
 add_proto qw/void vp8_loop_filter_simple_bv/, "unsigned char *y, int ystride, const unsigned char *blimit";
 specialize qw/vp8_loop_filter_simple_bv mmx sse2 media neon_asm/;
@@ -103,12 +100,12 @@
 $vp8_loop_filter_simple_bv_neon_asm=vp8_loop_filter_bvs_neon;
 
 add_proto qw/void vp8_loop_filter_simple_bh/, "unsigned char *y, int ystride, const unsigned char *blimit";
-specialize qw/vp8_loop_filter_simple_bh mmx sse2 media neon_asm/;
+specialize qw/vp8_loop_filter_simple_bh mmx sse2 media neon/;
 $vp8_loop_filter_simple_bh_c=vp8_loop_filter_bhs_c;
 $vp8_loop_filter_simple_bh_mmx=vp8_loop_filter_bhs_mmx;
 $vp8_loop_filter_simple_bh_sse2=vp8_loop_filter_bhs_sse2;
 $vp8_loop_filter_simple_bh_media=vp8_loop_filter_bhs_armv6;
-$vp8_loop_filter_simple_bh_neon_asm=vp8_loop_filter_bhs_neon;
+$vp8_loop_filter_simple_bh_neon=vp8_loop_filter_bhs_neon;
 
 #
 # IDCT
@@ -555,7 +552,7 @@
 # Denoiser filter
 #
 if (vpx_config("CONFIG_TEMPORAL_DENOISING") eq "yes") {
-    add_proto qw/int vp8_denoiser_filter/, "struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset";
+    add_proto qw/int vp8_denoiser_filter/, "unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
     specialize qw/vp8_denoiser_filter sse2 neon/;
 }
 
diff --git a/vp8/common/x86/variance_impl_mmx.asm b/vp8/common/x86/variance_impl_mmx.asm
index d9120d0..7d5e681 100644
--- a/vp8/common/x86/variance_impl_mmx.asm
+++ b/vp8/common/x86/variance_impl_mmx.asm
@@ -342,8 +342,8 @@
         movsxd      rdx, dword ptr arg(3) ;[recon_stride]
 
         ; Row 1
-        movq        mm0, [rax]                  ; Copy eight bytes to mm0
-        movq        mm1, [rbx]                  ; Copy eight bytes to mm1
+        movd        mm0, [rax]                  ; Copy four bytes to mm0
+        movd        mm1, [rbx]                  ; Copy four bytes to mm1
         punpcklbw   mm0, mm6                    ; unpack to higher prrcision
         punpcklbw   mm1, mm6
         psubsw      mm0, mm1                    ; A-B (low order) to MM0
@@ -351,12 +351,12 @@
         pmaddwd     mm0, mm0                    ; square and accumulate
         add         rbx,rdx                     ; Inc pointer into ref data
         add         rax,rcx                     ; Inc pointer into the new data
-        movq        mm1, [rbx]                  ; Copy eight bytes to mm1
+        movd        mm1, [rbx]                  ; Copy four bytes to mm1
         paddd       mm7, mm0                    ; accumulate in mm7
 
 
         ; Row 2
-        movq        mm0, [rax]                  ; Copy eight bytes to mm0
+        movd        mm0, [rax]                  ; Copy four bytes to mm0
         punpcklbw   mm0, mm6                    ; unpack to higher prrcision
         punpcklbw   mm1, mm6
         psubsw      mm0, mm1                    ; A-B (low order) to MM0
@@ -365,12 +365,12 @@
         pmaddwd     mm0, mm0                    ; square and accumulate
         add         rbx,rdx                     ; Inc pointer into ref data
         add         rax,rcx                     ; Inc pointer into the new data
-        movq        mm1, [rbx]                  ; Copy eight bytes to mm1
+        movd        mm1, [rbx]                  ; Copy four bytes to mm1
         paddd       mm7, mm0                    ; accumulate in mm7
 
         ; Row 3
-        movq        mm0, [rax]                  ; Copy eight bytes to mm0
-        punpcklbw   mm0, mm6                    ; unpack to higher prrcision
+        movd        mm0, [rax]                  ; Copy four bytes to mm0
+        punpcklbw   mm0, mm6                    ; unpack to higher precision
         punpcklbw   mm1, mm6
         psubsw      mm0, mm1                    ; A-B (low order) to MM0
         paddw       mm5, mm0                    ; accumulate differences in mm5
@@ -378,11 +378,11 @@
         pmaddwd     mm0, mm0                    ; square and accumulate
         add         rbx,rdx                     ; Inc pointer into ref data
         add         rax,rcx                     ; Inc pointer into the new data
-        movq        mm1, [rbx]                  ; Copy eight bytes to mm1
+        movd        mm1, [rbx]                  ; Copy four bytes to mm1
         paddd       mm7, mm0                    ; accumulate in mm7
 
         ; Row 4
-        movq        mm0, [rax]                  ; Copy eight bytes to mm0
+        movd        mm0, [rax]                  ; Copy four bytes to mm0
 
         punpcklbw   mm0, mm6                    ; unpack to higher prrcision
         punpcklbw   mm1, mm6
diff --git a/vp8/encoder/arm/neon/denoising_neon.c b/vp8/encoder/arm/neon/denoising_neon.c
index 23dc0a9..32ce65a 100644
--- a/vp8/encoder/arm/neon/denoising_neon.c
+++ b/vp8/encoder/arm/neon/denoising_neon.c
@@ -45,10 +45,13 @@
  *      [16, 255]       3               6                    7
  */
 
-int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
-                             YV12_BUFFER_CONFIG *running_avg,
-                             MACROBLOCK *signal, unsigned int motion_magnitude,
-                             int y_offset, int uv_offset) {
+int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y,
+                             int mc_running_avg_y_stride,
+                             unsigned char *running_avg_y,
+                             int running_avg_y_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
     /* If motion_magnitude is small, making the denoiser more aggressive by
      * increasing the adjustment for each level, level1 adjustment is
      * increased, the deltas stay the same.
@@ -60,14 +63,6 @@
     const uint8x16_t v_level1_threshold = vdupq_n_u8(4);
     const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
     const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
-
-    /* Local variables for array pointers and strides. */
-    unsigned char *sig = signal->thismb;
-    int            sig_stride = 16;
-    unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset;
-    int            mc_running_avg_y_stride = mc_running_avg->y_stride;
-    unsigned char *running_avg_y = running_avg->y_buffer + y_offset;
-    int            running_avg_y_stride = running_avg->y_stride;
     int64x2_t v_sum_diff_total = vdupq_n_s64(0);
 
     /* Go over lines. */
diff --git a/vp8/encoder/block.h b/vp8/encoder/block.h
index dd733e5..34879cf 100644
--- a/vp8/encoder/block.h
+++ b/vp8/encoder/block.h
@@ -125,6 +125,7 @@
 
     int optimize;
     int q_index;
+    int increase_denoising;
 
 #if CONFIG_TEMPORAL_DENOISING
     MB_PREDICTION_MODE best_sse_inter_mode;
diff --git a/vp8/encoder/denoising.c b/vp8/encoder/denoising.c
index 7819265..1e645fb 100644
--- a/vp8/encoder/denoising.c
+++ b/vp8/encoder/denoising.c
@@ -21,6 +21,7 @@
  */
 static const unsigned int SSE_DIFF_THRESHOLD = 16 * 16 * 20;
 static const unsigned int SSE_THRESHOLD = 16 * 16 * 40;
+static const unsigned int SSE_THRESHOLD_HIGH = 16 * 16 * 60;
 
 /*
  * The filter function was modified to reduce the computational complexity.
@@ -51,27 +52,32 @@
  * [16, 255]              6                                    7
  */
 
-int vp8_denoiser_filter_c(YV12_BUFFER_CONFIG *mc_running_avg,
-                          YV12_BUFFER_CONFIG *running_avg, MACROBLOCK *signal,
-                          unsigned int motion_magnitude, int y_offset,
-                          int uv_offset)
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride,
+                          unsigned char *running_avg_y, int avg_y_stride,
+                          unsigned char *sig, int sig_stride,
+                          unsigned int motion_magnitude,
+                          int increase_denoising)
 {
-    unsigned char *sig = signal->thismb;
-    int sig_stride = 16;
-    unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset;
-    int mc_avg_y_stride = mc_running_avg->y_stride;
-    unsigned char *running_avg_y = running_avg->y_buffer + y_offset;
-    int avg_y_stride = running_avg->y_stride;
-    int r, c, i;
+    unsigned char *running_avg_y_start = running_avg_y;
+    unsigned char *sig_start = sig;
+    int sum_diff_thresh;
+    int r, c;
     int sum_diff = 0;
     int adj_val[3] = {3, 4, 6};
-
+    int shift_inc1 = 0;
+    int shift_inc2 = 1;
     /* If motion_magnitude is small, making the denoiser more aggressive by
-     * increasing the adjustment for each level. */
+     * increasing the adjustment for each level. Add another increment for
+     * blocks that are labeled for increase denoising. */
     if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
     {
-        for (i = 0; i < 3; i++)
-            adj_val[i] += 1;
+      if (increase_denoising) {
+        shift_inc1 = 1;
+        shift_inc2 = 2;
+      }
+      adj_val[0] += shift_inc2;
+      adj_val[1] += shift_inc2;
+      adj_val[2] += shift_inc2;
     }
 
     for (r = 0; r < 16; ++r)
@@ -85,8 +91,9 @@
             diff = mc_running_avg_y[c] - sig[c];
             absdiff = abs(diff);
 
-            /* When |diff| < 4, use pixel value from last denoised raw. */
-            if (absdiff <= 3)
+            // When |diff| <= |3 + shift_inc1|, use pixel value from
+            // last denoised raw.
+            if (absdiff <= 3 + shift_inc1)
             {
                 running_avg_y[c] = mc_running_avg_y[c];
                 sum_diff += diff;
@@ -127,11 +134,12 @@
         running_avg_y += avg_y_stride;
     }
 
-    if (abs(sum_diff) > SUM_DIFF_THRESHOLD)
+    sum_diff_thresh= SUM_DIFF_THRESHOLD;
+    if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
+    if (abs(sum_diff) > sum_diff_thresh)
         return COPY_BLOCK;
 
-    vp8_copy_mem16x16(running_avg->y_buffer + y_offset, avg_y_stride,
-                      signal->thismb, sig_stride);
+    vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
     return FILTER_BLOCK;
 }
 
@@ -192,7 +200,7 @@
     int mv_row;
     int mv_col;
     unsigned int motion_magnitude2;
-
+    unsigned int sse_thresh;
     MV_REFERENCE_FRAME frame = x->best_reference_frame;
     MV_REFERENCE_FRAME zero_frame = x->best_zeromv_reference_frame;
 
@@ -277,7 +285,10 @@
     mv_row = x->best_sse_mv.as_mv.row;
     mv_col = x->best_sse_mv.as_mv.col;
     motion_magnitude2 = mv_row * mv_row + mv_col * mv_col;
-    if (best_sse > SSE_THRESHOLD || motion_magnitude2
+    sse_thresh = SSE_THRESHOLD;
+    if (x->increase_denoising) sse_thresh = SSE_THRESHOLD_HIGH;
+
+    if (best_sse > sse_thresh || motion_magnitude2
            > 8 * NOISE_MOTION_THRESHOLD)
     {
         decision = COPY_BLOCK;
@@ -285,12 +296,18 @@
 
     if (decision == FILTER_BLOCK)
     {
+        unsigned char *mc_running_avg_y =
+            denoiser->yv12_mc_running_avg.y_buffer + recon_yoffset;
+        int mc_avg_y_stride = denoiser->yv12_mc_running_avg.y_stride;
+        unsigned char *running_avg_y =
+            denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset;
+        int avg_y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride;
+
         /* Filter. */
-        decision = vp8_denoiser_filter(&denoiser->yv12_mc_running_avg,
-                                       &denoiser->yv12_running_avg[INTRA_FRAME],
-                                       x,
-                                       motion_magnitude2,
-                                       recon_yoffset, recon_uvoffset);
+        decision = vp8_denoiser_filter(mc_running_avg_y, mc_avg_y_stride,
+                                         running_avg_y, avg_y_stride,
+                                         x->thismb, 16, motion_magnitude2,
+                                         x->increase_denoising);
     }
     if (decision == COPY_BLOCK)
     {
diff --git a/vp8/encoder/denoising.h b/vp8/encoder/denoising.h
index cc9913a..ae744d2 100644
--- a/vp8/encoder/denoising.h
+++ b/vp8/encoder/denoising.h
@@ -18,6 +18,7 @@
 #endif
 
 #define SUM_DIFF_THRESHOLD (16 * 16 * 2)
+#define SUM_DIFF_THRESHOLD_HIGH (16 * 16 * 3)
 #define MOTION_MAGNITUDE_THRESHOLD (8*3)
 
 enum vp8_denoiser_decision
diff --git a/vp8/encoder/pickinter.c b/vp8/encoder/pickinter.c
index 39a3baf..cf6a82f 100644
--- a/vp8/encoder/pickinter.c
+++ b/vp8/encoder/pickinter.c
@@ -1177,6 +1177,7 @@
             x->best_reference_frame = best_mbmode.ref_frame;
             best_sse = best_rd_sse;
         }
+        x->increase_denoising = 0;
         vp8_denoiser_denoise_mb(&cpi->denoiser, x, best_sse, zero_mv_sse,
                                 recon_yoffset, recon_uvoffset);
 
diff --git a/vp8/encoder/x86/denoising_sse2.c b/vp8/encoder/x86/denoising_sse2.c
index cceb826..5112f89 100644
--- a/vp8/encoder/x86/denoising_sse2.c
+++ b/vp8/encoder/x86/denoising_sse2.c
@@ -22,26 +22,28 @@
     signed char e[16];
 };
 
-int vp8_denoiser_filter_sse2(YV12_BUFFER_CONFIG *mc_running_avg,
-                             YV12_BUFFER_CONFIG *running_avg,
-                             MACROBLOCK *signal, unsigned int motion_magnitude,
-                             int y_offset, int uv_offset)
+int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
+                             int mc_avg_y_stride,
+                             unsigned char *running_avg_y, int avg_y_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising)
 {
-    unsigned char *sig = signal->thismb;
-    int sig_stride = 16;
-    unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset;
-    int mc_avg_y_stride = mc_running_avg->y_stride;
-    unsigned char *running_avg_y = running_avg->y_buffer + y_offset;
-    int avg_y_stride = running_avg->y_stride;
+    unsigned char *running_avg_y_start = running_avg_y;
+    unsigned char *sig_start = sig;
+    int sum_diff_thresh;
     int r;
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 1 : 0;
     __m128i acc_diff = _mm_setzero_si128();
     const __m128i k_0 = _mm_setzero_si128();
-    const __m128i k_4 = _mm_set1_epi8(4);
+    const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
     const __m128i k_8 = _mm_set1_epi8(8);
     const __m128i k_16 = _mm_set1_epi8(16);
     /* Modify each level's adjustment according to motion_magnitude. */
     const __m128i l3 = _mm_set1_epi8(
-                      (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 : 6);
+                       (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ?
+                        7 + shift_inc : 6);
     /* Difference between level 3 and level 2 is 2. */
     const __m128i l32 = _mm_set1_epi8(2);
     /* Difference between level 2 and level 1 is 1. */
@@ -108,13 +110,14 @@
                  + s.e[6] + s.e[7] + s.e[8] + s.e[9] + s.e[10] + s.e[11]
                  + s.e[12] + s.e[13] + s.e[14] + s.e[15];
 
-        if (abs(sum_diff) > SUM_DIFF_THRESHOLD)
+        sum_diff_thresh = SUM_DIFF_THRESHOLD;
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
+        if (abs(sum_diff) > sum_diff_thresh)
         {
             return COPY_BLOCK;
         }
     }
 
-    vp8_copy_mem16x16(running_avg->y_buffer + y_offset, avg_y_stride,
-                      signal->thismb, sig_stride);
+    vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
     return FILTER_BLOCK;
 }
diff --git a/vp8/vp8_common.mk b/vp8/vp8_common.mk
index e58cdad..8282547 100644
--- a/vp8/vp8_common.mk
+++ b/vp8/vp8_common.mk
@@ -158,11 +158,13 @@
 VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/vp8_variance_halfpixvar16x16_hv_armv6$(ASM)
 
 # common (neon)
-VP8_COMMON_SRCS-$(ARCH_NEON_ASM)  += common/arm/reconintra_arm.c
+#VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/reconintra_arm.c
 VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/loopfilter_neon$(ASM)
 VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/loopfiltersimpleverticaledge_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/buildintrapredictorsmby_neon$(ASM)
+#VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/buildintrapredictorsmby_neon$(ASM)
 VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/idct_blk_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/idct_dequant_0_2x_neon$(ASM)
+VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/idct_dequant_full_2x_neon$(ASM)
 VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/vp8_subpixelvariance8x8_neon$(ASM)
 VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/vp8_subpixelvariance16x16_neon$(ASM)
 VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/vp8_subpixelvariance16x16s_neon$(ASM)
@@ -173,14 +175,12 @@
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/dc_only_idct_add_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/dequant_idct_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/dequantizeb_neon.c
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/idct_dequant_full_2x_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/iwalsh_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/loopfiltersimplehorizontaledge_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/mbloopfilter_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sad_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/shortidct4x4llm_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sixtappredict_neon.c
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/idct_dequant_0_2x_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/variance_neon.c
 
 $(eval $(call rtcd_h_template,vp8_rtcd,vp8/common/rtcd_defs.pl))
diff --git a/vp9/common/vp9_common.h b/vp9/common/vp9_common.h
index 2dccb70..04db7c0 100644
--- a/vp9/common/vp9_common.h
+++ b/vp9/common/vp9_common.h
@@ -45,7 +45,7 @@
     vpx_memcpy(dest, src, n * sizeof(*src)); \
   }
 
-#define vp9_zero(dest) vpx_memset(&dest, 0, sizeof(dest))
+#define vp9_zero(dest) vpx_memset(&(dest), 0, sizeof(dest))
 #define vp9_zero_array(dest, n) vpx_memset(dest, 0, n * sizeof(*dest))
 
 static INLINE uint8_t clip_pixel(int val) {
diff --git a/vp9/common/vp9_debugmodes.c b/vp9/common/vp9_debugmodes.c
index 8f150a4..d2522bb 100644
--- a/vp9/common/vp9_debugmodes.c
+++ b/vp9/common/vp9_debugmodes.c
@@ -24,10 +24,9 @@
  */
 static void print_mi_data(VP9_COMMON *cm, FILE *file, const char *descriptor,
                           size_t member_offset) {
-  int mi_row;
-  int mi_col;
+  int mi_row, mi_col;
   int mi_index = 0;
-  MODE_INFO **mi_8x8 = cm->mi_grid_visible;
+  MODE_INFO **mi = cm->mi_grid_visible;
   int rows = cm->mi_rows;
   int cols = cm->mi_cols;
   char prefix = descriptor[0];
@@ -38,7 +37,7 @@
     fprintf(file, "%c ", prefix);
     for (mi_col = 0; mi_col < cols; mi_col++) {
       fprintf(file, "%2d ",
-              *((int*) ((char *) (&mi_8x8[mi_index]->mbmi) +
+              *((int*) ((char *) (&mi[mi_index]->mbmi) +
                         member_offset)));
       mi_index++;
     }
@@ -52,7 +51,7 @@
   int mi_col;
   int mi_index = 0;
   FILE *mvs = fopen(file, "a");
-  MODE_INFO **mi_8x8 = cm->mi_grid_visible;
+  MODE_INFO **mi = cm->mi_grid_visible;
   int rows = cm->mi_rows;
   int cols = cm->mi_cols;
 
@@ -67,8 +66,8 @@
   for (mi_row = 0; mi_row < rows; mi_row++) {
     fprintf(mvs, "V ");
     for (mi_col = 0; mi_col < cols; mi_col++) {
-      fprintf(mvs, "%4d:%4d ", mi_8x8[mi_index]->mbmi.mv[0].as_mv.row,
-                               mi_8x8[mi_index]->mbmi.mv[0].as_mv.col);
+      fprintf(mvs, "%4d:%4d ", mi[mi_index]->mbmi.mv[0].as_mv.row,
+                               mi[mi_index]->mbmi.mv[0].as_mv.col);
       mi_index++;
     }
     fprintf(mvs, "\n");
diff --git a/vp9/common/vp9_filter.c b/vp9/common/vp9_filter.c
index 7474a88..afcdf22 100644
--- a/vp9/common/vp9_filter.c
+++ b/vp9/common/vp9_filter.c
@@ -32,7 +32,8 @@
 };
 
 // Lagrangian interpolation filter
-const InterpKernel vp9_sub_pel_filters_8[SUBPEL_SHIFTS] = {
+DECLARE_ALIGNED(256, const InterpKernel,
+                vp9_sub_pel_filters_8[SUBPEL_SHIFTS]) = {
   { 0,   0,   0, 128,   0,   0,   0,  0},
   { 0,   1,  -5, 126,   8,  -3,   1,  0},
   { -1,   3, -10, 122,  18,  -6,   2,  0},
@@ -52,7 +53,8 @@
 };
 
 // DCT based filter
-const InterpKernel vp9_sub_pel_filters_8s[SUBPEL_SHIFTS] = {
+DECLARE_ALIGNED(256, const InterpKernel,
+                vp9_sub_pel_filters_8s[SUBPEL_SHIFTS]) = {
   {0,   0,   0, 128,   0,   0,   0, 0},
   {-1,   3,  -7, 127,   8,  -3,   1, 0},
   {-2,   5, -13, 125,  17,  -6,   3, -1},
@@ -72,7 +74,8 @@
 };
 
 // freqmultiplier = 0.5
-const InterpKernel vp9_sub_pel_filters_8lp[SUBPEL_SHIFTS] = {
+DECLARE_ALIGNED(256, const InterpKernel,
+                vp9_sub_pel_filters_8lp[SUBPEL_SHIFTS]) = {
   { 0,  0,  0, 128,  0,  0,  0,  0},
   {-3, -1, 32,  64, 38,  1, -3,  0},
   {-2, -2, 29,  63, 41,  2, -3,  0},
diff --git a/vp9/common/vp9_filter.h b/vp9/common/vp9_filter.h
index 29d3867..8c359c7 100644
--- a/vp9/common/vp9_filter.h
+++ b/vp9/common/vp9_filter.h
@@ -41,12 +41,6 @@
 
 DECLARE_ALIGNED(256, extern const InterpKernel,
                 vp9_bilinear_filters[SUBPEL_SHIFTS]);
-DECLARE_ALIGNED(256, extern const InterpKernel,
-                vp9_sub_pel_filters_8[SUBPEL_SHIFTS]);
-DECLARE_ALIGNED(256, extern const InterpKernel,
-                vp9_sub_pel_filters_8s[SUBPEL_SHIFTS]);
-DECLARE_ALIGNED(256, extern const InterpKernel,
-                vp9_sub_pel_filters_8lp[SUBPEL_SHIFTS]);
 
 // The VP9_BILINEAR_FILTERS_2TAP macro returns a pointer to the bilinear
 // filter kernel as a 2 tap filter.
diff --git a/vp9/common/vp9_loopfilter.c b/vp9/common/vp9_loopfilter.c
index 5b43e23..efd0249 100644
--- a/vp9/common/vp9_loopfilter.c
+++ b/vp9/common/vp9_loopfilter.c
@@ -619,12 +619,12 @@
 // by mi_row, mi_col.
 // TODO(JBB): This function only works for yv12.
 void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
-                    MODE_INFO **mi_8x8, const int mode_info_stride,
+                    MODE_INFO **mi, const int mode_info_stride,
                     LOOP_FILTER_MASK *lfm) {
   int idx_32, idx_16, idx_8;
   const loop_filter_info_n *const lfi_n = &cm->lf_info;
-  MODE_INFO **mip = mi_8x8;
-  MODE_INFO **mip2 = mi_8x8;
+  MODE_INFO **mip = mi;
+  MODE_INFO **mip2 = mi;
 
   // These are offsets to the next mi in the 64x64 block. It is what gets
   // added to the mi ptr as we go through each loop.  It helps us to avoids
@@ -1192,32 +1192,33 @@
 }
 
 void vp9_loop_filter_rows(const YV12_BUFFER_CONFIG *frame_buffer,
-                          VP9_COMMON *cm, MACROBLOCKD *xd,
+                          VP9_COMMON *cm,
+                          struct macroblockd_plane planes[MAX_MB_PLANE],
                           int start, int stop, int y_only) {
   const int num_planes = y_only ? 1 : MAX_MB_PLANE;
-  int mi_row, mi_col;
+  const int use_420 = y_only || (planes[1].subsampling_y == 1 &&
+                                 planes[1].subsampling_x == 1);
   LOOP_FILTER_MASK lfm;
-  int use_420 = y_only || (xd->plane[1].subsampling_y == 1 &&
-      xd->plane[1].subsampling_x == 1);
+  int mi_row, mi_col;
 
   for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) {
-    MODE_INFO **mi_8x8 = cm->mi_grid_visible + mi_row * cm->mi_stride;
+    MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
 
     for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
       int plane;
 
-      vp9_setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
+      vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
 
       // TODO(JBB): Make setup_mask work for non 420.
       if (use_420)
-        vp9_setup_mask(cm, mi_row, mi_col, mi_8x8 + mi_col, cm->mi_stride,
+        vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
                        &lfm);
 
       for (plane = 0; plane < num_planes; ++plane) {
         if (use_420)
-          vp9_filter_block_plane(cm, &xd->plane[plane], mi_row, &lfm);
+          vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm);
         else
-          filter_block_plane_non420(cm, &xd->plane[plane], mi_8x8 + mi_col,
+          filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
                                     mi_row, mi_col);
       }
     }
@@ -1239,7 +1240,7 @@
   }
   end_mi_row = start_mi_row + mi_rows_to_filter;
   vp9_loop_filter_frame_init(cm, frame_filter_level);
-  vp9_loop_filter_rows(frame, cm, xd,
+  vp9_loop_filter_rows(frame, cm, xd->plane,
                        start_mi_row, end_mi_row,
                        y_only);
 }
@@ -1247,7 +1248,7 @@
 int vp9_loop_filter_worker(void *arg1, void *arg2) {
   LFWorkerData *const lf_data = (LFWorkerData*)arg1;
   (void)arg2;
-  vp9_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, &lf_data->xd,
+  vp9_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
                        lf_data->start, lf_data->stop, lf_data->y_only);
   return 1;
 }
diff --git a/vp9/common/vp9_loopfilter.h b/vp9/common/vp9_loopfilter.h
index 83463c5..6fa2773 100644
--- a/vp9/common/vp9_loopfilter.h
+++ b/vp9/common/vp9_loopfilter.h
@@ -112,15 +112,15 @@
 
 // Apply the loop filter to [start, stop) macro block rows in frame_buffer.
 void vp9_loop_filter_rows(const YV12_BUFFER_CONFIG *frame_buffer,
-                          struct VP9Common *cm, struct macroblockd *xd,
+                          struct VP9Common *cm,
+                          struct macroblockd_plane planes[MAX_MB_PLANE],
                           int start, int stop, int y_only);
 
 typedef struct LoopFilterWorkerData {
   const YV12_BUFFER_CONFIG *frame_buffer;
   struct VP9Common *cm;
-  struct macroblockd xd;  // TODO(jzern): most of this is unnecessary to the
-                          // loopfilter. the planes are necessary as their state
-                          // is changed during decode.
+  struct macroblockd_plane planes[MAX_MB_PLANE];
+
   int start;
   int stop;
   int y_only;
diff --git a/vp9/common/vp9_postproc.c b/vp9/common/vp9_postproc.c
index 5601a93..9f32104 100644
--- a/vp9/common/vp9_postproc.c
+++ b/vp9/common/vp9_postproc.c
@@ -24,61 +24,7 @@
 #include "vp9/common/vp9_systemdependent.h"
 #include "vp9/common/vp9_textblit.h"
 
