Merge "mips msa vp9 convolve8 vert optimization"
diff --git a/build/make/Android.mk b/build/make/Android.mk
index 3d3f57d..20f670a 100644
--- a/build/make/Android.mk
+++ b/build/make/Android.mk
@@ -158,8 +158,6 @@
LOCAL_MODULE := libvpx
-LOCAL_LDLIBS := -llog
-
ifeq ($(CONFIG_RUNTIME_CPU_DETECT),yes)
LOCAL_STATIC_LIBRARIES := cpufeatures
endif
diff --git a/test/android/Android.mk b/test/android/Android.mk
index 4e750b2..af85634 100644
--- a/test/android/Android.mk
+++ b/test/android/Android.mk
@@ -40,7 +40,13 @@
LOCAL_ARM_MODE := arm
LOCAL_MODULE := libvpx_test
LOCAL_STATIC_LIBRARIES := gtest libwebm
-LOCAL_SHARED_LIBRARIES := vpx
+
+ifeq ($(ENABLE_SHARED),1)
+ LOCAL_SHARED_LIBRARIES := vpx
+else
+ LOCAL_STATIC_LIBRARIES += vpx
+endif
+
include $(LOCAL_PATH)/test/test.mk
LOCAL_C_INCLUDES := $(BINDINGS_DIR)
FILTERED_SRC := $(sort $(filter %.cc %.c, $(LIBVPX_TEST_SRCS-yes)))
diff --git a/test/blockiness_test.cc b/test/blockiness_test.cc
new file mode 100644
index 0000000..92cce6a
--- /dev/null
+++ b/test/blockiness_test.cc
@@ -0,0 +1,229 @@
+/*
+ * Copyright (c) 2012 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 <string.h>
+#include <limits.h>
+#include <stdio.h>
+
+#include "./vpx_config.h"
+#if CONFIG_VP9_ENCODER
+#include "./vp9_rtcd.h"
+#endif
+
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "vpx_mem/vpx_mem.h"
+
+
+extern "C"
+double vp9_get_blockiness(const unsigned char *img1, int img1_pitch,
+ const unsigned char *img2, int img2_pitch,
+ int width, int height);
+
+using libvpx_test::ACMRandom;
+
+namespace {
+class BlockinessTestBase : public ::testing::Test {
+ public:
+ BlockinessTestBase(int width, int height) : width_(width), height_(height) {}
+
+ static void SetUpTestCase() {
+ source_data_ = reinterpret_cast<uint8_t*>(
+ vpx_memalign(kDataAlignment, kDataBufferSize));
+ reference_data_ = reinterpret_cast<uint8_t*>(
+ vpx_memalign(kDataAlignment, kDataBufferSize));
+ }
+
+ static void TearDownTestCase() {
+ vpx_free(source_data_);
+ source_data_ = NULL;
+ vpx_free(reference_data_);
+ reference_data_ = NULL;
+ }
+
+ virtual void TearDown() {
+ libvpx_test::ClearSystemState();
+ }
+
+ protected:
+ // Handle frames up to 640x480
+ static const int kDataAlignment = 16;
+ static const int kDataBufferSize = 640*480;
+
+ virtual void SetUp() {
+ source_stride_ = (width_ + 31) & ~31;
+ reference_stride_ = width_ * 2;
+ rnd_.Reset(ACMRandom::DeterministicSeed());
+ }
+
+ void FillConstant(uint8_t *data, int stride, uint8_t fill_constant,
+ int width, int height) {
+ for (int h = 0; h < height; ++h) {
+ for (int w = 0; w < width; ++w) {
+ data[h * stride + w] = fill_constant;
+ }
+ }
+ }
+
+ void FillConstant(uint8_t *data, int stride, uint8_t fill_constant) {
+ FillConstant(data, stride, fill_constant, width_, height_);
+ }
+
+ void FillRandom(uint8_t *data, int stride, int width, int height) {
+ for (int h = 0; h < height; ++h) {
+ for (int w = 0; w < width; ++w) {
+ data[h * stride + w] = rnd_.Rand8();
+ }
+ }
+ }
+
+ void FillRandom(uint8_t *data, int stride) {
+ FillRandom(data, stride, width_, height_);
+ }
+
+ void FillRandomBlocky(uint8_t *data, int stride) {
+ for (int h = 0; h < height_; h += 4) {
+ for (int w = 0; w < width_; w += 4) {
+ FillRandom(data + h * stride + w, stride, 4, 4);
+ }
+ }
+ }
+
+ void FillCheckerboard(uint8_t *data, int stride) {
+ for (int h = 0; h < height_; h += 4) {
+ for (int w = 0; w < width_; w += 4) {
+ if (((h/4) ^ (w/4)) & 1)
+ FillConstant(data + h * stride + w, stride, 255, 4, 4);
+ else
+ FillConstant(data + h * stride + w, stride, 0, 4, 4);
+ }
+ }
+ }
+
+ void Blur(uint8_t *data, int stride, int taps) {
+ int sum = 0;
+ int half_taps = taps / 2;
+ for (int h = 0; h < height_; ++h) {
+ for (int w = 0; w < taps; ++w) {
+ sum += data[w + h * stride];
+ }
+ for (int w = taps; w < width_; ++w) {
+ sum += data[w + h * stride] - data[w - taps + h * stride];
+ data[w - half_taps + h * stride] = (sum + half_taps) / taps;
+ }
+ }
+ for (int w = 0; w < width_; ++w) {
+ for (int h = 0; h < taps; ++h) {
+ sum += data[h + w * stride];
+ }
+ for (int h = taps; h < height_; ++h) {
+ sum += data[w + h * stride] - data[(h - taps) * stride + w];
+ data[(h - half_taps) * stride + w] = (sum + half_taps) / taps;
+ }
+ }
+ }
+ int width_, height_;
+ static uint8_t* source_data_;
+ int source_stride_;
+ static uint8_t* reference_data_;
+ int reference_stride_;
+
+ ACMRandom rnd_;
+};
+
+#if CONFIG_VP9_ENCODER
+typedef std::tr1::tuple<int, int> BlockinessParam;
+class BlockinessVP9Test
+ : public BlockinessTestBase,
+ public ::testing::WithParamInterface<BlockinessParam> {
+ public:
+ BlockinessVP9Test() : BlockinessTestBase(GET_PARAM(0), GET_PARAM(1)) {}
+
+ protected:
+ int CheckBlockiness() {
+ return vp9_get_blockiness(source_data_, source_stride_,
+ reference_data_, reference_stride_,
+ width_, height_);
+ }
+};
+#endif // CONFIG_VP9_ENCODER
+
+uint8_t* BlockinessTestBase::source_data_ = NULL;
+uint8_t* BlockinessTestBase::reference_data_ = NULL;
+
+#if CONFIG_VP9_ENCODER
+TEST_P(BlockinessVP9Test, SourceBlockierThanReference) {
+ // Source is blockier than reference.
+ FillRandomBlocky(source_data_, source_stride_);
+ FillConstant(reference_data_, reference_stride_, 128);
+ int super_blocky = CheckBlockiness();
+
+ EXPECT_EQ(0, super_blocky) << "Blocky source should produce 0 blockiness.";
+}
+
+TEST_P(BlockinessVP9Test, ReferenceBlockierThanSource) {
+ // Source is blockier than reference.
+ FillConstant(source_data_, source_stride_, 128);
+ FillRandomBlocky(reference_data_, reference_stride_);
+ int super_blocky = CheckBlockiness();
+
+ EXPECT_GT(super_blocky, 0.0)
+ << "Blocky reference should score high for blockiness.";
+}
+
+TEST_P(BlockinessVP9Test, BlurringDecreasesBlockiness) {
+ // Source is blockier than reference.
+ FillConstant(source_data_, source_stride_, 128);
+ FillRandomBlocky(reference_data_, reference_stride_);
+ int super_blocky = CheckBlockiness();
+
+ Blur(reference_data_, reference_stride_, 4);
+ int less_blocky = CheckBlockiness();
+
+ EXPECT_GT(super_blocky, less_blocky)
+ << "A straight blur should decrease blockiness.";
+}
+
+TEST_P(BlockinessVP9Test, WorstCaseBlockiness) {
+ // Source is blockier than reference.
+ FillConstant(source_data_, source_stride_, 128);
+ FillCheckerboard(reference_data_, reference_stride_);
+
+ int super_blocky = CheckBlockiness();
+
+ Blur(reference_data_, reference_stride_, 4);
+ int less_blocky = CheckBlockiness();
+
+ EXPECT_GT(super_blocky, less_blocky)
+ << "A straight blur should decrease blockiness.";
+}
+#endif // CONFIG_VP9_ENCODER
+
+
+using std::tr1::make_tuple;
+
+//------------------------------------------------------------------------------
+// C functions
+
+#if CONFIG_VP9_ENCODER
+const BlockinessParam c_vp9_tests[] = {
+ make_tuple(320, 240),
+ make_tuple(318, 242),
+ make_tuple(318, 238),
+};
+INSTANTIATE_TEST_CASE_P(C, BlockinessVP9Test, ::testing::ValuesIn(c_vp9_tests));
+#endif
+
+} // namespace
diff --git a/test/encode_test_driver.cc b/test/encode_test_driver.cc
index bdd71c6..ff39f1a 100644
--- a/test/encode_test_driver.cc
+++ b/test/encode_test_driver.cc
@@ -29,8 +29,6 @@
cfg_.g_timebase = video->timebase();
cfg_.rc_twopass_stats_in = stats_->buf();
- // Default to 1 thread.
- cfg_.g_threads = 1;
res = vpx_codec_enc_init(&encoder_, CodecInterface(), &cfg_,
init_flags_);
ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
diff --git a/test/encode_test_driver.h b/test/encode_test_driver.h
index 7b7dd31..e16cf9c 100644
--- a/test/encode_test_driver.h
+++ b/test/encode_test_driver.h
@@ -183,7 +183,10 @@
protected:
explicit EncoderTest(const CodecFactory *codec)
: codec_(codec), abort_(false), init_flags_(0), frame_flags_(0),
- last_pts_(0) {}
+ last_pts_(0) {
+ // Default to 1 thread.
