diff options
Diffstat (limited to 'src/dsp/x86/intra_edge_sse4.cc')
| -rw-r--r-- | src/dsp/x86/intra_edge_sse4.cc | 9 |
1 files changed, 6 insertions, 3 deletions
diff --git a/src/dsp/x86/intra_edge_sse4.cc b/src/dsp/x86/intra_edge_sse4.cc index d6af907..967be06 100644 --- a/src/dsp/x86/intra_edge_sse4.cc +++ b/src/dsp/x86/intra_edge_sse4.cc @@ -41,7 +41,8 @@ constexpr int kMaxEdgeBufferSize = 129; // This function applies the kernel [0, 4, 8, 4, 0] to 12 values. // Assumes |edge| has 16 packed byte values. Produces 12 filter outputs to // write as overlapping sets of 8-bytes. -inline void ComputeKernel1Store12(uint8_t* dest, const uint8_t* source) { +inline void ComputeKernel1Store12(uint8_t* LIBGAV1_RESTRICT dest, + const uint8_t* LIBGAV1_RESTRICT source) { const __m128i edge_lo = LoadUnaligned16(source); const __m128i edge_hi = _mm_srli_si128(edge_lo, 6); // Samples matched with the '4' tap, expanded to 16-bit. @@ -77,7 +78,8 @@ inline void ComputeKernel1Store12(uint8_t* dest, const uint8_t* source) { // This function applies the kernel [0, 5, 6, 5, 0] to 12 values. // Assumes |edge| has 8 packed byte values, and that the 2 invalid values will // be overwritten or safely discarded. -inline void ComputeKernel2Store12(uint8_t* dest, const uint8_t* source) { +inline void ComputeKernel2Store12(uint8_t* LIBGAV1_RESTRICT dest, + const uint8_t* LIBGAV1_RESTRICT source) { const __m128i edge_lo = LoadUnaligned16(source); const __m128i edge_hi = _mm_srli_si128(edge_lo, 6); const __m128i outers_lo = _mm_cvtepu8_epi16(edge_lo); @@ -115,7 +117,8 @@ inline void ComputeKernel2Store12(uint8_t* dest, const uint8_t* source) { } // This function applies the kernel [2, 4, 4, 4, 2] to 8 values. -inline void ComputeKernel3Store8(uint8_t* dest, const uint8_t* source) { +inline void ComputeKernel3Store8(uint8_t* LIBGAV1_RESTRICT dest, + const uint8_t* LIBGAV1_RESTRICT source) { const __m128i edge_lo = LoadUnaligned16(source); const __m128i edge_hi = _mm_srli_si128(edge_lo, 4); // Finish |edge_lo| life cycle quickly. |