-#define RGB_TO_YUV(t)                                            \
-  ( (0.257*(float)(t >> 16))  + (0.504*(float)(t >> 8 & 0xff)) + \
-    (0.098*(float)(t & 0xff)) + 16),                             \
-  (-(0.148*(float)(t >> 16))  - (0.291*(float)(t >> 8 & 0xff)) + \
-    (0.439*(float)(t & 0xff)) + 128),                            \
-  ( (0.439*(float)(t >> 16))  - (0.368*(float)(t >> 8 & 0xff)) - \
-    (0.071*(float)(t & 0xff)) + 128)
-
-/* global constants */
-#if 0 && CONFIG_POSTPROC_VISUALIZER
-static const unsigned char PREDICTION_MODE_colors[MB_MODE_COUNT][3] = {
-  { RGB_TO_YUV(0x98FB98) },   /* PaleGreen */
-  { RGB_TO_YUV(0x00FF00) },   /* Green */
-  { RGB_TO_YUV(0xADFF2F) },   /* GreenYellow */
-  { RGB_TO_YUV(0x8F0000) },   /* Dark Red */
-  { RGB_TO_YUV(0x008F8F) },   /* Dark Cyan */
-  { RGB_TO_YUV(0x008F8F) },   /* Dark Cyan */
-  { RGB_TO_YUV(0x008F8F) },   /* Dark Cyan */
-  { RGB_TO_YUV(0x8F0000) },   /* Dark Red */
-  { RGB_TO_YUV(0x8F0000) },   /* Dark Red */
-  { RGB_TO_YUV(0x228B22) },   /* ForestGreen */
-  { RGB_TO_YUV(0x006400) },   /* DarkGreen */
-  { RGB_TO_YUV(0x98F5FF) },   /* Cadet Blue */
-  { RGB_TO_YUV(0x6CA6CD) },   /* Sky Blue */
-  { RGB_TO_YUV(0x00008B) },   /* Dark blue */
-  { RGB_TO_YUV(0x551A8B) },   /* Purple */
-  { RGB_TO_YUV(0xFF0000) }    /* Red */
-  { RGB_TO_YUV(0xCC33FF) },   /* Magenta */
-};
-
-static const unsigned char B_PREDICTION_MODE_colors[INTRA_MODES][3] = {
-  { RGB_TO_YUV(0x6633ff) },   /* Purple */
-  { RGB_TO_YUV(0xcc33ff) },   /* Magenta */
-  { RGB_TO_YUV(0xff33cc) },   /* Pink */
-  { RGB_TO_YUV(0xff3366) },   /* Coral */
-  { RGB_TO_YUV(0x3366ff) },   /* Blue */
-  { RGB_TO_YUV(0xed00f5) },   /* Dark Blue */
-  { RGB_TO_YUV(0x2e00b8) },   /* Dark Purple */
-  { RGB_TO_YUV(0xff6633) },   /* Orange */
-  { RGB_TO_YUV(0x33ccff) },   /* Light Blue */
-  { RGB_TO_YUV(0x8ab800) },   /* Green */
-  { RGB_TO_YUV(0xffcc33) },   /* Light Orange */
-  { RGB_TO_YUV(0x33ffcc) },   /* Aqua */
-  { RGB_TO_YUV(0x66ff33) },   /* Light Green */
-  { RGB_TO_YUV(0xccff33) },   /* Yellow */
-};
-
-static const unsigned char MV_REFERENCE_FRAME_colors[MAX_REF_FRAMES][3] = {
-  { RGB_TO_YUV(0x00ff00) },   /* Blue */
-  { RGB_TO_YUV(0x0000ff) },   /* Green */
-  { RGB_TO_YUV(0xffff00) },   /* Yellow */
-  { RGB_TO_YUV(0xff0000) },   /* Red */
-};
-#endif
-
+#if CONFIG_VP9_POSTPROC
 static const short kernel5[] = {
   1, 1, 4, 1, 1
 };
@@ -448,163 +394,6 @@
   }
 }
 
-/* Blend the macro block with a solid colored square.  Leave the
- * edges unblended to give distinction to macro blocks in areas
- * filled with the same color block.
- */
-void vp9_blend_mb_inner_c(uint8_t *y, uint8_t *u, uint8_t *v,
-                          int y1, int u1, int v1, int alpha, int stride) {
-  int i, j;
-  int y1_const = y1 * ((1 << 16) - alpha);
-  int u1_const = u1 * ((1 << 16) - alpha);
-  int v1_const = v1 * ((1 << 16) - alpha);
-
-  y += 2 * stride + 2;
-  for (i = 0; i < 12; i++) {
-    for (j = 0; j < 12; j++) {
-      y[j] = (y[j] * alpha + y1_const) >> 16;
-    }
-    y += stride;
-  }
-
-  stride >>= 1;
-
-  u += stride + 1;
-  v += stride + 1;
-
-  for (i = 0; i < 6; i++) {
-    for (j = 0; j < 6; j++) {
-      u[j] = (u[j] * alpha + u1_const) >> 16;
-      v[j] = (v[j] * alpha + v1_const) >> 16;
-    }
-    u += stride;
-    v += stride;
-  }
-}
-
-/* Blend only the edge of the macro block.  Leave center
- * unblended to allow for other visualizations to be layered.
- */
-void vp9_blend_mb_outer_c(uint8_t *y, uint8_t *u, uint8_t *v,
-                          int y1, int u1, int v1, int alpha, int stride) {
-  int i, j;
-  int y1_const = y1 * ((1 << 16) - alpha);
-  int u1_const = u1 * ((1 << 16) - alpha);
-  int v1_const = v1 * ((1 << 16) - alpha);
-
-  for (i = 0; i < 2; i++) {
-    for (j = 0; j < 16; j++) {
-      y[j] = (y[j] * alpha + y1_const) >> 16;
-    }
-    y += stride;
-  }
-
-  for (i = 0; i < 12; i++) {
-    y[0]  = (y[0] * alpha  + y1_const) >> 16;
-    y[1]  = (y[1] * alpha  + y1_const) >> 16;
-    y[14] = (y[14] * alpha + y1_const) >> 16;
-    y[15] = (y[15] * alpha + y1_const) >> 16;
-    y += stride;
-  }
-
-  for (i = 0; i < 2; i++) {
-    for (j = 0; j < 16; j++) {
-      y[j] = (y[j] * alpha + y1_const) >> 16;
-    }
-    y += stride;
-  }
-
-  stride >>= 1;
-
-  for (j = 0; j < 8; j++) {
-    u[j] = (u[j] * alpha + u1_const) >> 16;
-    v[j] = (v[j] * alpha + v1_const) >> 16;
-  }
-  u += stride;
-  v += stride;
-
-  for (i = 0; i < 6; i++) {
-    u[0] = (u[0] * alpha + u1_const) >> 16;
-    v[0] = (v[0] * alpha + v1_const) >> 16;
-
-    u[7] = (u[7] * alpha + u1_const) >> 16;
-    v[7] = (v[7] * alpha + v1_const) >> 16;
-
-    u += stride;
-    v += stride;
-  }
-
-  for (j = 0; j < 8; j++) {
-    u[j] = (u[j] * alpha + u1_const) >> 16;
-    v[j] = (v[j] * alpha + v1_const) >> 16;
-  }
-}
-
-void vp9_blend_b_c(uint8_t *y, uint8_t *u, uint8_t *v,
-                   int y1, int u1, int v1, int alpha, int stride) {
-  int i, j;
-  int y1_const = y1 * ((1 << 16) - alpha);
-  int u1_const = u1 * ((1 << 16) - alpha);
-  int v1_const = v1 * ((1 << 16) - alpha);
-
-  for (i = 0; i < 4; i++) {
-    for (j = 0; j < 4; j++) {
-      y[j] = (y[j] * alpha + y1_const) >> 16;
-    }
-    y += stride;
-  }
-
-  stride >>= 1;
-
-  for (i = 0; i < 2; i++) {
-    for (j = 0; j < 2; j++) {
-      u[j] = (u[j] * alpha + u1_const) >> 16;
-      v[j] = (v[j] * alpha + v1_const) >> 16;
-    }
-    u += stride;
-    v += stride;
-  }
-}
-
-static void constrain_line(int x0, int *x1, int y0, int *y1,
-                           int width, int height) {
-  int dx;
-  int dy;
-
-  if (*x1 > width) {
-    dx = *x1 - x0;
-    dy = *y1 - y0;
-
-    *x1 = width;
-    if (dx)
-      *y1 = ((width - x0) * dy) / dx + y0;
-  }
-  if (*x1 < 0) {
-    dx = *x1 - x0;
-    dy = *y1 - y0;
-
-    *x1 = 0;
-    if (dx)
-      *y1 = ((0 - x0) * dy) / dx + y0;
-  }
-  if (*y1 > height) {
-    dx = *x1 - x0;
-    dy = *y1 - y0;
-
-    *y1 = height;
-    if (dy)
-      *x1 = ((height - y0) * dx) / dy + x0;
-  }
-  if (*y1 < 0) {
-    dx = *x1 - x0;
-    dy = *y1 - y0;
-
-    *y1 = 0;
-    if (dy)
-      *x1 = ((0 - y0) * dx) / dy + x0;
-  }
-}
-
 int vp9_post_proc_frame(struct VP9Common *cm,
                         YV12_BUFFER_CONFIG *dest, vp9_ppflags_t *ppflags) {
   const int q = MIN(63, cm->lf.filter_level * 10 / 6);
@@ -643,328 +432,6 @@
                         ppbuf->y_width, ppbuf->y_height, ppbuf->y_stride);
   }
 
-#if 0 && CONFIG_POSTPROC_VISUALIZER
-  if (flags & VP9D_DEBUG_TXT_FRAME_INFO) {
-    char message[512];
-    snprintf(message, sizeof(message) -1,
-             "F%1dG%1dQ%3dF%3dP%d_s%dx%d",
-             (cm->frame_type == KEY_FRAME),
-             cm->refresh_golden_frame,
-             cm->base_qindex,
-             cm->filter_level,
-             flags,
-             cm->mb_cols, cm->mb_rows);
-    vp9_blit_text(message, ppbuf->y_buffer, ppbuf->y_stride);
-  }
-
-  if (flags & VP9D_DEBUG_TXT_MBLK_MODES) {
-    int i, j;
-    uint8_t *y_ptr;
-    int mb_rows = ppbuf->y_height >> 4;
-    int mb_cols = ppbuf->y_width  >> 4;
-    int mb_index = 0;
-    MODE_INFO *mi = cm->mi;
-
-    y_ptr = post->y_buffer + 4 * post->y_stride + 4;
-
-    /* vp9_filter each macro block */
-    for (i = 0; i < mb_rows; i++) {
-      for (j = 0; j < mb_cols; j++) {
-        char zz[4];
-
-        snprintf(zz, sizeof(zz) - 1, "%c", mi[mb_index].mbmi.mode + 'a');
-
-        vp9_blit_text(zz, y_ptr, post->y_stride);
-        mb_index++;
-        y_ptr += 16;
-      }
-
-      mb_index++; /* border */
-      y_ptr += post->y_stride  * 16 - post->y_width;
-    }
-  }
-
-  if (flags & VP9D_DEBUG_TXT_DC_DIFF) {
-    int i, j;
-    uint8_t *y_ptr;
-    int mb_rows = ppbuf->y_height >> 4;
-    int mb_cols = ppbuf->y_width  >> 4;
-    int mb_index = 0;
-    MODE_INFO *mi = cm->mi;
-
-    y_ptr = post->y_buffer + 4 * post->y_stride + 4;
-
-    /* vp9_filter each macro block */
-    for (i = 0; i < mb_rows; i++) {
-      for (j = 0; j < mb_cols; j++) {
-        char zz[4];
-        int dc_diff = !(mi[mb_index].mbmi.mode != I4X4_PRED &&
-                        mi[mb_index].mbmi.mode != SPLITMV &&
-                        mi[mb_index].mbmi.skip);
-
-        if (cm->frame_type == KEY_FRAME)
-          snprintf(zz, sizeof(zz) - 1, "a");
-        else
-          snprintf(zz, sizeof(zz) - 1, "%c", dc_diff + '0');
-
-        vp9_blit_text(zz, y_ptr, post->y_stride);
-        mb_index++;
-        y_ptr += 16;
-      }
-
-      mb_index++; /* border */
-      y_ptr += post->y_stride  * 16 - post->y_width;
-    }
-  }
-
-  if (flags & VP9D_DEBUG_TXT_RATE_INFO) {
-    char message[512];
-    snprintf(message, sizeof(message),
-             "Bitrate: %10.2f framerate: %10.2f ",
-             cm->bitrate, cm->framerate);
-    vp9_blit_text(message, ppbuf->y_buffer, ppbuf->y_stride);
-  }
-
-  /* Draw motion vectors */
-  if ((flags & VP9D_DEBUG_DRAW_MV) && ppflags->display_mv_flag) {
-    int width  = ppbuf->y_width;
-    int height = ppbuf->y_height;
-    uint8_t *y_buffer = ppbuf->y_buffer;
-    int y_stride = ppbuf->y_stride;
-    MODE_INFO *mi = cm->mi;
-    int x0, y0;
-
-    for (y0 = 0; y0 < height; y0 += 16) {
-      for (x0 = 0; x0 < width; x0 += 16) {
-        int x1, y1;
-
-        if (!(ppflags->display_mv_flag & (1 << mi->mbmi.mode))) {
-          mi++;
-          continue;
-        }
-
-        if (mi->mbmi.mode == SPLITMV) {
-          switch (mi->mbmi.partitioning) {
-            case PARTITIONING_16X8 : {  /* mv_top_bottom */
-              union b_mode_info *bmi = &mi->bmi[0];
-              MV *mv = &bmi->mv.as_mv;
-
-              x1 = x0 + 8 + (mv->col >> 3);
-              y1 = y0 + 4 + (mv->row >> 3);
-
-              constrain_line(x0 + 8, &x1, y0 + 4, &y1, width, height);
-              vp9_blit_line(x0 + 8,  x1, y0 + 4,  y1, y_buffer, y_stride);
-
-              bmi = &mi->bmi[8];
-
-              x1 = x0 + 8 + (mv->col >> 3);
-              y1 = y0 + 12 + (mv->row >> 3);
-
-              constrain_line(x0 + 8, &x1, y0 + 12, &y1, width, height);
-              vp9_blit_line(x0 + 8,  x1, y0 + 12,  y1, y_buffer, y_stride);
-
-              break;
-            }
-            case PARTITIONING_8X16 : {  /* mv_left_right */
-              union b_mode_info *bmi = &mi->bmi[0];
-              MV *mv = &bmi->mv.as_mv;
-
-              x1 = x0 + 4 + (mv->col >> 3);
-              y1 = y0 + 8 + (mv->row >> 3);
-
-              constrain_line(x0 + 4, &x1, y0 + 8, &y1, width, height);
-              vp9_blit_line(x0 + 4,  x1, y0 + 8,  y1, y_buffer, y_stride);
-
-              bmi = &mi->bmi[2];
-
-              x1 = x0 + 12 + (mv->col >> 3);
-              y1 = y0 + 8 + (mv->row >> 3);
-
-              constrain_line(x0 + 12, &x1, y0 + 8, &y1, width, height);
-              vp9_blit_line(x0 + 12,  x1, y0 + 8,  y1, y_buffer, y_stride);
-
-              break;
-            }
-            case PARTITIONING_8X8 : {  /* mv_quarters   */
-              union b_mode_info *bmi = &mi->bmi[0];
-              MV *mv = &bmi->mv.as_mv;
-
-              x1 = x0 + 4 + (mv->col >> 3);
-              y1 = y0 + 4 + (mv->row >> 3);
-
-              constrain_line(x0 + 4, &x1, y0 + 4, &y1, width, height);
-              vp9_blit_line(x0 + 4,  x1, y0 + 4,  y1, y_buffer, y_stride);
-
-              bmi = &mi->bmi[2];
-
-              x1 = x0 + 12 + (mv->col >> 3);
-              y1 = y0 + 4 + (mv->row >> 3);
-
-              constrain_line(x0 + 12, &x1, y0 + 4, &y1, width, height);
-              vp9_blit_line(x0 + 12,  x1, y0 + 4,  y1, y_buffer, y_stride);
-
-              bmi = &mi->bmi[8];
-
-              x1 = x0 + 4 + (mv->col >> 3);
-              y1 = y0 + 12 + (mv->row >> 3);
-
-              constrain_line(x0 + 4, &x1, y0 + 12, &y1, width, height);
-              vp9_blit_line(x0 + 4,  x1, y0 + 12,  y1, y_buffer, y_stride);
-
-              bmi = &mi->bmi[10];
-
-              x1 = x0 + 12 + (mv->col >> 3);
-              y1 = y0 + 12 + (mv->row >> 3);
-
-              constrain_line(x0 + 12, &x1, y0 + 12, &y1, width, height);
-              vp9_blit_line(x0 + 12,  x1, y0 + 12,  y1, y_buffer, y_stride);
-              break;
-            }
-            case PARTITIONING_4X4:
-            default : {
-              union b_mode_info *bmi = mi->bmi;
-              int bx0, by0;
-
-              for (by0 = y0; by0 < (y0 + 16); by0 += 4) {
-                for (bx0 = x0; bx0 < (x0 + 16); bx0 += 4) {
-                  MV *mv = &bmi->mv.as_mv;
-
-                  x1 = bx0 + 2 + (mv->col >> 3);
-                  y1 = by0 + 2 + (mv->row >> 3);
-
-                  constrain_line(bx0 + 2, &x1, by0 + 2, &y1, width, height);
-                  vp9_blit_line(bx0 + 2,  x1, by0 + 2,  y1, y_buffer, y_stride);
-
-                  bmi++;
-                }
-              }
-            }
-          }
-        } else if (is_inter_mode(mi->mbmi.mode)) {
-          MV *mv = &mi->mbmi.mv.as_mv;
-          const int lx0 = x0 + 8;
-          const int ly0 = y0 + 8;
-
-          x1 = lx0 + (mv->col >> 3);
-          y1 = ly0 + (mv->row >> 3);
-
-          if (x1 != lx0 && y1 != ly0) {
-            constrain_line(lx0, &x1, ly0 - 1, &y1, width, height);
-            vp9_blit_line(lx0,  x1, ly0 - 1,  y1, y_buffer, y_stride);
-
-            constrain_line(lx0, &x1, ly0 + 1, &y1, width, height);
-            vp9_blit_line(lx0,  x1, ly0 + 1,  y1, y_buffer, y_stride);
-          } else {
-            vp9_blit_line(lx0,  x1, ly0,  y1, y_buffer, y_stride);
-          }
-        }
-
-        mi++;
-      }
-      mi++;
-    }
-  }
-
-  /* Color in block modes */
-  if ((flags & VP9D_DEBUG_CLR_BLK_MODES)
-      && (ppflags->display_mb_modes_flag || ppflags->display_b_modes_flag)) {
-    int y, x;
-    int width  = ppbuf->y_width;
-    int height = ppbuf->y_height;
-    uint8_t *y_ptr = ppbuf->y_buffer;
-    uint8_t *u_ptr = ppbuf->u_buffer;
-    uint8_t *v_ptr = ppbuf->v_buffer;
-    int y_stride = ppbuf->y_stride;
-    MODE_INFO *mi = cm->mi;
-
-    for (y = 0; y < height; y += 16) {
-      for (x = 0; x < width; x += 16) {
-        int Y = 0, U = 0, V = 0;
-
-        if (mi->mbmi.mode == I4X4_PRED &&
-            ((ppflags->display_mb_modes_flag & I4X4_PRED) ||
-             ppflags->display_b_modes_flag)) {
-          int by, bx;
-          uint8_t *yl, *ul, *vl;
-          union b_mode_info *bmi = mi->bmi;
-
-          yl = y_ptr + x;
-          ul = u_ptr + (x >> 1);
-          vl = v_ptr + (x >> 1);
-
-          for (by = 0; by < 16; by += 4) {
-            for (bx = 0; bx < 16; bx += 4) {
-              if ((ppflags->display_b_modes_flag & (1 << mi->mbmi.mode))
-                  || (ppflags->display_mb_modes_flag & I4X4_PRED)) {
-                Y = B_PREDICTION_MODE_colors[bmi->as_mode][0];
-                U = B_PREDICTION_MODE_colors[bmi->as_mode][1];
-                V = B_PREDICTION_MODE_colors[bmi->as_mode][2];
-
-                vp9_blend_b(yl + bx, ul + (bx >> 1), vl + (bx >> 1), Y, U, V,
-                    0xc000, y_stride);
-              }
-              bmi++;
-            }
-
-            yl += y_stride * 4;
-            ul += y_stride * 1;
-            vl += y_stride * 1;
-          }
-        } else if (ppflags->display_mb_modes_flag & (1 << mi->mbmi.mode)) {
-          Y = PREDICTION_MODE_colors[mi->mbmi.mode][0];
-          U = PREDICTION_MODE_colors[mi->mbmi.mode][1];
-          V = PREDICTION_MODE_colors[mi->mbmi.mode][2];
-
-          vp9_blend_mb_inner(y_ptr + x, u_ptr + (x >> 1), v_ptr + (x >> 1),
-                             Y, U, V, 0xc000, y_stride);
-        }
-
-        mi++;
-      }
-      y_ptr += y_stride * 16;
-      u_ptr += y_stride * 4;
-      v_ptr += y_stride * 4;
-
-      mi++;
-    }
-  }
-
-  /* Color in frame reference blocks */
-  if ((flags & VP9D_DEBUG_CLR_FRM_REF_BLKS) &&
-      ppflags->display_ref_frame_flag) {
-    int y, x;
-    int width  = ppbuf->y_width;
-    int height = ppbuf->y_height;
-    uint8_t *y_ptr = ppbuf->y_buffer;
-    uint8_t *u_ptr = ppbuf->u_buffer;
-    uint8_t *v_ptr = ppbuf->v_buffer;
-    int y_stride = ppbuf->y_stride;
-    MODE_INFO *mi = cm->mi;
-
-    for (y = 0; y < height; y += 16) {
-      for (x = 0; x < width; x += 16) {
-        int Y = 0, U = 0, V = 0;
-
-        if (ppflags->display_ref_frame_flag & (1 << mi->mbmi.ref_frame)) {
-          Y = MV_REFERENCE_FRAME_colors[mi->mbmi.ref_frame][0];
-          U = MV_REFERENCE_FRAME_colors[mi->mbmi.ref_frame][1];
-          V = MV_REFERENCE_FRAME_colors[mi->mbmi.ref_frame][2];
-
-          vp9_blend_mb_outer(y_ptr + x, u_ptr + (x >> 1), v_ptr + (x >> 1),
-                             Y, U, V, 0xc000, y_stride);
-        }
-
-        mi++;
-      }
-      y_ptr += y_stride * 16;
-      u_ptr += y_stride * 4;
-      v_ptr += y_stride * 4;
-
-      mi++;
-    }
-  }
-#endif
-
   *dest = *ppbuf;
 
   /* handle problem with extending borders */
@@ -975,3 +442,4 @@
 
   return 0;
 }
+#endif
diff --git a/vp9/common/vp9_ppflags.h b/vp9/common/vp9_ppflags.h
index e8b04d2..1644a1b 100644
--- a/vp9/common/vp9_ppflags.h
+++ b/vp9/common/vp9_ppflags.h
@@ -33,12 +33,6 @@
   int post_proc_flag;
   int deblocking_level;
   int noise_level;
-#if CONFIG_POSTPROC_VISUALIZER
-  int display_ref_frame_flag;
-  int display_mb_modes_flag;
-  int display_b_modes_flag;
-  int display_mv_flag;
-#endif  // CONFIG_POSTPROC_VISUALIZER
 } vp9_ppflags_t;
 
 #ifdef __cplusplus
diff --git a/vp9/common/vp9_reconinter.c b/vp9/common/vp9_reconinter.c
index e722d6a..edc36d7 100644
--- a/vp9/common/vp9_reconinter.c
+++ b/vp9/common/vp9_reconinter.c
@@ -409,7 +409,7 @@
   }
 }
 
-void vp9_setup_dst_planes(MACROBLOCKD *xd,
+void vp9_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
                           const YV12_BUFFER_CONFIG *src,
                           int mi_row, int mi_col) {
   uint8_t *const buffers[4] = {src->y_buffer, src->u_buffer, src->v_buffer,
@@ -419,7 +419,7 @@
   int i;
 
   for (i = 0; i < MAX_MB_PLANE; ++i) {
-    struct macroblockd_plane *const pd = &xd->plane[i];
+    struct macroblockd_plane *const pd = &planes[i];
     setup_pred_plane(&pd->dst, buffers[i], strides[i], mi_row, mi_col, NULL,
                      pd->subsampling_x, pd->subsampling_y);
   }
diff --git a/vp9/common/vp9_reconinter.h b/vp9/common/vp9_reconinter.h
index 86f3158..58c596e 100644
--- a/vp9/common/vp9_reconinter.h
+++ b/vp9/common/vp9_reconinter.h
@@ -57,7 +57,8 @@
   dst->stride = stride;
 }
 
-void vp9_setup_dst_planes(MACROBLOCKD *xd, const YV12_BUFFER_CONFIG *src,
+void vp9_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
                           int mi_row, int mi_col);
 
 void vp9_setup_pre_planes(MACROBLOCKD *xd, int idx,
diff --git a/vp9/common/vp9_reconintra.c b/vp9/common/vp9_reconintra.c
index 32e4551..403e105 100644
--- a/vp9/common/vp9_reconintra.c
+++ b/vp9/common/vp9_reconintra.c
@@ -31,6 +31,9 @@
   ADST_ADST,  // TM
 };
 
+// This serves as a wrapper function, so that all the prediction functions
+// can be unified and accessed as a pointer array. Note that the boundary
+// above and left are not necessarily used all the time.
 #define intra_pred_sized(type, size) \
   void vp9_##type##_predictor_##size##x##size##_c(uint8_t *dst, \
                                                   ptrdiff_t stride, \
@@ -48,7 +51,7 @@
 static INLINE void d207_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                   const uint8_t *above, const uint8_t *left) {
   int r, c;
-
+  (void) above;
   // first column
   for (r = 0; r < bs - 1; ++r)
     dst[r * stride] = ROUND_POWER_OF_TWO(left[r] + left[r + 1], 1);
@@ -77,6 +80,7 @@
 static INLINE void d63_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                  const uint8_t *above, const uint8_t *left) {
   int r, c;
+  (void) left;
   for (r = 0; r < bs; ++r) {
     for (c = 0; c < bs; ++c)
       dst[c] = r & 1 ? ROUND_POWER_OF_TWO(above[r/2 + c] +
@@ -92,6 +96,7 @@
 static INLINE void d45_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                  const uint8_t *above, const uint8_t *left) {
   int r, c;
+  (void) left;
   for (r = 0; r < bs; ++r) {
     for (c = 0; c < bs; ++c)
       dst[c] = r + c + 2 < bs * 2 ?  ROUND_POWER_OF_TWO(above[r + c] +
@@ -184,6 +189,7 @@
 static INLINE void v_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                const uint8_t *above, const uint8_t *left) {
   int r;
+  (void) left;
 
   for (r = 0; r < bs; r++) {
     vpx_memcpy(dst, above, bs);
@@ -195,6 +201,7 @@
 static INLINE void h_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                const uint8_t *above, const uint8_t *left) {
   int r;
+  (void) above;
 
   for (r = 0; r < bs; r++) {
     vpx_memset(dst, left[r], bs);
@@ -219,6 +226,8 @@
 static INLINE void dc_128_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                     const uint8_t *above, const uint8_t *left) {
   int r;
+  (void) above;
+  (void) left;
 
   for (r = 0; r < bs; r++) {
     vpx_memset(dst, 128, bs);
@@ -231,6 +240,7 @@
                                      const uint8_t *above,
                                      const uint8_t *left) {
   int i, r, expected_dc, sum = 0;
+  (void) above;
 
   for (i = 0; i < bs; i++)
     sum += left[i];
@@ -246,6 +256,7 @@
 static INLINE void dc_top_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                     const uint8_t *above, const uint8_t *left) {
   int i, r, expected_dc, sum = 0;
+  (void) left;
 
   for (i = 0; i < bs; i++)
     sum += above[i];
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 7754763..1827396 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -703,7 +703,7 @@
 # ENCODEMB INVOKE
 
 add_proto qw/int64_t vp9_block_error/, "const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size, int64_t *ssz";
-specialize qw/vp9_block_error/, "$sse2_x86inc";
+specialize qw/vp9_block_error avx2/, "$sse2_x86inc";
 
 add_proto qw/void vp9_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
 specialize qw/vp9_subtract_block/, "$sse2_x86inc";
@@ -772,7 +772,7 @@
 add_proto qw/int vp9_full_range_search/, "const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv";
 specialize qw/vp9_full_range_search/;
 
-add_proto qw/void vp9_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
+add_proto qw/void vp9_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
 specialize qw/vp9_temporal_filter_apply sse2/;
 
 }
diff --git a/vp9/common/vp9_tile_common.c b/vp9/common/vp9_tile_common.c
index 78909dd..8c4a303 100644
--- a/vp9/common/vp9_tile_common.c
+++ b/vp9/common/vp9_tile_common.c
@@ -21,13 +21,21 @@
   return MIN(offset, mis);
 }
 
-void vp9_tile_init(TileInfo *tile, const VP9_COMMON *cm, int row, int col) {
+void vp9_tile_set_row(TileInfo *tile, const VP9_COMMON *cm, int row) {
   tile->mi_row_start = get_tile_offset(row, cm->mi_rows, cm->log2_tile_rows);
   tile->mi_row_end = get_tile_offset(row + 1, cm->mi_rows, cm->log2_tile_rows);
+}
+
+void vp9_tile_set_col(TileInfo *tile, const VP9_COMMON *cm, int col) {
   tile->mi_col_start = get_tile_offset(col, cm->mi_cols, cm->log2_tile_cols);
   tile->mi_col_end = get_tile_offset(col + 1, cm->mi_cols, cm->log2_tile_cols);
 }
 
