Switch in replaceSOGLColor into using custom despatcher.

1 files changed, 327 insertions, 49 deletions

diff --git a/src/resources/dye/dyepalette_replacesoglcolor.cpp b/src/resources/dye/dyepalette_replacesoglcolor.cpp
index 57ffd6b50..005523b4b 100644
--- a/src/resources/dye/dyepalette_replacesoglcolor.cpp
+++ b/src/resources/dye/dyepalette_replacesoglcolor.cpp
@@ -35,26 +35,98 @@
 
 #include "debug.h"
 
-void DyePalette::replaceSOGLColor(uint32_t *restrict pixels,
-                                  const int bufSize) const restrict2
-{
-#ifdef SIMD_SUPPORTED
-    if (bufSize >= 8)
-        replaceSOGLColorSimd(pixels, bufSize);
-    else
-        replaceSOGLColorDefault(pixels, bufSize);
-#else  // SIMD_SUPPORTED
-#include "resources/dye/dyepalette_replacesoglcolor_default.hpp"
-#endif  // SIMD_SUPPORTED
-}
-
 void DyePalette::replaceSOGLColorDefault(uint32_t *restrict pixels,
                                          const int bufSize) const restrict2
 {
-#include "resources/dye/dyepalette_replacesoglcolor_default.hpp"
-}
+    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]) & 0xffffff00;
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+        const unsigned int data = (pixels[ptr]) & 0x00ffffff;
+#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[0] << 24)
+                | (col.value[1] << 16) | (col.value[2] << 8);
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+            const unsigned int coldata = (col.value[0])
+                | (col.value[1] << 8) | (col.value[2] << 16);
+#endif  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+            if (data == coldata)
+            {
+                p[0] = col2.value[0];
+                p[1] = col2.value[1];
+                p[2] = 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) & 0xffffff00;
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+        const unsigned int data = (*pixels) & 0x00ffffff;
+#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[0] << 24)
+                | (col.value[1] << 16) | (col.value[2] << 8);
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
 
+            const unsigned int coldata = (col.value[0])
+                | (col.value[1] << 8) | (col.value[2] << 16);
+#endif  // SDL_BYTEORDER == SDL_BIG_ENDIAN
 
+            if (data == coldata)
+            {
+                p[0] = col2.value[0];
+                p[1] = col2.value[1];
+                p[2] = col2.value[2];
+                break;
+            }
+
+            ++ it;
+        }
+    }
+#endif  // ENABLE_CILKPLUS
+}
 
 #ifdef SIMD_SUPPORTED
 /*
@@ -66,54 +138,260 @@ static void print256(const char *const text, const __m256i &val)
 */
 
 __attribute__ ((target ("sse2")))
-void DyePalette::replaceSOGLColorSimd(uint32_t *restrict pixels,
+void DyePalette::replaceSOGLColorSse2(uint32_t *restrict pixels,
                                       const int bufSize) const restrict2
 {
-#include "resources/dye/dyepalette_replacesoglcolor_sse2.hpp"
-}
+    std::vector<DyeColor>::const_iterator it_end = mColors.end();
+    const size_t sz = mColors.size();
+    if (!sz || !pixels)
+        return;
+    if (sz % 2)
+        -- it_end;
 
-__attribute__ ((target ("avx2")))
-void DyePalette::replaceSOGLColorSimd(uint32_t *restrict pixels,
-                                      const int bufSize) const restrict2
-{
-#include "resources/dye/dyepalette_replacesoglcolor_avx2.hpp"
-}
+    if (bufSize >= 8)
+    {
+        for (int ptr = 0; ptr < bufSize; ptr += 4)
+        {
+            __m128i mask = _mm_set1_epi32(0x00ffffff);
+//            __m128i base = _mm_load_si128(reinterpret_cast<__m128i*>(
+//             &pixels[ptr]));
+            __m128i base = _mm_loadu_si128(reinterpret_cast<__m128i*>(
+                &pixels[ptr]));
 
