AVX2 optimization of motion compensation functions
AVX2 implementation of av1_highbd_jnt_convolve_2d_copy_c,
av1_convolve_2d_copy_sr_c, av1_highbd_convolve_2d_copy_sr_c
and av1_highbd_jnt_convolve_2d_c have been added.
1) av1_highbd_jnt_convolve_2d_c - ~ 19.3 times faster than
C implementation.
2) av1_highbd_jnt_convolve_2d_copy_c - ~ 8.3 times faster
than C implementation.
3) av1_convolve_2d_copy_sr_c - ~ 1.3 times faster than
SSE2. The gains are higher for larger block sizes.
4) av1_highbd_convolve_2d_copy_sr_c - ~ 4.1 times faster
than C implementation.
Change-Id: Ia8fc9be66d79e5d6e99cc125e523a7821da65014
diff --git a/av1/av1.cmake b/av1/av1.cmake
index a5795aa..64ea170 100644
--- a/av1/av1.cmake
+++ b/av1/av1.cmake
@@ -271,6 +271,9 @@
set(AOM_AV1_COMMON_INTRIN_SSE4_1
${AOM_AV1_COMMON_INTRIN_SSE4_1}
"${AOM_ROOT}/av1/common/x86/highbd_convolve_2d_sse4.c")
+ set(AOM_AV1_COMMON_INTRIN_AVX2
+ ${AOM_AV1_COMMON_INTRIN_AVX2}
+ "${AOM_ROOT}/av1/common/x86/highbd_jnt_convolve_avx2.c")
endif ()
if (CONFIG_JNT_COMP)
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl
index c752948..5f848f4 100755
--- a/av1/common/av1_rtcd_defs.pl
+++ b/av1/common/av1_rtcd_defs.pl
@@ -487,7 +487,7 @@
add_proto qw/void av1_convolve_2d_copy/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params";
specialize qw/av1_convolve_2d_copy sse2/;
add_proto qw/void av1_convolve_2d_copy_sr/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params";
-specialize qw/av1_convolve_2d_copy_sr sse2/;
+specialize qw/av1_convolve_2d_copy_sr sse2 avx2/;
add_proto qw/void av1_convolve_x/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params";
specialize qw/av1_convolve_x sse2 avx2/;
add_proto qw/void av1_convolve_y/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params";
@@ -520,6 +520,7 @@
add_proto qw/void av1_highbd_convolve_2d_copy/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd";
add_proto qw/void av1_highbd_convolve_2d_copy_sr/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd";
+specialize qw/av1_highbd_convolve_2d_copy_sr avx2/;
add_proto qw/void av1_highbd_convolve_2d_sr/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd";
specialize qw/av1_highbd_convolve_2d_sr c avx2/;
@@ -542,13 +543,14 @@
if (aom_config("CONFIG_JNT_COMP") eq "yes") {
add_proto qw/void av1_highbd_jnt_convolve_2d/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd";
- specialize qw/av1_highbd_jnt_convolve_2d sse4_1/;
+ specialize qw/av1_highbd_jnt_convolve_2d sse4_1 avx2/;
add_proto qw/void av1_highbd_jnt_convolve_x/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd";
add_proto qw/void av1_highbd_jnt_convolve_y/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd";
add_proto qw/void av1_highbd_jnt_convolve_2d_copy/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd";
+ specialize qw/av1_highbd_jnt_convolve_2d_copy avx2/;
}
# INTRA_EDGE functions
diff --git a/av1/common/x86/convolve_2d_avx2.c b/av1/common/x86/convolve_2d_avx2.c
index 854a18b..20e78c1 100644
--- a/av1/common/x86/convolve_2d_avx2.c
+++ b/av1/common/x86/convolve_2d_avx2.c
@@ -293,3 +293,125 @@
}
}
}
+
+static INLINE void copy_128(const uint8_t *src, uint8_t *dst) {