+void vp9_tile_init(TileInfo *tile, const VP9_COMMON *cm, int row, int col) {
+  vp9_tile_set_row(tile, cm, row);
+  vp9_tile_set_col(tile, cm, col);
+}
+
 void vp9_get_tile_n_bits(int mi_cols,
                          int *min_log2_tile_cols, int *max_log2_tile_cols) {
   const int sb_cols = mi_cols_aligned_to_sb(mi_cols) >> MI_BLOCK_SIZE_LOG2;
diff --git a/vp9/common/vp9_tile_common.h b/vp9/common/vp9_tile_common.h
index a97719e..ae58805 100644
--- a/vp9/common/vp9_tile_common.h
+++ b/vp9/common/vp9_tile_common.h
@@ -27,6 +27,9 @@
 void vp9_tile_init(TileInfo *tile, const struct VP9Common *cm,
                    int row, int col);
 
+void vp9_tile_set_row(TileInfo *tile, const struct VP9Common *cm, int row);
+void vp9_tile_set_col(TileInfo *tile, const struct VP9Common *cm, int col);
+
 void vp9_get_tile_n_bits(int mi_cols,
                          int *min_log2_tile_cols, int *max_log2_tile_cols);
 
diff --git a/vp9/decoder/vp9_decodeframe.c b/vp9/decoder/vp9_decodeframe.c
index 9dc0cf1..de58939 100644
--- a/vp9/decoder/vp9_decodeframe.c
+++ b/vp9/decoder/vp9_decodeframe.c
@@ -316,7 +316,7 @@
   // as they are always compared to values that are in 1/8th pel units
   set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
 
-  vp9_setup_dst_planes(xd, get_frame_new_buffer(cm), mi_row, mi_col);
+  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
   return &xd->mi[0]->mbmi;
 }
 
@@ -675,64 +675,6 @@
   setup_display_size(cm, rb);
 }
 
-static void decode_tile(VP9Decoder *pbi, const TileInfo *const tile,
-                        int do_loopfilter_inline, vp9_reader *r) {
-  const int num_threads = pbi->max_threads;
-  VP9_COMMON *const cm = &pbi->common;
-  int mi_row, mi_col;
-  MACROBLOCKD *xd = &pbi->mb;
-
-  if (do_loopfilter_inline) {
-    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
-    lf_data->frame_buffer = get_frame_new_buffer(cm);
-    lf_data->cm = cm;
-    lf_data->xd = pbi->mb;
-    lf_data->stop = 0;
-    lf_data->y_only = 0;
-    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
-  }
-
-  for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
-       mi_row += MI_BLOCK_SIZE) {
-    // For a SB there are 2 left contexts, each pertaining to a MB row within
-    vp9_zero(xd->left_context);
-    vp9_zero(xd->left_seg_context);
-    for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
-         mi_col += MI_BLOCK_SIZE) {
-      decode_partition(cm, xd, tile, mi_row, mi_col, r, BLOCK_64X64);
-    }
-
-    if (do_loopfilter_inline) {
-      const int lf_start = mi_row - MI_BLOCK_SIZE;
-      LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
-
-      // delay the loopfilter by 1 macroblock row.
-      if (lf_start < 0) continue;
-
-      // decoding has completed: finish up the loop filter in this thread.
-      if (mi_row + MI_BLOCK_SIZE >= tile->mi_row_end) continue;
-
-      vp9_worker_sync(&pbi->lf_worker);
-      lf_data->start = lf_start;
-      lf_data->stop = mi_row;
-      if (num_threads > 1) {
-        vp9_worker_launch(&pbi->lf_worker);
-      } else {
-        vp9_worker_execute(&pbi->lf_worker);
-      }
-    }
-  }
-
-  if (do_loopfilter_inline) {
-    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
-
-    vp9_worker_sync(&pbi->lf_worker);
-    lf_data->start = lf_data->stop;
-    lf_data->stop = cm->mi_rows;
-    vp9_worker_execute(&pbi->lf_worker);
-  }
-}
-
 static void setup_tile_info(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
   int min_log2_tile_cols, max_log2_tile_cols, max_ones;
   vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
@@ -811,16 +753,35 @@
 
 static const uint8_t *decode_tiles(VP9Decoder *pbi,
                                    const uint8_t *data,
-                                   const uint8_t *data_end,
-                                   int do_loopfilter_inline) {
+                                   const uint8_t *data_end) {
   VP9_COMMON *const cm = &pbi->common;
   const int aligned_cols = mi_cols_aligned_to_sb(cm->mi_cols);
   const int tile_cols = 1 << cm->log2_tile_cols;
   const int tile_rows = 1 << cm->log2_tile_rows;
   TileBuffer tile_buffers[4][1 << 6];
   int tile_row, tile_col;
-  const uint8_t *end = NULL;
-  vp9_reader r;
+  int mi_row, mi_col;
+  TileData *tile_data = NULL;
+
+  if (cm->lf.filter_level && pbi->lf_worker.data1 == NULL) {
+    CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
+                    vpx_memalign(32, sizeof(LFWorkerData)));
+    pbi->lf_worker.hook = (VP9WorkerHook)vp9_loop_filter_worker;
+    if (pbi->max_threads > 1 && !vp9_worker_reset(&pbi->lf_worker)) {
+      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                         "Loop filter thread creation failed");
+    }
+  }
+
+  if (cm->lf.filter_level) {
+    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+    lf_data->frame_buffer = get_frame_new_buffer(cm);
+    lf_data->cm = cm;
+    vp9_copy(lf_data->planes, pbi->mb.plane);
+    lf_data->stop = 0;
+    lf_data->y_only = 0;
+    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
+  }
 
   assert(tile_rows <= 4);
   assert(tile_cols <= (1 << 6));
@@ -835,26 +796,88 @@
 
   get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
 
-  // Decode tiles using data from tile_buffers
+  if (pbi->tile_data == NULL ||
+      (tile_cols * tile_rows) != pbi->total_tiles) {
+    vpx_free(pbi->tile_data);
+    CHECK_MEM_ERROR(
+        cm,
+        pbi->tile_data,
+        vpx_malloc(tile_cols * tile_rows * (sizeof(*pbi->tile_data))));
+    pbi->total_tiles = tile_rows * tile_cols;
+  }
+
+  // Load all tile information into tile_data.
   for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
     for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
-      const int col = pbi->inv_tile_order ? tile_cols - tile_col - 1 : tile_col;
-      const int last_tile = tile_row == tile_rows - 1 &&
-                                 col == tile_cols - 1;
-      const TileBuffer *const buf = &tile_buffers[tile_row][col];
       TileInfo tile;
-
-      vp9_tile_init(&tile, cm, tile_row, col);
-      setup_token_decoder(buf->data, data_end, buf->size, &cm->error, &r,
-                          pbi->decrypt_cb, pbi->decrypt_state);
-      decode_tile(pbi, &tile, do_loopfilter_inline, &r);
-
-      if (last_tile)
-        end = vp9_reader_find_end(&r);
+      const TileBuffer *const buf = &tile_buffers[tile_row][tile_col];
+      tile_data = pbi->tile_data + tile_cols * tile_row + tile_col;
+      tile_data->cm = cm;
+      tile_data->xd = pbi->mb;
+      tile_data->xd.corrupted = 0;
+      vp9_tile_init(&tile, tile_data->cm, tile_row, tile_col);
+      setup_token_decoder(buf->data, data_end, buf->size, &cm->error,
+                          &tile_data->bit_reader, pbi->decrypt_cb,
+                          pbi->decrypt_state);
+      init_macroblockd(cm, &tile_data->xd);
+      vp9_zero(tile_data->xd.dqcoeff);
     }
   }
 
-  return end;
+  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
+    TileInfo tile;
+    vp9_tile_set_row(&tile, cm, tile_row);
+    for (mi_row = tile.mi_row_start; mi_row < tile.mi_row_end;
+         mi_row += MI_BLOCK_SIZE) {
+      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+        const int col = pbi->inv_tile_order ?
+                        tile_cols - tile_col - 1 : tile_col;
+        tile_data = pbi->tile_data + tile_cols * tile_row + col;
+        vp9_tile_set_col(&tile, tile_data->cm, col);
+        vp9_zero(tile_data->xd.left_context);
+        vp9_zero(tile_data->xd.left_seg_context);
+        for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
+             mi_col += MI_BLOCK_SIZE) {
+          decode_partition(tile_data->cm, &tile_data->xd, &tile, mi_row, mi_col,
+                           &tile_data->bit_reader, BLOCK_64X64);
+        }
+      }
+      // Loopfilter one row.
+      if (cm->lf.filter_level) {
+        const int lf_start = mi_row - MI_BLOCK_SIZE;
+        LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+
+        // delay the loopfilter by 1 macroblock row.
+        if (lf_start < 0) continue;
+
+        // decoding has completed: finish up the loop filter in this thread.
+        if (mi_row + MI_BLOCK_SIZE >= cm->mi_rows) continue;
+
+        vp9_worker_sync(&pbi->lf_worker);
+        lf_data->start = lf_start;
+        lf_data->stop = mi_row;
+        if (pbi->max_threads > 1) {
+          vp9_worker_launch(&pbi->lf_worker);
+        } else {
+          vp9_worker_execute(&pbi->lf_worker);
+        }
+      }
+    }
+  }
+
+  // Loopfilter remaining rows in the frame.
+  if (cm->lf.filter_level) {
+    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+    vp9_worker_sync(&pbi->lf_worker);
+    lf_data->start = lf_data->stop;
+    lf_data->stop = cm->mi_rows;
+    vp9_worker_execute(&pbi->lf_worker);
+  }
+
+  // Get last tile data.
+  tile_data = pbi->tile_data + tile_cols * tile_rows - 1;
+
+  return vp9_reader_find_end(&tile_data->bit_reader);
 }
 
 static int tile_worker_hook(void *arg1, void *arg2) {
@@ -1307,8 +1330,6 @@
   const int tile_rows = 1 << cm->log2_tile_rows;
   const int tile_cols = 1 << cm->log2_tile_cols;
   YV12_BUFFER_CONFIG *const new_fb = get_frame_new_buffer(cm);
-  const int do_loopfilter_inline = tile_rows == 1 && tile_cols == 1 &&
-                                   cm->lf.filter_level;
   xd->cur_buf = new_fb;
 
   if (!first_partition_size) {
@@ -1325,16 +1346,6 @@
     vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                        "Truncated packet or corrupt header length");
 
-  if (do_loopfilter_inline && pbi->lf_worker.data1 == NULL) {
-    CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
-                    vpx_memalign(32, sizeof(LFWorkerData)));
-    pbi->lf_worker.hook = (VP9WorkerHook)vp9_loop_filter_worker;
-    if (pbi->max_threads > 1 && !vp9_worker_reset(&pbi->lf_worker)) {
-      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
-                         "Loop filter thread creation failed");
-    }
-  }
-
   init_macroblockd(cm, &pbi->mb);
 
   if (cm->coding_use_prev_mi)
@@ -1357,9 +1368,11 @@
   if (pbi->max_threads > 1 && tile_rows == 1 && tile_cols > 1 &&
       cm->frame_parallel_decoding_mode) {
     *p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
+    // If multiple threads are used to decode tiles, then we use those threads
+    // to do parallel loopfiltering.
+    vp9_loop_filter_frame_mt(new_fb, pbi, cm, cm->lf.filter_level, 0);
   } else {
-    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end,
-                               do_loopfilter_inline);
+    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
   }
 
   new_fb->corrupted |= xd->corrupted;
@@ -1388,16 +1401,5 @@
   if (cm->refresh_frame_context)
     cm->frame_contexts[cm->frame_context_idx] = cm->fc;
 
-  // Loopfilter
-  if (!do_loopfilter_inline) {
-    // If multiple threads are used to decode tiles, then we use those threads
-    // to do parallel loopfiltering.
-    if (pbi->num_tile_workers) {
-      vp9_loop_filter_frame_mt(new_fb, pbi, cm, cm->lf.filter_level, 0, 0);
-    } else {
-      vp9_loop_filter_frame(new_fb, cm, &pbi->mb, cm->lf.filter_level, 0, 0);
-    }
-  }
-
   return 0;
 }
diff --git a/vp9/decoder/vp9_decoder.c b/vp9/decoder/vp9_decoder.c
index 80f0727..8902f17 100644
--- a/vp9/decoder/vp9_decoder.c
+++ b/vp9/decoder/vp9_decoder.c
@@ -32,7 +32,7 @@
 #include "vp9/decoder/vp9_detokenize.h"
 #include "vp9/decoder/vp9_dthread.h"
 
-void vp9_initialize_dec() {
+static void initialize_dec() {
   static int init_done = 0;
 
   if (!init_done) {
@@ -58,7 +58,7 @@
   }
 
   cm->error.setjmp = 1;
-  vp9_initialize_dec();
+  initialize_dec();
 
   vp9_rtcd();
 
@@ -90,6 +90,7 @@
   vp9_remove_common(cm);
   vp9_worker_end(&pbi->lf_worker);
   vpx_free(pbi->lf_worker.data1);
+  vpx_free(pbi->tile_data);
   for (i = 0; i < pbi->num_tile_workers; ++i) {
     VP9Worker *const worker = &pbi->tile_workers[i];
     vp9_worker_end(worker);
@@ -301,11 +302,14 @@
 int vp9_get_raw_frame(VP9Decoder *pbi, YV12_BUFFER_CONFIG *sd,
                       vp9_ppflags_t *flags) {
   int ret = -1;
+#if !CONFIG_VP9_POSTPROC
+  (void)*flags;
+#endif
 
   if (pbi->ready_for_new_data == 1)
     return ret;
 
-  /* ie no raw frame to show!!! */
+  /* no raw frame to show!!! */
   if (pbi->common.show_frame == 0)
     return ret;
 
@@ -314,8 +318,8 @@
 #if CONFIG_VP9_POSTPROC
   ret = vp9_post_proc_frame(&pbi->common, sd, flags);
 #else
-    *sd = *pbi->common.frame_to_show;
-    ret = 0;
+  *sd = *pbi->common.frame_to_show;
+  ret = 0;
 #endif /*!CONFIG_POSTPROC*/
   vp9_clear_system_state();
   return ret;
diff --git a/vp9/decoder/vp9_decoder.h b/vp9/decoder/vp9_decoder.h
index 12014a7..1a5576e 100644
--- a/vp9/decoder/vp9_decoder.h
+++ b/vp9/decoder/vp9_decoder.h
@@ -27,6 +27,13 @@
 extern "C" {
 #endif
 
+// TODO(hkuang): combine this with TileWorkerData.
+typedef struct TileData {
+  VP9_COMMON *cm;
+  vp9_reader bit_reader;
+  DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+} TileData;
+
 typedef struct VP9Decoder {
   DECLARE_ALIGNED(16, MACROBLOCKD, mb);
 
@@ -39,10 +46,12 @@
   int decoded_key_frame;
 
   VP9Worker lf_worker;
-
   VP9Worker *tile_workers;
   int num_tile_workers;
 
+  TileData *tile_data;
+  int total_tiles;
+
   VP9LfSync lf_row_sync;
 
   vpx_decrypt_cb decrypt_cb;
@@ -52,8 +61,6 @@
   int inv_tile_order;
 } VP9Decoder;
 
-void vp9_initialize_dec();
-
 int vp9_receive_compressed_data(struct VP9Decoder *pbi,
                                 size_t size, const uint8_t **dest);
 
diff --git a/vp9/decoder/vp9_dthread.c b/vp9/decoder/vp9_dthread.c
index 5fe5ed7..bc6c418 100644
--- a/vp9/decoder/vp9_dthread.c
+++ b/vp9/decoder/vp9_dthread.c
@@ -89,7 +89,8 @@
 
 // Implement row loopfiltering for each thread.
 static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer,
-                                VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                                VP9_COMMON *const cm,
+                                struct macroblockd_plane planes[MAX_MB_PLANE],
                                 int start, int stop, int y_only,
                                 VP9LfSync *const lf_sync, int num_lf_workers) {
   const int num_planes = y_only ? 1 : MAX_MB_PLANE;
@@ -107,11 +108,11 @@
 
       sync_read(lf_sync, r, c);
 
-      vp9_setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
+      vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
       vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, &lfm);
 
       for (plane = 0; plane < num_planes; ++plane) {
-        vp9_filter_block_plane(cm, &xd->plane[plane], mi_row, &lfm);
+        vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm);
       }
 
       sync_write(lf_sync, r, c, sb_cols);
@@ -124,7 +125,7 @@
   TileWorkerData *const tile_data = (TileWorkerData*)arg1;
   LFWorkerData *const lf_data = &tile_data->lfdata;
 
-  loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, &lf_data->xd,
+  loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
                       lf_data->start, lf_data->stop, lf_data->y_only,
                       lf_data->lf_sync, lf_data->num_lf_workers);
   return 1;
@@ -135,7 +136,7 @@
 void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
                               VP9Decoder *pbi, VP9_COMMON *cm,
                               int frame_filter_level,
-                              int y_only, int partial_frame) {
+                              int y_only) {
   VP9LfSync *const lf_sync = &pbi->lf_row_sync;
   // Number of superblock rows and cols
   const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
@@ -186,7 +187,7 @@
     // Loopfilter data
     lf_data->frame_buffer = frame;
     lf_data->cm = cm;
-    lf_data->xd = pbi->mb;
+    vp9_copy(lf_data->planes, pbi->mb.plane);
     lf_data->start = i;
     lf_data->stop = sb_rows;
     lf_data->y_only = y_only;   // always do all planes in decoder
diff --git a/vp9/decoder/vp9_dthread.h b/vp9/decoder/vp9_dthread.h
index c3b7a29..a727e2a 100644
--- a/vp9/decoder/vp9_dthread.h
+++ b/vp9/decoder/vp9_dthread.h
@@ -52,6 +52,6 @@
                               struct VP9Decoder *pbi,
                               struct VP9Common *cm,
                               int frame_filter_level,
-                              int y_only, int partial_frame);
+                              int y_only);
 
 #endif  // VP9_DECODER_VP9_DTHREAD_H_
diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c
index 35d2ecf..8ef2b2e 100644
--- a/vp9/encoder/vp9_bitstream.c
+++ b/vp9/encoder/vp9_bitstream.c
@@ -485,8 +485,8 @@
 }
 
 static void build_tree_distribution(VP9_COMP *cpi, TX_SIZE tx_size,
-                                    vp9_coeff_stats *coef_branch_ct) {
-  vp9_coeff_probs_model *coef_probs = cpi->frame_coef_probs[tx_size];
+                                    vp9_coeff_stats *coef_branch_ct,
+                                    vp9_coeff_probs_model *coef_probs) {
   vp9_coeff_count *coef_counts = cpi->coef_counts[tx_size];
   unsigned int (*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
       cpi->common.counts.eob_branch[tx_size];
@@ -513,10 +513,9 @@
 
 static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
                                      TX_SIZE tx_size,
-                                     vp9_coeff_stats *frame_branch_ct) {
-  vp9_coeff_probs_model *new_frame_coef_probs = cpi->frame_coef_probs[tx_size];
-  vp9_coeff_probs_model *old_frame_coef_probs =
-      cpi->common.fc.coef_probs[tx_size];
+                                     vp9_coeff_stats *frame_branch_ct,
+                                     vp9_coeff_probs_model *new_coef_probs) {
+  vp9_coeff_probs_model *old_coef_probs = cpi->common.fc.coef_probs[tx_size];
   const vp9_prob upd = DIFF_UPDATE_PROB;
   const int entropy_nodes_update = UNCONSTRAINED_NODES;
   int i, j, k, l, t;
@@ -530,14 +529,14 @@
           for (k = 0; k < COEF_BANDS; ++k) {
             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
               for (t = 0; t < entropy_nodes_update; ++t) {
-                vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
-                const vp9_prob oldp = old_frame_coef_probs[i][j][k][l][t];
+                vp9_prob newp = new_coef_probs[i][j][k][l][t];
+                const vp9_prob oldp = old_coef_probs[i][j][k][l][t];
                 int s;
                 int u = 0;
                 if (t == PIVOT_NODE)
                   s = vp9_prob_diff_update_savings_search_model(
                       frame_branch_ct[i][j][k][l][0],
-                      old_frame_coef_probs[i][j][k][l], &newp, upd);
+                      old_coef_probs[i][j][k][l], &newp, upd);
                 else
                   s = vp9_prob_diff_update_savings_search(
                       frame_branch_ct[i][j][k][l][t], oldp, &newp, upd);
@@ -567,15 +566,15 @@
             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
               // calc probs and branch cts for this frame only
               for (t = 0; t < entropy_nodes_update; ++t) {
-                vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
-                vp9_prob *oldp = old_frame_coef_probs[i][j][k][l] + t;
+                vp9_prob newp = new_coef_probs[i][j][k][l][t];
+                vp9_prob *oldp = old_coef_probs[i][j][k][l] + t;
                 const vp9_prob upd = DIFF_UPDATE_PROB;
                 int s;
                 int u = 0;
                 if (t == PIVOT_NODE)
                   s = vp9_prob_diff_update_savings_search_model(
                       frame_branch_ct[i][j][k][l][0],
-                      old_frame_coef_probs[i][j][k][l], &newp, upd);
+                      old_coef_probs[i][j][k][l], &newp, upd);
                 else
                   s = vp9_prob_diff_update_savings_search(
                       frame_branch_ct[i][j][k][l][t],
@@ -612,8 +611,8 @@
             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
               // calc probs and branch cts for this frame only
               for (t = 0; t < entropy_nodes_update; ++t) {
-                vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
-                vp9_prob *oldp = old_frame_coef_probs[i][j][k][l] + t;
+                vp9_prob newp = new_coef_probs[i][j][k][l][t];
+                vp9_prob *oldp = old_coef_probs[i][j][k][l] + t;
                 int s;
                 int u = 0;
                 if (l >= prev_coef_contexts_to_update ||
@@ -623,7 +622,7 @@
                   if (t == PIVOT_NODE)
                     s = vp9_prob_diff_update_savings_search_model(
                         frame_branch_ct[i][j][k][l][0],
-                        old_frame_coef_probs[i][j][k][l], &newp, upd);
+                        old_coef_probs[i][j][k][l], &newp, upd);
                   else
                     s = vp9_prob_diff_update_savings_search(
                         frame_branch_ct[i][j][k][l][t],
@@ -670,14 +669,17 @@
   const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
   TX_SIZE tx_size;
   vp9_coeff_stats frame_branch_ct[TX_SIZES][PLANE_TYPES];
+  vp9_coeff_probs_model frame_coef_probs[TX_SIZES][PLANE_TYPES];
 
   vp9_clear_system_state();
 
   for (tx_size = TX_4X4; tx_size <= TX_32X32; ++tx_size)
-    build_tree_distribution(cpi, tx_size, frame_branch_ct[tx_size]);
+    build_tree_distribution(cpi, tx_size, frame_branch_ct[tx_size],
+                            frame_coef_probs[tx_size]);
 
   for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
-    update_coef_probs_common(w, cpi, tx_size, frame_branch_ct[tx_size]);
+    update_coef_probs_common(w, cpi, tx_size, frame_branch_ct[tx_size],
+                             frame_coef_probs[tx_size]);
 }
 
 static void encode_loopfilter(struct loopfilter *lf,
diff --git a/vp9/encoder/vp9_block.h b/vp9/encoder/vp9_block.h
index 2ccf4f8..c406860 100644
--- a/vp9/encoder/vp9_block.h
+++ b/vp9/encoder/vp9_block.h
@@ -20,43 +20,6 @@
 extern "C" {
 #endif
 
-// Structure to hold snapshot of coding context during the mode picking process
-typedef struct {
-  MODE_INFO mic;
-  uint8_t *zcoeff_blk;
-  int16_t *coeff[MAX_MB_PLANE][3];
-  int16_t *qcoeff[MAX_MB_PLANE][3];
-  int16_t *dqcoeff[MAX_MB_PLANE][3];
-  uint16_t *eobs[MAX_MB_PLANE][3];
-
-  // dual buffer pointers, 0: in use, 1: best in store
-  int16_t *coeff_pbuf[MAX_MB_PLANE][3];
-  int16_t *qcoeff_pbuf[MAX_MB_PLANE][3];
-  int16_t *dqcoeff_pbuf[MAX_MB_PLANE][3];
-  uint16_t *eobs_pbuf[MAX_MB_PLANE][3];
-
-  int is_coded;
-  int num_4x4_blk;
-  int skip;
-  int_mv best_ref_mv[2];
-  int_mv ref_mvs[MAX_REF_FRAMES][MAX_MV_REF_CANDIDATES];
-  int rate;
-  int distortion;
-  int best_mode_index;
-  int rddiv;
-  int rdmult;
-  int hybrid_pred_diff;
-  int comp_pred_diff;
-  int single_pred_diff;
-  int64_t tx_rd_diff[TX_MODES];
-  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
-
-  // motion vector cache for adaptive motion search control in partition
-  // search loop
-  int_mv pred_mv[MAX_REF_FRAMES];
-  INTERP_FILTER pred_interp_filter;
-} PICK_MODE_CONTEXT;
-
 struct macroblock_plane {
   DECLARE_ALIGNED(16, int16_t, src_diff[64 * 64]);
   int16_t *qcoeff;
@@ -73,18 +36,6 @@
   // Zbin Over Quant value
   int16_t zbin_extra;
 };
-typedef struct PC_TREE {
-  int index;
-  PARTITION_TYPE partitioning;
-  BLOCK_SIZE block_size;
-  PICK_MODE_CONTEXT none;
-  PICK_MODE_CONTEXT horizontal[2];
-  PICK_MODE_CONTEXT vertical[2];
-  union {
-    struct PC_TREE *split[4];
-    PICK_MODE_CONTEXT *leaf_split[4];
-  };
-} PC_TREE;
 
 /* The [2] dimension is for whether we skip the EOB node (i.e. if previous
  * coefficient in this block was zero) or not. */
@@ -155,17 +106,11 @@
   int skip_encode;
 
   // Used to store sub partition's choices.
-  int_mv pred_mv[MAX_REF_FRAMES];
-
-  PICK_MODE_CONTEXT *leaf_tree;
-  PC_TREE *pc_tree;
-  PC_TREE *pc_root;
-  int partition_cost[PARTITION_CONTEXTS][PARTITION_TYPES];
+  MV pred_mv[MAX_REF_FRAMES];
 
   void (*fwd_txm4x4)(const int16_t *input, int16_t *output, int stride);
 };
 
-
 #ifdef __cplusplus
 }  // extern "C"
 #endif
diff --git a/vp9/encoder/vp9_context_tree.c b/vp9/encoder/vp9_context_tree.c
index 659935c..9b7a932 100644
--- a/vp9/encoder/vp9_context_tree.c
+++ b/vp9/encoder/vp9_context_tree.c
@@ -8,14 +8,14 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-
 #include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_encoder.h"
 
 static const BLOCK_SIZE square[] = {
-    BLOCK_8X8,
-    BLOCK_16X16,
-    BLOCK_32X32,
-    BLOCK_64X64,
+  BLOCK_8X8,
+  BLOCK_16X16,
+  BLOCK_32X32,
+  BLOCK_64X64,
 };
 
 static void alloc_mode_context(VP9_COMMON *cm, int num_4x4_blk,
@@ -62,30 +62,32 @@
     }
   }
 }
-static void free_tree_contexts(PC_TREE *this_pc) {
-  free_mode_context(&this_pc->none);
-  free_mode_context(&this_pc->horizontal[0]);
-  free_mode_context(&this_pc->horizontal[1]);
-  free_mode_context(&this_pc->vertical[0]);
-  free_mode_context(&this_pc->vertical[1]);
-}
-static void alloc_tree_contexts(VP9_COMMON *cm, PC_TREE *this_pc,
+
+static void alloc_tree_contexts(VP9_COMMON *cm, PC_TREE *tree,
                                 int num_4x4_blk) {
-  alloc_mode_context(cm, num_4x4_blk, &this_pc->none);
-  alloc_mode_context(cm, num_4x4_blk/2, &this_pc->horizontal[0]);
-  alloc_mode_context(cm, num_4x4_blk/2, &this_pc->vertical[0]);
+  alloc_mode_context(cm, num_4x4_blk, &tree->none);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->horizontal[0]);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->vertical[0]);
 
   /* TODO(Jbb): for 4x8 and 8x4 these allocated values are not used.
    * Figure out a better way to do this. */
-  alloc_mode_context(cm, num_4x4_blk/2, &this_pc->horizontal[1]);
-  alloc_mode_context(cm, num_4x4_blk/2, &this_pc->vertical[1]);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->horizontal[1]);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->vertical[1]);
+}
+
+static void free_tree_contexts(PC_TREE *tree) {
+  free_mode_context(&tree->none);
+  free_mode_context(&tree->horizontal[0]);
+  free_mode_context(&tree->horizontal[1]);
+  free_mode_context(&tree->vertical[0]);
+  free_mode_context(&tree->vertical[1]);
 }
 
 // This function sets up a tree of contexts such that at each square
 // partition level. There are contexts for none, horizontal, vertical, and
 // split.  Along with a block_size value and a selected block_size which
 // represents the state of our search.
-void vp9_setup_pc_tree(VP9_COMMON *cm, MACROBLOCK *x) {
+void vp9_setup_pc_tree(VP9_COMMON *cm, VP9_COMP *cpi) {
   int i, j;
   const int leaf_nodes = 64;
   const int tree_nodes = 64 + 16 + 4 + 1;
@@ -95,61 +97,62 @@
   int square_index = 1;
   int nodes;
 
