From f3ebfff841aafc0aac6186243606e1b2ea381e71 Mon Sep 17 00:00:00 2001 From: Andrei Karas Date: Fri, 23 Dec 2016 01:31:10 +0300 Subject: Add a color dye function for sse2 (software). --- src/resources/dye/dyepalette_replaceacolor.cpp | 189 +++++-------------------- 1 file changed, 37 insertions(+), 152 deletions(-) (limited to 'src/resources/dye/dyepalette_replaceacolor.cpp') diff --git a/src/resources/dye/dyepalette_replaceacolor.cpp b/src/resources/dye/dyepalette_replaceacolor.cpp index b2693912b..b5b77f770 100644 --- a/src/resources/dye/dyepalette_replaceacolor.cpp +++ b/src/resources/dye/dyepalette_replaceacolor.cpp @@ -39,100 +39,14 @@ void DyePalette::replaceAColor(uint32_t *restrict pixels, #ifdef SIMD_SUPPORTED replaceAColorSimd(pixels, bufSize); #else // SIMD_SUPPORTED - replaceAColorDefault(pixels, bufSize); +#include "resources/dye/dyepalette_replaceacolor_default.hpp" #endif // SIMD_SUPPORTED } void DyePalette::replaceAColorDefault(uint32_t *restrict pixels, const int bufSize) const restrict2 { - std::vector::const_iterator it_end = mColors.end(); - const size_t sz = mColors.size(); - if (!sz || !pixels) - return; - if (sz % 2) - -- it_end; - -#ifdef ENABLE_CILKPLUS - cilk_for (int ptr = 0; ptr < bufSize; ptr ++) - { - uint8_t *const p = reinterpret_cast(&pixels[ptr]); - const unsigned int data = pixels[ptr]; - - std::vector::const_iterator it = mColors.begin(); - while (it != it_end) - { - const DyeColor &col = *it; - ++ it; - const DyeColor &col2 = *it; - -#if SDL_BYTEORDER == SDL_BIG_ENDIAN - const unsigned int coldata = (col.value[3] << 24U) - | (col.value[2] << 16U) - | (col.value[1] << 8U) - | (col.value[0]); -#else // SDL_BYTEORDER == SDL_BIG_ENDIAN - - const unsigned int coldata = (col.value[3]) - | (col.value[2] << 8U) - | (col.value[1] << 16U) | - (col.value[0] << 24U); -#endif // SDL_BYTEORDER == SDL_BIG_ENDIAN - - if (data == coldata) - { - p[3] = col2.value[0]; - p[2] = col2.value[1]; - p[1] = col2.value[2]; - p[0] = col2.value[3]; - break; - } - - ++ it; - } - } - -#else // ENABLE_CILKPLUS - - for (const uint32_t *const p_end = pixels + CAST_SIZE(bufSize); - pixels != p_end; - ++pixels) - { - uint8_t *const p = reinterpret_cast(pixels); - const unsigned int data = *pixels; - - std::vector::const_iterator it = mColors.begin(); - while (it != it_end) - { - const DyeColor &col = *it; - ++ it; - const DyeColor &col2 = *it; - -#if SDL_BYTEORDER == SDL_BIG_ENDIAN - const unsigned int coldata = (col.value[3] << 24U) - | (col.value[2] << 16U) - | (col.value[1] << 8U) - | (col.value[0]); -#else // SDL_BYTEORDER == SDL_BIG_ENDIAN - const unsigned int coldata = (col.value[3]) - | (col.value[2] << 8U) - | (col.value[1] << 16U) | - (col.value[0] << 24U); -#endif // SDL_BYTEORDER == SDL_BIG_ENDIAN - - if (data == coldata) - { - p[3] = col2.value[0]; - p[2] = col2.value[1]; - p[1] = col2.value[2]; - p[0] = col2.value[3]; - break; - } - - ++ it; - } - } -#endif // ENABLE_CILKPLUS +#include "resources/dye/dyepalette_replaceacolor_default.hpp" } #ifdef SIMD_SUPPORTED @@ -144,75 +58,32 @@ static void print256(const char *const text, const __m256i &val) } */ -__attribute__ ((target ("avx2"))) +__attribute__ ((target ("sse2"))) void DyePalette::replaceAColorSimd(uint32_t *restrict pixels, const int bufSize) const restrict2 { - std::vector::const_iterator it_end = mColors.end(); - const size_t sz = mColors.size(); - if (!sz || !pixels) - return; - if (sz % 2) - -- it_end; - const int mod = bufSize % 8; - const int bufEnd = bufSize - mod; - - for (int ptr = 0; ptr < bufEnd; ptr += 8) - { -// __m256i base = _mm256_load_si256(reinterpret_cast<__m256i*>(pixels)); - __m256i base = _mm256_loadu_si256(reinterpret_cast<__m256i*>( - &pixels[ptr])); - - std::vector::const_iterator it = mColors.begin(); - while (it != it_end) - { - const DyeColor &col = *it; - ++ it; - const DyeColor &col2 = *it; - - __m256i newMask = _mm256_set1_epi32(col2.valueA); - __m256i cmpMask = _mm256_set1_epi32(col.valueA); - __m256i cmpRes = _mm256_cmpeq_epi32(base, cmpMask); - __m256i srcAnd = _mm256_andnot_si256(cmpRes, base); - __m256i dstAnd = _mm256_and_si256(cmpRes, newMask); - base = _mm256_or_si256(srcAnd, dstAnd); - - ++ it; - } -// _mm256_store_si256(reinterpret_cast<__m256i*>(pixels), base); - _mm256_storeu_si256(reinterpret_cast<__m256i*>(&pixels[ptr]), base); - } - - // complete end without simd - for (int ptr = bufSize - mod; ptr < bufSize; ptr ++) - { - uint8_t *const p = reinterpret_cast(&pixels[ptr]); - const unsigned int data = pixels[ptr]; - - std::vector::const_iterator it = mColors.begin(); - while (it != it_end) - { - const DyeColor &col = *it; - ++ it; - const DyeColor &col2 = *it; +#include "resources/dye/dyepalette_replaceacolor_sse2.hpp" +} - const unsigned int coldata = (col.value[3]) | - (col.value[2] << 8U) | - (col.value[1] << 16U) | - (col.value[0] << 24U); +__attribute__ ((target ("avx2"))) +void DyePalette::replaceAColorSimd(uint32_t *restrict pixels, + const int bufSize) const restrict2 +{ +#include "resources/dye/dyepalette_replaceacolor_avx2.hpp" +} - if (data == coldata) - { - p[3] = col2.value[0]; - p[2] = col2.value[1]; - p[1] = col2.value[2]; - p[0] = col2.value[3]; - break; - } +__attribute__ ((target ("sse2"))) +void DyePalette::replaceAColorSse2(uint32_t *restrict pixels, + const int bufSize) const restrict2 +{ +#include "resources/dye/dyepalette_replaceacolor_sse2.hpp" +} - ++ it; - } - } +__attribute__ ((target ("avx2"))) +void DyePalette::replaceAColorAvx2(uint32_t *restrict pixels, + const int bufSize) const restrict2 +{ +#include "resources/dye/dyepalette_replaceacolor_avx2.hpp" } #endif // SIMD_SUPPORTED @@ -221,5 +92,19 @@ FUNCTION_SIMD_DEFAULT void DyePalette::replaceAColorSimd(uint32_t *restrict pixels, const int bufSize) const restrict2 { - replaceAColorDefault(pixels, bufSize); +#include "resources/dye/dyepalette_replaceacolor_default.hpp" +} + +FUNCTION_SIMD_DEFAULT +void DyePalette::replaceAColorSse2(uint32_t *restrict pixels, + const int bufSize) const restrict2 +{ +#include "resources/dye/dyepalette_replaceacolor_default.hpp" +} + +FUNCTION_SIMD_DEFAULT +void DyePalette::replaceAColorAvx2(uint32_t *restrict pixels, + const int bufSize) const restrict2 +{ +#include "resources/dye/dyepalette_replaceacolor_default.hpp" } -- cgit v1.2.3-60-g2f50