+ cfg_.g_threads = 1;
+ }
virtual ~EncoderTest() {}
diff --git a/test/svc_test.cc b/test/svc_test.cc
index 67e83e3..0af3005 100644
--- a/test/svc_test.cc
+++ b/test/svc_test.cc
@@ -63,6 +63,9 @@
vpx_codec_dec_cfg_t dec_cfg = vpx_codec_dec_cfg_t();
VP9CodecFactory codec_factory;
decoder_ = codec_factory.CreateDecoder(dec_cfg, 0);
+
+ tile_columns_ = 0;
+ tile_rows_ = 0;
}
virtual void TearDown() {
@@ -75,6 +78,8 @@
vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
EXPECT_EQ(VPX_CODEC_OK, res);
vpx_codec_control(&codec_, VP8E_SET_CPUUSED, 4); // Make the test faster
+ vpx_codec_control(&codec_, VP9E_SET_TILE_COLUMNS, tile_columns_);
+ vpx_codec_control(&codec_, VP9E_SET_TILE_ROWS, tile_rows_);
codec_initialized_ = true;
}
@@ -108,7 +113,8 @@
codec_enc_.g_pass = VPX_RC_FIRST_PASS;
InitializeEncoder();
- libvpx_test::I420VideoSource video(test_file_name_, kWidth, kHeight,
+ libvpx_test::I420VideoSource video(test_file_name_,
+ codec_enc_.g_w, codec_enc_.g_h,
codec_enc_.g_timebase.den,
codec_enc_.g_timebase.num, 0, 30);
video.Begin();
@@ -176,7 +182,8 @@
}
InitializeEncoder();
- libvpx_test::I420VideoSource video(test_file_name_, kWidth, kHeight,
+ libvpx_test::I420VideoSource video(test_file_name_,
+ codec_enc_.g_w, codec_enc_.g_h,
codec_enc_.g_timebase.den,
codec_enc_.g_timebase.num, 0, 30);
video.Begin();
@@ -310,6 +317,8 @@
std::string test_file_name_;
bool codec_initialized_;
Decoder *decoder_;
+ int tile_columns_;
+ int tile_rows_;
};
TEST_F(SvcTest, SvcInit) {
@@ -737,4 +746,51 @@
FreeBitstreamBuffers(&outputs[0], 10);
}
+TEST_F(SvcTest, TwoPassEncode2TemporalLayersWithTiles) {
+ // First pass encode
+ std::string stats_buf;
+ vpx_svc_set_options(&svc_, "scale-factors=1/1");
+ svc_.temporal_layers = 2;
+ Pass1EncodeNFrames(10, 1, &stats_buf);
+
+ // Second pass encode
+ codec_enc_.g_pass = VPX_RC_LAST_PASS;
+ svc_.temporal_layers = 2;
+ vpx_svc_set_options(&svc_, "auto-alt-refs=1 scale-factors=1/1");
+ codec_enc_.g_w = 704;
+ codec_enc_.g_h = 144;
+ tile_columns_ = 1;
+ tile_rows_ = 1;
+ vpx_fixed_buf outputs[10];
+ memset(&outputs[0], 0, sizeof(outputs));
+ Pass2EncodeNFrames(&stats_buf, 10, 1, &outputs[0]);
+ DecodeNFrames(&outputs[0], 10);
+ FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest,
+ TwoPassEncode2TemporalLayersWithMultipleFrameContextsAndTiles) {
+ // First pass encode
+ std::string stats_buf;
+ vpx_svc_set_options(&svc_, "scale-factors=1/1");
+ svc_.temporal_layers = 2;
+ Pass1EncodeNFrames(10, 1, &stats_buf);
+
+ // Second pass encode
+ codec_enc_.g_pass = VPX_RC_LAST_PASS;
+ svc_.temporal_layers = 2;
+ codec_enc_.g_error_resilient = 0;
+ codec_enc_.g_w = 704;
+ codec_enc_.g_h = 144;
+ tile_columns_ = 1;
+ tile_rows_ = 1;
+ vpx_svc_set_options(&svc_, "auto-alt-refs=1 scale-factors=1/1 "
+ "multi-frame-contexts=1");
+ vpx_fixed_buf outputs[10];
+ memset(&outputs[0], 0, sizeof(outputs));
+ Pass2EncodeNFrames(&stats_buf, 10, 1, &outputs[0]);
+ DecodeNFrames(&outputs[0], 10);
+ FreeBitstreamBuffers(&outputs[0], 10);
+}
+
} // namespace
diff --git a/test/test.mk b/test/test.mk
index 91dd335..5baf234 100644
--- a/test/test.mk
+++ b/test/test.mk
@@ -150,6 +150,7 @@
ifeq ($(CONFIG_VP9_ENCODER),yes)
LIBVPX_TEST_SRCS-$(CONFIG_SPATIAL_SVC) += svc_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_INTERNAL_STATS) += blockiness_test.cc
endif
ifeq ($(CONFIG_VP9_ENCODER)$(CONFIG_VP9_TEMPORAL_DENOISING),yesyes)
diff --git a/test/test_libvpx.cc b/test/test_libvpx.cc
index dcf5fc5..30a5255 100644
--- a/test/test_libvpx.cc
+++ b/test/test_libvpx.cc
@@ -14,12 +14,12 @@
#endif
extern "C" {
#if CONFIG_VP8
-#include "./vp8_rtcd.h"
+extern void vp8_rtcd();
#endif // CONFIG_VP8
#if CONFIG_VP9
-#include "./vp9_rtcd.h"
+extern void vp9_rtcd();
#endif // CONFIG_VP9
-#include "./vpx_scale_rtcd.h"
+extern void vpx_scale_rtcd();
}
#include "third_party/googletest/src/include/gtest/gtest.h"
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index ced5dcf..e1aecd8 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -1114,6 +1114,9 @@
add_proto qw/unsigned int vp9_avg_4x4/, "const uint8_t *, int p";
specialize qw/vp9_avg_4x4 sse2/;
+add_proto qw/void vp9_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
+specialize qw/vp9_minmax_8x8 sse2/;
+
add_proto qw/void vp9_hadamard_8x8/, "int16_t const *src_diff, int src_stride, int16_t *coeff";
specialize qw/vp9_hadamard_8x8 sse2/, "$ssse3_x86_64";
@@ -1137,6 +1140,8 @@
specialize qw/vp9_highbd_avg_8x8/;
add_proto qw/unsigned int vp9_highbd_avg_4x4/, "const uint8_t *, int p";
specialize qw/vp9_highbd_avg_4x4/;
+ add_proto qw/unsigned int vp9_highbd_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
+ specialize qw/vp9_highbd_minmax_8x8/;
}
# ENCODEMB INVOKE
diff --git a/vp9/encoder/vp9_avg.c b/vp9/encoder/vp9_avg.c
index 58daa3a..e26a03c 100644
--- a/vp9/encoder/vp9_avg.c
+++ b/vp9/encoder/vp9_avg.c
@@ -155,6 +155,20 @@
return var;
}
+void vp9_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp,
+ int *min, int *max) {
+ int i, j;
+ *min = 255;
+ *max = 0;
+ for (i = 0; i < 8; ++i, s += p, d += dp) {
+ for (j = 0; j < 8; ++j) {
+ int diff = abs(s[j]-d[j]);
+ *min = diff < *min ? diff : *min;
+ *max = diff > *max ? diff : *max;
+ }
+ }
+}
+
#if CONFIG_VP9_HIGHBITDEPTH
unsigned int vp9_highbd_avg_8x8_c(const uint8_t *s8, int p) {
int i, j;
@@ -175,6 +189,22 @@
return (sum + 8) >> 4;
}
+
+void vp9_highbd_minmax_8x8_c(const uint8_t *s8, int p, const uint8_t *d8,
+ int dp, int *min, int *max) {
+ int i, j;
+ *min = 255;
+ *max = 0;
+ const uint16_t* s = CONVERT_TO_SHORTPTR(s8);
+ const uint16_t* d = CONVERT_TO_SHORTPTR(d8);
+ for (i = 0; i < 8; ++i, s += p, d += dp) {
+ for (j = 0; j < 8; ++j) {
+ int diff = abs(s[j]-d[j]);
+ *min = diff < *min ? diff : *min;
+ *max = diff > *max ? diff : *max;
+ }
+ }
+}
#endif // CONFIG_VP9_HIGHBITDEPTH
diff --git a/vp9/encoder/vp9_blockiness.c b/vp9/encoder/vp9_blockiness.c
new file mode 100644
index 0000000..b8629bd
--- /dev/null
+++ b/vp9/encoder/vp9_blockiness.c
@@ -0,0 +1,138 @@
+/*
+ * 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 "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_convolve.h"
+#include "vp9/common/vp9_filter.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+static int horizontal_filter(const uint8_t *s) {
+ return (s[1] - s[-2]) * 2 + (s[-1] - s[0]) * 6;
+}
+
+static int vertical_filter(const uint8_t *s, int p) {
+ return (s[p] - s[-2 * p]) * 2 + (s[-p] - s[0]) * 6;
+}
+
+static int variance(int sum, int sum_squared, int size) {
+ return sum_squared / size - (sum / size) * (sum / size);
+}
+// Calculate a blockiness level for a vertical block edge.