-  vpx_free(x->leaf_tree);
-  CHECK_MEM_ERROR(cm, x->leaf_tree, vpx_calloc(leaf_nodes,
-                                               sizeof(PICK_MODE_CONTEXT)));
-  vpx_free(x->pc_tree);
-  CHECK_MEM_ERROR(cm, x->pc_tree, vpx_calloc(tree_nodes, sizeof(PC_TREE)));
+  vpx_free(cpi->leaf_tree);
+  CHECK_MEM_ERROR(cm, cpi->leaf_tree, vpx_calloc(leaf_nodes,
+                                                 sizeof(*cpi->leaf_tree)));
+  vpx_free(cpi->pc_tree);
+  CHECK_MEM_ERROR(cm, cpi->pc_tree, vpx_calloc(tree_nodes,
+                                               sizeof(*cpi->pc_tree)));
 
-  this_pc = &x->pc_tree[0];
-  this_leaf = &x->leaf_tree[0];
+  this_pc = &cpi->pc_tree[0];
+  this_leaf = &cpi->leaf_tree[0];
 
   // 4x4 blocks smaller than 8x8 but in the same 8x8 block share the same
   // context so we only need to allocate 1 for each 8x8 block.
   for (i = 0; i < leaf_nodes; ++i)
-    alloc_mode_context(cm, 1, &x->leaf_tree[i]);
+    alloc_mode_context(cm, 1, &cpi->leaf_tree[i]);
 
   // Sets up all the leaf nodes in the tree.
   for (pc_tree_index = 0; pc_tree_index < leaf_nodes; ++pc_tree_index) {
-    x->pc_tree[pc_tree_index].block_size = square[0];
-    alloc_tree_contexts(cm, &x->pc_tree[pc_tree_index], 4);
-    x->pc_tree[pc_tree_index].leaf_split[0] = this_leaf++;
-    for (j = 1; j < 4; j++) {
-      x->pc_tree[pc_tree_index].leaf_split[j] =
-          x->pc_tree[pc_tree_index].leaf_split[0];
-    }
+    PC_TREE *const tree = &cpi->pc_tree[pc_tree_index];
+    tree->block_size = square[0];
+    alloc_tree_contexts(cm, tree, 4);
+    tree->leaf_split[0] = this_leaf++;
+    for (j = 1; j < 4; j++)
+      tree->leaf_split[j] = tree->leaf_split[0];
   }
 
   // Each node has 4 leaf nodes, fill each block_size level of the tree
   // from leafs to the root.
-  for (nodes = 16; nodes > 0; nodes >>= 2, ++square_index) {
-    for (i = 0; i < nodes; ++pc_tree_index,  ++i) {
-      alloc_tree_contexts(cm, &x->pc_tree[pc_tree_index],
-                          4 << (2 * square_index));
-      x->pc_tree[pc_tree_index].block_size = square[square_index];
-      for (j = 0; j < 4; j++) {
-        x->pc_tree[pc_tree_index].split[j] = this_pc++;
-      }
+  for (nodes = 16; nodes > 0; nodes >>= 2) {
+    for (i = 0; i < nodes; ++i) {
+      PC_TREE *const tree = &cpi->pc_tree[pc_tree_index];
+      alloc_tree_contexts(cm, tree, 4 << (2 * square_index));
+      tree->block_size = square[square_index];
+      for (j = 0; j < 4; j++)
+        tree->split[j] = this_pc++;
+      ++pc_tree_index;
     }
+    ++square_index;
   }
-  x->pc_root = &x->pc_tree[tree_nodes-1];
-  x->pc_root[0].none.best_mode_index = 2;
+  cpi->pc_root = &cpi->pc_tree[tree_nodes - 1];
+  cpi->pc_root[0].none.best_mode_index = 2;
 }
 
-void vp9_free_pc_tree(MACROBLOCK *m) {
+void vp9_free_pc_tree(VP9_COMP *cpi) {
   const int tree_nodes = 64 + 16 + 4 + 1;
   int i;
 
   // Set up all 4x4 mode contexts
   for (i = 0; i < 64; ++i)
-    free_mode_context(&m->leaf_tree[i]);
+    free_mode_context(&cpi->leaf_tree[i]);
 
   // Sets up all the leaf nodes in the tree.
-  for (i = 0; i < tree_nodes; i++) {
-    free_tree_contexts(&m->pc_tree[i]);
-  }
-  vpx_free(m->pc_tree);
-  m->pc_tree = 0;
-  vpx_free(m->leaf_tree);
-  m->leaf_tree = 0;
+  for (i = 0; i < tree_nodes; ++i)
+    free_tree_contexts(&cpi->pc_tree[i]);
+
+  vpx_free(cpi->pc_tree);
+  cpi->pc_tree = NULL;
+  vpx_free(cpi->leaf_tree);
+  cpi->leaf_tree = NULL;
 }
diff --git a/vp9/encoder/vp9_context_tree.h b/vp9/encoder/vp9_context_tree.h
index 66a6f00..bb384aa 100644
--- a/vp9/encoder/vp9_context_tree.h
+++ b/vp9/encoder/vp9_context_tree.h
@@ -11,9 +11,55 @@
 #ifndef VP9_ENCODER_VP9_CONTEXT_TREE_H_
 #define VP9_ENCODER_VP9_CONTEXT_TREE_H_
 
-#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/common/vp9_onyxc_int.h"
 
-void vp9_setup_pc_tree(VP9_COMMON *cm, MACROBLOCK *x);
-void vp9_free_pc_tree(MACROBLOCK *x);
+struct VP9_COMP;
+
+// Structure to hold snapshot of coding context during the mode picking process
+typedef struct {
+  MODE_INFO mic;
+  uint8_t *zcoeff_blk;
+  int16_t *coeff[MAX_MB_PLANE][3];
+  int16_t *qcoeff[MAX_MB_PLANE][3];
+  int16_t *dqcoeff[MAX_MB_PLANE][3];
+  uint16_t *eobs[MAX_MB_PLANE][3];
+
+  // dual buffer pointers, 0: in use, 1: best in store
+  int16_t *coeff_pbuf[MAX_MB_PLANE][3];
+  int16_t *qcoeff_pbuf[MAX_MB_PLANE][3];
+  int16_t *dqcoeff_pbuf[MAX_MB_PLANE][3];
+  uint16_t *eobs_pbuf[MAX_MB_PLANE][3];
+
+  int is_coded;
+  int num_4x4_blk;
+  int skip;
+  int best_mode_index;
+  int hybrid_pred_diff;
+  int comp_pred_diff;
+  int single_pred_diff;
+  int64_t tx_rd_diff[TX_MODES];
+  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+
+  // motion vector cache for adaptive motion search control in partition
+  // search loop
+  MV pred_mv[MAX_REF_FRAMES];
+  INTERP_FILTER pred_interp_filter;
+} PICK_MODE_CONTEXT;
+
+typedef struct PC_TREE {
+  int index;
+  PARTITION_TYPE partitioning;
+  BLOCK_SIZE block_size;
+  PICK_MODE_CONTEXT none;
+  PICK_MODE_CONTEXT horizontal[2];
+  PICK_MODE_CONTEXT vertical[2];
+  union {
+    struct PC_TREE *split[4];
+    PICK_MODE_CONTEXT *leaf_split[4];
+  };
+} PC_TREE;
+
+void vp9_setup_pc_tree(struct VP9Common *cm, struct VP9_COMP *cpi);
+void vp9_free_pc_tree(struct VP9_COMP *cpi);
 
 #endif /* VP9_ENCODER_VP9_CONTEXT_TREE_H_ */
diff --git a/vp9/encoder/vp9_dct.c b/vp9/encoder/vp9_dct.c
index d523239..5772767 100644
--- a/vp9/encoder/vp9_dct.c
+++ b/vp9/encoder/vp9_dct.c
@@ -445,20 +445,20 @@
         step3[7] = step1[7] + step2[4];
         // step 4
         temp1 = step3[1] *  -cospi_8_64 + step3[6] * cospi_24_64;
-        temp2 = step3[2] * -cospi_24_64 - step3[5] *  cospi_8_64;
+        temp2 = step3[2] * cospi_24_64 + step3[5] *  cospi_8_64;
         step2[1] = fdct_round_shift(temp1);
         step2[2] = fdct_round_shift(temp2);
-        temp1 = step3[2] * -cospi_8_64 + step3[5] * cospi_24_64;
+        temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
         temp2 = step3[1] * cospi_24_64 + step3[6] *  cospi_8_64;
         step2[5] = fdct_round_shift(temp1);
         step2[6] = fdct_round_shift(temp2);
         // step 5
         step1[0] = step3[0] + step2[1];
         step1[1] = step3[0] - step2[1];
-        step1[2] = step3[3] - step2[2];
-        step1[3] = step3[3] + step2[2];
-        step1[4] = step3[4] + step2[5];
-        step1[5] = step3[4] - step2[5];
+        step1[2] = step3[3] + step2[2];
+        step1[3] = step3[3] - step2[2];
+        step1[4] = step3[4] - step2[5];
+        step1[5] = step3[4] + step2[5];
         step1[6] = step3[7] - step2[6];
         step1[7] = step3[7] + step2[6];
         // step 6
@@ -755,10 +755,10 @@
 
   // step 4
   temp1 = step3[1] *  -cospi_8_64 + step3[6] * cospi_24_64;
-  temp2 = step3[2] * -cospi_24_64 - step3[5] *  cospi_8_64;
+  temp2 = step3[2] * cospi_24_64 + step3[5] *  cospi_8_64;
   step2[1] = fdct_round_shift(temp1);
   step2[2] = fdct_round_shift(temp2);
-  temp1 = step3[2] * -cospi_8_64 + step3[5] * cospi_24_64;
+  temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
   temp2 = step3[1] * cospi_24_64 + step3[6] *  cospi_8_64;
   step2[5] = fdct_round_shift(temp1);
   step2[6] = fdct_round_shift(temp2);
@@ -766,10 +766,10 @@
   // step 5
   step1[0] = step3[0] + step2[1];
   step1[1] = step3[0] - step2[1];
-  step1[2] = step3[3] - step2[2];
-  step1[3] = step3[3] + step2[2];
-  step1[4] = step3[4] + step2[5];
-  step1[5] = step3[4] - step2[5];
+  step1[2] = step3[3] + step2[2];
+  step1[3] = step3[3] - step2[2];
+  step1[4] = step3[4] - step2[5];
+  step1[5] = step3[4] + step2[5];
   step1[6] = step3[7] - step2[6];
   step1[7] = step3[7] + step2[6];
 
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index fba9465..ef33fca 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -70,6 +70,12 @@
   128, 128, 128, 128, 128, 128, 128, 128
 };
 
+typedef struct {
+  unsigned int sse;
+  int sum;
+  unsigned int var;
+} diff;
+
 static void get_sse_sum_8x8(const uint8_t *src, int src_stride,
                             const uint8_t *ref, int ref_stride,
                             unsigned int *sse, int *sum) {
@@ -201,7 +207,7 @@
   mbmi = &xd->mi[0]->mbmi;
 
   // Set up destination pointers.
-  vp9_setup_dst_planes(xd, get_frame_new_buffer(cm), mi_row, mi_col);
+  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
 
   // Set up limit values for MV components.
   // Mv beyond the range do not produce new/different prediction block.
@@ -1538,7 +1544,7 @@
       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
 
       if (none_rate < INT_MAX) {
-        none_rate += x->partition_cost[pl][PARTITION_NONE];
+        none_rate += cpi->partition_cost[pl][PARTITION_NONE];
         none_rd = RDCOST(x->rdmult, x->rddiv, none_rate, none_dist);
       }
 
@@ -1636,7 +1642,7 @@
 
   pl = partition_plane_context(xd, mi_row, mi_col, bsize);
   if (last_part_rate < INT_MAX) {
-    last_part_rate += x->partition_cost[pl][partition];
+    last_part_rate += cpi->partition_cost[pl][partition];
     last_part_rd = RDCOST(x->rdmult, x->rddiv, last_part_rate, last_part_dist);
   }
 
@@ -1689,11 +1695,11 @@
 
       pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
                                    split_subsize);
-      chosen_rate += x->partition_cost[pl][PARTITION_NONE];
+      chosen_rate += cpi->partition_cost[pl][PARTITION_NONE];
     }
     pl = partition_plane_context(xd, mi_row, mi_col, bsize);
     if (chosen_rate < INT_MAX) {
-      chosen_rate += x->partition_cost[pl][PARTITION_SPLIT];
+      chosen_rate += cpi->partition_cost[pl][PARTITION_SPLIT];
       chosen_rd = RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist);
     }
   }
@@ -1805,15 +1811,11 @@
                                     BLOCK_SIZE *max_block_size) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCKD *const xd = &cpi->mb.e_mbd;
-  MODE_INFO **mi_8x8 = xd->mi;
-  const int left_in_image = xd->left_available && mi_8x8[-1];
-  const int above_in_image = xd->up_available &&
-                             mi_8x8[-xd->mi_stride];
-  MODE_INFO **above_sb64_mi_8x8;
-  MODE_INFO **left_sb64_mi_8x8;
-
-  int row8x8_remaining = tile->mi_row_end - mi_row;
-  int col8x8_remaining = tile->mi_col_end - mi_col;
+  MODE_INFO **mi = xd->mi;
+  const int left_in_image = xd->left_available && mi[-1];
+  const int above_in_image = xd->up_available && mi[-xd->mi_stride];
+  const int row8x8_remaining = tile->mi_row_end - mi_row;
+  const int col8x8_remaining = tile->mi_col_end - mi_col;
   int bh, bw;
   BLOCK_SIZE min_size = BLOCK_4X4;
   BLOCK_SIZE max_size = BLOCK_64X64;
@@ -1833,15 +1835,13 @@
     }
     // Find the min and max partition sizes used in the left SB64
     if (left_in_image) {
-      left_sb64_mi_8x8 = &mi_8x8[-MI_BLOCK_SIZE];
-      get_sb_partition_size_range(cpi, left_sb64_mi_8x8,
-                                  &min_size, &max_size);
+      MODE_INFO **left_sb64_mi = &mi[-MI_BLOCK_SIZE];
+      get_sb_partition_size_range(cpi, left_sb64_mi, &min_size, &max_size);
     }
     // Find the min and max partition sizes used in the above SB64.
     if (above_in_image) {
-      above_sb64_mi_8x8 = &mi_8x8[-xd->mi_stride * MI_BLOCK_SIZE];
-      get_sb_partition_size_range(cpi, above_sb64_mi_8x8,
-                                  &min_size, &max_size);
+      MODE_INFO **above_sb64_mi = &mi[-xd->mi_stride * MI_BLOCK_SIZE];
+      get_sb_partition_size_range(cpi, above_sb64_mi, &min_size, &max_size);
     }
     // adjust observed min and max
     if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
@@ -2021,7 +2021,7 @@
     if (this_rate != INT_MAX) {
       if (bsize >= BLOCK_8X8) {
         pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-        this_rate += x->partition_cost[pl][PARTITION_NONE];
+        this_rate += cpi->partition_cost[pl][PARTITION_NONE];
       }
       sum_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_dist);
       if (sum_rd < best_rd) {
@@ -2109,7 +2109,7 @@
 
     if (sum_rd < best_rd && i == 4) {
       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-      sum_rate += x->partition_cost[pl][PARTITION_SPLIT];
+      sum_rate += cpi->partition_cost[pl][PARTITION_SPLIT];
       sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
       if (sum_rd < best_rd) {
         best_rate = sum_rate;
@@ -2163,7 +2163,7 @@
     }
     if (sum_rd < best_rd) {
       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-      sum_rate += x->partition_cost[pl][PARTITION_HORZ];
+      sum_rate += cpi->partition_cost[pl][PARTITION_HORZ];
       sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
       if (sum_rd < best_rd) {
         best_rd = sum_rd;
@@ -2212,7 +2212,7 @@
     }
     if (sum_rd < best_rd) {
       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-      sum_rate += x->partition_cost[pl][PARTITION_VERT];
+      sum_rate += cpi->partition_cost[pl][PARTITION_VERT];
       sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
       if (sum_rd < best_rd) {
         best_rate = sum_rate;
@@ -2274,17 +2274,16 @@
     int64_t dummy_dist;
 
     int i;
-    MACROBLOCK *x = &cpi->mb;
 
     if (sf->adaptive_pred_interp_filter) {
       for (i = 0; i < 64; ++i)
-        x->leaf_tree[i].pred_interp_filter = SWITCHABLE;
+        cpi->leaf_tree[i].pred_interp_filter = SWITCHABLE;
 
       for (i = 0; i < 64; ++i) {
-        x->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
-        x->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
-        x->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
-        x->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
+        cpi->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
+        cpi->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
+        cpi->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
+        cpi->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
       }
     }
 
@@ -2296,26 +2295,26 @@
          sf->partition_search_type == VAR_BASED_PARTITION ||
          sf->partition_search_type == VAR_BASED_FIXED_PARTITION) {
       const int idx_str = cm->mi_stride * mi_row + mi_col;
-      MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str;
-      MODE_INFO **prev_mi_8x8 = cm->prev_mi_grid_visible + idx_str;
+      MODE_INFO **mi = cm->mi_grid_visible + idx_str;
+      MODE_INFO **prev_mi = cm->prev_mi_grid_visible + idx_str;
       cpi->mb.source_variance = UINT_MAX;
       if (sf->partition_search_type == FIXED_PARTITION) {
         set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
-        set_fixed_partitioning(cpi, tile, mi_8x8, mi_row, mi_col,
+        set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col,
                                sf->always_this_block_size);
-        rd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                         &dummy_rate, &dummy_dist, 1, x->pc_root);
+        rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                         &dummy_rate, &dummy_dist, 1, cpi->pc_root);
       } else if (sf->partition_search_type == VAR_BASED_FIXED_PARTITION) {
         BLOCK_SIZE bsize;
         set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
         bsize = get_rd_var_based_fixed_partition(cpi, mi_row, mi_col);
-        set_fixed_partitioning(cpi, tile, mi_8x8, mi_row, mi_col, bsize);
-        rd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                         &dummy_rate, &dummy_dist, 1, x->pc_root);
+        set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col, bsize);
+        rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                         &dummy_rate, &dummy_dist, 1, cpi->pc_root);
       } else if (sf->partition_search_type == VAR_BASED_PARTITION) {
         choose_partitioning(cpi, tile, mi_row, mi_col);
-        rd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                         &dummy_rate, &dummy_dist, 1, x->pc_root);
+        rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                         &dummy_rate, &dummy_dist, 1, cpi->pc_root);
       } else {
         if ((cm->current_video_frame
             % sf->last_partitioning_redo_frequency) == 0
@@ -2325,7 +2324,7 @@
             || cpi->rc.is_src_frame_alt_ref
             || ((sf->use_lastframe_partitioning ==
                  LAST_FRAME_PARTITION_LOW_MOTION) &&
-                 sb_has_motion(cm, prev_mi_8x8))) {
+                 sb_has_motion(cm, prev_mi))) {
           // If required set upper and lower partition size limits
           if (sf->auto_min_max_partition_size) {
             set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
@@ -2334,16 +2333,17 @@
                                     &sf->max_partition_size);
           }
           rd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
-                            &dummy_rate, &dummy_dist, 1, INT64_MAX, x->pc_root);
+                            &dummy_rate, &dummy_dist, 1, INT64_MAX,
+                            cpi->pc_root);
         } else {
           if (sf->constrain_copy_partition &&
-              sb_has_motion(cm, prev_mi_8x8))
-            constrain_copy_partitioning(cpi, tile, mi_8x8, prev_mi_8x8,
+              sb_has_motion(cm, prev_mi))
+            constrain_copy_partitioning(cpi, tile, mi, prev_mi,
                                         mi_row, mi_col, BLOCK_16X16);
           else
-            copy_partitioning(cm, mi_8x8, prev_mi_8x8);
-          rd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                           &dummy_rate, &dummy_dist, 1, x->pc_root);
+            copy_partitioning(cm, mi, prev_mi);
+          rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                           &dummy_rate, &dummy_dist, 1, cpi->pc_root);
         }
       }
     } else {
@@ -2355,7 +2355,7 @@
                                 &sf->max_partition_size);
       }
       rd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
-                        &dummy_rate, &dummy_dist, 1, INT64_MAX, x->pc_root);
+                        &dummy_rate, &dummy_dist, 1, INT64_MAX, cpi->pc_root);
     }
   }
 }
@@ -2633,7 +2633,7 @@
 
     if (this_rate != INT_MAX) {
       int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-      this_rate += x->partition_cost[pl][PARTITION_NONE];
+      this_rate += cpi->partition_cost[pl][PARTITION_NONE];
       sum_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_dist);
       if (sum_rd < best_rd) {
         int64_t stop_thresh = 4096;
@@ -2671,7 +2671,7 @@
   sum_rd = 0;
   if (do_split) {
     int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-    sum_rate += x->partition_cost[pl][PARTITION_SPLIT];
+    sum_rate += cpi->partition_cost[pl][PARTITION_SPLIT];
     subsize = get_subsize(bsize, PARTITION_SPLIT);
     for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
       const int x_idx = (i & 1) * ms;
@@ -2730,7 +2730,7 @@
         sum_rd = INT64_MAX;
       } else {
         int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-        this_rate += x->partition_cost[pl][PARTITION_HORZ];
+        this_rate += cpi->partition_cost[pl][PARTITION_HORZ];
         sum_rate += this_rate;
         sum_dist += this_dist;
         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
@@ -2764,7 +2764,7 @@
         sum_rd = INT64_MAX;
       } else {
         int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-        this_rate += x->partition_cost[pl][PARTITION_VERT];
+        this_rate += cpi->partition_cost[pl][PARTITION_VERT];
         sum_rate += this_rate;
         sum_dist += this_dist;
         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
@@ -2822,7 +2822,7 @@
 
 static void nonrd_use_partition(VP9_COMP *cpi,
                                 const TileInfo *const tile,
-                                MODE_INFO **mi_8x8,
+                                MODE_INFO **mi,
                                 TOKENEXTRA **tp,
                                 int mi_row, int mi_col,
                                 BLOCK_SIZE bsize, int output_enabled,
@@ -2841,7 +2841,7 @@
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
-  subsize = (bsize >= BLOCK_8X8) ? mi_8x8[0]->mbmi.sb_type : BLOCK_4X4;
+  subsize = (bsize >= BLOCK_8X8) ? mi[0]->mbmi.sb_type : BLOCK_4X4;
   partition = partition_lookup[bsl][subsize];
 
   switch (partition) {
@@ -2869,7 +2869,7 @@
       if (mi_row + hbs < cm->mi_rows) {
         nonrd_pick_sb_modes(cpi, tile, mi_row + hbs, mi_col,
                             &rate, &dist, subsize);
-        pc_tree->horizontal[1].mic.mbmi = mi_8x8[0]->mbmi;
+        pc_tree->horizontal[1].mic.mbmi = mi[0]->mbmi;
         if (rate != INT_MAX && dist != INT64_MAX &&
             *totrate != INT_MAX && *totdist != INT64_MAX) {
           *totrate += rate;
@@ -2879,10 +2879,10 @@
       break;
     case PARTITION_SPLIT:
       subsize = get_subsize(bsize, PARTITION_SPLIT);
-      nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col,
+      nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col,
                           subsize, output_enabled, totrate, totdist,
                           pc_tree->split[0]);
-      nonrd_use_partition(cpi, tile, mi_8x8 + hbs, tp,
+      nonrd_use_partition(cpi, tile, mi + hbs, tp,
                           mi_row, mi_col + hbs, subsize, output_enabled,
                           &rate, &dist, pc_tree->split[1]);
       if (rate != INT_MAX && dist != INT64_MAX &&
@@ -2890,7 +2890,7 @@
         *totrate += rate;
         *totdist += dist;
       }
-      nonrd_use_partition(cpi, tile, mi_8x8 + hbs * mis, tp,
+      nonrd_use_partition(cpi, tile, mi + hbs * mis, tp,
                           mi_row + hbs, mi_col, subsize, output_enabled,
                           &rate, &dist, pc_tree->split[2]);
       if (rate != INT_MAX && dist != INT64_MAX &&
@@ -2898,7 +2898,7 @@
         *totrate += rate;
         *totdist += dist;
       }
-      nonrd_use_partition(cpi, tile, mi_8x8 + hbs * mis + hbs, tp,
+      nonrd_use_partition(cpi, tile, mi + hbs * mis + hbs, tp,
                           mi_row + hbs, mi_col + hbs, subsize, output_enabled,
                           &rate, &dist, pc_tree->split[3]);
       if (rate != INT_MAX && dist != INT64_MAX &&
@@ -2937,8 +2937,8 @@
     int dummy_rate = 0;
     int64_t dummy_dist = 0;
     const int idx_str = cm->mi_stride * mi_row + mi_col;
-    MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str;
-    MODE_INFO **prev_mi_8x8 = cm->prev_mi_grid_visible + idx_str;
+    MODE_INFO **mi = cm->mi_grid_visible + idx_str;
+    MODE_INFO **prev_mi = cm->prev_mi_grid_visible + idx_str;
     BLOCK_SIZE bsize;
 
     x->in_static_area = 0;
@@ -2949,22 +2949,22 @@
     switch (cpi->sf.partition_search_type) {
       case VAR_BASED_PARTITION:
         choose_partitioning(cpi, tile, mi_row, mi_col);
-        nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                            1, &dummy_rate, &dummy_dist, x->pc_root);
+        nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                            1, &dummy_rate, &dummy_dist, cpi->pc_root);
         break;
       case SOURCE_VAR_BASED_PARTITION:
-        set_source_var_based_partition(cpi, tile, mi_8x8, mi_row, mi_col);
-        nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                            1, &dummy_rate, &dummy_dist, x->pc_root);
+        set_source_var_based_partition(cpi, tile, mi, mi_row, mi_col);
+        nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                            1, &dummy_rate, &dummy_dist, cpi->pc_root);
         break;
       case VAR_BASED_FIXED_PARTITION:
       case FIXED_PARTITION:
         bsize = cpi->sf.partition_search_type == FIXED_PARTITION ?
                 cpi->sf.always_this_block_size :
                 get_nonrd_var_based_fixed_partition(cpi, mi_row, mi_col);
-        set_fixed_partitioning(cpi, tile, mi_8x8, mi_row, mi_col, bsize);
-        nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                            1, &dummy_rate, &dummy_dist, x->pc_root);
+        set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col, bsize);
+        nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                            1, &dummy_rate, &dummy_dist, cpi->pc_root);
         break;
       case REFERENCE_PARTITION:
         if (cpi->sf.partition_check ||
@@ -2975,12 +2975,12 @@
                                &cpi->sf.max_partition_size);
           nonrd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
                                &dummy_rate, &dummy_dist, 1, INT64_MAX,
-                               x->pc_root);
+                               cpi->pc_root);
         } else {
-          copy_partitioning(cm, mi_8x8, prev_mi_8x8);
-          nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col,
+          copy_partitioning(cm, mi, prev_mi);
+          nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col,
                               BLOCK_64X64, 1, &dummy_rate, &dummy_dist,
-                              x->pc_root);
+                              cpi->pc_root);
         }
         break;
       default:
@@ -3043,7 +3043,7 @@
     int i;
     struct macroblock_plane *const p = x->plane;
     struct macroblockd_plane *const pd = xd->plane;
-    PICK_MODE_CONTEXT *ctx = &x->pc_root->none;
+    PICK_MODE_CONTEXT *ctx = &cpi->pc_root->none;
 
     for (i = 0; i < MAX_MB_PLANE; ++i) {
       p[i].coeff = ctx->coeff_pbuf[i][0];
diff --git a/vp9/encoder/vp9_encodeframe.h b/vp9/encoder/vp9_encodeframe.h
index 131e932..72343cd 100644
--- a/vp9/encoder/vp9_encodeframe.h
+++ b/vp9/encoder/vp9_encodeframe.h
@@ -20,12 +20,6 @@
 struct yv12_buffer_config;
 struct VP9_COMP;
 
-typedef struct {
-  unsigned int sse;
-  int sum;
-  unsigned int var;
-} diff;
-
 void vp9_setup_src_planes(struct macroblock *x,
                           const struct yv12_buffer_config *src,
                           int mi_row, int mi_col);
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 6474763..1f68f03 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -187,7 +187,7 @@
   vpx_free(cpi->tok);
   cpi->tok = 0;
 
-  vp9_free_pc_tree(&cpi->mb);
+  vp9_free_pc_tree(cpi);
 
   for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
     LAYER_CONTEXT *const lc = &cpi->svc.layer_context[i];
@@ -455,7 +455,7 @@
     CHECK_MEM_ERROR(cm, cpi->tok, vpx_calloc(tokens, sizeof(*cpi->tok)));
   }
 
