High bit depth 32x32 inverse DCT_DCT transform, AVX2

- Witness the follow user-level speedup on AV1 baseline:
 Encoding time reduction: 4.26%
 Decoding time reduction: 25.35%

Change-Id: Ideaf3cd473ad45ed9256c80d5a5daed0a6e098cf

av1/av1_common.mk

@@ -144,6 +144,8 @@ AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/av1_fwd_txfm2d_sse4.c
 endif
 ifeq ($(CONFIG_AOM_HIGHBITDEPTH),yes)
 AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/highbd_txfm_utility_sse4.h
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/highbd_inv_txfm_sse4.c
+AV1_COMMON_SRCS-$(HAVE_AVX2) += common/x86/highbd_inv_txfm_avx2.c
 endif
 
 ifneq ($(CONFIG_AOM_HIGHBITDEPTH),yes)

av1/av1_cx.mk

@@ -140,7 +140,6 @@ AV1_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/dct_ssse3.c
 AV1_CX_SRCS-$(HAVE_AVX2) += encoder/x86/hybrid_fwd_txfm_avx2.c
 ifeq ($(CONFIG_AOM_HIGHBITDEPTH),yes)
 AV1_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/highbd_fwd_txfm_sse4.c
-AV1_CX_SRCS-$(HAVE_SSE4_1) += common/x86/highbd_inv_txfm_sse4.c
 AV1_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/av1_highbd_quantize_sse4.c
 endif
 

av1/common/av1_rtcd_defs.pl

@@ -565,7 +565,7 @@ if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void av1_inv_txfm2d_add_16x16/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
   specialize qw/av1_inv_txfm2d_add_16x16 sse4_1/;
   add_proto qw/void av1_inv_txfm2d_add_32x32/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
-  specialize qw/av1_inv_txfm2d_add_32x32/;
+  specialize qw/av1_inv_txfm2d_add_32x32 avx2/;
   add_proto qw/void av1_inv_txfm2d_add_64x64/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
   specialize qw/av1_inv_txfm2d_add_64x64/;
 }

test/av1_highbd_iht_test.cc

@@ -86,6 +86,8 @@ class AV1HighbdInvHTNxN : public ::testing::TestWithParam<IHbdHtParam> {
       return 8;
     } else if (256 == num_coeffs_) {
       return 16;
+    } else if (1024 == num_coeffs_) {
+      return 32;
     } else {
       return 0;
     }
@@ -216,4 +218,19 @@ INSTANTIATE_TEST_CASE_P(SSE4_1, AV1HighbdInvHTNxN,
                         ::testing::ValuesIn(kArrayIhtParam));
 #endif  // HAVE_SSE4_1 && CONFIG_AOM_HIGHBITDEPTH
 
+#if HAVE_AVX2 && CONFIG_AOM_HIGHBITDEPTH
+#define PARAM_LIST_32X32                                   \
+  &av1_fwd_txfm2d_32x32_c, &av1_inv_txfm2d_add_32x32_avx2, \
+      &av1_inv_txfm2d_add_32x32_c, 1024
+
+const IHbdHtParam kArrayIhtParam32x32[] = {
+  // 32x32
+  make_tuple(PARAM_LIST_32X32, DCT_DCT, 10),
+  make_tuple(PARAM_LIST_32X32, DCT_DCT, 12),
+};
+
+INSTANTIATE_TEST_CASE_P(AVX2, AV1HighbdInvHTNxN,
+                        ::testing::ValuesIn(kArrayIhtParam32x32));
+
+#endif  // HAVE_AVX2 && CONFIG_AOM_HIGHBITDEPTH
 }  // namespace