Fix bug 806

in the function sad32x32x4d and sad64x64x4d the source is aligned to 16 bytes
and not to 32 bytes - the load is now unaligned.

Change-Id: I922fdba56d0936b5cf72e4503519f185645a168c

diff --git a/test/sad_test.cc b/test/sad_test.cc
index f07a989..e63770b 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -640,19 +640,9 @@ INSTANTIATE_TEST_CASE_P(SSE3, SADTest, ::testing::Values(
 
 #if HAVE_AVX2
 #if CONFIG_VP9_ENCODER
-// TODO(jzern): these prototypes can be removed after the avx2 versions are
-// reenabled in vp9_rtcd_defs.pl.
-extern "C" {
-void vp9_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride,
-                          const uint8_t *const ref_ptr[], int ref_stride,
-                          unsigned int *sad_array);
-void vp9_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride,
-                          const uint8_t *const ref_ptr[], int ref_stride,
-                          unsigned int *sad_array);
-}
 const SadMxNx4Func sad_64x64x4d_avx2 = vp9_sad64x64x4d_avx2;
 const SadMxNx4Func sad_32x32x4d_avx2 = vp9_sad32x32x4d_avx2;
-INSTANTIATE_TEST_CASE_P(DISABLED_AVX2, SADx4Test, ::testing::Values(
+INSTANTIATE_TEST_CASE_P(AVX2, SADx4Test, ::testing::Values(
                         make_tuple(32, 32, sad_32x32x4d_avx2),
                         make_tuple(64, 64, sad_64x64x4d_avx2)));
 #endif  // CONFIG_VP9_ENCODER

diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 8d91791..36bdc8e 100644 (file)
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -653,7 +653,7 @@ add_proto qw/void vp9_sad4x4x8/, "const uint8_t *src_ptr, int  src_stride, const
 specialize qw/vp9_sad4x4x8 sse4/;
 
 add_proto qw/void vp9_sad64x64x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
-specialize qw/vp9_sad64x64x4d sse2/;
+specialize qw/vp9_sad64x64x4d sse2 avx2/;
 
 add_proto qw/void vp9_sad32x64x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
 specialize qw/vp9_sad32x64x4d sse2/;
@@ -668,7 +668,7 @@ add_proto qw/void vp9_sad16x32x4d/, "const uint8_t *src_ptr, int  src_stride, co
 specialize qw/vp9_sad16x32x4d sse2/;
 
 add_proto qw/void vp9_sad32x32x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
-specialize qw/vp9_sad32x32x4d sse2/;
+specialize qw/vp9_sad32x32x4d sse2 avx2/;
 
 add_proto qw/void vp9_sad16x16x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
 specialize qw/vp9_sad16x16x4d sse2/;

diff --git a/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c b/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c
index f31b176..1feed62 100644 (file)
--- a/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c
+++ b/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c
@@ -31,7 +31,7 @@ void vp9_sad32x32x4d_avx2(uint8_t *src,
   sum_ref3 = _mm256_set1_epi16(0);
   for (i = 0; i < 32 ; i++) {
     // load src and all refs
-    src_reg = _mm256_load_si256((__m256i *)(src));
+    src_reg = _mm256_loadu_si256((__m256i *)(src));
     ref0_reg = _mm256_loadu_si256((__m256i *) (ref0));
     ref1_reg = _mm256_loadu_si256((__m256i *) (ref1));
     ref2_reg = _mm256_loadu_si256((__m256i *) (ref2));
@@ -103,8 +103,8 @@ void vp9_sad64x64x4d_avx2(uint8_t *src,
   sum_ref3 = _mm256_set1_epi16(0);
   for (i = 0; i < 64 ; i++) {
     // load 64 bytes from src and all refs
-    src_reg = _mm256_load_si256((__m256i *)(src));
-    srcnext_reg = _mm256_load_si256((__m256i *)(src + 32));
+    src_reg = _mm256_loadu_si256((__m256i *)(src));
+    srcnext_reg = _mm256_loadu_si256((__m256i *)(src + 32));
     ref0_reg = _mm256_loadu_si256((__m256i *) (ref0));
     ref0next_reg = _mm256_loadu_si256((__m256i *) (ref0 + 32));
     ref1_reg = _mm256_loadu_si256((__m256i *) (ref1));