AVX2 implementation for convolve_2d

Can be up to 40% faster with bit exact results

Change-Id: Ia67ba154222fdfb6173bf8942275649e511abe43

av1/av1.cmake

@@ -301,6 +301,10 @@ if (CONFIG_CONVOLVE_ROUND)
       ${AOM_AV1_COMMON_INTRIN_SSE2}
       "${AOM_ROOT}/av1/common/x86/convolve_2d_sse2.c")
 
+  set(AOM_AV1_COMMON_INTRIN_AVX2
+      ${AOM_AV1_COMMON_INTRIN_AVX2}
+      "${AOM_ROOT}/av1/common/x86/convolve_2d_avx2.c")
+
   if (CONFIG_HIGHBITDEPTH)
     set(AOM_AV1_COMMON_INTRIN_SSSE3
         ${AOM_AV1_COMMON_INTRIN_SSSE3}

av1/av1_common.mk

@@ -162,6 +162,7 @@ endif
 ifeq ($(CONFIG_CONVOLVE_ROUND),yes)
 AV1_COMMON_SRCS-$(HAVE_SSE2) += common/x86/convolve_2d_sse2.c
 AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/convolve_2d_sse4.c
+AV1_COMMON_SRCS-$(HAVE_AVX2) += common/x86/convolve_2d_avx2.c
 ifeq ($(CONFIG_HIGHBITDEPTH),yes)
 AV1_COMMON_SRCS-$(HAVE_SSSE3) += common/x86/highbd_convolve_2d_ssse3.c
 endif

av1/common/av1_rtcd_defs.pl

@@ -566,7 +566,7 @@ if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
 
 if (aom_config("CONFIG_CONVOLVE_ROUND") eq "yes") {
   add_proto qw/void av1_convolve_2d/, "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_q4, const int subpel_y_q4, ConvolveParams *conv_params";
-  specialize qw/av1_convolve_2d sse2/;
+  specialize qw/av1_convolve_2d sse2 avx2/;
   add_proto qw/void av1_convolve_rounding/, "const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits";
   specialize qw/av1_convolve_rounding avx2/;
 

test/av1_convolve_2d_test.cc

@@ -37,6 +37,9 @@ TEST_P(AV1Convolve2DTest, CheckOutput) { RunCheckOutput(GET_PARAM(2)); }
 INSTANTIATE_TEST_CASE_P(
     SSE2, AV1Convolve2DTest,
     libaom_test::AV1Convolve2D::BuildParams(av1_convolve_2d_sse2));
+INSTANTIATE_TEST_CASE_P(
+    AVX2, AV1Convolve2DTest,
+    libaom_test::AV1Convolve2D::BuildParams(av1_convolve_2d_avx2));
 #endif  // CONFIG_JNT_COMP
 
 #if CONFIG_HIGHBITDEPTH && HAVE_SSSE3