-  vp9_setup_pc_tree(&cpi->common, &cpi->mb);
+  vp9_setup_pc_tree(&cpi->common, cpi);
 }
 
 static void update_frame_size(VP9_COMP *cpi) {
@@ -1414,6 +1414,7 @@
                             dst->alpha_buffer};
   const int dst_strides[4] = {dst->y_stride, dst->uv_stride, dst->uv_stride,
                               dst->alpha_stride};
+  const InterpKernel *const kernel = vp9_get_interp_kernel(EIGHTTAP);
   int x, y, i;
 
   for (y = 0; y < dst_h; y += 16) {
@@ -1429,8 +1430,8 @@
         uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
 
         vp9_convolve8(src_ptr, src_stride, dst_ptr, dst_stride,
-                      vp9_sub_pel_filters_8[x_q4 & 0xf], 16 * src_w / dst_w,
-                      vp9_sub_pel_filters_8[y_q4 & 0xf], 16 * src_h / dst_h,
+                      kernel[x_q4 & 0xf], 16 * src_w / dst_w,
+                      kernel[y_q4 & 0xf], 16 * src_h / dst_h,
                       16 / factor, 16 / factor);
       }
     }
@@ -1725,8 +1726,6 @@
 #endif
 
 static void encode_without_recode_loop(VP9_COMP *cpi,
-                                       size_t *size,
-                                       uint8_t *dest,
                                        int q) {
   VP9_COMMON *const cm = &cpi->common;
   vp9_clear_system_state();
@@ -2162,7 +2161,7 @@
   }
 
   if (cpi->sf.recode_loop == DISALLOW_RECODE) {
-    encode_without_recode_loop(cpi, size, dest, q);
+    encode_without_recode_loop(cpi, q);
   } else {
     encode_with_recode_loop(cpi, size, dest, q, bottom_index, top_index);
   }
@@ -2779,6 +2778,9 @@
 int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
                               vp9_ppflags_t *flags) {
   VP9_COMMON *cm = &cpi->common;
+#if !CONFIG_VP9_POSTPROC
+  (void)flags;
+#endif
 
   if (!cm->show_frame) {
     return -1;
@@ -2787,7 +2789,6 @@
 #if CONFIG_VP9_POSTPROC
     ret = vp9_post_proc_frame(cm, dest, flags);
 #else
-
     if (cm->frame_to_show) {
       *dest = *cm->frame_to_show;
       dest->y_width = cm->width;
@@ -2798,64 +2799,13 @@
     } else {
       ret = -1;
     }
-
 #endif  // !CONFIG_VP9_POSTPROC
     vp9_clear_system_state();
     return ret;
   }
 }
 
-int vp9_set_roimap(VP9_COMP *cpi, unsigned char *map, unsigned int rows,
-                   unsigned int cols, int delta_q[MAX_SEGMENTS],
-                   int delta_lf[MAX_SEGMENTS],
-                   unsigned int threshold[MAX_SEGMENTS]) {
-  signed char feature_data[SEG_LVL_MAX][MAX_SEGMENTS];
-  struct segmentation *seg = &cpi->common.seg;
-  const VP9_COMMON *const cm = &cpi->common;
-  int i;
-
-  if (cm->mb_rows != rows || cm->mb_cols != cols)
-    return -1;
-
-  if (!map) {
-    vp9_disable_segmentation(seg);
-    return 0;
-  }
-
-  vpx_memcpy(cpi->segmentation_map, map, cm->mi_rows * cm->mi_cols);
-
-  // Activate segmentation.
-  vp9_enable_segmentation(seg);
-
-  // Set up the quant, LF and breakout threshold segment data
-  for (i = 0; i < MAX_SEGMENTS; i++) {
-    feature_data[SEG_LVL_ALT_Q][i] = delta_q[i];
-    feature_data[SEG_LVL_ALT_LF][i] = delta_lf[i];
-    cpi->segment_encode_breakout[i] = threshold[i];
-  }
-
-  // Enable the loop and quant changes in the feature mask
-  for (i = 0; i < MAX_SEGMENTS; i++) {
-    if (delta_q[i])
-      vp9_enable_segfeature(seg, i, SEG_LVL_ALT_Q);
-    else
-      vp9_disable_segfeature(seg, i, SEG_LVL_ALT_Q);
-
-    if (delta_lf[i])
-      vp9_enable_segfeature(seg, i, SEG_LVL_ALT_LF);
-    else
-      vp9_disable_segfeature(seg, i, SEG_LVL_ALT_LF);
-  }
-
-  // Initialize the feature data structure
-  // SEGMENT_DELTADATA    0, SEGMENT_ABSDATA      1
-  vp9_set_segment_data(seg, &feature_data[0][0], SEGMENT_DELTADATA);
-
-  return 0;
-}
-
-int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map,
-                       unsigned int rows, unsigned int cols) {
+int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols) {
   if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols) {
     if (map) {
       vpx_memcpy(cpi->active_map, map, rows * cols);
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h
index f48909e..47c9019 100644
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -24,6 +24,7 @@
 #include "vp9/common/vp9_onyxc_int.h"
 
 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
+#include "vp9/encoder/vp9_context_tree.h"
 #include "vp9/encoder/vp9_encodemb.h"
 #include "vp9/encoder/vp9_firstpass.h"
 #include "vp9/encoder/vp9_lookahead.h"
@@ -391,7 +392,6 @@
   RATE_CONTROL rc;
 
   vp9_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
-  vp9_coeff_probs_model frame_coef_probs[TX_SIZES][PLANE_TYPES];
 
   struct vpx_codec_pkt_list  *output_pkt_list;
 
@@ -411,8 +411,8 @@
   // Default value is 1. From first pass stats, encode_breakout may be disabled.
   ENCODE_BREAKOUT_TYPE allow_encode_breakout;
 
-  // Get threshold from external input. In real time mode, it can be
-  // overwritten according to encoding speed.
+  // Get threshold from external input. A suggested threshold is 800 for HD
+  // clips, and 300 for < HD clips.
   int encode_breakout;
 
   unsigned char *segmentation_map;
@@ -438,7 +438,7 @@
   uint64_t time_pick_lpf;
   uint64_t time_encode_sb_row;
 
-  struct twopass_rc twopass;
+  TWO_PASS twopass;
 
   YV12_BUFFER_CONFIG alt_ref_buffer;
   YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS];
@@ -503,6 +503,11 @@
   int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
   int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
 
+  PICK_MODE_CONTEXT *leaf_tree;
+  PC_TREE *pc_tree;
+  PC_TREE *pc_root;
+  int partition_cost[PARTITION_CONTEXTS][PARTITION_TYPES];
+
 #if CONFIG_MULTIPLE_ARF
   // ARF tracking variables.
   int multi_arf_enabled;
@@ -552,14 +557,7 @@
 
 int vp9_update_entropy(VP9_COMP *cpi, int update);
 
-int vp9_set_roimap(VP9_COMP *cpi, unsigned char *map,
-                   unsigned int rows, unsigned int cols,
-                   int delta_q[MAX_SEGMENTS],
-                   int delta_lf[MAX_SEGMENTS],
-                   unsigned int threshold[MAX_SEGMENTS]);
-
-int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map,
-                       unsigned int rows, unsigned int cols);
+int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols);
 
 int vp9_set_internal_size(VP9_COMP *cpi,
                           VPX_SCALING horiz_mode, VPX_SCALING vert_mode);
diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c
index 43ae7c1..9929ae1 100644
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -46,6 +46,9 @@
 #define GF_RMAX    96.0
 #define ERR_DIVISOR   150.0
 #define MIN_DECAY_FACTOR 0.1
+#define SVC_FACTOR_PT_LOW 0.45
+#define FACTOR_PT_LOW 0.5
+#define FACTOR_PT_HIGH 0.9
 
 #define KF_MB_INTRA_MIN 150
 #define GF_MB_INTRA_MIN 100
@@ -61,6 +64,7 @@
 #define MIN_GF_INTERVAL             4
 #endif
 
+
 // #define LONG_TERM_VBR_CORRECTION
 
 static void swap_yv12(YV12_BUFFER_CONFIG *a, YV12_BUFFER_CONFIG *b) {
@@ -78,12 +82,12 @@
 
 // Resets the first pass file to the given position using a relative seek from
 // the current position.
-static void reset_fpf_position(struct twopass_rc *p,
+static void reset_fpf_position(TWO_PASS *p,
                                const FIRSTPASS_STATS *position) {
   p->stats_in = position;
 }
 
-static int lookup_next_frame_stats(const struct twopass_rc *p,
+static int lookup_next_frame_stats(const TWO_PASS *p,
                                    FIRSTPASS_STATS *next_frame) {
   if (p->stats_in >= p->stats_in_end)
     return EOF;
@@ -94,7 +98,7 @@
 
 
 // Read frame stats at an offset from the current position.
-static int read_frame_stats(const struct twopass_rc *p,
+static int read_frame_stats(const TWO_PASS *p,
                             FIRSTPASS_STATS *frame_stats, int offset) {
   const FIRSTPASS_STATS *fps_ptr = p->stats_in;
 
@@ -111,7 +115,7 @@
   return 1;
 }
 
-static int input_stats(struct twopass_rc *p, FIRSTPASS_STATS *fps) {
+static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
   if (p->stats_in >= p->stats_in_end)
     return EOF;
 
@@ -257,7 +261,7 @@
 // harder frames.
 static double calculate_modified_err(const VP9_COMP *cpi,
                                      const FIRSTPASS_STATS *this_frame) {
-  const struct twopass_rc *twopass = &cpi->twopass;
+  const TWO_PASS *twopass = &cpi->twopass;
   const SVC *const svc = &cpi->svc;
   const FIRSTPASS_STATS *stats;
   double av_err;
@@ -475,7 +479,7 @@
   TileInfo tile;
   struct macroblock_plane *const p = x->plane;
   struct macroblockd_plane *const pd = xd->plane;
-  const PICK_MODE_CONTEXT *ctx = &x->pc_root->none;
+  const PICK_MODE_CONTEXT *ctx = &cpi->pc_root->none;
   int i;
 
   int recon_yoffset, recon_uvoffset;
@@ -500,7 +504,7 @@
   int new_mv_count = 0;
   int sum_in_vectors = 0;
   uint32_t lastmv_as_int = 0;
-  struct twopass_rc *twopass = &cpi->twopass;
+  TWO_PASS *twopass = &cpi->twopass;
   const MV zero_mv = {0, 0};
   const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
 
@@ -540,7 +544,7 @@
 
   vp9_setup_src_planes(x, cpi->Source, 0, 0);
   vp9_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
-  vp9_setup_dst_planes(xd, new_yv12, 0, 0);
+  vp9_setup_dst_planes(xd->plane, new_yv12, 0, 0);
 
   xd->mi = cm->mi_grid_visible;
   xd->mi[0] = cm->mi;
@@ -937,8 +941,8 @@
     for (q = rc->best_quality; q < rc->worst_quality; ++q) {
       const double factor =
           calc_correction_factor(err_per_mb, ERR_DIVISOR,
-                                 is_svc_upper_layer ? 0.8 : 0.5,
-                                 is_svc_upper_layer ? 1.0 : 0.90, q);
+                                 is_svc_upper_layer ? SVC_FACTOR_PT_LOW :
+                                 FACTOR_PT_LOW, FACTOR_PT_HIGH, q);
       const int bits_per_mb = vp9_rc_bits_per_mb(INTER_FRAME, q,
                                                  factor * speed_term);
       if (bits_per_mb <= target_norm_bits_per_mb)
@@ -959,7 +963,7 @@
   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   const int is_spatial_svc = (svc->number_spatial_layers > 1) &&
                              (svc->number_temporal_layers == 1);
-  struct twopass_rc *const twopass = is_spatial_svc ?
+  TWO_PASS *const twopass = is_spatial_svc ?
       &svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
   double frame_rate;
   FIRSTPASS_STATS *stats;
@@ -1007,25 +1011,6 @@
   // This variable monitors how far behind the second ref update is lagging.
   twopass->sr_update_lag = 1;
 
-  // Scan the first pass file and calculate an average Intra / Inter error
-  // score ratio for the sequence.
-  {
-    const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
-    FIRSTPASS_STATS this_frame;
-    double sum_iiratio = 0.0;
-
-    while (input_stats(twopass, &this_frame) != EOF) {
-      const double iiratio = this_frame.intra_error /
-                                 DOUBLE_DIVIDE_CHECK(this_frame.coded_error);
-      sum_iiratio += fclamp(iiratio, 1.0, 20.0);
-    }
-
-    twopass->avg_iiratio = sum_iiratio /
-                               DOUBLE_DIVIDE_CHECK((double)stats->count);
-
-    reset_fpf_position(twopass, start_pos);
-  }
-
   // Scan the first pass file and calculate a modified total error based upon
   // the bias/power function used to allocate bits.
   {
@@ -1072,7 +1057,7 @@
 // Function to test for a condition where a complex transition is followed
 // by a static section. For example in slide shows where there is a fade
 // between slides. This is to help with more optimal kf and gf positioning.
-static int detect_transition_to_still(struct twopass_rc *twopass,
+static int detect_transition_to_still(TWO_PASS *twopass,
                                       int frame_interval, int still_interval,
                                       double loop_decay_rate,
                                       double last_decay_rate) {
@@ -1110,7 +1095,7 @@
 // This function detects a flash through the high relative pcnt_second_ref
 // score in the frame following a flash frame. The offset passed in should
 // reflect this.
-static int detect_flash(const struct twopass_rc *twopass, int offset) {
+static int detect_flash(const TWO_PASS *twopass, int offset) {
   FIRSTPASS_STATS next_frame;
 
   int flash_detected = 0;
@@ -1196,7 +1181,7 @@
                           int f_frames, int b_frames,
                           int *f_boost, int *b_boost) {
   FIRSTPASS_STATS this_frame;
-  struct twopass_rc *const twopass = &cpi->twopass;
+  TWO_PASS *const twopass = &cpi->twopass;
   int i;
   double boost_score = 0.0;
   double mv_ratio_accumulator = 0.0;
@@ -1418,34 +1403,29 @@
 #endif
 
 // Calculate a section intra ratio used in setting max loop filter.
-static void calculate_section_intra_ratio(struct twopass_rc *twopass,
-                                          const FIRSTPASS_STATS *start_pos,
-                                          int section_length) {
-  FIRSTPASS_STATS next_frame = { 0 };
-  FIRSTPASS_STATS sectionstats = { 0 };
-  int i;
+static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
+                                         const FIRSTPASS_STATS *end,
+                                         int section_length) {
+  const FIRSTPASS_STATS *s = begin;
+  double intra_error = 0.0;
+  double coded_error = 0.0;
+  int i = 0;
 
-  reset_fpf_position(twopass, start_pos);
-
-  for (i = 0; i < section_length; ++i) {
-    input_stats(twopass, &next_frame);
-    accumulate_stats(&sectionstats, &next_frame);
+  while (s < end && i < section_length) {
+    intra_error += s->intra_error;
+    coded_error += s->coded_error;
+    ++s;
+    ++i;
   }
 
-  avg_stats(&sectionstats);
-
-  twopass->section_intra_rating =
-    (int)(sectionstats.intra_error /
-          DOUBLE_DIVIDE_CHECK(sectionstats.coded_error));
-
-  reset_fpf_position(twopass, start_pos);
+  return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
 }
 
 // Calculate the total bits to allocate in this GF/ARF group.
 static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
                                              double gf_group_err) {
   const RATE_CONTROL *const rc = &cpi->rc;
-  const struct twopass_rc *const twopass = &cpi->twopass;
+  const TWO_PASS *const twopass = &cpi->twopass;
   const int max_bits = frame_max_bits(rc, &cpi->oxcf);
   int64_t total_group_bits;
 
@@ -1469,13 +1449,36 @@
   return total_group_bits;
 }
 
+// Calculate the number bits extra to assign to boosted frames in a group.
+static int calculate_boost_bits(int frame_count,
+                                int boost, int64_t total_group_bits) {
+  int allocation_chunks;
+
+  // return 0 for invalid inputs (could arise e.g. through rounding errors)
+  if (!boost || (total_group_bits <= 0) || (frame_count <= 0) )
+    return 0;
+
+  allocation_chunks = (frame_count * 100) + boost;
+
+  // Prevent overflow.
+  if (boost > 1023) {
+    int divisor = boost >> 10;
+    boost /= divisor;
+    allocation_chunks /= divisor;
+  }
+
+  // Calculate the number of extra bits for use in the boosted frame or frames.
+  return MAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks), 0);
+}
+
+
 // Analyse and define a gf/arf group.
 static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   RATE_CONTROL *const rc = &cpi->rc;
   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
-  struct twopass_rc *const twopass = &cpi->twopass;
-  FIRSTPASS_STATS next_frame = { 0 };
-  const FIRSTPASS_STATS *start_pos;
+  TWO_PASS *const twopass = &cpi->twopass;
+  FIRSTPASS_STATS next_frame;
+  const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
   int i;
   double boost_score = 0.0;
   double old_boost_score = 0.0;
@@ -1501,11 +1504,10 @@
   int flash_detected;
   int active_max_gf_interval;
 
-  twopass->gf_group_bits = 0;
-
   vp9_clear_system_state();
+  vp9_zero(next_frame);
 
-  start_pos = twopass->stats_in;
+  twopass->gf_group_bits = 0;
 
   // Load stats for the current frame.
   mod_frame_err = calculate_modified_err(cpi, this_frame);
@@ -1707,127 +1709,65 @@
   }
 #endif
 #endif
+  // Reset the file position.
+  reset_fpf_position(twopass, start_pos);
 
   // Calculate the bits to be allocated to the gf/arf group as a whole
   twopass->gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
 
-  // Reset the file position.
-  reset_fpf_position(twopass, start_pos);
-
-  // Assign  bits to the arf or gf.
-  for (i = 0; i <= (rc->source_alt_ref_pending &&
-                    cpi->common.frame_type != KEY_FRAME); ++i) {
-    int allocation_chunks;
+  // Calculate the extra bits to be used for boosted frame(s)
+  {
     int q = rc->last_q[INTER_FRAME];
-    int gf_bits;
-
     int boost = (rc->gfu_boost * gfboost_qadjust(q)) / 100;
 
     // Set max and minimum boost and hence minimum allocation.
     boost = clamp(boost, 125, (rc->baseline_gf_interval + 1) * 200);
 
-    if (rc->source_alt_ref_pending && i == 0)
-      allocation_chunks = ((rc->baseline_gf_interval + 1) * 100) + boost;
-    else
-      allocation_chunks = (rc->baseline_gf_interval * 100) + (boost - 100);
+    // Calculate the extra bits to be used for boosted frame(s)
+    twopass->gf_bits = calculate_boost_bits(rc->baseline_gf_interval,
+                                            boost, twopass->gf_group_bits);
 
-    // Prevent overflow.
-    if (boost > 1023) {
-      int divisor = boost >> 10;
-      boost /= divisor;
-      allocation_chunks /= divisor;
-    }
 
-    // Calculate the number of bits to be spent on the gf or arf based on
-    // the boost number.
-    gf_bits = (int)((double)boost * (twopass->gf_group_bits /
-                  (double)allocation_chunks));
-
-    // If the frame that is to be boosted is simpler than the average for
-    // the gf/arf group then use an alternative calculation
-    // based on the error score of the frame itself.
-    if (rc->baseline_gf_interval < 1 ||
-        mod_frame_err < gf_group_err / (double)rc->baseline_gf_interval) {
-      double alt_gf_grp_bits = (double)twopass->kf_group_bits  *
-        (mod_frame_err * (double)rc->baseline_gf_interval) /
-        DOUBLE_DIVIDE_CHECK(twopass->kf_group_error_left);
-
-      int alt_gf_bits = (int)((double)boost * (alt_gf_grp_bits /
-                                           (double)allocation_chunks));
-
-      if (gf_bits > alt_gf_bits)
-        gf_bits = alt_gf_bits;
-    } else {
-      // If it is harder than other frames in the group make sure it at
-      // least receives an allocation in keeping with its relative error
-      // score, otherwise it may be worse off than an "un-boosted" frame.
-      int alt_gf_bits = (int)((double)twopass->kf_group_bits *
-                        mod_frame_err /
-                        DOUBLE_DIVIDE_CHECK(twopass->kf_group_error_left));
-
-      if (alt_gf_bits > gf_bits)
-        gf_bits = alt_gf_bits;
-    }
-
-    // Don't allow a negative value for gf_bits.
-    if (gf_bits < 0)
-      gf_bits = 0;
-
-    if (i == 0) {
-      twopass->gf_bits = gf_bits;
-    }
-    if (i == 1 ||
-        (!rc->source_alt_ref_pending && cpi->common.frame_type != KEY_FRAME &&
-         !vp9_is_upper_layer_key_frame(cpi))) {
-      // Calculate the per frame bit target for this frame.
-      vp9_rc_set_frame_target(cpi, gf_bits);
+    // For key frames the frame target rate is set already.
+    // NOTE: We dont bother to check for the special case of ARF overlay
+    // frames here, as there is clamping code for this in the function
+    // vp9_rc_clamp_pframe_target_size(), which applies to one and two pass
+    // encodes.
+    if (cpi->common.frame_type != KEY_FRAME &&
+        !vp9_is_upper_layer_key_frame(cpi)) {
+      vp9_rc_set_frame_target(cpi, twopass->gf_bits);
     }
   }
 
-  {
-    // Adjust KF group bits and error remaining.
-    twopass->kf_group_error_left -= (int64_t)gf_group_err;
+  // Adjust KF group bits and error remaining.
+  twopass->kf_group_error_left -= (int64_t)gf_group_err;
 
-    // If this is an arf update we want to remove the score for the overlay
-    // frame at the end which will usually be very cheap to code.
-    // The overlay frame has already, in effect, been coded so we want to spread
-    // the remaining bits among the other frames.
-    // For normal GFs remove the score for the GF itself unless this is
-    // also a key frame in which case it has already been accounted for.
-    if (rc->source_alt_ref_pending) {
-      twopass->gf_group_error_left = (int64_t)(gf_group_err - mod_frame_err);
-    } else if (cpi->common.frame_type != KEY_FRAME) {
-      twopass->gf_group_error_left = (int64_t)(gf_group_err
-                                                   - gf_first_frame_err);
-    } else {
-      twopass->gf_group_error_left = (int64_t)gf_group_err;
-    }
-
-    // This condition could fail if there are two kfs very close together
-    // despite MIN_GF_INTERVAL and would cause a divide by 0 in the
-    // calculation of alt_extra_bits.
-    if (rc->baseline_gf_interval >= 3) {
-      const int boost = rc->source_alt_ref_pending ? b_boost : rc->gfu_boost;
-
-      if (boost >= 150) {
-        const int pct_extra = MIN(20, (boost - 100) / 50);
-        const int alt_extra_bits = (int)((
-            MAX(twopass->gf_group_bits - twopass->gf_bits, 0) *
-            pct_extra) / 100);
-        twopass->gf_group_bits -= alt_extra_bits;
-      }
-    }
+  // If this is an arf update we want to remove the score for the overlay
+  // frame at the end which will usually be very cheap to code.
+  // The overlay frame has already, in effect, been coded so we want to spread
+  // the remaining bits among the other frames.
+  // For normal GFs remove the score for the GF itself unless this is
+  // also a key frame in which case it has already been accounted for.
+  if (rc->source_alt_ref_pending) {
+    twopass->gf_group_error_left = (int64_t)(gf_group_err - mod_frame_err);
+  } else if (cpi->common.frame_type != KEY_FRAME) {
+    twopass->gf_group_error_left = (int64_t)(gf_group_err
+                                                 - gf_first_frame_err);
+  } else {
+    twopass->gf_group_error_left = (int64_t)gf_group_err;
   }
 
   // Calculate a section intra ratio used in setting max loop filter.
   if (cpi->common.frame_type != KEY_FRAME) {
-    calculate_section_intra_ratio(twopass, start_pos, rc->baseline_gf_interval);
+    twopass->section_intra_rating =
+        calculate_section_intra_ratio(start_pos, twopass->stats_in_end,
+                                      rc->baseline_gf_interval);
   }
 }
 
 // Allocate bits to a normal frame that is neither a gf an arf or a key frame.
 static void assign_std_frame_bits(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
-  struct twopass_rc *twopass = &cpi->twopass;
+  TWO_PASS *twopass = &cpi->twopass;
   // For a single frame.
   const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
   // Calculate modified prediction error used in bit allocation.
@@ -1856,7 +1796,7 @@
   vp9_rc_set_frame_target(cpi, target_frame_size);
 }
 
-static int test_candidate_kf(struct twopass_rc *twopass,
+static int test_candidate_kf(TWO_PASS *twopass,
                              const FIRSTPASS_STATS *last_frame,
                              const FIRSTPASS_STATS *this_frame,
                              const FIRSTPASS_STATS *next_frame) {
@@ -1936,9 +1876,9 @@
 static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   int i, j;
   RATE_CONTROL *const rc = &cpi->rc;
-  struct twopass_rc *const twopass = &cpi->twopass;
+  TWO_PASS *const twopass = &cpi->twopass;
   const FIRSTPASS_STATS first_frame = *this_frame;
-  const FIRSTPASS_STATS *start_position = twopass->stats_in;
+  const FIRSTPASS_STATS *const start_position = twopass->stats_in;
   FIRSTPASS_STATS next_frame;
   FIRSTPASS_STATS last_frame;
   double decay_accumulator = 1.0;
@@ -2072,15 +2012,15 @@
   } else {
     twopass->kf_group_bits = 0;
   }
+  twopass->kf_group_bits = MAX(0, twopass->kf_group_bits);
+
   // Reset the first pass file position.
   reset_fpf_position(twopass, start_position);
 
-  // Determine how big to make this keyframe based on how well the subsequent
-  // frames use inter blocks.
+  // Scan through the kf group collating various stats used to deteermine
+  // how many bits to spend on it.
   decay_accumulator = 1.0;
   boost_score = 0.0;
-
-  // Scan through the kf group collating various stats.
   for (i = 0; i < rc->frames_to_key; ++i) {
     if (EOF == input_stats(twopass, &next_frame))
       break;
@@ -2117,84 +2057,29 @@
     }
   }
 
+  // Store the zero motion percentage
+  twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
+
   // Calculate a section intra ratio used in setting max loop filter.
-  calculate_section_intra_ratio(twopass, start_position, rc->frames_to_key);
+  twopass->section_intra_rating =
+      calculate_section_intra_ratio(start_position, twopass->stats_in_end,
+                                    rc->frames_to_key);
 
   // Work out how many bits to allocate for the key frame itself.
-  if (1) {
-    int kf_boost = (int)boost_score;
-    int allocation_chunks;
+  rc->kf_boost = (int)boost_score;
 
-    if (kf_boost < (rc->frames_to_key * 3))
-      kf_boost = (rc->frames_to_key * 3);
+  if (rc->kf_boost  < (rc->frames_to_key * 3))
+    rc->kf_boost  = (rc->frames_to_key * 3);
+  if (rc->kf_boost   < MIN_KF_BOOST)
+    rc->kf_boost = MIN_KF_BOOST;
 
-    if (kf_boost < MIN_KF_BOOST)
-      kf_boost = MIN_KF_BOOST;
+  twopass->kf_bits = calculate_boost_bits((rc->frames_to_key - 1),
+                                          rc->kf_boost, twopass->kf_group_bits);
 
-    // Make a note of baseline boost and the zero motion
-    // accumulator value for use elsewhere.
-    rc->kf_boost = kf_boost;
-    twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
+  twopass->kf_group_bits -= twopass->kf_bits;
 
-    // Key frame size depends on:
-    // (1) the error score for the whole key frame group,
-    // (2) the key frames' own error if this is smaller than the
-    //     average for the group (optional),
-    // (3) insuring that the frame receives at least the allocation it would
-    //     have received based on its own error score vs the error score
-    //     remaining.
-    // Special case:
-    // If the sequence appears almost totally static we want to spend almost
-    // all of the bits on the key frame.
-    //
-    // We use (cpi->rc.frames_to_key - 1) below because the key frame itself is
-    // taken care of by kf_boost.
-    if (zero_motion_accumulator >= 0.99) {
-      allocation_chunks = ((rc->frames_to_key - 1) * 10) + kf_boost;
-    } else {
-      allocation_chunks = ((rc->frames_to_key - 1) * 100) + kf_boost;
-    }
-
-    // Prevent overflow.
-    if (kf_boost > 1028) {
-      const int divisor = kf_boost >> 10;
-      kf_boost /= divisor;
-      allocation_chunks /= divisor;
-    }
-
-    twopass->kf_group_bits = MAX(0, twopass->kf_group_bits);
-    // Calculate the number of bits to be spent on the key frame.
-    twopass->kf_bits = (int)((double)kf_boost *
-        ((double)twopass->kf_group_bits / allocation_chunks));
-
-    // If the key frame is actually easier than the average for the
-    // kf group (which does sometimes happen, e.g. a blank intro frame)
-    // then use an alternate calculation based on the kf error score
-    // which should give a smaller key frame.
-    if (kf_mod_err < kf_group_err / rc->frames_to_key) {
-      double alt_kf_grp_bits = ((double)twopass->bits_left *
-         (kf_mod_err * (double)rc->frames_to_key) /
-         DOUBLE_DIVIDE_CHECK(twopass->modified_error_left));
-
-      const int alt_kf_bits = (int)((double)kf_boost *
-                          (alt_kf_grp_bits / (double)allocation_chunks));
-
-      if (twopass->kf_bits > alt_kf_bits)
-        twopass->kf_bits = alt_kf_bits;
-    } else {
-      // Else if it is much harder than other frames in the group make sure
-      // it at least receives an allocation in keeping with its relative
-      // error score.
-      const int alt_kf_bits = (int)((double)twopass->bits_left * (kf_mod_err /
-               DOUBLE_DIVIDE_CHECK(twopass->modified_error_left)));
-
-      if (alt_kf_bits > twopass->kf_bits)
-        twopass->kf_bits = alt_kf_bits;
-    }
-    twopass->kf_group_bits -= twopass->kf_bits;
-    // Per frame bit target for this frame.
-    vp9_rc_set_frame_target(cpi, twopass->kf_bits);
-  }
+  // Per frame bit target for this frame.
+  vp9_rc_set_frame_target(cpi, twopass->kf_bits);
 
