hadamard highbd ssse3: use tran_low_t for coeff

BUG=webm:1365

Change-Id: I374dfc08732932382043905f128e928b08cb4f57

diff --git a/test/hadamard_test.cc b/test/hadamard_test.cc
index 3b19b23..7e43c69 100644 (file)
--- a/test/hadamard_test.cc
+++ b/test/hadamard_test.cc
@@ -150,20 +150,18 @@ INSTANTIATE_TEST_CASE_P(SSE2, Hadamard8x8Test,
                         ::testing::Values(&vpx_hadamard_8x8_sse2));
 #endif  // HAVE_SSE2
 
-// TODO(jingning): Remove highbitdepth flag when the SIMD functions are
-// in place and turn on the unit test.
-#if !CONFIG_VP9_HIGHBITDEPTH
 #if HAVE_SSSE3 && ARCH_X86_64
 INSTANTIATE_TEST_CASE_P(SSSE3, Hadamard8x8Test,
                         ::testing::Values(&vpx_hadamard_8x8_ssse3));
 #endif  // HAVE_SSSE3 && ARCH_X86_64
-#endif  // !CONFIG_VP9_HIGHBITDEPTH
 
 #if HAVE_NEON
 INSTANTIATE_TEST_CASE_P(NEON, Hadamard8x8Test,
                         ::testing::Values(&vpx_hadamard_8x8_neon));
 #endif  // HAVE_NEON
 
+// TODO(jingning): Remove highbitdepth flag when the SIMD functions are
+// in place and turn on the unit test.
 #if !CONFIG_VP9_HIGHBITDEPTH
 #if HAVE_MSA
 INSTANTIATE_TEST_CASE_P(MSA, Hadamard8x8Test,

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index cf85cc8..d23a5e7 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -888,7 +888,7 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
 
   if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
     add_proto qw/void vpx_hadamard_8x8/, "const int16_t *src_diff, int src_stride, tran_low_t *coeff";
-    specialize qw/vpx_hadamard_8x8 sse2 neon/;
+    specialize qw/vpx_hadamard_8x8 sse2 neon/, "$ssse3_x86_64";
 
     add_proto qw/void vpx_hadamard_16x16/, "const int16_t *src_diff, int src_stride, tran_low_t *coeff";
     specialize qw/vpx_hadamard_16x16 sse2 neon/;

diff --git a/vpx_dsp/x86/avg_ssse3_x86_64.asm b/vpx_dsp/x86/avg_ssse3_x86_64.asm
index 36d38da..d170a44 100644 (file)
--- a/vpx_dsp/x86/avg_ssse3_x86_64.asm
+++ b/vpx_dsp/x86/avg_ssse3_x86_64.asm
@@ -8,8 +8,6 @@
 ;  be found in the AUTHORS file in the root of the source tree.
 ;
 
-%define private_prefix vpx
-
 %include "third_party/x86inc/x86inc.asm"
 
 SECTION .text
@@ -96,6 +94,21 @@ SECTION .text
   SWAP               7, 9
 %endmacro
 
+%if CONFIG_VP9_HIGHBITDEPTH
+; store %1 to outputq + %2
+; uses m8-m10 as scratch registers
+%macro STORE_TRAN_LOW 2
+  pxor                           m8, m8
+  mova                           m9, m%1
+  mova                          m10, m%1
+  pcmpgtw                        m8, m%1
+  punpcklwd                      m9, m8
+  punpckhwd                     m10, m8
+  mova               [outputq + %2], m9
+  mova          [outputq + %2 + 16], m10
+%endmacro
+%endif
+
 INIT_XMM ssse3
 cglobal hadamard_8x8, 3, 5, 11, input, stride, output
   lea                r3, [2 * strideq]
@@ -117,6 +130,16 @@ cglobal hadamard_8x8, 3, 5, 11, input, stride, output
   TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9, 10
   HMD8_1D
 
+%if CONFIG_VP9_HIGHBITDEPTH
+  STORE_TRAN_LOW                  0, 0
+  STORE_TRAN_LOW                  1, 32
+  STORE_TRAN_LOW                  2, 64
+  STORE_TRAN_LOW                  3, 96
+  STORE_TRAN_LOW                  4, 128
+  STORE_TRAN_LOW                  5, 160
+  STORE_TRAN_LOW                  6, 192
+  STORE_TRAN_LOW                  7, 224
+%else
   mova              [outputq +   0], m0
   mova              [outputq +  16], m1
   mova              [outputq +  32], m2
@@ -125,6 +148,7 @@ cglobal hadamard_8x8, 3, 5, 11, input, stride, output
   mova              [outputq +  80], m5
   mova              [outputq +  96], m6
   mova              [outputq + 112], m7
+%endif
 
   RET
 %endif