Index: source/libvpx/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c |
=================================================================== |
--- source/libvpx/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c (revision 290053) |
+++ source/libvpx/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c (working copy) |
@@ -31,7 +31,7 @@ |
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 @@ |
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)); |