   // Note the total error score of the kf group minus the key frame itself.
   twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
@@ -2238,17 +2123,15 @@
 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
   RATE_CONTROL *const rc = &cpi->rc;
-  struct twopass_rc *const twopass = &cpi->twopass;
+  TWO_PASS *const twopass = &cpi->twopass;
   int frames_left;
   FIRSTPASS_STATS this_frame;
   FIRSTPASS_STATS this_frame_copy;
 
-  double this_frame_intra_error;
-  double this_frame_coded_error;
   int target;
   LAYER_CONTEXT *lc = NULL;
-  int is_spatial_svc = (cpi->use_svc && cpi->svc.number_temporal_layers == 1);
-
+  const int is_spatial_svc = (cpi->use_svc &&
+                              cpi->svc.number_temporal_layers == 1);
   if (is_spatial_svc) {
     lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
     frames_left = (int)(twopass->total_stats.count -
@@ -2298,27 +2181,27 @@
   if (EOF == input_stats(twopass, &this_frame))
     return;
 
-  this_frame_intra_error = this_frame.intra_error;
-  this_frame_coded_error = this_frame.coded_error;
-
   // Keyframe and section processing.
   if (rc->frames_to_key == 0 ||
       (cpi->frame_flags & FRAMEFLAGS_KEY)) {
     // Define next KF group and assign bits to it.
     this_frame_copy = this_frame;
     find_next_key_frame(cpi, &this_frame_copy);
-    // Don't place key frame in any enhancement layers in spatial svc
-    if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
-      lc->is_key_frame = 1;
-      if (cpi->svc.spatial_layer_id > 0) {
-        cm->frame_type = INTER_FRAME;
+  } else {
+    cm->frame_type = INTER_FRAME;
+  }
+
+  if (is_spatial_svc) {
+    if (cpi->svc.spatial_layer_id == 0) {
+      lc->is_key_frame = (cm->frame_type == KEY_FRAME);
+    } else {
+      cm->frame_type = INTER_FRAME;
+      lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
+
+      if (lc->is_key_frame) {
+        cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
       }
     }
-  } else {
-    if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
-      lc->is_key_frame = 0;
-    }
-    cm->frame_type = INTER_FRAME;
   }
 
   // Is this frame a GF / ARF? (Note: a key frame is always also a GF).
@@ -2355,9 +2238,6 @@
     assign_std_frame_bits(cpi, &this_frame_copy);
   }
 
-  // Keep a globally available copy of this and the next frame's iiratio.
-  twopass->this_iiratio = (int)(this_frame_intra_error /
-                              DOUBLE_DIVIDE_CHECK(this_frame_coded_error));
   {
     FIRSTPASS_STATS next_frame;
     if (lookup_next_frame_stats(twopass, &next_frame) != EOF) {
@@ -2384,6 +2264,7 @@
 }
 
 void vp9_twopass_postencode_update(VP9_COMP *cpi) {
+  TWO_PASS *const twopass = &cpi->twopass;
   RATE_CONTROL *const rc = &cpi->rc;
 #ifdef LONG_TERM_VBR_CORRECTION
   // In this experimental mode, the VBR correction is done exclusively through
@@ -2405,14 +2286,13 @@
   // vs. actual bitrate gradually as we progress towards the end of the
   // sequence in order to mitigate this effect.
   const double progress =
-      (double)(cpi->twopass.stats_in - cpi->twopass.stats_in_start) /
-              (cpi->twopass.stats_in_end - cpi->twopass.stats_in_start);
+      (double)(twopass->stats_in - twopass->stats_in_start) /
+              (twopass->stats_in_end - twopass->stats_in_start);
   const int bits_used = (int)(progress * rc->this_frame_target +
                              (1.0 - progress) * rc->projected_frame_size);
 #endif
 
-  cpi->twopass.bits_left -= bits_used;
-  cpi->twopass.bits_left = MAX(cpi->twopass.bits_left, 0);
+  twopass->bits_left = MAX(twopass->bits_left - bits_used, 0);
 
 #ifdef LONG_TERM_VBR_CORRECTION
   if (cpi->common.frame_type != KEY_FRAME &&
@@ -2422,12 +2302,12 @@
       vp9_is_upper_layer_key_frame(cpi)) {
     // For key frames kf_group_bits already had the target bits subtracted out.
     // So now update to the correct value based on the actual bits used.
-    cpi->twopass.kf_group_bits += cpi->rc.this_frame_target - bits_used;
+    twopass->kf_group_bits += rc->this_frame_target - bits_used;
   } else {
 #endif
-    cpi->twopass.kf_group_bits -= bits_used;
-    cpi->twopass.gf_group_bits -= bits_used;
-    cpi->twopass.gf_group_bits = MAX(cpi->twopass.gf_group_bits, 0);
+    twopass->kf_group_bits -= bits_used;
+    twopass->gf_group_bits -= bits_used;
+    twopass->gf_group_bits = MAX(twopass->gf_group_bits, 0);
   }
-  cpi->twopass.kf_group_bits = MAX(cpi->twopass.kf_group_bits, 0);
+  twopass->kf_group_bits = MAX(twopass->kf_group_bits, 0);
 }
diff --git a/vp9/encoder/vp9_firstpass.h b/vp9/encoder/vp9_firstpass.h
index f7ba423..d84793e 100644
--- a/vp9/encoder/vp9_firstpass.h
+++ b/vp9/encoder/vp9_firstpass.h
@@ -38,10 +38,9 @@
   int64_t spatial_layer_id;
 } FIRSTPASS_STATS;
 
-struct twopass_rc {
+typedef struct {
   unsigned int section_intra_rating;
   unsigned int next_iiratio;
-  unsigned int this_iiratio;
   FIRSTPASS_STATS total_stats;
   FIRSTPASS_STATS this_frame_stats;
   const FIRSTPASS_STATS *stats_in;
@@ -50,8 +49,6 @@
   FIRSTPASS_STATS total_left_stats;
   int first_pass_done;
   int64_t bits_left;
-  int64_t clip_bits_total;
-  double avg_iiratio;
   double modified_error_min;
   double modified_error_max;
   double modified_error_total;
@@ -72,7 +69,6 @@
   int64_t gf_group_bits;
   // Bits for the golden frame or ARF - 2 pass only
   int gf_bits;
-  int alt_extra_bits;
 
   int sr_update_lag;
 
@@ -80,7 +76,7 @@
   int gf_zeromotion_pct;
 
   int active_worst_quality;
-};
+} TWO_PASS;
 
 struct VP9_COMP;
 
diff --git a/vp9/encoder/vp9_mbgraph.c b/vp9/encoder/vp9_mbgraph.c
index e7dcc7a..041e583 100644
--- a/vp9/encoder/vp9_mbgraph.c
+++ b/vp9/encoder/vp9_mbgraph.c
@@ -20,7 +20,6 @@
 #include "vp9/common/vp9_systemdependent.h"
 
 
-
 static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
                                               const MV *ref_mv,
                                               MV *dst_mv,
@@ -236,9 +235,10 @@
 
   int mb_col, mb_row, offset = 0;
   int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
-  MV arf_top_mv = {0, 0}, gld_top_mv = {0, 0};
-  MODE_INFO mi_local = { { 0 } };
+  MV gld_top_mv = {0, 0};
+  MODE_INFO mi_local;
 
+  vp9_zero(mi_local);
   // Set up limit values for motion vectors to prevent them extending outside
   // the UMV borders.
   x->mv_row_min     = -BORDER_MV_PIXELS_B16;
@@ -253,7 +253,7 @@
   mi_local.mbmi.ref_frame[1] = NONE;
 
   for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
-    MV arf_left_mv = arf_top_mv, gld_left_mv = gld_top_mv;
+    MV gld_left_mv = gld_top_mv;
     int mb_y_in_offset  = mb_y_offset;
     int arf_y_in_offset = arf_y_offset;
     int gld_y_in_offset = gld_y_offset;
@@ -270,10 +270,8 @@
       update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset,
                               golden_ref, &gld_left_mv, alt_ref,
                               mb_row, mb_col);
-      arf_left_mv = mb_stats->ref[ALTREF_FRAME].m.mv.as_mv;
       gld_left_mv = mb_stats->ref[GOLDEN_FRAME].m.mv.as_mv;
       if (mb_col == 0) {
-        arf_top_mv = arf_left_mv;
         gld_top_mv = gld_left_mv;
       }
       xd->left_available = 1;
diff --git a/vp9/encoder/vp9_mcomp.c b/vp9/encoder/vp9_mcomp.c
index 43c8ab8..4f7d6f1 100644
--- a/vp9/encoder/vp9_mcomp.c
+++ b/vp9/encoder/vp9_mcomp.c
@@ -886,6 +886,10 @@
   int r, c, i;
   int start_col, end_col, start_row, end_row;
 
+  // The cfg and search_param parameters are not used in this search variant
+  (void)cfg;
+  (void)search_param;
+
   clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
   *best_mv = *ref_mv;
   *num00 = 11;
@@ -1551,7 +1555,7 @@
                              int search_range,
                              const vp9_variance_fn_ptr_t *fn_ptr,
                              const MV *center_mv,
-                             const uint8_t *second_pred, int w, int h) {
+                             const uint8_t *second_pred) {
   const MV neighbors[8] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0},
                            {-1, -1}, {1, -1}, {-1, 1}, {1, 1}};
   const MACROBLOCKD *const xd = &x->e_mbd;
diff --git a/vp9/encoder/vp9_mcomp.h b/vp9/encoder/vp9_mcomp.h
index 827957d..873edf3 100644
--- a/vp9/encoder/vp9_mcomp.h
+++ b/vp9/encoder/vp9_mcomp.h
@@ -144,8 +144,7 @@
                              MV *ref_mv, int error_per_bit,
                              int search_range,
                              const vp9_variance_fn_ptr_t *fn_ptr,
-                             const MV *center_mv, const uint8_t *second_pred,
-                             int w, int h);
+                             const MV *center_mv, const uint8_t *second_pred);
 #ifdef __cplusplus
 }  // extern "C"
 #endif
diff --git a/vp9/encoder/vp9_pickmode.c b/vp9/encoder/vp9_pickmode.c
index 78fba73..437b680 100644
--- a/vp9/encoder/vp9_pickmode.c
+++ b/vp9/encoder/vp9_pickmode.c
@@ -31,7 +31,7 @@
                                     int_mv *tmp_mv, int *rate_mv) {
   MACROBLOCKD *xd = &x->e_mbd;
   MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
-  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
   int step_param;
   int sadpb = x->sadperbit16;
   MV mvp_full;
@@ -79,7 +79,7 @@
   if (x->mv_best_ref_index[ref] < 2)
     mvp_full = mbmi->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
   else
-    mvp_full = x->pred_mv[ref].as_mv;
+    mvp_full = x->pred_mv[ref];
 
   mvp_full.col >>= 3;
   mvp_full.row >>= 3;
@@ -110,7 +110,7 @@
                                     MV *tmp_mv) {
   MACROBLOCKD *xd = &x->e_mbd;
   MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
-  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
   int ref = mbmi->ref_frame[0];
   MV ref_mv = mbmi->ref_mvs[ref][0].as_mv;
   int dis;
@@ -143,12 +143,13 @@
       xd->plane[i].pre[0] = backup_yv12[i];
   }
 
-  x->pred_mv[ref].as_mv = *tmp_mv;
+  x->pred_mv[ref] = *tmp_mv;
 }
 
 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
                               MACROBLOCK *x, MACROBLOCKD *xd,
-                              int *out_rate_sum, int64_t *out_dist_sum) {
+                              int *out_rate_sum, int64_t *out_dist_sum,
+                              unsigned int *var_y, unsigned int *sse_y) {
   // Note our transform coeffs are 8 times an orthogonal transform.
   // Hence quantizer step is also 8 times. To get effective quantizer
   // we need to divide by 8 before sending to modeling function.
@@ -162,6 +163,9 @@
   unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
                                            pd->dst.buf, pd->dst.stride, &sse);
 
+  *var_y = var;
+  *sse_y = sse;
+
   // TODO(jingning) This is a temporary solution to account for frames with
   // light changes. Need to customize the rate-distortion modeling for non-RD
   // mode decision.
@@ -198,6 +202,9 @@
 
   int rate = INT_MAX;
   int64_t dist = INT64_MAX;
+  // var_y and sse_y are saved to be used in skipping checking
+  unsigned int var_y = UINT_MAX;
+  unsigned int sse_y = UINT_MAX;
 
   VP9_COMMON *cm = &cpi->common;
   int intra_cost_penalty = 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q);
@@ -219,8 +226,7 @@
   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
 
   x->skip = 0;
-  if (!x->in_active_map)
-    x->skip = 1;
+
   // initialize mode decisions
   *returnrate = INT_MAX;
   *returndistortion = INT64_MAX;
@@ -318,54 +324,37 @@
           pred_filter_search &&
           ((mbmi->mv[0].as_mv.row & 0x07) != 0 ||
            (mbmi->mv[0].as_mv.col & 0x07) != 0)) {
-        int64_t tmp_rdcost1 = INT64_MAX;
-        int64_t tmp_rdcost2 = INT64_MAX;
-        int64_t tmp_rdcost3 = INT64_MAX;
         int pf_rate[3];
         int64_t pf_dist[3];
+        unsigned int pf_var[3];
+        unsigned int pf_sse[3];
+        int64_t best_cost = INT64_MAX;
+        INTERP_FILTER best_filter = SWITCHABLE, filter;
 
-        mbmi->interp_filter = EIGHTTAP;
-        vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
-        model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[EIGHTTAP],
-                          &pf_dist[EIGHTTAP]);
-        tmp_rdcost1 = RDCOST(x->rdmult, x->rddiv,
-                             vp9_get_switchable_rate(cpi) + pf_rate[EIGHTTAP],
-                             pf_dist[EIGHTTAP]);
-
-        mbmi->interp_filter = EIGHTTAP_SHARP;
-        vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
-        model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[EIGHTTAP_SHARP],
-                          &pf_dist[EIGHTTAP_SHARP]);
-        tmp_rdcost2 = RDCOST(x->rdmult, x->rddiv, vp9_get_switchable_rate(cpi) +
-                                 pf_rate[EIGHTTAP_SHARP],
-                             pf_dist[EIGHTTAP_SHARP]);
-
-        mbmi->interp_filter = EIGHTTAP_SMOOTH;
-        vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
-        model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[EIGHTTAP_SMOOTH],
-                          &pf_dist[EIGHTTAP_SMOOTH]);
-        tmp_rdcost3 = RDCOST(x->rdmult, x->rddiv, vp9_get_switchable_rate(cpi) +
-                                 pf_rate[EIGHTTAP_SMOOTH],
-                             pf_dist[EIGHTTAP_SMOOTH]);
-
-        if (tmp_rdcost2 < tmp_rdcost1) {
-          if (tmp_rdcost2 < tmp_rdcost3)
-            mbmi->interp_filter = EIGHTTAP_SHARP;
-          else
-            mbmi->interp_filter = EIGHTTAP_SMOOTH;
-        } else {
-          if (tmp_rdcost1 < tmp_rdcost3)
-            mbmi->interp_filter = EIGHTTAP;
-          else
-            mbmi->interp_filter = EIGHTTAP_SMOOTH;
+        for (filter = EIGHTTAP; filter <= EIGHTTAP_SHARP; ++filter) {
+          int64_t cost;
+          mbmi->interp_filter = filter;
+          vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
+          model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter],
+                            &pf_dist[filter], &pf_var[filter], &pf_sse[filter]);
+          cost = RDCOST(x->rdmult, x->rddiv,
+                        vp9_get_switchable_rate(cpi) + pf_rate[filter],
+                        pf_dist[filter]);
+          if (cost < best_cost) {
+              best_filter = filter;
+              best_cost = cost;
+          }
         }
 
+        mbmi->interp_filter = best_filter;
         rate = pf_rate[mbmi->interp_filter];
         dist = pf_dist[mbmi->interp_filter];
+        var_y = pf_var[mbmi->interp_filter];
+        sse_y = pf_sse[mbmi->interp_filter];
       } else {
         mbmi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP: filter_ref;
         vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
-        model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist);
+        model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
       }
 
       rate += rate_mv;
@@ -373,7 +362,78 @@
                                 [INTER_OFFSET(this_mode)];
       this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
 
-      if (this_rd < best_rd) {
+      // Skipping checking: test to see if this block can be reconstructed by
+      // prediction only.
+      if (!x->in_active_map) {
+        x->skip = 1;
+      } else if (cpi->allow_encode_breakout && x->encode_breakout) {
+        const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
+        unsigned int var = var_y, sse = sse_y;
+        // Skipping threshold for ac.
+        unsigned int thresh_ac;
+        // Skipping threshold for dc.
+        unsigned int thresh_dc;
+        // Set a maximum for threshold to avoid big PSNR loss in low bit rate
+        // case. Use extreme low threshold for static frames to limit skipping.
+        const unsigned int max_thresh = 36000;
+        // The encode_breakout input
+        const unsigned int min_thresh =
+            MIN(((unsigned int)x->encode_breakout << 4), max_thresh);
+
+        // Calculate threshold according to dequant value.
+        thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
+        thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
+
+        // Adjust ac threshold according to partition size.
+        thresh_ac >>= 8 - (b_width_log2_lookup[bsize] +
+            b_height_log2_lookup[bsize]);
+
+        thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
+
+        // Y skipping condition checking for ac and dc.
+        if (var <= thresh_ac && (sse - var) <= thresh_dc) {
+          unsigned int sse_u, sse_v;
+          unsigned int var_u, var_v;
+
+          // Skip u v prediction for less calculation, that won't affect
+          // result much.
+          var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
+                                          x->plane[1].src.stride,
+                                          xd->plane[1].dst.buf,
+                                          xd->plane[1].dst.stride, &sse_u);
+
+          // U skipping condition checking
+          if ((var_u * 4 <= thresh_ac) && (sse_u - var_u <= thresh_dc)) {
+            var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
+                                            x->plane[2].src.stride,
+                                            xd->plane[2].dst.buf,
+                                            xd->plane[2].dst.stride, &sse_v);
+
+            // V skipping condition checking
+            if ((var_v * 4 <= thresh_ac) && (sse_v - var_v <= thresh_dc)) {
+              x->skip = 1;
+
+              // The cost of skip bit needs to be added.
+              rate = rate_mv;
+              rate += cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
+                                           [INTER_OFFSET(this_mode)];
+
+              // More on this part of rate
+              // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+
+              // Scaling factor for SSE from spatial domain to frequency
+              // domain is 16. Adjust distortion accordingly.
+              // TODO(yunqingwang): In this function, only y-plane dist is
+              // calculated.
+              dist = (sse << 4);  // + ((sse_u + sse_v) << 4);
+              this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
+              // *disable_skip = 1;
+            }
+          }
+        }
+      }
+
+      if (this_rd < best_rd || x->skip) {
         best_rd = this_rd;
         *returnrate = rate;
         *returndistortion = dist;
@@ -381,6 +441,9 @@
         best_pred_filter = mbmi->interp_filter;
         best_ref_frame = ref_frame;
       }
+
+      if (x->skip)
+        break;
     }
   }
 
@@ -392,14 +455,15 @@
 
   // Perform intra prediction search, if the best SAD is above a certain
   // threshold.
-  if (best_rd > inter_mode_thresh && bsize < cpi->sf.max_intra_bsize) {
+  if (!x->skip && best_rd > inter_mode_thresh &&
+      bsize < cpi->sf.max_intra_bsize) {
     for (this_mode = DC_PRED; this_mode <= DC_PRED; ++this_mode) {
       vp9_predict_intra_block(xd, 0, b_width_log2(bsize),
                               mbmi->tx_size, this_mode,
                               &p->src.buf[0], p->src.stride,
                               &pd->dst.buf[0], pd->dst.stride, 0, 0, 0);
 
-      model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist);
+      model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
       rate += cpi->mbmode_cost[this_mode];
       rate += intra_cost_penalty;
       this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
diff --git a/vp9/encoder/vp9_quantize.c b/vp9/encoder/vp9_quantize.c
index 5206bb6..4d3086d 100644
--- a/vp9/encoder/vp9_quantize.c
+++ b/vp9/encoder/vp9_quantize.c
@@ -32,6 +32,7 @@
                          zbin_ptr[1] + zbin_oq_value };
   const int nzbins[2] = { zbins[0] * -1,
                           zbins[1] * -1 };
+  (void)iscan;
 
   vpx_memset(qcoeff_ptr, 0, count * sizeof(int16_t));
   vpx_memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
@@ -87,6 +88,7 @@
   int idx = 0;
   int idx_arr[1024];
   int i, eob = -1;
+  (void)iscan;
 
   vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(int16_t));
   vpx_memset(dqcoeff_ptr, 0, n_coeffs * sizeof(int16_t));
diff --git a/vp9/encoder/vp9_ratectrl.c b/vp9/encoder/vp9_ratectrl.c
index 0f4c4da..a04622c 100644
--- a/vp9/encoder/vp9_ratectrl.c
+++ b/vp9/encoder/vp9_ratectrl.c
@@ -48,6 +48,7 @@
 static int arfgf_low_motion_minq[QINDEX_RANGE];
 static int arfgf_high_motion_minq[QINDEX_RANGE];
 static int inter_minq[QINDEX_RANGE];
+static int rtc_minq[QINDEX_RANGE];
 static int gf_high = 2000;
 static int gf_low = 400;
 static int kf_high = 5000;
@@ -84,6 +85,7 @@
     arfgf_low_motion_minq[i] = get_minq_index(maxq, 0.0000015, -0.0009, 0.30);
     arfgf_high_motion_minq[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.50);
     inter_minq[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.90);
+    rtc_minq[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.70);
   }
 }
 
@@ -549,14 +551,14 @@
     // Use the lower of active_worst_quality and recent/average Q.
     if (cm->current_video_frame > 1) {
       if (rc->avg_frame_qindex[INTER_FRAME] < active_worst_quality)
-        active_best_quality = inter_minq[rc->avg_frame_qindex[INTER_FRAME]];
+        active_best_quality = rtc_minq[rc->avg_frame_qindex[INTER_FRAME]];
       else
-        active_best_quality = inter_minq[active_worst_quality];
+        active_best_quality = rtc_minq[active_worst_quality];
     } else {
       if (rc->avg_frame_qindex[KEY_FRAME] < active_worst_quality)
-        active_best_quality = inter_minq[rc->avg_frame_qindex[KEY_FRAME]];
+        active_best_quality = rtc_minq[rc->avg_frame_qindex[KEY_FRAME]];
       else
-        active_best_quality = inter_minq[active_worst_quality];
+        active_best_quality = rtc_minq[active_worst_quality];
     }
   }
 
@@ -1145,10 +1147,6 @@
   cpi->rc.frames_to_key--;
 }
 
-static int test_for_kf_one_pass(VP9_COMP *cpi) {
-  // Placeholder function for auto key frame
-  return 0;
-}
 // Use this macro to turn on/off use of alt-refs in one-pass mode.
 #define USE_ALTREF_FOR_ONE_PASS   1
 
@@ -1180,11 +1178,12 @@
   VP9_COMMON *const cm = &cpi->common;
   RATE_CONTROL *const rc = &cpi->rc;
   int target;
+  // TODO(yaowu): replace the "auto_key && 0" below with proper decision logic.
   if (!cpi->refresh_alt_ref_frame &&
       (cm->current_video_frame == 0 ||
        (cpi->frame_flags & FRAMEFLAGS_KEY) ||
        rc->frames_to_key == 0 ||
-       (cpi->oxcf.auto_key && test_for_kf_one_pass(cpi)))) {
+       (cpi->oxcf.auto_key && 0))) {
     cm->frame_type = KEY_FRAME;
     rc->this_key_frame_forced = cm->current_video_frame != 0 &&
                                 rc->frames_to_key == 0;
@@ -1311,10 +1310,11 @@
   VP9_COMMON *const cm = &cpi->common;
   RATE_CONTROL *const rc = &cpi->rc;
   int target;
+  // TODO(yaowu): replace the "auto_key && 0" below with proper decision logic.
   if ((cm->current_video_frame == 0 ||
       (cpi->frame_flags & FRAMEFLAGS_KEY) ||
       rc->frames_to_key == 0 ||
-      (cpi->oxcf.auto_key && test_for_kf_one_pass(cpi)))) {
+      (cpi->oxcf.auto_key && 0))) {
     cm->frame_type = KEY_FRAME;
     rc->this_key_frame_forced = cm->current_video_frame != 0 &&
                                 rc->frames_to_key == 0;
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index 22e19fe..601e64d 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -242,6 +242,31 @@
   cpi->mb.sadperbit4 = sad_per_bit4lut[qindex];
 }
 
+static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
+                           int m, int n, int min_plane, int max_plane) {
+  int i;
+
+  for (i = min_plane; i < max_plane; ++i) {
+    struct macroblock_plane *const p = &x->plane[i];
+    struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
+
+    p->coeff    = ctx->coeff_pbuf[i][m];
+    p->qcoeff   = ctx->qcoeff_pbuf[i][m];
+    pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
+    p->eobs     = ctx->eobs_pbuf[i][m];
+
+    ctx->coeff_pbuf[i][m]   = ctx->coeff_pbuf[i][n];
+    ctx->qcoeff_pbuf[i][m]  = ctx->qcoeff_pbuf[i][n];
+    ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
+    ctx->eobs_pbuf[i][m]    = ctx->eobs_pbuf[i][n];
+
+    ctx->coeff_pbuf[i][n]   = p->coeff;
+    ctx->qcoeff_pbuf[i][n]  = p->qcoeff;
+    ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
+    ctx->eobs_pbuf[i][n]    = p->eobs;
+  }
+}
+
 static void set_block_thresholds(const VP9_COMMON *cm, RD_OPT *rd) {
   int i, bsize, segment_id;
 
@@ -297,7 +322,7 @@
     fill_token_costs(x->token_costs, cm->fc.coef_probs);
 
     for (i = 0; i < PARTITION_CONTEXTS; i++)
-      vp9_cost_tokens(x->partition_cost[i], get_partition_probs(cm, i),
+      vp9_cost_tokens(cpi->partition_cost[i], get_partition_probs(cm, i),
                       vp9_partition_tree);
   }
 
@@ -745,7 +770,8 @@
                              int use_fast_coef_casting) {
   MACROBLOCKD *const xd = &x->e_mbd;
   const struct macroblockd_plane *const pd = &xd->plane[plane];
-  struct rdcost_block_args args = { 0 };
+  struct rdcost_block_args args;
+  vp9_zero(args);
   args.x = x;
   args.best_rd = ref_best_rd;
   args.use_fast_coef_costing = use_fast_coef_casting;
@@ -1387,27 +1413,8 @@
       *rate_tokenonly = this_rate_tokenonly;
       *distortion     = this_distortion;
       *skippable      = s;
-      if (!x->select_txfm_size) {
-        int i;
-        struct macroblock_plane *const p = x->plane;
-        struct macroblockd_plane *const pd = xd->plane;
-        for (i = 1; i < MAX_MB_PLANE; ++i) {
-          p[i].coeff    = ctx->coeff_pbuf[i][2];
-          p[i].qcoeff   = ctx->qcoeff_pbuf[i][2];
-          pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
-          p[i].eobs    = ctx->eobs_pbuf[i][2];
-
-          ctx->coeff_pbuf[i][2]   = ctx->coeff_pbuf[i][0];
-          ctx->qcoeff_pbuf[i][2]  = ctx->qcoeff_pbuf[i][0];
-          ctx->dqcoeff_pbuf[i][2] = ctx->dqcoeff_pbuf[i][0];
-          ctx->eobs_pbuf[i][2]    = ctx->eobs_pbuf[i][0];
-
-          ctx->coeff_pbuf[i][0]   = p[i].coeff;
-          ctx->qcoeff_pbuf[i][0]  = p[i].qcoeff;
-          ctx->dqcoeff_pbuf[i][0] = pd[i].dqcoeff;
-          ctx->eobs_pbuf[i][0]    = p[i].eobs;
-        }
-      }
+      if (!x->select_txfm_size)
+        swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
     }
   }
 
@@ -1843,8 +1850,8 @@
           mvp_full.col = bsi->mvp.as_mv.col >> 3;
 
           if (cpi->sf.adaptive_motion_search && cm->show_frame) {
-            mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].as_mv.row >> 3;
-            mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].as_mv.col >> 3;
+            mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3;
+            mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3;
             step_param = MAX(step_param, 8);
           }
 