+ __m256i s[4];
+ s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 32));
+ s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 32));
+ s[2] = _mm256_loadu_si256((__m256i *)(src + 2 * 32));
+ s[3] = _mm256_loadu_si256((__m256i *)(src + 3 * 32));
+ _mm256_storeu_si256((__m256i *)(dst + 0 * 32), s[0]);
+ _mm256_storeu_si256((__m256i *)(dst + 1 * 32), s[1]);
+ _mm256_storeu_si256((__m256i *)(dst + 2 * 32), s[2]);
+ _mm256_storeu_si256((__m256i *)(dst + 3 * 32), s[3]);
+}
+
+void av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride,
+ uint8_t *dst, int dst_stride, int w, int h,
+ InterpFilterParams *filter_params_x,
+ InterpFilterParams *filter_params_y,
+ const int subpel_x_q4, const int subpel_y_q4,
+ ConvolveParams *conv_params) {
+ (void)filter_params_x;
+ (void)filter_params_y;
+ (void)subpel_x_q4;
+ (void)subpel_y_q4;
+ (void)conv_params;
+
+ if (w >= 16) {
+ assert(!((intptr_t)dst % 16));
+ assert(!(dst_stride % 16));
+ }
+
+ if (w == 2) {
+ do {
+ __m128i s = _mm_loadl_epi64((__m128i *)src);
+ *(uint16_t *)dst = _mm_cvtsi128_si32(s);
+ src += src_stride;
+ dst += dst_stride;
+ s = _mm_loadl_epi64((__m128i *)src);
+ *(uint16_t *)dst = _mm_cvtsi128_si32(s);
+ src += src_stride;
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else if (w == 4) {
+ do {
+ __m128i s = _mm_loadl_epi64((__m128i *)src);
+ *(uint32_t *)dst = _mm_cvtsi128_si32(s);
+ src += src_stride;
+ dst += dst_stride;
+ s = _mm_loadl_epi64((__m128i *)src);
+ *(uint32_t *)dst = _mm_cvtsi128_si32(s);
+ src += src_stride;
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else if (w == 8) {
+ do {
+ __m128i s[2];
+ s[0] = _mm_loadl_epi64((__m128i *)src);
+ src += src_stride;
+ s[1] = _mm_loadl_epi64((__m128i *)src);
+ src += src_stride;
+ _mm_storel_epi64((__m128i *)dst, s[0]);
+ dst += dst_stride;
+ _mm_storel_epi64((__m128i *)dst, s[1]);
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else if (w == 16) {
+ do {
+ __m128i s[2];
+ s[0] = _mm_loadu_si128((__m128i *)src);
+ src += src_stride;
+ s[1] = _mm_loadu_si128((__m128i *)src);
+ src += src_stride;
+ _mm_store_si128((__m128i *)dst, s[0]);
+ dst += dst_stride;
+ _mm_store_si128((__m128i *)dst, s[1]);
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else if (w == 32) {
+ do {
+ __m256i s[2];
+ s[0] = _mm256_loadu_si256((__m256i *)src);
+ src += src_stride;
+ s[1] = _mm256_loadu_si256((__m256i *)src);
+ src += src_stride;
+ _mm256_storeu_si256((__m256i *)dst, s[0]);
+ dst += dst_stride;
+ _mm256_storeu_si256((__m256i *)dst, s[1]);
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else if (w == 64) {
+ do {
+ __m256i s[4];
+ s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 32));
+ s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 32));
+ src += src_stride;
+ s[2] = _mm256_loadu_si256((__m256i *)(src + 0 * 32));
+ s[3] = _mm256_loadu_si256((__m256i *)(src + 1 * 32));
+ src += src_stride;
+ _mm256_storeu_si256((__m256i *)(dst + 0 * 32), s[0]);
+ _mm256_storeu_si256((__m256i *)(dst + 1 * 32), s[1]);
+ dst += dst_stride;
+ _mm256_storeu_si256((__m256i *)(dst + 0 * 32), s[2]);
+ _mm256_storeu_si256((__m256i *)(dst + 1 * 32), s[3]);