-__attribute__ ((target ("sse2")))
-void DyePalette::replaceSOGLColorSse2(uint32_t *restrict pixels,
-                                      const int bufSize) const restrict2
-{
-#include "resources/dye/dyepalette_replacesoglcolor_sse2.hpp"
+            std::vector<DyeColor>::const_iterator it = mColors.begin();
+            while (it != it_end)
+            {
+                const DyeColor &col = *it;
+                ++ it;
+                const DyeColor &col2 = *it;
+
+                __m128i base2 = _mm_and_si128(mask, base);
+                __m128i newMask = _mm_set1_epi32(col2.valueSOgl);
+                __m128i cmpMask = _mm_set1_epi32(col.valueSOgl);
+                __m128i cmpRes = _mm_cmpeq_epi32(base2, cmpMask);
+                cmpRes = _mm_and_si128(mask, cmpRes);
+                __m128i srcAnd = _mm_andnot_si128(cmpRes, base);
+                __m128i dstAnd = _mm_and_si128(cmpRes, newMask);
+                base = _mm_or_si128(srcAnd, dstAnd);
+                ++ it;
+            }
+//            _mm_store_si128(reinterpret_cast<__m128i*>(&pixels[ptr]), base);
+            _mm_storeu_si128(reinterpret_cast<__m128i*>(&pixels[ptr]), base);
+        }
+    }
+    else
+    {
+#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]) & 0xffffff00;
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+            const unsigned int data = (pixels[ptr]) & 0x00ffffff;
+#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[0] << 24)
+                    | (col.value[1] << 16) | (col.value[2] << 8);
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+                const unsigned int coldata = (col.value[0])
+                    | (col.value[1] << 8) | (col.value[2] << 16);
+#endif  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+                if (data == coldata)
+                {
+                    p[0] = col2.value[0];
+                    p[1] = col2.value[1];
+                    p[2] = 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) & 0xffffff00;
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+            const unsigned int data = (*pixels) & 0x00ffffff;
+#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[0] << 24)
+                    | (col.value[1] << 16) | (col.value[2] << 8);
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+                const unsigned int coldata = (col.value[0])
+                    | (col.value[1] << 8) | (col.value[2] << 16);
+#endif  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+                if (data == coldata)
+                {
+                    p[0] = col2.value[0];
+                    p[1] = col2.value[1];
+                    p[2] = col2.value[2];
+                    break;
+                }
+
+                ++ it;
+            }
+        }
+#endif  // ENABLE_CILKPLUS
+    }
 }
 
 __attribute__ ((target ("avx2")))
 void DyePalette::replaceSOGLColorAvx2(uint32_t *restrict pixels,
                                       const int bufSize) const restrict2
 {
-#include "resources/dye/dyepalette_replacesoglcolor_avx2.hpp"
-}
+    std::vector<DyeColor>::const_iterator it_end = mColors.end();
+    const size_t sz = mColors.size();
+    if (!sz || !pixels)
+        return;
+    if (sz % 2)
+        -- it_end;
 
-#endif  // SIMD_SUPPORTED
+    if (bufSize >= 8)
+    {
+        for (int ptr = 0; ptr < bufSize; ptr += 8)
+        {
+            __m256i mask = _mm256_set1_epi32(0x00ffffff);
+//          __m256i base = _mm256_load_si256(reinterpret_cast<__m256i*>(
+//              &pixels[ptr]));
+            __m256i base = _mm256_loadu_si256(reinterpret_cast<__m256i*>(
+                &pixels[ptr]));
 
-FUNCTION_SIMD_DEFAULT
-void DyePalette::replaceSOGLColorSimd(uint32_t *restrict pixels,
-                                      const int bufSize) const restrict2
-{
-#include "resources/dye/dyepalette_replacesoglcolor_default.hpp"
-}
+            std::vector<DyeColor>::const_iterator it = mColors.begin();
+            while (it != it_end)
+            {
+                const DyeColor &col = *it;
+                ++ it;
+                const DyeColor &col2 = *it;
 
-FUNCTION_SIMD_DEFAULT
-void DyePalette::replaceSOGLColorSse2(uint32_t *restrict pixels,
-                                      const int bufSize) const restrict2
-{
-#include "resources/dye/dyepalette_replacesoglcolor_default.hpp"
-}
+                __m256i base2 = _mm256_and_si256(mask, base);
+                __m256i newMask = _mm256_set1_epi32(col2.valueSOgl);
+                __m256i cmpMask = _mm256_set1_epi32(col.valueSOgl);
+                __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[ptr]), base);
+            _mm256_storeu_si256(reinterpret_cast<__m256i*>(&pixels[ptr]), base);
+        }
+    }
+    else
+    {
+#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]) & 0xffffff00;
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
 
-FUNCTION_SIMD_DEFAULT
-void DyePalette::replaceSOGLColorAvx2(uint32_t *restrict pixels,
-                                      const int bufSize) const restrict2
-{
-#include "resources/dye/dyepalette_replacesoglcolor_default.hpp"
+            const unsigned int data = (pixels[ptr]) & 0x00ffffff;
+#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[0] << 24)
+                    | (col.value[1] << 16) | (col.value[2] << 8);
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+                const unsigned int coldata = (col.value[0])
+                    | (col.value[1] << 8) | (col.value[2] << 16);
+#endif  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+                if (data == coldata)
+                {
+                    p[0] = col2.value[0];
+                    p[1] = col2.value[1];
+                    p[2] = 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) & 0xffffff00;
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+            const unsigned int data = (*pixels) & 0x00ffffff;
+#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[0] << 24)
+                    | (col.value[1] << 16) | (col.value[2] << 8);
+#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+                const unsigned int coldata = (col.value[0])
+                    | (col.value[1] << 8) | (col.value[2] << 16);
+#endif  // SDL_BYTEORDER == SDL_BIG_ENDIAN
+
+                if (data == coldata)
+                {
+                    p[0] = col2.value[0];
+                    p[1] = col2.value[1];
+                    p[2] = col2.value[2];
+                    break;
+                }
+
+                ++ it;
+            }
+        }
+#endif  // ENABLE_CILKPLUS
+    }
 }
 
+#endif  // SIMD_SUPPORTED
 #endif  // USE_OPENGL