@@ -1895,7 +1902,7 @@
           }
 
           if (cpi->sf.adaptive_motion_search)
-            x->pred_mv[mbmi->ref_frame[0]].as_mv = *new_mv;
+            x->pred_mv[mbmi->ref_frame[0]] = *new_mv;
 
           // restore src pointers
           mi_buf_restore(x, orig_src, orig_pre);
@@ -2101,14 +2108,14 @@
                      cpi->common.show_frame &&
                      block_size < cpi->sf.max_partition_size);
 
-  int_mv pred_mv[3];
-  pred_mv[0] = mbmi->ref_mvs[ref_frame][0];
-  pred_mv[1] = mbmi->ref_mvs[ref_frame][1];
+  MV pred_mv[3];
+  pred_mv[0] = mbmi->ref_mvs[ref_frame][0].as_mv;
+  pred_mv[1] = mbmi->ref_mvs[ref_frame][1].as_mv;
   pred_mv[2] = x->pred_mv[ref_frame];
 
   // Get the sad for each candidate reference mv
   for (i = 0; i < num_mv_refs; i++) {
-    this_mv.as_int = pred_mv[i].as_int;
+    this_mv.as_mv = pred_mv[i];
 
     max_mv = MAX(max_mv,
                  MAX(abs(this_mv.as_mv.row), abs(this_mv.as_mv.col)) >> 3);
@@ -2215,10 +2222,6 @@
   ctx->skip = x->skip;
   ctx->best_mode_index = mode_index;
   ctx->mic = *xd->mi[0];
-
-  ctx->best_ref_mv[0].as_int = ref_mv->as_int;
-  ctx->best_ref_mv[1].as_int = second_ref_mv->as_int;
-
   ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
   ctx->comp_pred_diff   = (int)comp_pred_diff[COMPOUND_REFERENCE];
   ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
@@ -2312,7 +2315,7 @@
   MACROBLOCKD *xd = &x->e_mbd;
   const VP9_COMMON *cm = &cpi->common;
   MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
-  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
   int bestsme = INT_MAX;
   int step_param;
   int sadpb = x->sadperbit16;
@@ -2331,7 +2334,7 @@
   MV pred_mv[3];
   pred_mv[0] = mbmi->ref_mvs[ref][0].as_mv;
   pred_mv[1] = mbmi->ref_mvs[ref][1].as_mv;
-  pred_mv[2] = x->pred_mv[ref].as_mv;
+  pred_mv[2] = x->pred_mv[ref];
 
   if (scaled_ref_frame) {
     int i;
@@ -2376,7 +2379,8 @@
 
     for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
       if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
-        x->pred_mv[ref].as_int = 0;
+        x->pred_mv[ref].row = 0;
+        x->pred_mv[ref].col = 0;
         tmp_mv->as_int = INVALID_MV;
 
         if (scaled_ref_frame) {
@@ -2417,7 +2421,7 @@
                              x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
 
   if (cpi->sf.adaptive_motion_search && cm->show_frame)
-    x->pred_mv[ref].as_int = tmp_mv->as_int;
+    x->pred_mv[ref] = tmp_mv->as_mv;
 
   if (scaled_ref_frame) {
     int i;
@@ -2514,8 +2518,7 @@
     bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
                                        search_range,
                                        &cpi->fn_ptr[bsize],
-                                       &ref_mv[id].as_mv, second_pred,
-                                       pw, ph);
+                                       &ref_mv[id].as_mv, second_pred);
     if (bestsme < INT_MAX)
       bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
                                       second_pred, &cpi->fn_ptr[bsize], 1);
@@ -2931,30 +2934,6 @@
   return this_rd;  // if 0, this will be re-calculated by caller
 }
 
-static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
-                           int max_plane) {
-  struct macroblock_plane *const p = x->plane;
-  struct macroblockd_plane *const pd = x->e_mbd.plane;
-  int i;
-
-  for (i = 0; i < max_plane; ++i) {
-    p[i].coeff    = ctx->coeff_pbuf[i][1];
-    p[i].qcoeff  = ctx->qcoeff_pbuf[i][1];
-    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
-    p[i].eobs    = ctx->eobs_pbuf[i][1];
-
-    ctx->coeff_pbuf[i][1]   = ctx->coeff_pbuf[i][0];
-    ctx->qcoeff_pbuf[i][1]  = ctx->qcoeff_pbuf[i][0];
-    ctx->dqcoeff_pbuf[i][1] = ctx->dqcoeff_pbuf[i][0];
-    ctx->eobs_pbuf[i][1]    = ctx->eobs_pbuf[i][0];
-
-    ctx->coeff_pbuf[i][0]   = p[i].coeff;
-    ctx->qcoeff_pbuf[i][0]  = p[i].qcoeff;
-    ctx->dqcoeff_pbuf[i][0] = pd[i].dqcoeff;
-    ctx->eobs_pbuf[i][0]    = p[i].eobs;
-  }
-}
-
 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
                                int *returnrate, int64_t *returndist,
                                BLOCK_SIZE bsize,
@@ -3069,7 +3048,7 @@
   int64_t best_pred_rd[REFERENCE_MODES];
   int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
-  MB_MODE_INFO best_mbmode = { 0 };
+  MB_MODE_INFO best_mbmode;
   int mode_index, best_mode_index = -1;
   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
   vp9_prob comp_mode_p;
@@ -3095,7 +3074,7 @@
   const int intra_y_mode_mask =
       cpi->sf.intra_y_mode_mask[max_txsize_lookup[bsize]];
   int disable_inter_mode_mask = cpi->sf.disable_inter_mode_mask[bsize];
-
+  vp9_zero(best_mbmode);
   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
 
   estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
@@ -3476,7 +3455,7 @@
         best_mbmode = *mbmi;
         best_skip2 = this_skip2;
         if (!x->select_txfm_size)
-          swap_block_ptr(x, ctx, max_plane);
+          swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
         vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
                    sizeof(uint8_t) * ctx->num_4x4_blk);
 
@@ -3678,7 +3657,7 @@
   int64_t best_pred_rd[REFERENCE_MODES];
   int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
-  MB_MODE_INFO best_mbmode = { 0 };
+  MB_MODE_INFO best_mbmode;
   int ref_index, best_ref_index = 0;
   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
   vp9_prob comp_mode_p;
@@ -3698,6 +3677,7 @@
 
   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
   vpx_memset(x->zcoeff_blk[TX_4X4], 0, 4);
+  vp9_zero(best_mbmode);
 
   for (i = 0; i < 4; i++) {
     int j;
@@ -4130,7 +4110,7 @@
         best_mbmode = *mbmi;
         best_skip2 = this_skip2;
         if (!x->select_txfm_size)
-          swap_block_ptr(x, ctx, max_plane);
+          swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
         vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
                    sizeof(uint8_t) * ctx->num_4x4_blk);
 
diff --git a/vp9/encoder/vp9_segmentation.c b/vp9/encoder/vp9_segmentation.c
index 7537d1b..574df62 100644
--- a/vp9/encoder/vp9_segmentation.c
+++ b/vp9/encoder/vp9_segmentation.c
@@ -109,7 +109,7 @@
 }
 
 static void count_segs(VP9_COMP *cpi, const TileInfo *const tile,
-                       MODE_INFO **mi_8x8,
+                       MODE_INFO **mi,
                        int *no_pred_segcounts,
                        int (*temporal_predictor_count)[2],
                        int *t_unpred_seg_counts,
@@ -121,7 +121,7 @@
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
-  xd->mi = mi_8x8;
+  xd->mi = mi;
   segment_id = xd->mi[0]->mbmi.segment_id;
 
   set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
@@ -131,7 +131,7 @@
 
   // Temporal prediction not allowed on key frames
   if (cm->frame_type != KEY_FRAME) {
-    const BLOCK_SIZE bsize = mi_8x8[0]->mbmi.sb_type;
+    const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
     // Test to see if the segment id matches the predicted value.
     const int pred_segment_id = vp9_get_segment_id(cm, cm->last_frame_seg_map,
                                                    bsize, mi_row, mi_col);
@@ -143,14 +143,14 @@
     xd->mi[0]->mbmi.seg_id_predicted = pred_flag;
     temporal_predictor_count[pred_context][pred_flag]++;
 
+    // Update the "unpredicted" segment count
     if (!pred_flag)
-      // Update the "unpredicted" segment count
       t_unpred_seg_counts[segment_id]++;
   }
 }
 
 static void count_segs_sb(VP9_COMP *cpi, const TileInfo *const tile,
-                          MODE_INFO **mi_8x8,
+                          MODE_INFO **mi,
                           int *no_pred_segcounts,
                           int (*temporal_predictor_count)[2],
                           int *t_unpred_seg_counts,
@@ -164,22 +164,22 @@
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
-  bw = num_8x8_blocks_wide_lookup[mi_8x8[0]->mbmi.sb_type];
-  bh = num_8x8_blocks_high_lookup[mi_8x8[0]->mbmi.sb_type];
+  bw = num_8x8_blocks_wide_lookup[mi[0]->mbmi.sb_type];
+  bh = num_8x8_blocks_high_lookup[mi[0]->mbmi.sb_type];
 
   if (bw == bs && bh == bs) {
-    count_segs(cpi, tile, mi_8x8, no_pred_segcounts, temporal_predictor_count,
+    count_segs(cpi, tile, mi, no_pred_segcounts, temporal_predictor_count,
                t_unpred_seg_counts, bs, bs, mi_row, mi_col);
   } else if (bw == bs && bh < bs) {
-    count_segs(cpi, tile, mi_8x8, no_pred_segcounts, temporal_predictor_count,
+    count_segs(cpi, tile, mi, no_pred_segcounts, temporal_predictor_count,
                t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
-    count_segs(cpi, tile, mi_8x8 + hbs * mis, no_pred_segcounts,
+    count_segs(cpi, tile, mi + hbs * mis, no_pred_segcounts,
                temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
                mi_row + hbs, mi_col);
   } else if (bw < bs && bh == bs) {
-    count_segs(cpi, tile, mi_8x8, no_pred_segcounts, temporal_predictor_count,
+    count_segs(cpi, tile, mi, no_pred_segcounts, temporal_predictor_count,
                t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
-    count_segs(cpi, tile, mi_8x8 + hbs,
+    count_segs(cpi, tile, mi + hbs,
                no_pred_segcounts, temporal_predictor_count, t_unpred_seg_counts,
                hbs, bs, mi_row, mi_col + hbs);
   } else {
@@ -192,7 +192,7 @@
       const int mi_dc = hbs * (n & 1);
       const int mi_dr = hbs * (n >> 1);
 
-      count_segs_sb(cpi, tile, &mi_8x8[mi_dr * mis + mi_dc],
+      count_segs_sb(cpi, tile, &mi[mi_dr * mis + mi_dc],
                     no_pred_segcounts, temporal_predictor_count,
                     t_unpred_seg_counts,
                     mi_row + mi_dr, mi_col + mi_dc, subsize);
@@ -217,9 +217,6 @@
   vp9_prob t_pred_tree[SEG_TREE_PROBS];
   vp9_prob t_nopred_prob[PREDICTION_PROBS];
 
-  const int mis = cm->mi_stride;
-  MODE_INFO **mi_ptr, **mi;
-
   // Set default state for the segment tree probabilities and the
   // temporal coding probabilities
   vpx_memset(seg->tree_probs, 255, sizeof(seg->tree_probs));
@@ -229,12 +226,13 @@
   // predicts this one
   for (tile_col = 0; tile_col < 1 << cm->log2_tile_cols; tile_col++) {
     TileInfo tile;
-
+    MODE_INFO **mi_ptr;
     vp9_tile_init(&tile, cm, 0, tile_col);
+
     mi_ptr = cm->mi_grid_visible + tile.mi_col_start;
     for (mi_row = 0; mi_row < cm->mi_rows;
-         mi_row += 8, mi_ptr += 8 * mis) {
-      mi = mi_ptr;
+         mi_row += 8, mi_ptr += 8 * cm->mi_stride) {
+      MODE_INFO **mi = mi_ptr;
       for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
            mi_col += 8, mi += 8)
         count_segs_sb(cpi, &tile, mi, no_pred_segcounts,
diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c
index 93e23ee..7b2d1e2 100644
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -146,7 +146,6 @@
                                  int speed) {
   sf->static_segmentation = 0;
   sf->adaptive_rd_thresh = 1;
-  sf->encode_breakout_thresh = 1;
   sf->use_fast_coef_costing = 1;
 
   if (speed == 1) {
@@ -169,7 +168,6 @@
     sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
     sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
     sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
-    sf->encode_breakout_thresh = 8;
   }
 
   if (speed >= 2) {
@@ -208,7 +206,6 @@
     sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
     sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
     sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
-    sf->encode_breakout_thresh = 200;
   }
 
   if (speed >= 3) {
@@ -226,7 +223,6 @@
     sf->optimize_coefficients = 0;
     sf->disable_split_mask = DISABLE_ALL_SPLIT;
     sf->lpf_pick = LPF_PICK_FROM_Q;
-    sf->encode_breakout_thresh = 700;
   }
 
   if (speed >= 4) {
@@ -245,7 +241,6 @@
     }
     sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_ONLY;
     sf->frame_parameter_update = 0;
-    sf->encode_breakout_thresh = 1000;
     sf->search_method = FAST_HEX;
     sf->disable_inter_mode_mask[BLOCK_32X32] = 1 << INTER_OFFSET(ZEROMV);
     sf->disable_inter_mode_mask[BLOCK_32X64] = ~(1 << INTER_OFFSET(NEARESTMV));
@@ -338,7 +333,6 @@
   sf->use_fast_coef_costing = 0;
   sf->mode_skip_start = MAX_MODES;  // Mode index at which mode skip mask set
   sf->use_nonrd_pick_mode = 0;
-  sf->encode_breakout_thresh = 0;
   for (i = 0; i < BLOCK_SIZES; ++i)
     sf->disable_inter_mode_mask[i] = 0;
   sf->max_intra_bsize = BLOCK_64X64;
@@ -384,10 +378,6 @@
 
   cpi->mb.optimize = sf->optimize_coefficients == 1 && cpi->pass != 1;
 
-  if (cpi->encode_breakout && oxcf->mode == REALTIME &&
-      sf->encode_breakout_thresh > cpi->encode_breakout)
-    cpi->encode_breakout = sf->encode_breakout_thresh;
-
   if (sf->disable_split_mask == DISABLE_ALL_SPLIT)
     sf->adaptive_pred_interp_filter = 0;
 
diff --git a/vp9/encoder/vp9_speed_features.h b/vp9/encoder/vp9_speed_features.h
index 46806c9..d8c1a8b 100644
--- a/vp9/encoder/vp9_speed_features.h
+++ b/vp9/encoder/vp9_speed_features.h
@@ -321,10 +321,6 @@
   // This flag controls the use of non-RD mode decision.
   int use_nonrd_pick_mode;
 
-  // This variable sets the encode_breakout threshold. Currently, it is only
-  // enabled in real time mode.
-  int encode_breakout_thresh;
-
   // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
   // modes are disabled in order from LSB to MSB for each BLOCK_SIZE.
   int disable_inter_mode_mask[BLOCK_SIZES];
diff --git a/vp9/encoder/vp9_svc_layercontext.c b/vp9/encoder/vp9_svc_layercontext.c
index 2e98fa7..dd28496 100644
--- a/vp9/encoder/vp9_svc_layercontext.c
+++ b/vp9/encoder/vp9_svc_layercontext.c
@@ -203,7 +203,7 @@
   int i;
 
   for (i = 0; i < svc->number_spatial_layers; ++i) {
-    struct twopass_rc *const twopass = &svc->layer_context[i].twopass;
+    TWO_PASS *const twopass = &svc->layer_context[i].twopass;
 
     svc->spatial_layer_id = i;
     vp9_init_second_pass(cpi);
diff --git a/vp9/encoder/vp9_svc_layercontext.h b/vp9/encoder/vp9_svc_layercontext.h
index 74d9c1c..6881ce1 100644
--- a/vp9/encoder/vp9_svc_layercontext.h
+++ b/vp9/encoder/vp9_svc_layercontext.h
@@ -27,7 +27,7 @@
   int64_t maximum_buffer_size;
   double framerate;
   int avg_frame_size;
-  struct twopass_rc twopass;
+  TWO_PASS twopass;
   struct vpx_fixed_buf rc_twopass_stats_in;
   unsigned int current_video_frame_in_layer;
   int is_key_frame;
diff --git a/vp9/encoder/vp9_temporal_filter.c b/vp9/encoder/vp9_temporal_filter.c
index 6eff200..f501971 100644
--- a/vp9/encoder/vp9_temporal_filter.c
+++ b/vp9/encoder/vp9_temporal_filter.c
@@ -34,7 +34,8 @@
                                             uint8_t *u_mb_ptr,
                                             uint8_t *v_mb_ptr,
                                             int stride,
-                                            int uv_block_size,
+                                            int uv_block_width,
+                                            int uv_block_height,
                                             int mv_row,
                                             int mv_col,
                                             uint8_t *pred,
@@ -47,7 +48,7 @@
 
   enum mv_precision mv_precision_uv;
   int uv_stride;
-  if (uv_block_size == 8) {
+  if (uv_block_width == 8) {
     uv_stride = (stride + 1) >> 1;
     mv_precision_uv = MV_PRECISION_Q4;
   } else {
@@ -64,18 +65,18 @@
                             kernel, MV_PRECISION_Q3, x, y);
 
   vp9_build_inter_predictor(u_mb_ptr, uv_stride,
-                            &pred[256], uv_block_size,
+                            &pred[256], uv_block_width,
                             &mv,
                             scale,
-                            uv_block_size, uv_block_size,
+                            uv_block_width, uv_block_height,
                             which_mv,
                             kernel, mv_precision_uv, x, y);
 
   vp9_build_inter_predictor(v_mb_ptr, uv_stride,
-                            &pred[512], uv_block_size,
+                            &pred[512], uv_block_width,
                             &mv,
                             scale,
-                            uv_block_size, uv_block_size,
+                            uv_block_width, uv_block_height,
                             which_mv,
                             kernel, mv_precision_uv, x, y);
 }
@@ -91,7 +92,8 @@
 void vp9_temporal_filter_apply_c(uint8_t *frame1,
                                  unsigned int stride,
                                  uint8_t *frame2,
-                                 unsigned int block_size,
+                                 unsigned int block_width,
+                                 unsigned int block_height,
                                  int strength,
                                  int filter_weight,
                                  unsigned int *accumulator,
@@ -101,8 +103,8 @@
   int byte = 0;
   const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
 
-  for (i = 0, k = 0; i < block_size; i++) {
-    for (j = 0; j < block_size; j++, k++) {
+  for (i = 0, k = 0; i < block_height; i++) {
+    for (j = 0; j < block_width; j++, k++) {
       int src_byte = frame1[byte];
       int pixel_value = *frame2++;
 
@@ -127,7 +129,7 @@
       byte++;
     }
 
-    byte += stride - block_size;
+    byte += stride - block_width;
   }
 }
 
@@ -204,14 +206,12 @@
   uint8_t *dst1, *dst2;
   DECLARE_ALIGNED_ARRAY(16, uint8_t,  predictor, 16 * 16 * 3);
   const int mb_uv_height = 16 >> mbd->plane[1].subsampling_y;
+  const int mb_uv_width  = 16 >> mbd->plane[1].subsampling_x;
 
   // Save input state
   uint8_t* input_buffer[MAX_MB_PLANE];
   int i;
 
-  // TODO(aconverse): Add 4:2:2 support
-  assert(mbd->plane[1].subsampling_x == mbd->plane[1].subsampling_y);
-
   for (i = 0; i < MAX_MB_PLANE; i++)
     input_buffer[i] = mbd->plane[i].pre[0].buf;
 
@@ -275,7 +275,7 @@
               cpi->frames[frame]->u_buffer + mb_uv_offset,
               cpi->frames[frame]->v_buffer + mb_uv_offset,
               cpi->frames[frame]->y_stride,
-              mb_uv_height,
+              mb_uv_width, mb_uv_height,
               mbd->mi[0]->bmi[0].as_mv[0].as_mv.row,
               mbd->mi[0]->bmi[0].as_mv[0].as_mv.col,
               predictor, scale,
@@ -283,16 +283,17 @@
 
           // Apply the filter (YUV)
           vp9_temporal_filter_apply(f->y_buffer + mb_y_offset, f->y_stride,
-                                    predictor, 16, strength, filter_weight,
+                                    predictor, 16, 16,
+                                    strength, filter_weight,
                                     accumulator, count);
-
           vp9_temporal_filter_apply(f->u_buffer + mb_uv_offset, f->uv_stride,
-                                    predictor + 256, mb_uv_height, strength,
+                                    predictor + 256,
+                                    mb_uv_width, mb_uv_height, strength,
                                     filter_weight, accumulator + 256,
                                     count + 256);
-
           vp9_temporal_filter_apply(f->v_buffer + mb_uv_offset, f->uv_stride,
-                                    predictor + 512, mb_uv_height, strength,
+                                    predictor + 512,
+                                    mb_uv_width, mb_uv_height, strength,
                                     filter_weight, accumulator + 512,
                                     count + 512);
         }
@@ -321,7 +322,7 @@
       stride = cpi->alt_ref_buffer.uv_stride;
       byte = mb_uv_offset;
       for (i = 0, k = 256; i < mb_uv_height; i++) {
-        for (j = 0; j < mb_uv_height; j++, k++) {
+        for (j = 0; j < mb_uv_width; j++, k++) {
           int m = k + 256;
 
           // U
@@ -339,13 +340,13 @@
           // move to next pixel
           byte++;
         }
-        byte += stride - mb_uv_height;
+        byte += stride - mb_uv_width;
       }
       mb_y_offset += 16;
-      mb_uv_offset += mb_uv_height;
+      mb_uv_offset += mb_uv_width;
     }
     mb_y_offset += 16 * (f->y_stride - mb_cols);
-    mb_uv_offset += mb_uv_height * (f->uv_stride - mb_cols);
+    mb_uv_offset += mb_uv_height * f->uv_stride - mb_uv_width * mb_cols;
   }
 
   // Restore input state
diff --git a/vp9/encoder/vp9_tokenize.c b/vp9/encoder/vp9_tokenize.c
index 8ce98d9..dcca92d 100644
--- a/vp9/encoder/vp9_tokenize.c
+++ b/vp9/encoder/vp9_tokenize.c
@@ -232,7 +232,6 @@
       cpi->common.fc.coef_probs[tx_size][type][ref];
   unsigned int (*const eob_branch)[COEFF_CONTEXTS] =
       cpi->common.counts.eob_branch[tx_size][type][ref];
-
   const uint8_t *const band = get_band_translate(tx_size);
   const int seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
 
@@ -289,14 +288,17 @@
   MACROBLOCK *x;
   int *skippable;
 };
-
 static void is_skippable(int plane, int block,
                          BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                          void *argv) {
   struct is_skippable_args *args = argv;
+  (void)plane_bsize;
+  (void)tx_size;
   args->skippable[0] &= (!args->x->plane[plane].eobs[block]);
 }
 
+// TODO(yaowu): rewrite and optimize this function to remove the usage of
+//              vp9_foreach_transform_block() and simplify is_skippable().
 int vp9_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
   int result = 1;
   struct is_skippable_args args = {x, &result};
diff --git a/vp9/encoder/x86/vp9_dct_sse2.c b/vp9/encoder/x86/vp9_dct_sse2.c
index 6865822..1f58d87 100644
--- a/vp9/encoder/x86/vp9_dct_sse2.c
+++ b/vp9/encoder/x86/vp9_dct_sse2.c
@@ -1187,7 +1187,7 @@
   const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
   const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
   const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
   const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
   const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
   const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
@@ -1513,8 +1513,8 @@
           const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
           const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m08_p24);
           const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m08_p24);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m24_m08);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m24_m08);
+          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p24_p08);
           // dct_const_round_shift
           const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
           const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
@@ -1535,8 +1535,8 @@
           const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
           const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p24_p08);
           const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p24_p08);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m08_p24);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p08_m24);
+          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p08_m24);
           // dct_const_round_shift
           const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
           const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
@@ -1554,10 +1554,10 @@
         {
           step1_0 = _mm_add_epi16(step3_0, step2_1);
           step1_1 = _mm_sub_epi16(step3_0, step2_1);
-          step1_2 = _mm_sub_epi16(step3_3, step2_2);
-          step1_3 = _mm_add_epi16(step3_3, step2_2);
-          step1_4 = _mm_add_epi16(step3_4, step2_5);
-          step1_5 = _mm_sub_epi16(step3_4, step2_5);
+          step1_2 = _mm_add_epi16(step3_3, step2_2);
+          step1_3 = _mm_sub_epi16(step3_3, step2_2);
+          step1_4 = _mm_sub_epi16(step3_4, step2_5);
+          step1_5 = _mm_add_epi16(step3_4, step2_5);
           step1_6 = _mm_sub_epi16(step3_7, step2_6);
           step1_7 = _mm_add_epi16(step3_7, step2_6);
         }
@@ -1848,7 +1848,7 @@
   const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
   const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
   const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
   const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
   const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
   const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
@@ -2052,10 +2052,10 @@
 
   v[0] = _mm_madd_epi16(u[0], k__cospi_m08_p24);
   v[1] = _mm_madd_epi16(u[1], k__cospi_m08_p24);
-  v[2] = _mm_madd_epi16(u[2], k__cospi_m24_m08);
-  v[3] = _mm_madd_epi16(u[3], k__cospi_m24_m08);
-  v[4] = _mm_madd_epi16(u[2], k__cospi_m08_p24);
-  v[5] = _mm_madd_epi16(u[3], k__cospi_m08_p24);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_p24_p08);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_p24_p08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p08_m24);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p08_m24);
   v[6] = _mm_madd_epi16(u[0], k__cospi_p24_p08);
   v[7] = _mm_madd_epi16(u[1], k__cospi_p24_p08);
 