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else {
+ do {
+ copy_128(src, dst);
+ src += src_stride;
+ dst += dst_stride;
+ copy_128(src, dst);
+ src += src_stride;
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ }
+}
diff --git a/av1/common/x86/highbd_convolve_2d_avx2.c b/av1/common/x86/highbd_convolve_2d_avx2.c
index 750d4a8..9c84e74 100644
--- a/av1/common/x86/highbd_convolve_2d_avx2.c
+++ b/av1/common/x86/highbd_convolve_2d_avx2.c
@@ -416,3 +416,144 @@
}
}
}
+
+static INLINE void copy_64(const uint16_t *src, uint16_t *dst) {
+ __m256i s[4];
+ s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 16));
+ s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 16));
+ s[2] = _mm256_loadu_si256((__m256i *)(src + 2 * 16));
+ s[3] = _mm256_loadu_si256((__m256i *)(src + 3 * 16));
+ _mm256_storeu_si256((__m256i *)(dst + 0 * 16), s[0]);
+ _mm256_storeu_si256((__m256i *)(dst + 1 * 16), s[1]);
+ _mm256_storeu_si256((__m256i *)(dst + 2 * 16), s[2]);
+ _mm256_storeu_si256((__m256i *)(dst + 3 * 16), s[3]);
+}
+
+static INLINE void copy_128(const uint16_t *src, uint16_t *dst) {
+ __m256i s[8];
+ s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 16));
+ s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 16));
+ s[2] = _mm256_loadu_si256((__m256i *)(src + 2 * 16));
+ s[3] = _mm256_loadu_si256((__m256i *)(src + 3 * 16));
+ s[4] = _mm256_loadu_si256((__m256i *)(src + 4 * 16));
+ s[5] = _mm256_loadu_si256((__m256i *)(src + 5 * 16));
+ s[6] = _mm256_loadu_si256((__m256i *)(src + 6 * 16));
+ s[7] = _mm256_loadu_si256((__m256i *)(src + 7 * 16));
+
+ _mm256_storeu_si256((__m256i *)(dst + 0 * 16), s[0]);
+ _mm256_storeu_si256((__m256i *)(dst + 1 * 16), s[1]);
+ _mm256_storeu_si256((__m256i *)(dst + 2 * 16), s[2]);
+ _mm256_storeu_si256((__m256i *)(dst + 3 * 16), s[3]);
+ _mm256_storeu_si256((__m256i *)(dst + 4 * 16), s[4]);
+ _mm256_storeu_si256((__m256i *)(dst + 5 * 16), s[5]);
+ _mm256_storeu_si256((__m256i *)(dst + 6 * 16), s[6]);
+ _mm256_storeu_si256((__m256i *)(dst + 7 * 16), s[7]);
+}
+
+void av1_highbd_convolve_2d_copy_sr_avx2(
+ const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w,
+ int h, InterpFilterParams *filter_params_x,
+ InterpFilterParams *filter_params_y, const int subpel_x_q4,
+ const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+ (void)filter_params_x;
+ (void)filter_params_y;
+ (void)subpel_x_q4;
+ (void)subpel_y_q4;
+ (void)conv_params;
+ (void)bd;
+
+ if (w >= 16) {
+ assert(!((intptr_t)dst % 16));
+ assert(!(dst_stride % 16));
+ }
+
+ if (w == 2) {
+ do {
+ __m128i s = _mm_loadl_epi64((__m128i *)src);
+ *(uint32_t *)dst = _mm_cvtsi128_si32(s);
+ src += src_stride;
+ dst += dst_stride;
+ s = _mm_loadl_epi64((__m128i *)src);
+ *(uint32_t *)dst = _mm_cvtsi128_si32(s);
+ src += src_stride;
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else if (w == 4) {
+ do {
+ __m128i s[2];
+ s[0] = _mm_loadl_epi64((__m128i *)src);
+ src += src_stride;
+ s[1] = _mm_loadl_epi64((__m128i *)src);
+ src += src_stride;
+ _mm_storel_epi64((__m128i *)dst, s[0]);
+ dst += dst_stride;
+ _mm_storel_epi64((__m128i *)dst, s[1]);