+// This function returns a new blockiness metric that's defined as
+
+// p0 p1 p2 p3
+// q0 q1 q2 q3
+// block edge ->
+// r0 r1 r2 r3
+// s0 s1 s2 s3
+
+// blockiness = p0*-2+q0*6+r0*-6+s0*2 +
+// p1*-2+q1*6+r1*-6+s1*2 +
+// p2*-2+q2*6+r2*-6+s2*2 +
+// p3*-2+q3*6+r3*-6+s3*2 ;
+
+// reconstructed_blockiness = abs(blockiness from reconstructed buffer -
+// blockiness from source buffer,0)
+//
+// I make the assumption that flat blocks are much more visible than high
+// contrast blocks. As such, I scale the result of the blockiness calc
+// by dividing the blockiness by the variance of the pixels on either side
+// of the edge as follows:
+// var_0 = (q0^2+q1^2+q2^2+q3^2) - ((q0 + q1 + q2 + q3) / 4 )^2
+// var_1 = (r0^2+r1^2+r2^2+r3^2) - ((r0 + r1 + r2 + r3) / 4 )^2
+// The returned blockiness is the scaled value
+// Reconstructed blockiness / ( 1 + var_0 + var_1 ) ;
+int blockiness_vertical(const uint8_t *s, int sp, const uint8_t *r, int rp,
+ int size) {
+ int s_blockiness = 0;
+ int r_blockiness = 0;
+ int sum_0 = 0;
+ int sum_sq_0 = 0;
+ int sum_1 = 0;
+ int sum_sq_1 = 0;
+ int i;
+ int var_0;
+ int var_1;
+ for (i = 0; i < size; ++i, s += sp, r += rp) {
+ s_blockiness += horizontal_filter(s);
+ r_blockiness += horizontal_filter(r);
+ sum_0 += s[0];
+ sum_sq_0 += s[0]*s[0];
+ sum_1 += s[-1];
+ sum_sq_1 += s[-1]*s[-1];
+ }
+ var_0 = variance(sum_0, sum_sq_0, size);
+ var_1 = variance(sum_1, sum_sq_1, size);
+ r_blockiness = abs(r_blockiness);
+ s_blockiness = abs(s_blockiness);
+
+ if (r_blockiness > s_blockiness)
+ return (r_blockiness - s_blockiness) / (1 + var_0 + var_1);
+ else
+ return 0;
+}
+
+// Calculate a blockiness level for a horizontal block edge
+// same as above.
+int blockiness_horizontal(const uint8_t *s, int sp, const uint8_t *r, int rp,
+ int size) {
+ int s_blockiness = 0;
+ int r_blockiness = 0;
+ int sum_0 = 0;
+ int sum_sq_0 = 0;
+ int sum_1 = 0;
+ int sum_sq_1 = 0;
+ int i;
+ int var_0;
+ int var_1;
+ for (i = 0; i < size; ++i, ++s, ++r) {
+ s_blockiness += vertical_filter(s, sp);
+ r_blockiness += vertical_filter(r, rp);
+ sum_0 += s[0];
+ sum_sq_0 += s[0] * s[0];
+ sum_1 += s[-sp];
+ sum_sq_1 += s[-sp] * s[-sp];
+ }
+ var_0 = variance(sum_0, sum_sq_0, size);
+ var_1 = variance(sum_1, sum_sq_1, size);
+ r_blockiness = abs(r_blockiness);
+ s_blockiness = abs(s_blockiness);
+
+ if (r_blockiness > s_blockiness)
+ return (r_blockiness - s_blockiness) / (1 + var_0 + var_1);
+ else
+ return 0;
+}
+
+// This function returns the blockiness for the entire frame currently by
+// looking at all borders in steps of 4.
+double vp9_get_blockiness(const unsigned char *img1, int img1_pitch,
+ const unsigned char *img2, int img2_pitch,
+ int width, int height ) {
+ double blockiness = 0;
+ int i, j;
+ vp9_clear_system_state();
+ for (i = 0; i < height; i += 4, img1 += img1_pitch * 4,
+ img2 += img2_pitch * 4) {
+ for (j = 0; j < width; j += 4) {
+ if (i > 0 && i < height && j > 0 && j < width) {
+ blockiness += blockiness_vertical(img1 + j, img1_pitch,
+ img2 + j, img2_pitch, 4);
+ blockiness += blockiness_horizontal(img1 + j, img1_pitch,
+ img2 + j, img2_pitch, 4);
+ }
+ }
+ }
+ blockiness /= width * height / 16;
+ return blockiness;
+}
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index e59d2c2..d428175 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -390,18 +390,21 @@
variance_node vt;
const int block_width = num_8x8_blocks_wide_lookup[bsize];
const int block_height = num_8x8_blocks_high_lookup[bsize];
+ const int low_res = (cm->width <= 352 && cm->height <= 288);
assert(block_height == block_width);
tree_to_node(data, bsize, &vt);
- if (force_split)
+ if (force_split == 1)
return 0;
// For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
// variance is below threshold, otherwise split will be selected.
// No check for vert/horiz split as too few samples for variance.
if (bsize == bsize_min) {
- get_variance(&vt.part_variances->none);
+ // Variance already computed to set the force_split.
+ if (low_res || cm->frame_type == KEY_FRAME)
+ get_variance(&vt.part_variances->none);
if (mi_col + block_width / 2 < cm->mi_cols &&
mi_row + block_height / 2 < cm->mi_rows &&
vt.part_variances->none.variance < threshold) {
@@ -410,11 +413,10 @@
}
return 0;
} else if (bsize > bsize_min) {
- // Variance is already computed for 32x32 blocks to set the force_split.
- if (bsize != BLOCK_32X32)
+ // Variance already computed to set the force_split.
+ if (low_res || cm->frame_type == KEY_FRAME)
get_variance(&vt.part_variances->none);
- // For key frame or low_res: for bsize above 32X32 or very high variance,
- // take split.
+ // For key frame: take split for bsize above 32X32 or very high variance.
if (cm->frame_type == KEY_FRAME &&
(bsize > BLOCK_32X32 ||
vt.part_variances->none.variance > (threshold << 4))) {
@@ -483,21 +485,68 @@
cpi->vbp_thresholds[1] = threshold_base >> 2;
cpi->vbp_thresholds[2] = threshold_base >> 2;
cpi->vbp_thresholds[3] = threshold_base << 2;
+ cpi->vbp_threshold_sad = 0;
cpi->vbp_bsize_min = BLOCK_8X8;
} else {
cpi->vbp_thresholds[1] = threshold_base;
if (cm->width <= 352 && cm->height <= 288) {
cpi->vbp_thresholds[0] = threshold_base >> 2;
cpi->vbp_thresholds[2] = threshold_base << 3;
+ cpi->vbp_threshold_sad = 100;
} else {
cpi->vbp_thresholds[0] = threshold_base;
+ cpi->vbp_thresholds[1] = (5 * threshold_base) >> 2;
cpi->vbp_thresholds[2] = threshold_base << cpi->oxcf.speed;
+ cpi->vbp_threshold_sad = 1000;
}
cpi->vbp_bsize_min = BLOCK_16X16;
}
+ cpi->vbp_threshold_minmax = 15 + (q >> 3);
}
}
+// Compute the minmax over the 8x8 subblocks.
+static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
+ int dp, int x16_idx, int y16_idx,
+#if CONFIG_VP9_HIGHBITDEPTH
+ int highbd_flag,
+#endif
+ int pixels_wide,
+ int pixels_high) {
+ int k;
+ int minmax_max = 0;
+ int minmax_min = 255;
+ // Loop over the 4 8x8 subblocks.
+ for (k = 0; k < 4; k++) {
+ int x8_idx = x16_idx + ((k & 1) << 3);
+ int y8_idx = y16_idx + ((k >> 1) << 3);
+ int min = 0;
+ int max = 0;
+ if (x8_idx < pixels_wide && y8_idx < pixels_high) {
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
+ vp9_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
+ d + y8_idx * dp + x8_idx, dp,
+ &min, &max);
+ } else {
+ vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
+ d + y8_idx * dp + x8_idx, dp,
+ &min, &max);
+ }
+#else
+ vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
+ d + y8_idx * dp + x8_idx, dp,
+ &min, &max);
+#endif
+ if ((max - min) > minmax_max)
+ minmax_max = (max - min);
+ if ((max - min) < minmax_min)
+ minmax_min = (max - min);
+ }
+ }
+ return (minmax_max - minmax_min);
+}
+
static void modify_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q) {
VP9_COMMON *const cm = &cpi->common;
const int64_t threshold_base = (int64_t)(cpi->y_dequant[q][1]);
@@ -510,6 +559,7 @@
thresholds[2] = threshold_base << 3;
} else {
thresholds[0] = threshold_base;
+ thresholds[1] = (5 * threshold_base) >> 2;
thresholds[2] = threshold_base << cpi->oxcf.speed;
}
}
@@ -594,7 +644,7 @@
// This function chooses partitioning based on the variance between source and
// reconstructed last, where variance is computed for down-sampled inputs.
-static void choose_partitioning(VP9_COMP *cpi,
+static int choose_partitioning(VP9_COMP *cpi,
const TileInfo *const tile,
MACROBLOCK *x,
int mi_row, int mi_col) {
@@ -603,7 +653,7 @@
int i, j, k, m;
v64x64 vt;
v16x16 vt2[16];
- int force_split[5];
+ int force_split[21];
uint8_t *s;
const uint8_t *d;
int sp;
@@ -699,6 +749,19 @@
d = xd->plane[0].dst.buf;
dp = xd->plane[0].dst.stride;
+
+ // If the y_sad is very small, take 64x64 as partition and exit.
+ // Don't check on boosted segment for now, as 64x64 is suppressed there.
+ if (segment_id == CR_SEGMENT_ID_BASE &&
+ y_sad < cpi->vbp_threshold_sad) {
+ const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
+ const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
+ if (mi_col + block_width / 2 < cm->mi_cols &&
+ mi_row + block_height / 2 < cm->mi_rows) {
+ set_block_size(cpi, xd, mi_row, mi_col, BLOCK_64X64);
+ return 0;
+ }
+ }
} else {
d = VP9_VAR_OFFS;
dp = 0;
@@ -721,6 +784,7 @@
}
// Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
+ // 5-20 for the 16x16 blocks.
force_split[0] = 0;
// Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
// for splits.
