Merge "fdct32x32 neon implementation"

test/dct32x32_test.cc

@@ -383,14 +383,14 @@ INSTANTIATE_TEST_CASE_P(C, PartialTrans32x32Test,
                                                      VPX_BITS_8)));
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 
-#if HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+#if HAVE_NEON && !CONFIG_EMULATE_HARDWARE
 INSTANTIATE_TEST_CASE_P(
     NEON, Trans32x32Test,
-    ::testing::Values(make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_neon,
-                                 0, VPX_BITS_8),
+    ::testing::Values(make_tuple(&vpx_fdct32x32_neon,
+                                 &vpx_idct32x32_1024_add_neon, 0, VPX_BITS_8),
                       make_tuple(&vpx_fdct32x32_rd_c,
                                  &vpx_idct32x32_1024_add_neon, 1, VPX_BITS_8)));
-#endif  // HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+#endif  // HAVE_NEON && !CONFIG_EMULATE_HARDWARE
 
 #if HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
 INSTANTIATE_TEST_CASE_P(

test/dct_test.cc

@@ -527,7 +527,10 @@ INSTANTIATE_TEST_CASE_P(
 #if !CONFIG_EMULATE_HARDWARE
 INSTANTIATE_TEST_CASE_P(
     NEON, TransDCT,
-    ::testing::Values(make_tuple(&vpx_fdct16x16_neon,
+    ::testing::Values(make_tuple(&vpx_fdct32x32_neon,
+                                 &vpx_idct32x32_1024_add_neon, 32, 0,
+                                 VPX_BITS_8),
+                      make_tuple(&vpx_fdct16x16_neon,
                                  &vpx_idct16x16_256_add_neon, 16, 0,
                                  VPX_BITS_8),
                       make_tuple(&vpx_fdct8x8_neon, &vpx_idct8x8_64_add_neon, 8,

vpx_dsp/vpx_dsp.mk

@@ -195,6 +195,7 @@ DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_txfm_avx2.c
 DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_dct32x32_impl_avx2.h
 DSP_SRCS-$(HAVE_NEON)   += arm/fdct_neon.c
 DSP_SRCS-$(HAVE_NEON)   += arm/fdct16x16_neon.c
+DSP_SRCS-$(HAVE_NEON)   += arm/fdct32x32_neon.c
 DSP_SRCS-$(HAVE_NEON)   += arm/fwd_txfm_neon.c
 DSP_SRCS-$(HAVE_MSA)    += mips/fwd_txfm_msa.h
 DSP_SRCS-$(HAVE_MSA)    += mips/fwd_txfm_msa.c

vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -502,7 +502,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vpx_fdct16x16_1 sse2/;
 
   add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vpx_fdct32x32 sse2/;
+  specialize qw/vpx_fdct32x32 neon sse2/;
 
   add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
   specialize qw/vpx_fdct32x32_rd sse2/;
@@ -550,7 +550,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vpx_fdct16x16_1 sse2 msa/;
 
   add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vpx_fdct32x32 sse2 avx2 msa/;
+  specialize qw/vpx_fdct32x32 neon sse2 avx2 msa/;
 
   add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
   specialize qw/vpx_fdct32x32_rd sse2 avx2 msa/;