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else if (w == 8) {
+ do {
+ __m128i s[2];
+ s[0] = _mm_loadu_si128((__m128i *)src);
+ src += src_stride;
+ s[1] = _mm_loadu_si128((__m128i *)src);
+ src += src_stride;
+ _mm_store_si128((__m128i *)dst, s[0]);
+ dst += dst_stride;
+ _mm_store_si128((__m128i *)dst, s[1]);
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else if (w == 16) {
+ do {
+ __m256i s[2];
+ s[0] = _mm256_loadu_si256((__m256i *)src);
+ src += src_stride;
+ s[1] = _mm256_loadu_si256((__m256i *)src);
+ src += src_stride;
+ _mm256_storeu_si256((__m256i *)dst, s[0]);
+ dst += dst_stride;
+ _mm256_storeu_si256((__m256i *)dst, s[1]);
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else if (w == 32) {
+ do {
+ __m256i s[4];
+ s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 16));
+ s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 16));
+ src += src_stride;
+ s[2] = _mm256_loadu_si256((__m256i *)(src + 0 * 16));
+ s[3] = _mm256_loadu_si256((__m256i *)(src + 1 * 16));
+ src += src_stride;
+ _mm256_storeu_si256((__m256i *)(dst + 0 * 16), s[0]);
+ _mm256_storeu_si256((__m256i *)(dst + 1 * 16), s[1]);
+ dst += dst_stride;
+ _mm256_storeu_si256((__m256i *)(dst + 0 * 16), s[2]);
+ _mm256_storeu_si256((__m256i *)(dst + 1 * 16), s[3]);
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else if (w == 64) {
+ do {
+ copy_64(src, dst);
+ src += src_stride;
+ dst += dst_stride;
+ copy_64(src, dst);
+ src += src_stride;
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ } else {
+ do {
+ copy_128(src, dst);
+ src += src_stride;
+ dst += dst_stride;
+ copy_128(src, dst);
+ src += src_stride;
+ dst += dst_stride;
+ h -= 2;
+ } while (h);
+ }
+}
diff --git a/av1/common/x86/highbd_jnt_convolve_avx2.c b/av1/common/x86/highbd_jnt_convolve_avx2.c
new file mode 100644
index 0000000..54b557f
--- /dev/null
+++ b/av1/common/x86/highbd_jnt_convolve_avx2.c
@@ -0,0 +1,277 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>
+#include <assert.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/x86/convolve_avx2.h"
+#include "aom_dsp/x86/convolve_common_intrin.h"
+#include "aom_dsp/x86/convolve_sse4_1.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "av1/common/convolve.h"
+
+#if CONFIG_JNT_COMP
+void av1_highbd_jnt_convolve_2d_copy_avx2(
+ const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w,
+ int h, InterpFilterParams *filter_params_x,
+ InterpFilterParams *filter_params_y, const int subpel_x_q4,
+ const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+ CONV_BUF_TYPE *dst = conv_params->dst;
+ int dst_stride = conv_params->dst_stride;
+ (void)filter_params_x;
+ (void)filter_params_y;
+ (void)subpel_x_q4;
+ (void)subpel_y_q4;
+ (void)dst0;
+ (void)dst_stride0;
+ (void)bd;
+
+ const int bits =
+ FILTER_BITS * 2 - conv_params->round_1 - conv_params->round_0;
+ const __m128i left_shift = _mm_cvtsi32_si128(bits);
+ const int do_average = conv_params->do_average;
+ const int w0 = conv_params->fwd_offset;
+ const int w1 = conv_params->bck_offset;
+ const __m256i wt0 = _mm256_set1_epi32(w0);
+ const __m256i wt1 = _mm256_set1_epi32(w1);
+ int i, j;
+
+ assert(bits <= 4);
+
+ if (!(w % 16)) {