@@ -732,7 +796,9 @@
for (j = 0; j < 4; j++) {
const int x16_idx = x32_idx + ((j & 1) << 4);
const int y16_idx = y32_idx + ((j >> 1) << 4);
+ const int split_index = 5 + i2 + j;
v16x16 *vst = &vt.split[i].split[j];
+ force_split[split_index] = 0;
variance4x4downsample[i2 + j] = 0;
if (!is_key_frame) {
fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
@@ -743,15 +809,36 @@
pixels_high,
is_key_frame);
fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
- // For low-resolution, compute the variance based on 8x8 down-sampling,
- // and if it is large (above the threshold) we go down for 4x4.
- // For key frame we always go down to 4x4.
- if (low_res)
- get_variance(&vt.split[i].split[j].part_variances.none);
+ get_variance(&vt.split[i].split[j].part_variances.none);
+ if (vt.split[i].split[j].part_variances.none.variance >
+ thresholds[2]) {
+ // 16X16 variance is above threshold for split, so force split to 8x8
+ // for this 16x16 block (this also forces splits for upper levels).
+ force_split[split_index] = 1;
+ force_split[i + 1] = 1;
+ force_split[0] = 1;
+ } else if (vt.split[i].split[j].part_variances.none.variance >
+ thresholds[1] &&
+ !cyclic_refresh_segment_id_boosted(segment_id)) {
+ // We have some nominal amount of 16x16 variance (based on average),
+ // compute the minmax over the 8x8 sub-blocks, and if above threshold,
+ // force split to 8x8 block for this 16x16 block.
+ int minmax = compute_minmax_8x8(s, sp, d, dp, x16_idx, y16_idx,
+#if CONFIG_VP9_HIGHBITDEPTH
+ xd->cur_buf->flags,
+#endif
+ pixels_wide, pixels_high);
+ if (minmax > cpi->vbp_threshold_minmax) {
+ force_split[split_index] = 1;
+ force_split[i + 1] = 1;
+ force_split[0] = 1;
+ }
+ }
}
if (is_key_frame || (low_res &&
vt.split[i].split[j].part_variances.none.variance >
(thresholds[1] << 1))) {
+ force_split[split_index] = 0;
// Go down to 4x4 down-sampling for variance.
variance4x4downsample[i2 + j] = 1;
for (k = 0; k < 4; k++) {
@@ -786,16 +873,20 @@
fill_variance_tree(&vt.split[i], BLOCK_32X32);
// If variance of this 32x32 block is above the threshold, force the block
// to split. This also forces a split on the upper (64x64) level.
- get_variance(&vt.split[i].part_variances.none);
- if (vt.split[i].part_variances.none.variance > thresholds[1]) {
- force_split[i + 1] = 1;
- force_split[0] = 1;
+ if (!force_split[i + 1]) {
+ get_variance(&vt.split[i].part_variances.none);
+ if (vt.split[i].part_variances.none.variance > thresholds[1]) {
+ force_split[i + 1] = 1;
+ force_split[0] = 1;
+ }
}
}
- if (!force_split[0])
+ if (!force_split[0]) {
fill_variance_tree(&vt, BLOCK_64X64);
+ get_variance(&vt.part_variances.none);
+ }
- // Now go through the entire structure, splitting every block size until
+ // Now go through the entire structure, splitting every block size until
// we get to one that's got a variance lower than our threshold.
if ( mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
!set_vt_partitioning(cpi, xd, &vt, BLOCK_64X64, mi_row, mi_col,
@@ -820,7 +911,9 @@
if (!set_vt_partitioning(cpi, xd, vtemp, BLOCK_16X16,
mi_row + y32_idx + y16_idx,
mi_col + x32_idx + x16_idx,
- thresholds[2], cpi->vbp_bsize_min, 0)) {
+ thresholds[2],
+ cpi->vbp_bsize_min,
+ force_split[5 + i2 + j])) {
for (k = 0; k < 4; ++k) {
const int x8_idx = (k & 1);
const int y8_idx = (k >> 1);
@@ -847,6 +940,7 @@
}
}
}
+ return 0;
}
static void update_state(VP9_COMP *cpi, ThreadData *td,
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 8a7ae8e..c6bc6aa 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -694,9 +694,14 @@
int min_log2_tile_cols, max_log2_tile_cols;
vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
- cm->log2_tile_cols = clamp(cpi->oxcf.tile_columns,
- min_log2_tile_cols, max_log2_tile_cols);
- cm->log2_tile_rows = cpi->oxcf.tile_rows;
+ if (is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING) {
+ cm->log2_tile_cols = 0;
+ cm->log2_tile_rows = 0;
+ } else {
+ cm->log2_tile_cols = clamp(cpi->oxcf.tile_columns,
+ min_log2_tile_cols, max_log2_tile_cols);
+ cm->log2_tile_rows = cpi->oxcf.tile_rows;
+ }
}
static void init_buffer_indices(VP9_COMP *cpi) {
@@ -1612,6 +1617,8 @@
cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
#if CONFIG_INTERNAL_STATS
cpi->b_calculate_ssimg = 0;
+ cpi->b_calculate_blockiness = 1;
+
cpi->count = 0;
cpi->bytes = 0;
@@ -1644,6 +1651,23 @@
cpi->total_ssimg_v = 0;
cpi->total_ssimg_all = 0;
}
+ cpi->total_fastssim_y = 0;
+ cpi->total_fastssim_u = 0;
+ cpi->total_fastssim_v = 0;
+ cpi->total_fastssim_all = 0;
+
+ cpi->total_psnrhvs_y = 0;
+ cpi->total_psnrhvs_u = 0;
+ cpi->total_psnrhvs_v = 0;
+ cpi->total_psnrhvs_all = 0;
+
+ if (cpi->b_calculate_blockiness) {
+ cpi->total_blockiness = 0;
+ }
+
+ if (cpi->b_calculate_blockiness) {
+ cpi->total_blockiness = 0;
+ }
#endif
@@ -1852,7 +1876,6 @@
if (cpi && (cm->current_video_frame > 0)) {
#if CONFIG_INTERNAL_STATS
-
vp9_clear_system_state();
// printf("\n8x8-4x4:%d-%d\n", cpi->t8x8_count, cpi->t4x4_count);
@@ -1875,17 +1898,31 @@
(double)cpi->totalp_sq_error);
const double total_ssim = 100 * pow(cpi->summed_quality /
cpi->summed_weights, 8.0);
- const double totalp_ssim = 100 * pow(cpi->summedp_quality /
- cpi->summedp_weights, 8.0);
-
- fprintf(f, "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
- "VPXSSIM\tVPSSIMP\t Time(ms)\n");
- fprintf(f, "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t%8.0f\n",
- dr, cpi->total / cpi->count, total_psnr,
- cpi->totalp / cpi->count, totalp_psnr, total_ssim, totalp_ssim,
+ if (cpi->b_calculate_blockiness) {
+ fprintf(f, "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
+ "VPXSSIM\tVPSSIMP\tFASTSSIM\tPSNRHVS\tTime(ms)\n");
+ fprintf(f, "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
+ "%7.3f\t%7.3f\t%8.0f\n",
+ dr, cpi->total / cpi->count, total_psnr,
+ cpi->totalp / cpi->count, totalp_psnr, total_ssim,
+ cpi->total_fastssim_all / cpi->count,
+ cpi->total_psnrhvs_all / cpi->count,
total_encode_time);
+ } else {
+ fprintf(f, "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
+ "VPXSSIM\tVPSSIMP\tBlockiness\tFASTSSIM\tPSNRHVS\tTime(ms)\n");
+ fprintf(f, "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
+ "%7.3f\t%7.3f\t%7.3f\t%8.0f\n",
+ dr, cpi->total / cpi->count, total_psnr,
+ cpi->totalp / cpi->count, totalp_psnr, total_ssim,
+ cpi->total_blockiness / cpi->count,
+ cpi->total_fastssim_all / cpi->count,
+ cpi->total_psnrhvs_all / cpi->count,
+ total_encode_time);
+ }
}
+
if (cpi->b_calculate_ssimg) {
fprintf(f, "BitRate\tSSIM_Y\tSSIM_U\tSSIM_V\tSSIM_A\t Time(ms)\n");
fprintf(f, "%7.2f\t%6.4f\t%6.4f\t%6.4f\t%6.4f\t%8.0f\n", dr,
@@ -3802,6 +3839,12 @@
}
}
+#if CONFIG_INTERNAL_STATS
+extern double vp9_get_blockiness(const unsigned char *img1, int img1_pitch,
+ const unsigned char *img2, int img2_pitch,
+ int width, int height);
+#endif
+
int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
size_t *size, uint8_t *dest,
int64_t *time_stamp, int64_t *time_end, int flush) {
@@ -4151,7 +4194,12 @@
#endif
}
}
-
+ if (cpi->b_calculate_blockiness)
+ cpi->total_blockiness +=
+ vp9_get_blockiness(cpi->Source->y_buffer, cpi->Source->y_stride,
+ cm->frame_to_show->y_buffer,
+ cm->frame_to_show->y_stride,
+ cpi->Source->y_width, cpi->Source->y_height);
if (cpi->b_calculate_ssimg) {
double y, u, v, frame_all;
@@ -4171,6 +4219,26 @@
cpi->total_ssimg_v += v;
cpi->total_ssimg_all += frame_all;
}
+ {
+ double y, u, v, frame_all;
+ frame_all = vp9_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u,
+ &v);
+
+ cpi->total_fastssim_y += y;
+ cpi->total_fastssim_u += u;
+ cpi->total_fastssim_v += v;
+ cpi->total_fastssim_all += frame_all;
+ /* TODO(JBB): add 10/12 bit support */
+ }
+ {
+ double y, u, v, frame_all;
+ frame_all = vp9_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v);
+
+ cpi->total_psnrhvs_y += y;
+ cpi->total_psnrhvs_u += u;
+ cpi->total_psnrhvs_v += v;
+ cpi->total_psnrhvs_all += frame_all;
+ }
}
}
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h
index a5342ad..267c796 100644
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -402,6 +402,8 @@
uint64_t totalp_sq_error;
uint64_t totalp_samples;
+ double total_blockiness;
+
int bytes;
double summed_quality;
double summed_weights;
@@ -415,7 +417,18 @@
double total_ssimg_v;
double total_ssimg_all;
+ double total_fastssim_y;
+ double total_fastssim_u;
+ double total_fastssim_v;
+ double total_fastssim_all;
+
+ double total_psnrhvs_y;
+ double total_psnrhvs_u;
+ double total_psnrhvs_v;
+ double total_psnrhvs_all;
+
int b_calculate_ssimg;
+ int b_calculate_blockiness;
#endif
int b_calculate_psnr;
@@ -463,6 +476,8 @@
// 0 - threshold_64x64; 1 - threshold_32x32;
// 2 - threshold_16x16; 3 - vbp_threshold_8x8;
int64_t vbp_thresholds[4];
+ int64_t vbp_threshold_minmax;
+ int64_t vbp_threshold_sad;
BLOCK_SIZE vbp_bsize_min;
// Multi-threading
diff --git a/vp9/encoder/vp9_fastssim.c b/vp9/encoder/vp9_fastssim.c
new file mode 100644
index 0000000..f1d408c
--- /dev/null
+++ b/vp9/encoder/vp9_fastssim.c
@@ -0,0 +1,465 @@
+/*
+ * 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.
