diff options
Diffstat (limited to 'files/source/compare_win.cc')
| -rw-r--r-- | files/source/compare_win.cc | 37 |
1 files changed, 28 insertions, 9 deletions
diff --git a/files/source/compare_win.cc b/files/source/compare_win.cc index b17fc8e1..d57d3d9d 100644 --- a/files/source/compare_win.cc +++ b/files/source/compare_win.cc @@ -13,16 +13,35 @@ #include "libyuv/compare_row.h" #include "libyuv/row.h" +#if defined(_MSC_VER) +#include <intrin.h> // For __popcnt +#endif + #ifdef __cplusplus namespace libyuv { extern "C" { #endif // This module is for 32 bit Visual C x86 and clangcl -#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) +#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER) + +uint32_t HammingDistance_SSE42(const uint8_t* src_a, + const uint8_t* src_b, + int count) { + uint32_t diff = 0u; + + int i; + for (i = 0; i < count - 3; i += 4) { + uint32_t x = *((uint32_t*)src_a) ^ *((uint32_t*)src_b); // NOLINT + src_a += 4; + src_b += 4; + diff += __popcnt(x); + } + return diff; +} -__declspec(naked) uint32 - SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) { +__declspec(naked) uint32_t + SumSquareError_SSE2(const uint8_t* src_a, const uint8_t* src_b, int count) { __asm { mov eax, [esp + 4] // src_a mov edx, [esp + 8] // src_b @@ -62,8 +81,8 @@ __declspec(naked) uint32 #if _MSC_VER >= 1700 // C4752: found Intel(R) Advanced Vector Extensions; consider using /arch:AVX. #pragma warning(disable : 4752) -__declspec(naked) uint32 - SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) { +__declspec(naked) uint32_t + SumSquareError_AVX2(const uint8_t* src_a, const uint8_t* src_b, int count) { __asm { mov eax, [esp + 4] // src_a mov edx, [esp + 8] // src_b @@ -127,8 +146,8 @@ uvec32 kHashMul3 = { 0x00000001, // 33 ^ 0 }; -__declspec(naked) uint32 - HashDjb2_SSE41(const uint8* src, int count, uint32 seed) { +__declspec(naked) uint32_t + HashDjb2_SSE41(const uint8_t* src, int count, uint32_t seed) { __asm { mov eax, [esp + 4] // src mov ecx, [esp + 8] // count @@ -178,8 +197,8 @@ __declspec(naked) uint32 // Visual C 2012 required for AVX2. #if _MSC_VER >= 1700 -__declspec(naked) uint32 - HashDjb2_AVX2(const uint8* src, int count, uint32 seed) { +__declspec(naked) uint32_t + HashDjb2_AVX2(const uint8_t* src, int count, uint32_t seed) { __asm { mov eax, [esp + 4] // src mov ecx, [esp + 8] // count |