+ for (i = 0; i < h; i += 1) {
+ for (j = 0; j < w; j += 16) {
+ const __m256i src_16bit =
+ _mm256_loadu_si256((__m256i *)(&src[i * src_stride + j]));
+
+ const __m256i res = _mm256_sll_epi16(src_16bit, left_shift);
+ const __m256i res_lo =
+ _mm256_cvtepu16_epi32(_mm256_castsi256_si128(res));
+ const __m256i res_hi =
+ _mm256_cvtepu16_epi32(_mm256_extracti128_si256(res, 1));
+
+ if (conv_params->use_jnt_comp_avg) {
+ mult_add_store_aligned_256(&dst[i * dst_stride + j], &res_lo, &wt0,
+ &wt1, do_average);
+ mult_add_store_aligned_256(&dst[i * dst_stride + j + 8], &res_hi,
+ &wt0, &wt1, do_average);
+ } else {
+ add_store_aligned_256(&dst[i * dst_stride + j], &res_lo, do_average);
+ add_store_aligned_256(&dst[i * dst_stride + j + 8], &res_hi,
+ do_average);
+ }
+ }
+ }
+ } else if (!(w % 4)) {
+ for (i = 0; i < h; i += 2) {
+ for (j = 0; j < w; j += 8) {
+ const __m128i src_row_0 =
+ _mm_loadu_si128((__m128i *)(&src[i * src_stride + j]));
+ const __m128i src_row_1 =
+ _mm_loadu_si128((__m128i *)(&src[i * src_stride + j + src_stride]));
+ // since not all compilers yet support _mm256_set_m128i()
+ const __m256i src_10 = _mm256_insertf128_si256(
+ _mm256_castsi128_si256(src_row_0), src_row_1, 1);
+
+ const __m256i res = _mm256_sll_epi16(src_10, left_shift);
+
+ const __m256i res_lo =
+ _mm256_cvtepu16_epi32(_mm256_castsi256_si128(res));
+ const __m256i res_hi =
+ _mm256_cvtepu16_epi32(_mm256_extracti128_si256(res, 1));
+
+ if (conv_params->use_jnt_comp_avg) {
+ mult_add_store_aligned_256(&dst[i * dst_stride + j], &res_lo, &wt0,
+ &wt1, do_average);
+ mult_add_store_aligned_256(&dst[i * dst_stride + j + dst_stride],
+ &res_hi, &wt0, &wt1, do_average);
+ } else {
+ add_store_aligned_256(&dst[i * dst_stride + j], &res_lo, do_average);
+ add_store_aligned_256(&dst[i * dst_stride + j + dst_stride], &res_hi,
+ do_average);
+ }
+ }
+ }
+ }
+}
+#endif
+
+#if CONFIG_JNT_COMP
+void av1_highbd_jnt_convolve_2d_avx2(
+ const uint16_t *src, int src_stride, CONV_BUF_TYPE *dst0, int dst_stride0,
+ int w, int h, InterpFilterParams *filter_params_x,
+ InterpFilterParams *filter_params_y, const int subpel_x_q4,
+ const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+ DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]);
+ CONV_BUF_TYPE *dst = conv_params->dst;
+ int dst_stride = conv_params->dst_stride;
+ int im_h = h + filter_params_y->taps - 1;
+ int im_stride = 8;
+ int i, j;
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const int fo_horiz = filter_params_x->taps / 2 - 1;
+ const uint16_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
+ (void)dst0;
+ (void)dst_stride0;
+
+ // Check that, even with 12-bit input, the intermediate values will fit
+ // into an unsigned 16-bit intermediate array.
+ assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16);
+
+ __m256i s[8], coeffs_y[4], coeffs_x[4];
+ const int do_average = conv_params->do_average;
+
+ const int w0 = conv_params->fwd_offset;
+ const int w1 = conv_params->bck_offset;
+ const __m256i wt0 = _mm256_set1_epi32(w0);
+ const __m256i wt1 = _mm256_set1_epi32(w1);
+ const __m128i wt0_128 = _mm256_castsi256_si128(wt0);