+ *
+ * This code was originally written by: Nathan E. Egge, at the Daala
+ * project.
+ */
+#include <math.h>
+#include <string.h>
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "vp9/encoder/vp9_ssim.h"
+/* TODO(jbb): High bit depth version of this code needed */
+typedef struct fs_level fs_level;
+typedef struct fs_ctx fs_ctx;
+
+#define SSIM_C1 (255 * 255 * 0.01 * 0.01)
+#define SSIM_C2 (255 * 255 * 0.03 * 0.03)
+
+#define FS_MINI(_a, _b) ((_a) < (_b) ? (_a) : (_b))
+#define FS_MAXI(_a, _b) ((_a) > (_b) ? (_a) : (_b))
+
+struct fs_level {
+ uint16_t *im1;
+ uint16_t *im2;
+ double *ssim;
+ int w;
+ int h;
+};
+
+struct fs_ctx {
+ fs_level *level;
+ int nlevels;
+ unsigned *col_buf;
+};
+
+static void fs_ctx_init(fs_ctx *_ctx, int _w, int _h, int _nlevels) {
+ unsigned char *data;
+ size_t data_size;
+ int lw;
+ int lh;
+ int l;
+ lw = (_w + 1) >> 1;
+ lh = (_h + 1) >> 1;
+ data_size = _nlevels * sizeof(fs_level)
+ + 2 * (lw + 8) * 8 * sizeof(*_ctx->col_buf);
+ for (l = 0; l < _nlevels; l++) {
+ size_t im_size;
+ size_t level_size;
+ im_size = lw * (size_t) lh;
+ level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
+ level_size += sizeof(*_ctx->level[l].ssim) - 1;
+ level_size /= sizeof(*_ctx->level[l].ssim);
+ level_size += im_size;
+ level_size *= sizeof(*_ctx->level[l].ssim);
+ data_size += level_size;
+ lw = (lw + 1) >> 1;
+ lh = (lh + 1) >> 1;
+ }
+ data = (unsigned char *) malloc(data_size);
+ _ctx->level = (fs_level *) data;
+ _ctx->nlevels = _nlevels;
+ data += _nlevels * sizeof(*_ctx->level);
+ lw = (_w + 1) >> 1;
+ lh = (_h + 1) >> 1;
+ for (l = 0; l < _nlevels; l++) {
+ size_t im_size;
+ size_t level_size;
+ _ctx->level[l].w = lw;
+ _ctx->level[l].h = lh;
+ im_size = lw * (size_t) lh;
+ level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
+ level_size += sizeof(*_ctx->level[l].ssim) - 1;
+ level_size /= sizeof(*_ctx->level[l].ssim);
+ level_size *= sizeof(*_ctx->level[l].ssim);
+ _ctx->level[l].im1 = (uint16_t *) data;
+ _ctx->level[l].im2 = _ctx->level[l].im1 + im_size;
+ data += level_size;
+ _ctx->level[l].ssim = (double *) data;
+ data += im_size * sizeof(*_ctx->level[l].ssim);
+ lw = (lw + 1) >> 1;
+ lh = (lh + 1) >> 1;
+ }
+ _ctx->col_buf = (unsigned *) data;
+}
+
+static void fs_ctx_clear(fs_ctx *_ctx) {
+ free(_ctx->level);
+}
+
+static void fs_downsample_level(fs_ctx *_ctx, int _l) {
+ const uint16_t *src1;
+ const uint16_t *src2;
+ uint16_t *dst1;
+ uint16_t *dst2;
+ int w2;
+ int h2;
+ int w;
+ int h;
+ int i;
+ int j;
+ w = _ctx->level[_l].w;
+ h = _ctx->level[_l].h;
+ dst1 = _ctx->level[_l].im1;
+ dst2 = _ctx->level[_l].im2;
+ w2 = _ctx->level[_l - 1].w;
+ h2 = _ctx->level[_l - 1].h;
+ src1 = _ctx->level[_l - 1].im1;
+ src2 = _ctx->level[_l - 1].im2;
+ for (j = 0; j < h; j++) {
+ int j0offs;
+ int j1offs;
+ j0offs = 2 * j * w2;
+ j1offs = FS_MINI(2 * j + 1, h2) * w2;
+ for (i = 0; i < w; i++) {
+ int i0;
+ int i1;
+ i0 = 2 * i;
+ i1 = FS_MINI(i0 + 1, w2);
+ dst1[j * w + i] = src1[j0offs + i0] + src1[j0offs + i1]
+ + src1[j1offs + i0] + src1[j1offs + i1];
+ dst2[j * w + i] = src2[j0offs + i0] + src2[j0offs + i1]
+ + src2[j1offs + i0] + src2[j1offs + i1];
+ }
+ }
+}
+
+static void fs_downsample_level0(fs_ctx *_ctx, const unsigned char *_src1,
+ int _s1ystride, const unsigned char *_src2,
+ int _s2ystride, int _w, int _h) {
+ uint16_t *dst1;
+ uint16_t *dst2;
+ int w;
+ int h;
+ int i;
+ int j;
+ w = _ctx->level[0].w;
+ h = _ctx->level[0].h;
+ dst1 = _ctx->level[0].im1;
+ dst2 = _ctx->level[0].im2;
+ for (j = 0; j < h; j++) {
+ int j0;
+ int j1;
+ j0 = 2 * j;
+ j1 = FS_MINI(j0 + 1, _h);
+ for (i = 0; i < w; i++) {
+ int i0;
+ int i1;
+ i0 = 2 * i;
+ i1 = FS_MINI(i0 + 1, _w);
+ dst1[j * w + i] = _src1[j0 * _s1ystride + i0]
+ + _src1[j0 * _s1ystride + i1] + _src1[j1 * _s1ystride + i0]
+ + _src1[j1 * _s1ystride + i1];
+ dst2[j * w + i] = _src2[j0 * _s2ystride + i0]
+ + _src2[j0 * _s2ystride + i1] + _src2[j1 * _s2ystride + i0]
+ + _src2[j1 * _s2ystride + i1];
+ }
+ }
+}
+
+static void fs_apply_luminance(fs_ctx *_ctx, int _l) {
+ unsigned *col_sums_x;
+ unsigned *col_sums_y;
+ uint16_t *im1;
+ uint16_t *im2;
+ double *ssim;
+ double c1;
+ int w;
+ int h;
+ int j0offs;
+ int j1offs;
+ int i;
+ int j;
+ w = _ctx->level[_l].w;
+ h = _ctx->level[_l].h;
+ col_sums_x = _ctx->col_buf;
+ col_sums_y = col_sums_x + w;
+ im1 = _ctx->level[_l].im1;
+ im2 = _ctx->level[_l].im2;
+ for (i = 0; i < w; i++)
+ col_sums_x[i] = 5 * im1[i];
+ for (i = 0; i < w; i++)
+ col_sums_y[i] = 5 * im2[i];
+ for (j = 1; j < 4; j++) {
+ j1offs = FS_MINI(j, h - 1) * w;
+ for (i = 0; i < w; i++)
+ col_sums_x[i] += im1[j1offs + i];
+ for (i = 0; i < w; i++)
+ col_sums_y[i] += im2[j1offs + i];
+ }
+ ssim = _ctx->level[_l].ssim;
+ c1 = (double) (SSIM_C1 * 4096 * (1 << 4 * _l));
+ for (j = 0; j < h; j++) {
+ unsigned mux;
+ unsigned muy;
+ int i0;
+ int i1;
+ mux = 5 * col_sums_x[0];
+ muy = 5 * col_sums_y[0];
+ for (i = 1; i < 4; i++) {
+ i1 = FS_MINI(i, w - 1);
+ mux += col_sums_x[i1];
+ muy += col_sums_y[i1];
+ }
+ for (i = 0; i < w; i++) {
+ ssim[j * w + i] *= (2 * mux * (double) muy + c1)
+ / (mux * (double) mux + muy * (double) muy + c1);
+ if (i + 1 < w) {
+ i0 = FS_MAXI(0, i - 4);
+ i1 = FS_MINI(i + 4, w - 1);
+ mux += col_sums_x[i1] - col_sums_x[i0];
+ muy += col_sums_x[i1] - col_sums_x[i0];
+ }
+ }
+ if (j + 1 < h) {
+ j0offs = FS_MAXI(0, j - 4) * w;
+ for (i = 0; i < w; i++)
+ col_sums_x[i] -= im1[j0offs + i];
+ for (i = 0; i < w; i++)
+ col_sums_y[i] -= im2[j0offs + i];
+ j1offs = FS_MINI(j + 4, h - 1) * w;
+ for (i = 0; i < w; i++)
+ col_sums_x[i] += im1[j1offs + i];
+ for (i = 0; i < w; i++)
+ col_sums_y[i] += im2[j1offs + i];
+ }
+ }
+}
+
+#define FS_COL_SET(_col, _joffs, _ioffs) \
+ do { \
+ unsigned gx; \
+ unsigned gy; \
+ gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+ gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+ col_sums_gx2[(_col)] = gx * (double)gx; \
+ col_sums_gy2[(_col)] = gy * (double)gy; \
+ col_sums_gxgy[(_col)] = gx * (double)gy; \
+ } \
+ while (0)
+
+#define FS_COL_ADD(_col, _joffs, _ioffs) \
+ do { \
+ unsigned gx; \
+ unsigned gy; \
+ gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+ gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+ col_sums_gx2[(_col)] += gx * (double)gx; \
