diff options
Diffstat (limited to 'src/resources/dye/dyepalette_replacescolor.cpp')
-rw-r--r-- | src/resources/dye/dyepalette_replacescolor.cpp | 199 |
1 files changed, 39 insertions, 160 deletions
diff --git a/src/resources/dye/dyepalette_replacescolor.cpp b/src/resources/dye/dyepalette_replacescolor.cpp index d6a058773..0dc78185b 100644 --- a/src/resources/dye/dyepalette_replacescolor.cpp +++ b/src/resources/dye/dyepalette_replacescolor.cpp @@ -1,5 +1,7 @@ /* * The ManaPlus Client + * Copyright (C) 2007-2009 The Mana World Development Team + * Copyright (C) 2009-2010 The Mana Developers * Copyright (C) 2011-2016 The ManaPlus Developers * * This file is part of The ManaPlus Client. @@ -37,98 +39,14 @@ void DyePalette::replaceSColor(uint32_t *restrict pixels, #ifdef SIMD_SUPPORTED replaceSColorSimd(pixels, bufSize); #else // SIMD_SUPPORTED - replaceSColorDefault(pixels, bufSize); +#include "resources/dye/dyepalette_replacescolor_default.hpp" #endif // SIMD_SUPPORTED } void DyePalette::replaceSColorDefault(uint32_t *restrict pixels, const int bufSize) const restrict2 { - std::vector<DyeColor>::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<uint8_t *>(&pixels[ptr]); -#if SDL_BYTEORDER == SDL_BIG_ENDIAN - const unsigned int data = pixels[ptr] & 0x00ffffff; -#else // SDL_BYTEORDER == SDL_BIG_ENDIAN - - const unsigned int data = pixels[ptr] & 0xffffff00; -#endif // SDL_BYTEORDER == SDL_BIG_ENDIAN - - std::vector<DyeColor>::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[2] << 16U) - | (col.value[1] << 8U) | (col.value[0]); -#else // SDL_BYTEORDER == SDL_BIG_ENDIAN - const unsigned int coldata = (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]; - 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<uint8_t *>(pixels); -#if SDL_BYTEORDER == SDL_BIG_ENDIAN - const unsigned int data = (*pixels) & 0x00ffffff; -#else // SDL_BYTEORDER == SDL_BIG_ENDIAN - - const unsigned int data = (*pixels) & 0xffffff00; -#endif // SDL_BYTEORDER == SDL_BIG_ENDIAN - - std::vector<DyeColor>::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[2] << 16U) - | (col.value[1] << 8U) | (col.value[0]); -#else // SDL_BYTEORDER == SDL_BIG_ENDIAN - - const unsigned int coldata = (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]; - break; - } - - ++ it; - } - } -#endif // ENABLE_CILKPLUS +#include "resources/dye/dyepalette_replacescolor_default.hpp" } #ifdef SIMD_SUPPORTED @@ -140,85 +58,32 @@ static void print256(const char *const text, const __m256i &val) } */ -__attribute__ ((target ("avx2"))) +__attribute__ ((target ("sse2"))) void DyePalette::replaceSColorSimd(uint32_t *restrict pixels, const int bufSize) const restrict2 { - std::vector<DyeColor>::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 mask = _mm256_set1_epi32(0xffffff00); -// __m256i base = _mm256_load_si256(reinterpret_cast<__m256i*>(pixels)); - __m256i base = _mm256_loadu_si256(reinterpret_cast<__m256i*>( - &pixels[ptr])); - - std::vector<DyeColor>::const_iterator it = mColors.begin(); - while (it != it_end) - { - const DyeColor &col = *it; - ++ it; - const DyeColor &col2 = *it; - - __m256i base2 = _mm256_and_si256(mask, base); - __m256i newMask = _mm256_set1_epi32(col2.valueS); - __m256i cmpMask = _mm256_set1_epi32(col.valueS); - __m256i cmpRes = _mm256_cmpeq_epi32(base2, cmpMask); - cmpRes = _mm256_and_si256(mask, cmpRes); - __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<uint8_t *>(&pixels[ptr]); -#if SDL_BYTEORDER == SDL_BIG_ENDIAN - const unsigned int data = pixels[ptr] & 0x00ffffff; -#else // SDL_BYTEORDER == SDL_BIG_ENDIAN - - const unsigned int data = pixels[ptr] & 0xffffff00; -#endif // SDL_BYTEORDER == SDL_BIG_ENDIAN - - std::vector<DyeColor>::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[2] << 16U) - | (col.value[1] << 8U) | (col.value[0]); -#else // SDL_BYTEORDER == SDL_BIG_ENDIAN +#include "resources/dye/dyepalette_replacescolor_sse2.hpp" +} - const unsigned int coldata = (col.value[2] << 8U) - | (col.value[1] << 16U) | (col.value[0] << 24U); -#endif // SDL_BYTEORDER == SDL_BIG_ENDIAN +__attribute__ ((target ("avx2"))) +void DyePalette::replaceSColorSimd(uint32_t *restrict pixels, + const int bufSize) const restrict2 +{ +#include "resources/dye/dyepalette_replacescolor_avx2.hpp" +} - if (data == coldata) - { - p[3] = col2.value[0]; - p[2] = col2.value[1]; - p[1] = col2.value[2]; - break; - } +__attribute__ ((target ("sse2"))) +void DyePalette::replaceSColorSse2(uint32_t *restrict pixels, + const int bufSize) const restrict2 +{ +#include "resources/dye/dyepalette_replacescolor_sse2.hpp" +} - ++ it; - } - } +__attribute__ ((target ("avx2"))) +void DyePalette::replaceSColorAvx2(uint32_t *restrict pixels, + const int bufSize) const restrict2 +{ +#include "resources/dye/dyepalette_replacescolor_avx2.hpp" } #endif // SIMD_SUPPORTED @@ -227,5 +92,19 @@ FUNCTION_SIMD_DEFAULT void DyePalette::replaceSColorSimd(uint32_t *restrict pixels, const int bufSize) const restrict2 { - replaceSColorDefault(pixels, bufSize); +#include "resources/dye/dyepalette_replacescolor_default.hpp" +} + +FUNCTION_SIMD_DEFAULT +void DyePalette::replaceSColorSse2(uint32_t *restrict pixels, + const int bufSize) const restrict2 +{ +#include "resources/dye/dyepalette_replacescolor_default.hpp" +} + +FUNCTION_SIMD_DEFAULT +void DyePalette::replaceSColorAvx2(uint32_t *restrict pixels, + const int bufSize) const restrict2 +{ +#include "resources/dye/dyepalette_replacescolor_default.hpp" } |