+ const __m128i wt1_128 = _mm256_castsi256_si128(wt1);
+
+ const __m256i round_const_x = _mm256_set1_epi32(
+ ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1)));
+ const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0);
+
+ const __m256i round_const_y = _mm256_set1_epi32(
+ ((1 << conv_params->round_1) >> 1) -
+ (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)));
+ const __m128i round_shift_y = _mm_cvtsi32_si128(conv_params->round_1);
+
+ prepare_coeffs(filter_params_x, subpel_x_q4, coeffs_x);
+ prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y);
+
+ for (j = 0; j < w; j += 8) {
+ /* Horizontal filter */
+ {
+ for (i = 0; i < im_h; i += 2) {
+ const __m256i row0 =
+ _mm256_loadu_si256((__m256i *)&src_ptr[i * src_stride + j]);
+ __m256i row1 = _mm256_set1_epi16(0);
+ if (i + 1 < im_h)
+ row1 =
+ _mm256_loadu_si256((__m256i *)&src_ptr[(i + 1) * src_stride + j]);
+
+ const __m256i r0 = _mm256_permute2x128_si256(row0, row1, 0x20);
+ const __m256i r1 = _mm256_permute2x128_si256(row0, row1, 0x31);
+
+ // even pixels
+ s[0] = _mm256_alignr_epi8(r1, r0, 0);
+ s[1] = _mm256_alignr_epi8(r1, r0, 4);
+ s[2] = _mm256_alignr_epi8(r1, r0, 8);
+ s[3] = _mm256_alignr_epi8(r1, r0, 12);
+
+ __m256i res_even = convolve(s, coeffs_x);
+ res_even = _mm256_sra_epi32(_mm256_add_epi32(res_even, round_const_x),
+ round_shift_x);
+
+ // odd pixels
+ s[0] = _mm256_alignr_epi8(r1, r0, 2);
+ s[1] = _mm256_alignr_epi8(r1, r0, 6);
+ s[2] = _mm256_alignr_epi8(r1, r0, 10);
+ s[3] = _mm256_alignr_epi8(r1, r0, 14);
+
+ __m256i res_odd = convolve(s, coeffs_x);
+ res_odd = _mm256_sra_epi32(_mm256_add_epi32(res_odd, round_const_x),
+ round_shift_x);
+
+ __m256i res_even1 = _mm256_packs_epi32(res_even, res_even);
+ __m256i res_odd1 = _mm256_packs_epi32(res_odd, res_odd);
+ __m256i res = _mm256_unpacklo_epi16(res_even1, res_odd1);
+
+ _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
+ }
+ }
+
+ /* Vertical filter */
+ {
+ __m256i s0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride));
+ __m256i s1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride));
+ __m256i s2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride));
+ __m256i s3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride));
+ __m256i s4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride));
+ __m256i s5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride));
+
+ s[0] = _mm256_unpacklo_epi16(s0, s1);
+ s[1] = _mm256_unpacklo_epi16(s2, s3);
+ s[2] = _mm256_unpacklo_epi16(s4, s5);
+
+ s[4] = _mm256_unpackhi_epi16(s0, s1);
+ s[5] = _mm256_unpackhi_epi16(s2, s3);
+ s[6] = _mm256_unpackhi_epi16(s4, s5);
+
+ for (i = 0; i < h; i += 2) {
+ const int16_t *data = &im_block[i * im_stride];
+
+ const __m256i s6 =
+ _mm256_loadu_si256((__m256i *)(data + 6 * im_stride));
+ const __m256i s7 =
+ _mm256_loadu_si256((__m256i *)(data + 7 * im_stride));
+
+ s[3] = _mm256_unpacklo_epi16(s6, s7);
+ s[7] = _mm256_unpackhi_epi16(s6, s7);
+
+ const __m256i res_a = convolve(s, coeffs_y);
+
+ const __m256i res_a_round = _mm256_sra_epi32(
+ _mm256_add_epi32(res_a, round_const_y), round_shift_y);
+
+ if (w - j > 4) {
+ const __m256i res_b = convolve(s + 4, coeffs_y);