+ col_sums_gy2[(_col)] += gy * (double)gy; \
+ col_sums_gxgy[(_col)] += gx * (double)gy; \
+ } \
+ while (0)
+
+#define FS_COL_SUB(_col, _joffs, _ioffs) \
+ do { \
+ unsigned gx; \
+ unsigned gy; \
+ gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+ gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+ col_sums_gx2[(_col)] -= gx * (double)gx; \
+ col_sums_gy2[(_col)] -= gy * (double)gy; \
+ col_sums_gxgy[(_col)] -= gx * (double)gy; \
+ } \
+ while (0)
+
+#define FS_COL_COPY(_col1, _col2) \
+ do { \
+ col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)]; \
+ col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)]; \
+ col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)]; \
+ } \
+ while (0)
+
+#define FS_COL_HALVE(_col1, _col2) \
+ do { \
+ col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 0.5; \
+ col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 0.5; \
+ col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 0.5; \
+ } \
+ while (0)
+
+#define FS_COL_DOUBLE(_col1, _col2) \
+ do { \
+ col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 2; \
+ col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 2; \
+ col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 2; \
+ } \
+ while (0)
+
+static void fs_calc_structure(fs_ctx *_ctx, int _l) {
+ uint16_t *im1;
+ uint16_t *im2;
+ unsigned *gx_buf;
+ unsigned *gy_buf;
+ double *ssim;
+ double col_sums_gx2[8];
+ double col_sums_gy2[8];
+ double col_sums_gxgy[8];
+ double c2;
+ int stride;
+ int w;
+ int h;
+ int i;
+ int j;
+ w = _ctx->level[_l].w;
+ h = _ctx->level[_l].h;
+ im1 = _ctx->level[_l].im1;
+ im2 = _ctx->level[_l].im2;
+ ssim = _ctx->level[_l].ssim;
+ gx_buf = _ctx->col_buf;
+ stride = w + 8;
+ gy_buf = gx_buf + 8 * stride;
+ memset(gx_buf, 0, 2 * 8 * stride * sizeof(*gx_buf));
+ c2 = SSIM_C2 * (1 << 4 * _l) * 16 * 104;
+ for (j = 0; j < h + 4; j++) {
+ if (j < h - 1) {
+ for (i = 0; i < w - 1; i++) {
+ unsigned g1;
+ unsigned g2;
+ unsigned gx;
+ unsigned gy;
+ g1 = abs(im1[(j + 1) * w + i + 1] - im1[j * w + i]);
+ g2 = abs(im1[(j + 1) * w + i] - im1[j * w + i + 1]);
+ gx = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
+ g1 = abs(im2[(j + 1) * w + i + 1] - im2[j * w + i]);
+ g2 = abs(im2[(j + 1) * w + i] - im2[j * w + i + 1]);
+ gy = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
+ gx_buf[(j & 7) * stride + i + 4] = gx;
+ gy_buf[(j & 7) * stride + i + 4] = gy;
+ }
+ } else {
+ memset(gx_buf + (j & 7) * stride, 0, stride * sizeof(*gx_buf));
+ memset(gy_buf + (j & 7) * stride, 0, stride * sizeof(*gy_buf));
+ }
+ if (j >= 4) {
+ int k;
+ col_sums_gx2[3] = col_sums_gx2[2] = col_sums_gx2[1] = col_sums_gx2[0] = 0;
+ col_sums_gy2[3] = col_sums_gy2[2] = col_sums_gy2[1] = col_sums_gy2[0] = 0;
+ col_sums_gxgy[3] = col_sums_gxgy[2] = col_sums_gxgy[1] =
+ col_sums_gxgy[0] = 0;
+ for (i = 4; i < 8; i++) {
+ FS_COL_SET(i, -1, 0);
+ FS_COL_ADD(i, 0, 0);
+ for (k = 1; k < 8 - i; k++) {
+ FS_COL_DOUBLE(i, i);
+ FS_COL_ADD(i, -k - 1, 0);
+ FS_COL_ADD(i, k, 0);
+ }
+ }
+ for (i = 0; i < w; i++) {
+ double mugx2;
+ double mugy2;
+ double mugxgy;
+ mugx2 = col_sums_gx2[0];
+ for (k = 1; k < 8; k++)
+ mugx2 += col_sums_gx2[k];
+ mugy2 = col_sums_gy2[0];
+ for (k = 1; k < 8; k++)
+ mugy2 += col_sums_gy2[k];
+ mugxgy = col_sums_gxgy[0];
+ for (k = 1; k < 8; k++)
+ mugxgy += col_sums_gxgy[k];
+ ssim[(j - 4) * w + i] = (2 * mugxgy + c2) / (mugx2 + mugy2 + c2);
+ if (i + 1 < w) {
+ FS_COL_SET(0, -1, 1);
+ FS_COL_ADD(0, 0, 1);
+ FS_COL_SUB(2, -3, 2);
+ FS_COL_SUB(2, 2, 2);
+ FS_COL_HALVE(1, 2);
+ FS_COL_SUB(3, -4, 3);
+ FS_COL_SUB(3, 3, 3);
+ FS_COL_HALVE(2, 3);
+ FS_COL_COPY(3, 4);
+ FS_COL_DOUBLE(4, 5);
+ FS_COL_ADD(4, -4, 5);
+ FS_COL_ADD(4, 3, 5);
+ FS_COL_DOUBLE(5, 6);
+ FS_COL_ADD(5, -3, 6);
+ FS_COL_ADD(5, 2, 6);
+ FS_COL_DOUBLE(6, 7);
+ FS_COL_ADD(6, -2, 7);
+ FS_COL_ADD(6, 1, 7);
+ FS_COL_SET(7, -1, 8);
+ FS_COL_ADD(7, 0, 8);
+ }
+ }
+ }
+ }
+}
+
+#define FS_NLEVELS (4)
+
+/*These weights were derived from the default weights found in Wang's original
+ Matlab implementation: {0.0448, 0.2856, 0.2363, 0.1333}.
+ We drop the finest scale and renormalize the rest to sum to 1.*/
+
+static const double FS_WEIGHTS[FS_NLEVELS] = {0.2989654541015625,
+ 0.3141326904296875, 0.2473602294921875, 0.1395416259765625};
+
+static double fs_average(fs_ctx *_ctx, int _l) {
+ double *ssim;
+ double ret;
+ int w;
+ int h;
+ int i;
+ int j;
+ w = _ctx->level[_l].w;
+ h = _ctx->level[_l].h;
+ ssim = _ctx->level[_l].ssim;
+ ret = 0;
+ for (j = 0; j < h; j++)
+ for (i = 0; i < w; i++)
+ ret += ssim[j * w + i];
+ return pow(ret / (w * h), FS_WEIGHTS[_l]);
+}
+
+static double calc_ssim(const unsigned char *_src, int _systride,
+ const unsigned char *_dst, int _dystride, int _w, int _h) {
+ fs_ctx ctx;
+ double ret;
+ int l;
+ ret = 1;
+ fs_ctx_init(&ctx, _w, _h, FS_NLEVELS);
+ fs_downsample_level0(&ctx, _src, _systride, _dst, _dystride, _w, _h);
+ for (l = 0; l < FS_NLEVELS - 1; l++) {
+ fs_calc_structure(&ctx, l);
+ ret *= fs_average(&ctx, l);
+ fs_downsample_level(&ctx, l + 1);
+ }
+ fs_calc_structure(&ctx, l);
+ fs_apply_luminance(&ctx, l);
+ ret *= fs_average(&ctx, l);
+ fs_ctx_clear(&ctx);
+ return ret;
+}
+
+static double convert_ssim_db(double _ssim, double _weight) {
+ return 10 * (log10(_weight) - log10(_weight - _ssim));
+}
+
+double vp9_calc_fastssim(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
+ double *ssim_y, double *ssim_u, double *ssim_v) {
+ double ssimv;
+ vp9_clear_system_state();
+
+ *ssim_y = calc_ssim(source->y_buffer, source->y_stride, dest->y_buffer,
+ dest->y_stride, source->y_crop_width,
+ source->y_crop_height);
+
+ *ssim_u = calc_ssim(source->u_buffer, source->uv_stride, dest->u_buffer,
+ dest->uv_stride, source->uv_crop_width,
+ source->uv_crop_height);
+
+ *ssim_v = calc_ssim(source->v_buffer, source->uv_stride, dest->v_buffer,
+ dest->uv_stride, source->uv_crop_width,
+ source->uv_crop_height);
+ ssimv = (*ssim_y) * .8 + .1 * ((*ssim_u) + (*ssim_v));
+
+ return convert_ssim_db(ssimv, 1.0);
+}
diff --git a/vp9/encoder/vp9_psnrhvs.c b/vp9/encoder/vp9_psnrhvs.c
new file mode 100644
index 0000000..6c034aa
--- /dev/null
+++ b/vp9/encoder/vp9_psnrhvs.c
@@ -0,0 +1,227 @@
+/*
+ * 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.