@@ -2085,10 +2085,10 @@
   // stage 5
   s[0] = _mm_add_epi16(p[0], t[1]);
   s[1] = _mm_sub_epi16(p[0], t[1]);
-  s[2] = _mm_sub_epi16(p[3], t[2]);
-  s[3] = _mm_add_epi16(p[3], t[2]);
-  s[4] = _mm_add_epi16(p[4], t[5]);
-  s[5] = _mm_sub_epi16(p[4], t[5]);
+  s[2] = _mm_add_epi16(p[3], t[2]);
+  s[3] = _mm_sub_epi16(p[3], t[2]);
+  s[4] = _mm_sub_epi16(p[4], t[5]);
+  s[5] = _mm_add_epi16(p[4], t[5]);
   s[6] = _mm_sub_epi16(p[7], t[6]);
   s[7] = _mm_add_epi16(p[7], t[6]);
 
diff --git a/vp9/encoder/x86/vp9_error_intrin_avx2.c b/vp9/encoder/x86/vp9_error_intrin_avx2.c
new file mode 100644
index 0000000..c67490f
--- /dev/null
+++ b/vp9/encoder/x86/vp9_error_intrin_avx2.c
@@ -0,0 +1,72 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Usee of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+#include "vpx/vpx_integer.h"
+
+
+int64_t vp9_block_error_avx2(const int16_t *coeff,
+                             const int16_t *dqcoeff,
+                             intptr_t block_size,
+                             int64_t *ssz) {
+  __m256i sse_reg, ssz_reg, coeff_reg, dqcoeff_reg;
+  __m256i exp_dqcoeff_lo, exp_dqcoeff_hi, exp_coeff_lo, exp_coeff_hi;
+  __m256i sse_reg_64hi, ssz_reg_64hi;
+  __m128i sse_reg128, ssz_reg128;
+  int64_t sse;
+  int i;
+  const __m256i zero_reg = _mm256_set1_epi16(0);
+
+  // init sse and ssz registerd to zero
+  sse_reg = _mm256_set1_epi16(0);
+  ssz_reg = _mm256_set1_epi16(0);
+
+  for (i = 0 ; i < block_size ; i+= 16) {
+    // load 32 bytes from coeff and dqcoeff
+    coeff_reg = _mm256_loadu_si256((const __m256i *)(coeff + i));
+    dqcoeff_reg = _mm256_loadu_si256((const __m256i *)(dqcoeff + i));
+    // dqcoeff - coeff
+    dqcoeff_reg = _mm256_sub_epi16(dqcoeff_reg, coeff_reg);
+    // madd (dqcoeff - coeff)
+    dqcoeff_reg = _mm256_madd_epi16(dqcoeff_reg, dqcoeff_reg);
+    // madd coeff
+    coeff_reg = _mm256_madd_epi16(coeff_reg, coeff_reg);
+    // expand each double word of madd (dqcoeff - coeff) to quad word
+    exp_dqcoeff_lo = _mm256_unpacklo_epi32(dqcoeff_reg, zero_reg);
+    exp_dqcoeff_hi = _mm256_unpackhi_epi32(dqcoeff_reg, zero_reg);
+    // expand each double word of madd (coeff) to quad word
+    exp_coeff_lo = _mm256_unpacklo_epi32(coeff_reg, zero_reg);
+    exp_coeff_hi = _mm256_unpackhi_epi32(coeff_reg, zero_reg);
+    // add each quad word of madd (dqcoeff - coeff) and madd (coeff)
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_lo);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_lo);
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_hi);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_hi);
+  }
+  // save the higher 64 bit of each 128 bit lane
+  sse_reg_64hi = _mm256_srli_si256(sse_reg, 8);
+  ssz_reg_64hi = _mm256_srli_si256(ssz_reg, 8);
+  // add the higher 64 bit to the low 64 bit
+  sse_reg = _mm256_add_epi64(sse_reg, sse_reg_64hi);
+  ssz_reg = _mm256_add_epi64(ssz_reg, ssz_reg_64hi);
+
+  // add each 64 bit from each of the 128 bit lane of the 256 bit
+  sse_reg128 = _mm_add_epi64(_mm256_castsi256_si128(sse_reg),
+                             _mm256_extractf128_si256(sse_reg, 1));
+
+  ssz_reg128 = _mm_add_epi64(_mm256_castsi256_si128(ssz_reg),
+                             _mm256_extractf128_si256(ssz_reg, 1));
+
+  // store the results
+  _mm_storel_epi64((__m128i*)(&sse), sse_reg128);
+
+  _mm_storel_epi64((__m128i*)(ssz), ssz_reg128);
+  return sse;
+}
diff --git a/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm b/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
index d2d13b3..673e0b3 100644
--- a/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
+++ b/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
@@ -15,41 +15,45 @@
 ;  (unsigned char  *frame1,           |  0
 ;   unsigned int    stride,           |  1
 ;   unsigned char  *frame2,           |  2
-;   unsigned int    block_size,       |  3
-;   int             strength,         |  4
-;   int             filter_weight,    |  5
-;   unsigned int   *accumulator,      |  6
-;   unsigned short *count)            |  7
+;   unsigned int    block_width,      |  3
+;   unsigned int    block_height,     |  4
+;   int             strength,         |  5
+;   int             filter_weight,    |  6
+;   unsigned int   *accumulator,      |  7
+;   unsigned short *count)            |  8
 global sym(vp9_temporal_filter_apply_sse2) PRIVATE
 sym(vp9_temporal_filter_apply_sse2):
 
     push        rbp
     mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 8
+    SHADOW_ARGS_TO_STACK 9
     SAVE_XMM 7
     GET_GOT     rbx
     push        rsi
     push        rdi
     ALIGN_STACK 16, rax
-    %define block_size    0
-    %define strength      16
-    %define filter_weight 32
-    %define rounding_bit  48
-    %define rbp_backup    64
-    %define stack_size    80
+    %define block_width    0
+    %define block_height  16
+    %define strength      32
+    %define filter_weight 48
+    %define rounding_bit  64
+    %define rbp_backup    80
+    %define stack_size    96
     sub         rsp,           stack_size
     mov         [rsp + rbp_backup], rbp
     ; end prolog
 
         mov         rdx,            arg(3)
-        mov         [rsp + block_size], rdx
-        movd        xmm6,            arg(4)
+        mov         [rsp + block_width], rdx
+        mov         rdx,            arg(4)
+        mov         [rsp + block_height], rdx
+        movd        xmm6,           arg(5)
         movdqa      [rsp + strength], xmm6 ; where strength is used, all 16 bytes are read
 
         ; calculate the rounding bit outside the loop
         ; 0x8000 >> (16 - strength)
         mov         rdx,            16
-        sub         rdx,            arg(4) ; 16 - strength
+        sub         rdx,            arg(5) ; 16 - strength
         movq        xmm4,           rdx    ; can't use rdx w/ shift
         movdqa      xmm5,           [GLOBAL(_const_top_bit)]
         psrlw       xmm5,           xmm4
@@ -57,11 +61,11 @@
 
         mov         rsi,            arg(0) ; src/frame1
         mov         rdx,            arg(2) ; predictor frame
-        mov         rdi,            arg(6) ; accumulator
-        mov         rax,            arg(7) ; count
+        mov         rdi,            arg(7) ; accumulator
+        mov         rax,            arg(8) ; count
 
         ; dup the filter weight and store for later
-        movd        xmm0,           arg(5) ; filter_weight
+        movd        xmm0,           arg(6) ; filter_weight
         pshuflw     xmm0,           xmm0, 0
         punpcklwd   xmm0,           xmm0
         movdqa      [rsp + filter_weight], xmm0
@@ -69,10 +73,11 @@
         mov         rbp,            arg(1) ; stride
         pxor        xmm7,           xmm7   ; zero for extraction
 
-        lea         rcx,            [rdx + 16*16*1]
-        cmp         dword ptr [rsp + block_size], 8
+        mov         rcx,            [rsp + block_width]
+        imul        rcx,            [rsp + block_height]
+        add         rcx,            rdx
+        cmp         dword ptr [rsp + block_width], 8
         jne         .temporal_filter_apply_load_16
-        lea         rcx,            [rdx + 8*8*1]
 
 .temporal_filter_apply_load_8:
         movq        xmm0,           [rsi]  ; first row
@@ -178,7 +183,7 @@
         cmp         rdx,            rcx
         je          .temporal_filter_apply_epilog
         pxor        xmm7,           xmm7   ; zero for extraction
-        cmp         dword ptr [rsp + block_size], 16
+        cmp         dword ptr [rsp + block_width], 16
         je          .temporal_filter_apply_load_16
         jmp         .temporal_filter_apply_load_8
 
diff --git a/vp9/vp9_cx_iface.c b/vp9/vp9_cx_iface.c
index c4058bb..5a8a4f4 100644
--- a/vp9/vp9_cx_iface.c
+++ b/vp9/vp9_cx_iface.c
@@ -211,8 +211,8 @@
       ERROR("Option --tune=ssim is not currently supported in VP9.");
 
   if (cfg->g_pass == VPX_RC_LAST_PASS) {
-    size_t           packet_sz = sizeof(FIRSTPASS_STATS);
-    int              n_packets = (int)(cfg->rc_twopass_stats_in.sz / packet_sz);
+    const size_t packet_sz = sizeof(FIRSTPASS_STATS);
+    const int n_packets = (int)(cfg->rc_twopass_stats_in.sz / packet_sz);
     const FIRSTPASS_STATS *stats;
 
     if (cfg->rc_twopass_stats_in.buf == NULL)
@@ -464,7 +464,7 @@
 
 static vpx_codec_err_t ctrl_get_param(vpx_codec_alg_priv_t *ctx, int ctrl_id,
                                  va_list args) {
-  void *arg = va_arg(args, void *);
+  void *const arg = va_arg(args, void *);
 
 #define MAP(id, var) case id: *(RECAST(id, arg)) = var; break
 
@@ -525,6 +525,7 @@
 static vpx_codec_err_t encoder_init(vpx_codec_ctx_t *ctx,
                                     vpx_codec_priv_enc_mr_cfg_t *data) {
   vpx_codec_err_t res = VPX_CODEC_OK;
+  (void)data;
 
   if (ctx->priv == NULL) {
     int i;
@@ -880,14 +881,15 @@
   return res;
 }
 
-static const vpx_codec_cx_pkt_t *encoder_get_cxdata(vpx_codec_alg_priv_t  *ctx,
+static const vpx_codec_cx_pkt_t *encoder_get_cxdata(vpx_codec_alg_priv_t *ctx,
                                                     vpx_codec_iter_t *iter) {
   return vpx_codec_pkt_list_get(&ctx->pkt_list.head, iter);
 }
 
 static vpx_codec_err_t ctrl_set_reference(vpx_codec_alg_priv_t *ctx,
-                                          int ctr_id, va_list args) {
+                                          int ctrl_id, va_list args) {
   vpx_ref_frame_t *const frame = va_arg(args, vpx_ref_frame_t *);
+  (void)ctrl_id;
 
   if (frame != NULL) {
     YV12_BUFFER_CONFIG sd;
@@ -902,8 +904,9 @@
 }
 
 static vpx_codec_err_t ctrl_copy_reference(vpx_codec_alg_priv_t *ctx,
-                                           int ctr_id, va_list args) {
+                                           int ctrl_id, va_list args) {
   vpx_ref_frame_t *const frame = va_arg(args, vpx_ref_frame_t *);
+  (void)ctrl_id;
 
   if (frame != NULL) {
     YV12_BUFFER_CONFIG sd;
@@ -918,11 +921,12 @@
 }
 
 static vpx_codec_err_t ctrl_get_reference(vpx_codec_alg_priv_t *ctx,
-                                          int ctr_id, va_list args) {
-  vp9_ref_frame_t *frame = va_arg(args, vp9_ref_frame_t *);
+                                          int ctrl_id, va_list args) {
+  vp9_ref_frame_t *const frame = va_arg(args, vp9_ref_frame_t *);
+  (void)ctrl_id;
 
   if (frame != NULL) {
-    YV12_BUFFER_CONFIG* fb;
+    YV12_BUFFER_CONFIG *fb;
 
     vp9_get_reference_enc(ctx->cpi, frame->idx, &fb);
     yuvconfig2image(&frame->img, fb, NULL);
@@ -955,7 +959,8 @@
 
 static vpx_image_t *encoder_get_preview(vpx_codec_alg_priv_t *ctx) {
   YV12_BUFFER_CONFIG sd;
-  vp9_ppflags_t flags = {0};
+  vp9_ppflags_t flags;
+  vp9_zero(flags);
 
   if (ctx->preview_ppcfg.post_proc_flag) {
     flags.post_proc_flag   = ctx->preview_ppcfg.post_proc_flag;
@@ -972,39 +977,51 @@
 }
 
 static vpx_codec_err_t ctrl_update_entropy(vpx_codec_alg_priv_t *ctx,
-                                           int ctr_id, va_list args) {
+                                           int ctrl_id, va_list args) {
   const int update = va_arg(args, int);
+  (void)ctrl_id;
+
   vp9_update_entropy(ctx->cpi, update);
   return VPX_CODEC_OK;
 }
 
 static vpx_codec_err_t ctrl_update_reference(vpx_codec_alg_priv_t *ctx,
-                                             int ctr_id, va_list args) {
+                                             int ctrl_id, va_list args) {
   const int ref_frame_flags = va_arg(args, int);
+  (void)ctrl_id;
+
   vp9_update_reference(ctx->cpi, ref_frame_flags);
   return VPX_CODEC_OK;
 }
 
 static vpx_codec_err_t ctrl_use_reference(vpx_codec_alg_priv_t *ctx,
-                                          int ctr_id, va_list args) {
+                                          int ctrl_id, va_list args) {
   const int reference_flag = va_arg(args, int);
+  (void)ctrl_id;
+
   vp9_use_as_reference(ctx->cpi, reference_flag);
   return VPX_CODEC_OK;
 }
 
 static vpx_codec_err_t ctrl_set_roi_map(vpx_codec_alg_priv_t *ctx,
-                                        int ctr_id, va_list args) {
+                                        int ctrl_id, va_list args) {
+  (void)ctx;
+  (void)ctrl_id;
+  (void)args;
+
   // TODO(yaowu): Need to re-implement and test for VP9.
   return VPX_CODEC_INVALID_PARAM;
 }
 
 
 static vpx_codec_err_t ctrl_set_active_map(vpx_codec_alg_priv_t *ctx,
-                                           int ctr_id, va_list args) {
+                                           int ctrl_id, va_list args) {
   vpx_active_map_t *const map = va_arg(args, vpx_active_map_t *);
+  (void)ctrl_id;
 
   if (map) {
-    if (!vp9_set_active_map(ctx->cpi, map->active_map, map->rows, map->cols))
+    if (!vp9_set_active_map(ctx->cpi, map->active_map,
+                            (int)map->rows, (int)map->cols))
       return VPX_CODEC_OK;
     else
       return VPX_CODEC_INVALID_PARAM;
@@ -1014,8 +1031,9 @@
 }
 
 static vpx_codec_err_t ctrl_set_scale_mode(vpx_codec_alg_priv_t *ctx,
-                                           int ctr_id, va_list args) {
+                                           int ctrl_id, va_list args) {
   vpx_scaling_mode_t *const mode = va_arg(args, vpx_scaling_mode_t *);
+  (void)ctrl_id;
 
   if (mode) {
     const int res = vp9_set_internal_size(ctx->cpi,
@@ -1027,10 +1045,12 @@
   }
 }
 
-static vpx_codec_err_t ctrl_set_svc(vpx_codec_alg_priv_t *ctx, int ctr_id,
+static vpx_codec_err_t ctrl_set_svc(vpx_codec_alg_priv_t *ctx, int ctrl_id,
                                     va_list args) {
   int data = va_arg(args, int);
   const vpx_codec_enc_cfg_t *cfg = &ctx->cfg;
+  (void)ctrl_id;
+
   vp9_set_svc(ctx->cpi, data);
   // CBR or two pass mode for SVC with both temporal and spatial layers
   // not yet supported.
@@ -1046,11 +1066,12 @@
 }
 
 static vpx_codec_err_t ctrl_set_svc_layer_id(vpx_codec_alg_priv_t *ctx,
-                                             int ctr_id,
-                                             va_list args) {
+                                             int ctrl_id, va_list args) {
   vpx_svc_layer_id_t *const data = va_arg(args, vpx_svc_layer_id_t *);
   VP9_COMP *const cpi = (VP9_COMP *)ctx->cpi;
   SVC *const svc = &cpi->svc;
+  (void)ctrl_id;
+
   svc->spatial_layer_id = data->spatial_layer_id;
   svc->temporal_layer_id = data->temporal_layer_id;
   // Checks on valid layer_id input.
@@ -1066,9 +1087,10 @@
 }
 
 static vpx_codec_err_t ctrl_set_svc_parameters(vpx_codec_alg_priv_t *ctx,
-                                               int ctr_id, va_list args) {
+                                               int ctrl_id, va_list args) {
   VP9_COMP *const cpi = ctx->cpi;
   vpx_svc_parameters_t *const params = va_arg(args, vpx_svc_parameters_t *);
+  (void)ctrl_id;
 
   if (params == NULL)
     return VPX_CODEC_INVALID_PARAM;
@@ -1213,6 +1235,7 @@
     NOT_IMPLEMENTED,  // vpx_codec_get_si_fn_t
     NOT_IMPLEMENTED,  // vpx_codec_decode_fn_t
     NOT_IMPLEMENTED,  // vpx_codec_frame_get_fn_t
+    NOT_IMPLEMENTED   // vpx_codec_set_fb_fn_t
   },
   {  // NOLINT
     encoder_usage_cfg_map,  // vpx_codec_enc_cfg_map_t
@@ -1221,6 +1244,6 @@
     encoder_set_config,     // vpx_codec_enc_config_set_fn_t
     NOT_IMPLEMENTED,        // vpx_codec_get_global_headers_fn_t
     encoder_get_preview,    // vpx_codec_get_preview_frame_fn_t
-    NOT_IMPLEMENTED ,       // vpx_codec_enc_mr_get_mem_loc_fn_t
+    NOT_IMPLEMENTED         // vpx_codec_enc_mr_get_mem_loc_fn_t
   }
 };
diff --git a/vp9/vp9_dx_iface.c b/vp9/vp9_dx_iface.c
index 9ba4d56..328b98f 100644
--- a/vp9/vp9_dx_iface.c
+++ b/vp9/vp9_dx_iface.c
@@ -32,21 +32,12 @@
   vpx_codec_priv_t        base;
   vpx_codec_dec_cfg_t     cfg;
   vp9_stream_info_t       si;
-  int                     decoder_init;
   struct VP9Decoder *pbi;
   int                     postproc_cfg_set;
   vp8_postproc_cfg_t      postproc_cfg;
-#if CONFIG_POSTPROC_VISUALIZER
-  unsigned int            dbg_postproc_flag;
-  int                     dbg_color_ref_frame_flag;
-  int                     dbg_color_mb_modes_flag;
-  int                     dbg_color_b_modes_flag;
-  int                     dbg_display_mv_flag;
-#endif
   vpx_decrypt_cb          decrypt_cb;
   void                   *decrypt_state;
   vpx_image_t             img;
-  int                     img_setup;
   int                     img_avail;
   int                     invert_tile_order;
 
@@ -226,22 +217,10 @@
 static void set_ppflags(const vpx_codec_alg_priv_t *ctx,
                         vp9_ppflags_t *flags) {
   flags->post_proc_flag =
-#if CONFIG_POSTPROC_VISUALIZER
-      (ctx->dbg_color_ref_frame_flag ? VP9D_DEBUG_CLR_FRM_REF_BLKS : 0) |
-      (ctx->dbg_color_mb_modes_flag ? VP9D_DEBUG_CLR_BLK_MODES : 0) |
-      (ctx->dbg_color_b_modes_flag ? VP9D_DEBUG_CLR_BLK_MODES : 0) |
-      (ctx->dbg_display_mv_flag ? VP9D_DEBUG_DRAW_MV : 0) |
-#endif
       ctx->postproc_cfg.post_proc_flag;
 
   flags->deblocking_level = ctx->postproc_cfg.deblocking_level;
   flags->noise_level = ctx->postproc_cfg.noise_level;
-#if CONFIG_POSTPROC_VISUALIZER
-  flags->display_ref_frame_flag = ctx->dbg_color_ref_frame_flag;
-  flags->display_mb_modes_flag = ctx->dbg_color_mb_modes_flag;
-  flags->display_b_modes_flag = ctx->dbg_color_b_modes_flag;
-  flags->display_mv_flag = ctx->dbg_display_mv_flag;
-#endif
 }
 
 static void init_decoder(vpx_codec_alg_priv_t *ctx) {
@@ -252,8 +231,6 @@
   ctx->pbi->max_threads = ctx->cfg.threads;
   ctx->pbi->inv_tile_order = ctx->invert_tile_order;
 
-  vp9_initialize_dec();
-
   // If postprocessing was enabled by the application and a
   // configuration has not been provided, default it.
   if (!ctx->postproc_cfg_set &&
@@ -286,12 +263,10 @@
   }
 
   // Initialize the decoder instance on the first frame
-  if (!ctx->decoder_init) {
+  if (ctx->pbi == NULL) {
     init_decoder(ctx);
     if (ctx->pbi == NULL)
       return VPX_CODEC_ERROR;
-
-    ctx->decoder_init = 1;
   }
 
   // Set these even if already initialized.  The caller may have changed the
@@ -540,22 +515,7 @@
 
 static vpx_codec_err_t ctrl_set_dbg_options(vpx_codec_alg_priv_t *ctx,
                                             int ctrl_id, va_list args) {
-#if CONFIG_POSTPROC_VISUALIZER && CONFIG_POSTPROC
-  int data = va_arg(args, int);
-
-#define MAP(id, var) case id: var = data; break;
-
-  switch (ctrl_id) {
-      MAP(VP8_SET_DBG_COLOR_REF_FRAME,   ctx->dbg_color_ref_frame_flag);
-      MAP(VP8_SET_DBG_COLOR_MB_MODES,    ctx->dbg_color_mb_modes_flag);
-      MAP(VP8_SET_DBG_COLOR_B_MODES,     ctx->dbg_color_b_modes_flag);
-      MAP(VP8_SET_DBG_DISPLAY_MV,        ctx->dbg_display_mv_flag);
-  }
-
-  return VPX_CODEC_OK;
-#else
   return VPX_CODEC_INCAPABLE;
-#endif
 }
 
 static vpx_codec_err_t ctrl_get_last_ref_updates(vpx_codec_alg_priv_t *ctx,
diff --git a/vp9/vp9cx.mk b/vp9/vp9cx.mk
index bc9a478..6e5c521 100644
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -103,6 +103,7 @@
 ifeq ($(CONFIG_USE_X86INC),yes)
 VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_dct_mmx.asm
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_error_sse2.asm
+VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_error_intrin_avx2.c
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_sad_sse2.asm
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_subtract_sse2.asm
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_variance_sse2.c
diff --git a/vpx/internal/vpx_codec_internal.h b/vpx/internal/vpx_codec_internal.h
index 51ca65e..83c3308 100644
--- a/vpx/internal/vpx_codec_internal.h
+++ b/vpx/internal/vpx_codec_internal.h
@@ -170,8 +170,8 @@
  * \ref MUST be non-zero.
  */
 typedef const struct vpx_codec_ctrl_fn_map {
-  int                    ctrl_id;
-  vpx_codec_control_fn_t   fn;
+  int ctrl_id;
+  vpx_codec_control_fn_t fn;
 } vpx_codec_ctrl_fn_map_t;
 
 /*!\brief decode data function pointer prototype
diff --git a/vpx/src/svc_encodeframe.c b/vpx/src/svc_encodeframe.c
index b874be7..4009a8a 100644
--- a/vpx/src/svc_encodeframe.c
+++ b/vpx/src/svc_encodeframe.c
@@ -496,7 +496,6 @@
 vpx_codec_err_t vpx_svc_init(SvcContext *svc_ctx, vpx_codec_ctx_t *codec_ctx,
                              vpx_codec_iface_t *iface,
                              vpx_codec_enc_cfg_t *enc_cfg) {
-  int max_intra_size_pct;
   vpx_codec_err_t res;
   SvcInternal *const si = get_svc_internal(svc_ctx);
   if (svc_ctx == NULL || codec_ctx == NULL || iface == NULL ||
@@ -575,7 +574,6 @@
   // modify encoder configuration
   enc_cfg->ss_number_layers = si->layers;
   enc_cfg->ts_number_layers = 1;  // Temporal layers not used in this encoder.
-  enc_cfg->kf_mode = VPX_KF_DISABLED;
   // Lag in frames not currently supported
   enc_cfg->g_lag_in_frames = 0;
 
@@ -605,16 +603,8 @@
   }
 
   vpx_codec_control(codec_ctx, VP9E_SET_SVC, 1);
-  vpx_codec_control(codec_ctx, VP8E_SET_CPUUSED, 1);
-  vpx_codec_control(codec_ctx, VP8E_SET_STATIC_THRESHOLD, 1);
-  vpx_codec_control(codec_ctx, VP8E_SET_NOISE_SENSITIVITY, 1);
   vpx_codec_control(codec_ctx, VP8E_SET_TOKEN_PARTITIONS, 1);
 
-  max_intra_size_pct =
-      (int)(((double)enc_cfg->rc_buf_optimal_sz * 0.5) *
-            ((double)enc_cfg->g_timebase.den / enc_cfg->g_timebase.num) / 10.0);
-  vpx_codec_control(codec_ctx, VP8E_SET_MAX_INTRA_BITRATE_PCT,
-                    max_intra_size_pct);
   return VPX_CODEC_OK;
 }
 
@@ -869,8 +859,7 @@
   si->rc_stats_buf_used = 0;
 
   si->layers = svc_ctx->spatial_layers;
-  if (si->frame_within_gop >= si->kf_dist ||
-      si->encode_frame_count == 0) {
+  if (si->encode_frame_count == 0) {
     si->frame_within_gop = 0;
   }
   si->is_keyframe = (si->frame_within_gop == 0);
diff --git a/vpx/vpx_codec.mk b/vpx/vpx_codec.mk
index 98d1d56..a1ad3c5 100644
--- a/vpx/vpx_codec.mk
+++ b/vpx/vpx_codec.mk
@@ -15,8 +15,10 @@
 API_SRCS-$(CONFIG_VP8_ENCODER) += vp8cx.h
 API_DOC_SRCS-$(CONFIG_VP8_ENCODER) += vp8.h
 API_DOC_SRCS-$(CONFIG_VP8_ENCODER) += vp8cx.h
-API_SRCS-$(CONFIG_VP9_ENCODER) += src/svc_encodeframe.c
-API_SRCS-$(CONFIG_VP9_ENCODER) += svc_context.h
+ifeq ($(CONFIG_VP9_ENCODER),yes)
+  API_SRCS-$(CONFIG_SPATIAL_SVC) += src/svc_encodeframe.c
+  API_SRCS-$(CONFIG_SPATIAL_SVC) += svc_context.h
+endif
 
 API_SRCS-$(CONFIG_VP8_DECODER) += vp8.h
 API_SRCS-$(CONFIG_VP8_DECODER) += vp8dx.h
diff --git a/vpxdec.c b/vpxdec.c
index ed37c70..127e65f 100644
--- a/vpxdec.c
+++ b/vpxdec.c
@@ -873,8 +873,16 @@
         }
 
         if (img->d_w != scaled_img->d_w || img->d_h != scaled_img->d_h) {
+#if CONFIG_LIBYUV
           vpx_image_scale(img, scaled_img, kFilterBox);
           img = scaled_img;
+#else
+          fprintf(stderr, "Failed  to scale output frame: %s.\n"
+                  "Scaling is disabled in this configuration. "
+                  "To enable scaling, configure with --enable-libyuv\n",
+                  vpx_codec_error(&decoder));
+          return EXIT_FAILURE;
+#endif
         }
       }
 
diff --git a/vpxenc.c b/vpxenc.c
index 96a7ab6..d46a83e 100644
--- a/vpxenc.c
+++ b/vpxenc.c
@@ -755,7 +755,7 @@
       input->height = input->y4m.pic_h;
       input->framerate.numerator = input->y4m.fps_n;
       input->framerate.denominator = input->y4m.fps_d;
-      input->use_i420 = 0;
+      input->fmt = input->y4m.vpx_fmt;
     } else
       fatal("Unsupported Y4M stream.");
   } else if (input->detect.buf_read == 4 && fourcc_is_ivf(input->detect.buf)) {
@@ -1059,6 +1059,23 @@
   }
 }
 
+static const char* file_type_to_string(enum VideoFileType t) {
+  switch (t) {
+    case FILE_TYPE_RAW: return "RAW";
+    case FILE_TYPE_Y4M: return "Y4M";
+    default: return "Other";
+  }
+}
+
+static const char* image_format_to_string(vpx_img_fmt_t f) {
+  switch (f) {
+    case VPX_IMG_FMT_I420: return "I420";
+    case VPX_IMG_FMT_I422: return "I422";
+    case VPX_IMG_FMT_I444: return "I444";
+    case VPX_IMG_FMT_YV12: return "YV12";
+    default: return "Other";
+  }
+}
 
 static void show_stream_config(struct stream_state *stream,
                                struct VpxEncoderConfig *global,
@@ -1070,8 +1087,10 @@
   if (stream->index == 0) {
     fprintf(stderr, "Codec: %s\n",
             vpx_codec_iface_name(global->codec->interface()));
-    fprintf(stderr, "Source file: %s Format: %s\n", input->filename,
-            input->use_i420 ? "I420" : "YV12");
+    fprintf(stderr, "Source file: %s File Type: %s Format: %s\n",
+            input->filename,
+            file_type_to_string(input->file_type),
+            image_format_to_string(input->fmt));
   }
   if (stream->next || stream->index)
     fprintf(stderr, "\nStream Index: %d\n", stream->index);
@@ -1245,6 +1264,11 @@
 
   /* Scale if necessary */
   if (img && (img->d_w != cfg->g_w || img->d_h != cfg->g_h)) {
+    if (img->fmt != VPX_IMG_FMT_I420 && img->fmt != VPX_IMG_FMT_YV12) {
+      fprintf(stderr, "%s can only scale 4:2:0 8bpp inputs\n", exec_name);
+      exit(EXIT_FAILURE);
+    }
+#if CONFIG_LIBYUV
     if (!stream->img)
       stream->img = vpx_img_alloc(NULL, VPX_IMG_FMT_I420,
                                   cfg->g_w, cfg->g_h, 16);
@@ -1260,8 +1284,15 @@
               stream->img->stride[VPX_PLANE_V],
               stream->img->d_w, stream->img->d_h,
               kFilterBox);
-
     img = stream->img;
+#else
+    stream->encoder.err = 1;
+    ctx_exit_on_error(&stream->encoder,
+                      "Stream %d: Failed to encode frame.\n"
+                      "Scaling disabled in this configuration. \n"
+                      "To enable, configure with --enable-libyuv\n",
+                      stream->index);
+#endif
   }
 
   vpx_usec_timer_start(&timer);
@@ -1501,7 +1532,6 @@
   /* Setup default input stream settings */
   input.framerate.numerator = 30;
   input.framerate.denominator = 1;
-  input.use_i420 = 1;
   input.only_i420 = 1;
 
   /* First parse the global configuration values, because we want to apply
@@ -1511,6 +1541,7 @@
   argv = argv_dup(argc - 1, argv_ + 1);
   parse_global_config(&global, argv);
 
+  input.fmt = global.use_i420 ? VPX_IMG_FMT_I420 : VPX_IMG_FMT_YV12;
 
   {
     /* Now parse each stream's parameters. Using a local scope here
@@ -1611,10 +1642,7 @@
            frames.*/
         memset(&raw, 0, sizeof(raw));
       else
-        vpx_img_alloc(&raw,
-                      input.use_i420 ? VPX_IMG_FMT_I420
-                      : VPX_IMG_FMT_YV12,
-                      input.width, input.height, 32);
+        vpx_img_alloc(&raw, input.fmt, input.width, input.height, 32);
 
       FOREACH_STREAM(stream->rate_hist =
                          init_rate_histogram(&stream->config.cfg,