AVX2 optimization of motion compensation functions

AVX2 implementation of av1_convolve_x_sr, av1_convolve_y_sr and
av1_convolve_2d_sr have been added.

Improvements have been made to av1_convolve_x_avx2, av1_convolve_y_avx2
and av1_convolve_2d_avx2.

Change-Id: I62a699dd9dcf42de94dd72cc2d43affc0dc31404

aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c

@@ -13,29 +13,9 @@
 
 #include "./aom_dsp_rtcd.h"
 #include "aom_dsp/x86/convolve.h"
+#include "aom_dsp/x86/convolve_avx2.h"
 #include "aom_ports/mem.h"
 
-// filters for 16_h8 and 16_v8
-DECLARE_ALIGNED(32, static const uint8_t, filt1_global_avx2[32]) = {
-  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
-  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
-};
-
-DECLARE_ALIGNED(32, static const uint8_t, filt2_global_avx2[32]) = {
-  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10,
-  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
-};
-
-DECLARE_ALIGNED(32, static const uint8_t, filt3_global_avx2[32]) = {
-  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12,
-  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
-};
-
-DECLARE_ALIGNED(32, static const uint8_t, filt4_global_avx2[32]) = {
-  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14,
-  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
-};
-
 #if defined(__clang__)
 #if (__clang_major__ > 0 && __clang_major__ < 3) ||            \
     (__clang_major__ == 3 && __clang_minor__ <= 3) ||          \

aom_dsp/x86/convolve_avx2.h

+/*
+* Copyright (c) 2018, 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.
+*/
+
+#ifndef AOM_DSP_X86_CONVOLVE_AVX2_H_
+#define AOM_DSP_X86_CONVOLVE_AVX2_H_
+
+// filters for 16
+DECLARE_ALIGNED(32, static const uint8_t, filt1_global_avx2[32]) = {
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt2_global_avx2[32]) = {
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10,
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt3_global_avx2[32]) = {
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12,
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt4_global_avx2[32]) = {
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14,
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
+};
+
+#endif

av1/common/av1_rtcd_defs.pl

@@ -497,9 +497,9 @@ 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";
 specialize qw/av1_convolve_y sse2 avx2/;
 add_proto qw/void av1_convolve_x_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_x_sr c sse2/;
+specialize qw/av1_convolve_x_sr c sse2 avx2/;
 add_proto qw/void av1_convolve_y_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_y_sr c sse2/;
+specialize qw/av1_convolve_y_sr c sse2 avx2/;
 
 add_proto qw/void av1_convolve_2d_scale/, "const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params";
 specialize qw/av1_convolve_2d_scale sse4_1/;

av1/common/convolve.c

@@ -401,6 +401,7 @@ void av1_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst0,
     for (int x = 0; x < w; ++x) {
       int32_t sum = (1 << (bd + FILTER_BITS - 1));
       for (int k = 0; k < filter_params_x->taps; ++k) {
+        assert((x_filter[k] % 2) == 0);
         sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
       }
       assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
@@ -418,6 +419,7 @@ void av1_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst0,
     for (int x = 0; x < w; ++x) {
       CONV_BUF_TYPE sum = 1 << offset_bits;
       for (int k = 0; k < filter_params_y->taps; ++k) {
+        assert((y_filter[k] % 2) == 0);
         sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
       }
       assert(0 <= sum && sum < (1 << (offset_bits + 2)));
@@ -454,6 +456,7 @@ void av1_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst0,
     for (int x = 0; x < w; ++x) {
       CONV_BUF_TYPE res = 0;
       for (int k = 0; k < filter_params_y->taps; ++k) {
+        assert((y_filter[k] % 2) == 0);
         res += y_filter[k] * src[(y - fo_vert + k) * src_stride + x];
       }
       res *= (1 << bits);
@@ -487,6 +490,7 @@ void av1_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst0,
     for (int x = 0; x < w; ++x) {
       CONV_BUF_TYPE res = 0;
       for (int k = 0; k < filter_params_x->taps; ++k) {
+        assert((x_filter[k] % 2) == 0);
         res += x_filter[k] * src[y * src_stride + x - fo_horiz + k];
       }
       res = (1 << bits) * ROUND_POWER_OF_TWO(res, conv_params->round_0);
@@ -550,6 +554,7 @@ void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst,
     for (int x = 0; x < w; ++x) {
       int32_t sum = (1 << (bd + FILTER_BITS - 1));
       for (int k = 0; k < filter_params_x->taps; ++k) {
+        assert((x_filter[k] % 2) == 0);
         sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
       }
       assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
@@ -567,6 +572,7 @@ void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst,
     for (int x = 0; x < w; ++x) {
       CONV_BUF_TYPE sum = 1 << offset_bits;
       for (int k = 0; k < filter_params_y->taps; ++k) {
+        assert((y_filter[k] % 2) == 0);
         sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
       }
       assert(0 <= sum && sum < (1 << (offset_bits + 2)));
@@ -596,6 +602,7 @@ void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst,
     for (int x = 0; x < w; ++x) {
       CONV_BUF_TYPE res = 0;
       for (int k = 0; k < filter_params_y->taps; ++k) {
+        assert((y_filter[k] % 2) == 0);
         res += y_filter[k] * src[(y - fo_vert + k) * src_stride + x];
       }
       dst[y * dst_stride + x] =
@@ -623,6 +630,7 @@ void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst,
     for (int x = 0; x < w; ++x) {
       CONV_BUF_TYPE res = 0;
       for (int k = 0; k < filter_params_x->taps; ++k) {
+        assert((x_filter[k] % 2) == 0);
         res += x_filter[k] * src[y * src_stride + x - fo_horiz + k];
       }
       res = ROUND_POWER_OF_TWO(res, conv_params->round_0);

test/av1_convolve_2d_test.cc

@@ -94,6 +94,14 @@ INSTANTIATE_TEST_CASE_P(
     libaom_test::AV1Convolve2D::BuildParams(av1_convolve_2d_sr_sse2, 1, 1, 0));
 
 #if HAVE_AVX2
+INSTANTIATE_TEST_CASE_P(
+    AVX2_X, AV1Convolve2DSrTest,
+    libaom_test::AV1Convolve2D::BuildParams(av1_convolve_x_sr_avx2, 1, 0, 0));
+
+INSTANTIATE_TEST_CASE_P(
+    AVX2_Y, AV1Convolve2DSrTest,
+    libaom_test::AV1Convolve2D::BuildParams(av1_convolve_y_sr_avx2, 0, 1, 0));
+
 INSTANTIATE_TEST_CASE_P(
     AVX2, AV1Convolve2DSrTest,
     libaom_test::AV1Convolve2D::BuildParams(av1_convolve_2d_sr_avx2, 1, 1, 0));