+ *
+ * This code was originally written by: Gregory Maxwell, at the Daala
+ * project.
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <math.h>
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "vp9/encoder/vp9_ssim.h"
+
+#if !defined(M_PI)
+# define M_PI (3.141592653589793238462643)
+#endif
+#include <string.h>
+
+typedef int16_t od_coeff;
+typedef int16_t tran_low_t;
+extern void vp9_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+
+void od_bin_fdct8x8(od_coeff *y, int ystride, const od_coeff *x, int xstride) {
+ (void) xstride;
+ vp9_fdct8x8_c(x, y, ystride);
+}
+
+/* Normalized inverse quantization matrix for 8x8 DCT at the point of
+ * transparency. This is not the JPEG based matrix from the paper,
+ this one gives a slightly higher MOS agreement.*/
+float csf_y[8][8] = {{1.6193873005, 2.2901594831, 2.08509755623, 1.48366094411,
+ 1.00227514334, 0.678296995242, 0.466224900598, 0.3265091542}, {2.2901594831,
+ 1.94321815382, 2.04793073064, 1.68731108984, 1.2305666963, 0.868920337363,
+ 0.61280991668, 0.436405793551}, {2.08509755623, 2.04793073064,
+ 1.34329019223, 1.09205635862, 0.875748795257, 0.670882927016,
+ 0.501731932449, 0.372504254596}, {1.48366094411, 1.68731108984,
+ 1.09205635862, 0.772819797575, 0.605636379554, 0.48309405692,
+ 0.380429446972, 0.295774038565}, {1.00227514334, 1.2305666963,
+ 0.875748795257, 0.605636379554, 0.448996256676, 0.352889268808,
+ 0.283006984131, 0.226951348204}, {0.678296995242, 0.868920337363,
+ 0.670882927016, 0.48309405692, 0.352889268808, 0.27032073436,
+ 0.215017739696, 0.17408067321}, {0.466224900598, 0.61280991668,
+ 0.501731932449, 0.380429446972, 0.283006984131, 0.215017739696,
+ 0.168869545842, 0.136153931001}, {0.3265091542, 0.436405793551,
+ 0.372504254596, 0.295774038565, 0.226951348204, 0.17408067321,
+ 0.136153931001, 0.109083846276}};
+float csf_cb420[8][8] = {
+ {1.91113096927, 2.46074210438, 1.18284184739, 1.14982565193, 1.05017074788,
+ 0.898018824055, 0.74725392039, 0.615105596242}, {2.46074210438,
+ 1.58529308355, 1.21363250036, 1.38190029285, 1.33100189972,
+ 1.17428548929, 0.996404342439, 0.830890433625}, {1.18284184739,
+ 1.21363250036, 0.978712413627, 1.02624506078, 1.03145147362,
+ 0.960060382087, 0.849823426169, 0.731221236837}, {1.14982565193,
+ 1.38190029285, 1.02624506078, 0.861317501629, 0.801821139099,
+ 0.751437590932, 0.685398513368, 0.608694761374}, {1.05017074788,
+ 1.33100189972, 1.03145147362, 0.801821139099, 0.676555426187,
+ 0.605503172737, 0.55002013668, 0.495804539034}, {0.898018824055,
+ 1.17428548929, 0.960060382087, 0.751437590932, 0.605503172737,
+ 0.514674450957, 0.454353482512, 0.407050308965}, {0.74725392039,
+ 0.996404342439, 0.849823426169, 0.685398513368, 0.55002013668,
+ 0.454353482512, 0.389234902883, 0.342353999733}, {0.615105596242,
+ 0.830890433625, 0.731221236837, 0.608694761374, 0.495804539034,
+ 0.407050308965, 0.342353999733, 0.295530605237}};
+float csf_cr420[8][8] = {
+ {2.03871978502, 2.62502345193, 1.26180942886, 1.11019789803, 1.01397751469,
+ 0.867069376285, 0.721500455585, 0.593906509971}, {2.62502345193,
+ 1.69112867013, 1.17180569821, 1.3342742857, 1.28513006198,
+ 1.13381474809, 0.962064122248, 0.802254508198}, {1.26180942886,
+ 1.17180569821, 0.944981930573, 0.990876405848, 0.995903384143,
+ 0.926972725286, 0.820534991409, 0.706020324706}, {1.11019789803,
+ 1.3342742857, 0.990876405848, 0.831632933426, 0.77418706195,
+ 0.725539939514, 0.661776842059, 0.587716619023}, {1.01397751469,
+ 1.28513006198, 0.995903384143, 0.77418706195, 0.653238524286,
+ 0.584635025748, 0.531064164893, 0.478717061273}, {0.867069376285,
+ 1.13381474809, 0.926972725286, 0.725539939514, 0.584635025748,
+ 0.496936637883, 0.438694579826, 0.393021669543}, {0.721500455585,
+ 0.962064122248, 0.820534991409, 0.661776842059, 0.531064164893,
+ 0.438694579826, 0.375820256136, 0.330555063063}, {0.593906509971,
+ 0.802254508198, 0.706020324706, 0.587716619023, 0.478717061273,
+ 0.393021669543, 0.330555063063, 0.285345396658}};
+
+static double convert_score_db(double _score, double _weight) {
+ return 10 * (log10(255 * 255) - log10(_weight * _score));
+}
+
+static double calc_psnrhvs(const unsigned char *_src, int _systride,
+ const unsigned char *_dst, int _dystride,
+ double _par, int _w, int _h, int _step,
+ float _csf[8][8]) {
+ float ret;
+ od_coeff dct_s[8 * 8];
+ od_coeff dct_d[8 * 8];
+ float mask[8][8];
+ int pixels;
+ int x;
+ int y;
+ (void) _par;
+ ret = pixels = 0;
+ /*In the PSNR-HVS-M paper[1] the authors describe the construction of
+ their masking table as "we have used the quantization table for the
+ color component Y of JPEG [6] that has been also obtained on the
+ basis of CSF. Note that the values in quantization table JPEG have
+ been normalized and then squared." Their CSF matrix (from PSNR-HVS)
+ was also constructed from the JPEG matrices. I can not find any obvious
+ scheme of normalizing to produce their table, but if I multiply their
+ CSF by 0.38857 and square the result I get their masking table.
+ I have no idea where this constant comes from, but deviating from it
+ too greatly hurts MOS agreement.
+
+ [1] Nikolay Ponomarenko, Flavia Silvestri, Karen Egiazarian, Marco Carli,
+ Jaakko Astola, Vladimir Lukin, "On between-coefficient contrast masking
+ of DCT basis functions", CD-ROM Proceedings of the Third
+ International Workshop on Video Processing and Quality Metrics for Consumer
+ Electronics VPQM-07, Scottsdale, Arizona, USA, 25-26 January, 2007, 4 p.*/
+ for (x = 0; x < 8; x++)
+ for (y = 0; y < 8; y++)
+ mask[x][y] = (_csf[x][y] * 0.3885746225901003)
+ * (_csf[x][y] * 0.3885746225901003);
+ for (y = 0; y < _h - 7; y += _step) {
+ for (x = 0; x < _w - 7; x += _step) {
+ int i;
+ int j;
+ float s_means[4];
+ float d_means[4];
+ float s_vars[4];
+ float d_vars[4];
+ float s_gmean = 0;
+ float d_gmean = 0;
+ float s_gvar = 0;
+ float d_gvar = 0;
+ float s_mask = 0;
+ float d_mask = 0;
+ for (i = 0; i < 4; i++)
+ s_means[i] = d_means[i] = s_vars[i] = d_vars[i] = 0;
+ for (i = 0; i < 8; i++) {
+ for (j = 0; j < 8; j++) {
+ int sub = ((i & 12) >> 2) + ((j & 12) >> 1);
+ dct_s[i * 8 + j] = _src[(y + i) * _systride + (j + x)];
+ dct_d[i * 8 + j] = _dst[(y + i) * _dystride + (j + x)];
+ s_gmean += dct_s[i * 8 + j];
+ d_gmean += dct_d[i * 8 + j];
+ s_means[sub] += dct_s[i * 8 + j];
+ d_means[sub] += dct_d[i * 8 + j];
+ }
+ }
+ s_gmean /= 64.f;
+ d_gmean /= 64.f;
+ for (i = 0; i < 4; i++)
+ s_means[i] /= 16.f;
+ for (i = 0; i < 4; i++)
+ d_means[i] /= 16.f;
+ for (i = 0; i < 8; i++) {
+ for (j = 0; j < 8; j++) {
+ int sub = ((i & 12) >> 2) + ((j & 12) >> 1);
+ s_gvar += (dct_s[i * 8 + j] - s_gmean) * (dct_s[i * 8 + j] - s_gmean);
+ d_gvar += (dct_d[i * 8 + j] - d_gmean) * (dct_d[i * 8 + j] - d_gmean);
+ s_vars[sub] += (dct_s[i * 8 + j] - s_means[sub])
+ * (dct_s[i * 8 + j] - s_means[sub]);
+ d_vars[sub] += (dct_d[i * 8 + j] - d_means[sub])
+ * (dct_d[i * 8 + j] - d_means[sub]);
+ }
+ }
+ s_gvar *= 1 / 63.f * 64;
+ d_gvar *= 1 / 63.f * 64;
+ for (i = 0; i < 4; i++)
+ s_vars[i] *= 1 / 15.f * 16;
+ for (i = 0; i < 4; i++)
+ d_vars[i] *= 1 / 15.f * 16;
+ if (s_gvar > 0)
+ s_gvar = (s_vars[0] + s_vars[1] + s_vars[2] + s_vars[3]) / s_gvar;
+ if (d_gvar > 0)
+ d_gvar = (d_vars[0] + d_vars[1] + d_vars[2] + d_vars[3]) / d_gvar;
+ od_bin_fdct8x8(dct_s, 8, dct_s, 8);
+ od_bin_fdct8x8(dct_d, 8, dct_d, 8);
+ for (i = 0; i < 8; i++)
+ for (j = (i == 0); j < 8; j++)
+ s_mask += dct_s[i * 8 + j] * dct_s[i * 8 + j] * mask[i][j];
+ for (i = 0; i < 8; i++)
+ for (j = (i == 0); j < 8; j++)
+ d_mask += dct_d[i * 8 + j] * dct_d[i * 8 + j] * mask[i][j];
+ s_mask = sqrt(s_mask * s_gvar) / 32.f;