+ const __m256i res_b_round = _mm256_sra_epi32(
+ _mm256_add_epi32(res_b, round_const_y), round_shift_y);
+ const __m256i res_ax =
+ _mm256_permute2x128_si256(res_a_round, res_b_round, 0x20);
+ const __m256i res_bx =
+ _mm256_permute2x128_si256(res_a_round, res_b_round, 0x31);
+
+ if (conv_params->use_jnt_comp_avg) {
+ mult_add_store_aligned_256(&dst[i * dst_stride + j], &res_ax, &wt0,
+ &wt1, do_average);
+ mult_add_store_aligned_256(&dst[i * dst_stride + j + dst_stride],
+ &res_bx, &wt0, &wt1, do_average);
+ } else {
+ add_store_aligned_256(&dst[i * dst_stride + j], &res_ax,
+ do_average);
+ add_store_aligned_256(&dst[i * dst_stride + j + dst_stride],
+ &res_bx, do_average);
+ }
+ } else {
+ const __m128i res_ax = _mm256_castsi256_si128(res_a_round);
+ const __m128i res_bx = _mm256_extracti128_si256(res_a_round, 1);
+
+ if (conv_params->use_jnt_comp_avg) {
+ mult_add_store(&dst[i * dst_stride + j], &res_ax, &wt0_128,
+ &wt1_128, do_average);
+ mult_add_store(&dst[i * dst_stride + j + dst_stride], &res_bx,
+ &wt0_128, &wt1_128, do_average);
+ } else {
+ add_store(&dst[i * dst_stride + j], &res_ax, do_average);
+ add_store(&dst[i * dst_stride + j + dst_stride], &res_bx,
+ do_average);
+ }
+ }
+
+ s[0] = s[1];
+ s[1] = s[2];
+ s[2] = s[3];
+
+ s[4] = s[5];
+ s[5] = s[6];
+ s[6] = s[7];
+ }
+ }
+ }
+}
+#endif
diff --git a/test/av1_convolve_2d_test.cc b/test/av1_convolve_2d_test.cc
index 8b4cc19..90512b0 100644
--- a/test/av1_convolve_2d_test.cc
+++ b/test/av1_convolve_2d_test.cc
@@ -78,10 +78,18 @@
TEST_P(AV1Convolve2DSrTest, CheckOutput) { RunCheckOutput(GET_PARAM(0)); }
+INSTANTIATE_TEST_CASE_P(C_COPY, AV1Convolve2DSrTest,
+ libaom_test::AV1Convolve2D::BuildParams(
+ av1_convolve_2d_copy_sr_c, 0, 0, 1));
+
INSTANTIATE_TEST_CASE_P(SSE2_COPY, AV1Convolve2DSrTest,
libaom_test::AV1Convolve2D::BuildParams(
av1_convolve_2d_copy_sr_sse2, 0, 0, 1));
-
+#if HAVE_AVX2
+INSTANTIATE_TEST_CASE_P(AVX2_COPY, AV1Convolve2DSrTest,
+ libaom_test::AV1Convolve2D::BuildParams(
+ av1_convolve_2d_copy_sr_avx2, 0, 0, 1));
+#endif
INSTANTIATE_TEST_CASE_P(
C_X, AV1Convolve2DSrTest,
libaom_test::AV1Convolve2D::BuildParams(av1_convolve_x_sr_c, 1, 0, 0));
@@ -213,6 +221,9 @@
INSTANTIATE_TEST_CASE_P(AVX2_Y, AV1HighbdConvolve2DSrTest,
libaom_test::AV1HighbdConvolve2D::BuildParams(
av1_highbd_convolve_y_sr_avx2, 0, 1, 0));
+INSTANTIATE_TEST_CASE_P(AVX2_COPY, AV1HighbdConvolve2DSrTest,
+ libaom_test::AV1HighbdConvolve2D::BuildParams(
+ av1_highbd_convolve_2d_copy_sr_avx2, 0, 0, 0));
#endif
#if CONFIG_JNT_COMP && HAVE_SSE4_1
TEST_P(AV1HighbdJntConvolve2DTest, CheckOutput) {
@@ -234,6 +245,14 @@
INSTANTIATE_TEST_CASE_P(C_COPY, AV1HighbdJntConvolve2DTest,
libaom_test::AV1HighbdConvolve2D::BuildParams(
av1_highbd_jnt_convolve_2d_copy_c, 0, 0, 1));
+#if HAVE_AVX2
+INSTANTIATE_TEST_CASE_P(AVX2_COPY, AV1HighbdJntConvolve2DTest,
+ libaom_test::AV1HighbdConvolve2D::BuildParams(
+ av1_highbd_jnt_convolve_2d_copy_avx2, 0, 0, 1));
+INSTANTIATE_TEST_CASE_P(AVX2, AV1HighbdJntConvolve2DTest,
+ libaom_test::AV1HighbdConvolve2D::BuildParams(
+ av1_highbd_jnt_convolve_2d_avx2, 1, 1, 1));
+#endif
#endif // CONFIG_JNT_COMP
#endif