+ d_mask = sqrt(d_mask * d_gvar) / 32.f;
+ if (d_mask > s_mask)
+ s_mask = d_mask;
+ for (i = 0; i < 8; i++) {
+ for (j = 0; j < 8; j++) {
+ float err;
+ err = fabs(dct_s[i * 8 + j] - dct_d[i * 8 + j]);
+ if (i != 0 || j != 0)
+ err = err < s_mask / mask[i][j] ? 0 : err - s_mask / mask[i][j];
+ ret += (err * _csf[i][j]) * (err * _csf[i][j]);
+ pixels++;
+ }
+ }
+ }
+ }
+ ret /= pixels;
+ return ret;
+}
+double vp9_psnrhvs(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
+ double *y_psnrhvs, double *u_psnrhvs, double *v_psnrhvs) {
+ double psnrhvs;
+ double par = 1.0;
+ int step = 7;
+ vp9_clear_system_state();
+ *y_psnrhvs = calc_psnrhvs(source->y_buffer, source->y_stride, dest->y_buffer,
+ dest->y_stride, par, source->y_crop_width,
+ source->y_crop_height, step, csf_y);
+
+ *u_psnrhvs = calc_psnrhvs(source->u_buffer, source->uv_stride, dest->u_buffer,
+ dest->uv_stride, par, source->uv_crop_width,
+ source->uv_crop_height, step, csf_cb420);
+
+ *v_psnrhvs = calc_psnrhvs(source->v_buffer, source->uv_stride, dest->v_buffer,
+ dest->uv_stride, par, source->uv_crop_width,
+ source->uv_crop_height, step, csf_cr420);
+ psnrhvs = (*y_psnrhvs) * .8 + .1 * ((*u_psnrhvs) + (*v_psnrhvs));
+
+ return convert_score_db(psnrhvs, 1.0);
+}
diff --git a/vp9/encoder/vp9_ssim.h b/vp9/encoder/vp9_ssim.h
index e75623b..ed1bb83 100644
--- a/vp9/encoder/vp9_ssim.h
+++ b/vp9/encoder/vp9_ssim.h
@@ -23,6 +23,12 @@
double vp9_calc_ssimg(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
double *ssim_y, double *ssim_u, double *ssim_v);
+double vp9_calc_fastssim(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
+ double *ssim_y, double *ssim_u, double *ssim_v);
+
+double vp9_psnrhvs(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
+ double *ssim_y, double *ssim_u, double *ssim_v);
+
#if CONFIG_VP9_HIGHBITDEPTH
double vp9_highbd_calc_ssim(YV12_BUFFER_CONFIG *source,
YV12_BUFFER_CONFIG *dest,
diff --git a/vp9/encoder/vp9_tokenize.c b/vp9/encoder/vp9_tokenize.c
index 799109b..9d2595b 100644
--- a/vp9/encoder/vp9_tokenize.c
+++ b/vp9/encoder/vp9_tokenize.c
@@ -613,7 +613,6 @@
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
- TOKENEXTRA *t_backup = *t;
const int ctx = vp9_get_skip_context(xd);
const int skip_inc = !vp9_segfeature_active(&cm->seg, mbmi->segment_id,
SEG_LVL_SKIP);
@@ -622,8 +621,6 @@
if (!dry_run)
td->counts->skip[ctx][1] += skip_inc;
reset_skip_context(xd, bsize);
- if (dry_run)
- *t = t_backup;
return;
}
@@ -632,6 +629,5 @@
vp9_foreach_transformed_block(xd, bsize, tokenize_b, &arg);
} else {
vp9_foreach_transformed_block(xd, bsize, set_entropy_context_b, &arg);
- *t = t_backup;
}
}
diff --git a/vp9/encoder/x86/vp9_avg_intrin_sse2.c b/vp9/encoder/x86/vp9_avg_intrin_sse2.c
index ecd6ce9..4672aa6 100644
--- a/vp9/encoder/x86/vp9_avg_intrin_sse2.c
+++ b/vp9/encoder/x86/vp9_avg_intrin_sse2.c
@@ -11,6 +11,83 @@
#include <emmintrin.h>
#include "vpx_ports/mem.h"
+void vp9_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp,
+ int *min, int *max) {
+ __m128i u0, s0, d0, diff, maxabsdiff, minabsdiff, negdiff, absdiff0, absdiff;
+ u0 = _mm_setzero_si128();
+ // Row 0
+ s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+ d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d)), u0);
+ diff = _mm_subs_epi16(s0, d0);
+ negdiff = _mm_subs_epi16(u0, diff);
+ absdiff0 = _mm_max_epi16(diff, negdiff);
+ // Row 1
+ s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+ d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + dp)), u0);
+ diff = _mm_subs_epi16(s0, d0);
+ negdiff = _mm_subs_epi16(u0, diff);
+ absdiff = _mm_max_epi16(diff, negdiff);
+ maxabsdiff = _mm_max_epi16(absdiff0, absdiff);
+ minabsdiff = _mm_min_epi16(absdiff0, absdiff);
+ // Row 2
+ s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+ d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 2 * dp)), u0);
+ diff = _mm_subs_epi16(s0, d0);
+ negdiff = _mm_subs_epi16(u0, diff);
+ absdiff = _mm_max_epi16(diff, negdiff);
+ maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+ minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+ // Row 3
+ s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+ d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 3 * dp)), u0);
+ diff = _mm_subs_epi16(s0, d0);
+ negdiff = _mm_subs_epi16(u0, diff);
+ absdiff = _mm_max_epi16(diff, negdiff);
+ maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+ minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+ // Row 4
+ s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0);
+ d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 4 * dp)), u0);
+ diff = _mm_subs_epi16(s0, d0);
+ negdiff = _mm_subs_epi16(u0, diff);
+ absdiff = _mm_max_epi16(diff, negdiff);
+ maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+ minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+ // Row 5
+ s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0);
+ d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 5 * dp)), u0);
+ diff = _mm_subs_epi16(s0, d0);
+ negdiff = _mm_subs_epi16(u0, diff);
+ absdiff = _mm_max_epi16(diff, negdiff);
+ maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+ minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+ // Row 6
+ s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0);
+ d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 6 * dp)), u0);
+ diff = _mm_subs_epi16(s0, d0);
+ negdiff = _mm_subs_epi16(u0, diff);
+ absdiff = _mm_max_epi16(diff, negdiff);
+ maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+ minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+ // Row 7
+ s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0);
+ d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 7 * dp)), u0);
+ diff = _mm_subs_epi16(s0, d0);
+ negdiff = _mm_subs_epi16(u0, diff);
+ absdiff = _mm_max_epi16(diff, negdiff);
+ maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+ minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+
+ maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_si128(maxabsdiff, 8));
+ maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 32));
+ maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 16));
+ *max = _mm_extract_epi16(maxabsdiff, 0);
+
+ minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_si128(minabsdiff, 8));
+ minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 32));
+ minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 16));
+ *min = _mm_extract_epi16(minabsdiff, 0);
+}
unsigned int vp9_avg_8x8_sse2(const uint8_t *s, int p) {
__m128i s0, s1, u0;
diff --git a/vp9/vp9cx.mk b/vp9/vp9cx.mk
index f8da734..f5f9be8 100644
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -34,6 +34,7 @@
VP9_CX_SRCS-yes += encoder/vp9_ethread.h
VP9_CX_SRCS-yes += encoder/vp9_ethread.c
VP9_CX_SRCS-yes += encoder/vp9_extend.c
+VP9_CX_SRCS-$(CONFIG_INTERNAL_STATS) += encoder/vp9_fastssim.c
VP9_CX_SRCS-yes += encoder/vp9_firstpass.c
VP9_CX_SRCS-yes += encoder/vp9_block.h
VP9_CX_SRCS-yes += encoder/vp9_writer.h
@@ -62,6 +63,7 @@
VP9_CX_SRCS-yes += encoder/vp9_encoder.c
VP9_CX_SRCS-yes += encoder/vp9_picklpf.c
VP9_CX_SRCS-yes += encoder/vp9_picklpf.h
+VP9_CX_SRCS-$(CONFIG_INTERNAL_STATS) += encoder/vp9_psnrhvs.c
VP9_CX_SRCS-yes += encoder/vp9_quantize.c
VP9_CX_SRCS-yes += encoder/vp9_ratectrl.c
VP9_CX_SRCS-yes += encoder/vp9_rd.c
@@ -79,6 +81,8 @@
VP9_CX_SRCS-yes += encoder/vp9_resize.h
VP9_CX_SRCS-$(CONFIG_INTERNAL_STATS) += encoder/vp9_ssim.c
VP9_CX_SRCS-$(CONFIG_INTERNAL_STATS) += encoder/vp9_ssim.h
+VP9_CX_SRCS-$(CONFIG_INTERNAL_STATS) += encoder/vp9_blockiness.c
+
VP9_CX_SRCS-yes += encoder/vp9_tokenize.c
VP9_CX_SRCS-yes += encoder/vp9_treewriter.c
VP9_CX_SRCS-yes += encoder/vp9_variance.c
