Use precalculated values for A and S dyes.

6 files changed, 35 insertions, 47 deletions

diff --git a/src/resources/dye/dyecolor.h b/src/resources/dye/dyecolor.h
index 11e12a038..11257fc22 100644
--- a/src/resources/dye/dyecolor.h
+++ b/src/resources/dye/dyecolor.h
@@ -27,42 +27,61 @@
 
 #include "localconsts.h"
 
+#define buildHex(a, b, c, d) \
+    (a) * 16777216U + (b) * 65536U + (c) * 256U + CAST_U32(d)
+
 struct DyeColor final
 {
-    DyeColor() noexcept2
+    DyeColor() noexcept2 :
+        valueA(buildHex(0, 0, 0, 255)),
+        valueS(buildHex(0, 0, 0, 0)),
+        valueSOgl(buildHex(0, 0, 0, 0))
     {
         value[3] = 255;
     }
 
     DyeColor(const uint8_t r,
              const uint8_t g,
-             const uint8_t b) noexcept2
+             const uint8_t b) noexcept2 :
+        valueA(buildHex(r, g, b, 255)),
+        valueS(buildHex(r, g, b, 0)),
+        valueSOgl(buildHex(0, b, g, r))
     {
         value[0] = r;
         value[1] = g;
         value[2] = b;
         value[3] = 255;
-//        value2 = buildHex(r, g, b, 255);
     }
 
     DyeColor(const uint8_t r,
              const uint8_t g,
              const uint8_t b,
-             const uint8_t a) noexcept2
+             const uint8_t a) noexcept2 :
+        valueA(buildHex(r, g, b, a)),
+        valueS(buildHex(r, g, b, 0)),
+        valueSOgl(buildHex(0, b, g, r))
     {
         value[0] = r;
         value[1] = g;
         value[2] = b;
         value[3] = a;
-//        value2 = buildHex(r, g, b, a);
+    }
+
+    void update()
+    {
+        valueA = buildHex(value[0], value[1], value[2], value[3]);
+        valueS = buildHex(value[0], value[1], value[2], 0);
+        valueSOgl = buildHex(0, value[2], value[1], value[0]);
     }
 
     union
     {
         uint8_t value[4];
-        uint32_t value1;
+        uint32_t valueAOgl;
     };
-//    uint32_t value2;
+    uint32_t valueA;
+    uint32_t valueS;
+    uint32_t valueSOgl;
 };
 
 #endif  // RESOURCES_DYE_DYECOLOR_H
diff --git a/src/resources/dye/dyepalette.cpp b/src/resources/dye/dyepalette.cpp
index 76f82b439..28efd76ba 100644
--- a/src/resources/dye/dyepalette.cpp
+++ b/src/resources/dye/dyepalette.cpp
@@ -109,6 +109,7 @@ void DyePalette::hexToColor(const std::string &restrict hexStr,
             hexDecode(hexStr[i]) << 4)
             + hexDecode(hexStr[i + 1]));
     }
+    color.update();
 }
 
 unsigned int DyePalette::hexDecode(const signed char c) noexcept2
diff --git a/src/resources/dye/dyepalette_replaceacolor.cpp b/src/resources/dye/dyepalette_replaceacolor.cpp
index f98c52343..7c24d6c1f 100644
--- a/src/resources/dye/dyepalette_replaceacolor.cpp
+++ b/src/resources/dye/dyepalette_replaceacolor.cpp
@@ -31,14 +31,6 @@
 #ifdef __x86_64__
 // avx2
 #include "immintrin.h"
-
-#if SDL_BYTEORDER == SDL_BIG_ENDIAN
-#define buildHex(a, b, c, d) \
-    (d) * 16777216U + (c) * 65536U + (b) * 256U + CAST_U32(a)
-#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
-#define buildHex(a, b, c, d) \
-    (a) * 16777216U + (b) * 65536U + (c) * 256U + CAST_U32(d)
-#endif  // SDL_BYTEORDER == SDL_BIG_ENDIAN
 #endif  // __x86_64__
 
 #include "debug.h"
@@ -177,8 +169,8 @@ void DyePalette::replaceAColorSimd(uint32_t *restrict pixels,
             ++ it;
             const DyeColor &col2 = *it;
 
-            __m256i newMask = _mm256_set1_epi32(buildHex(col2.value[0], col2.value[1], col2.value[2], col2.value[3]));
-            __m256i cmpMask = _mm256_set1_epi32(buildHex(col.value[0], col.value[1], col.value[2],  col.value[3]));
+            __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);
diff --git a/src/resources/dye/dyepalette_replaceaoglcolor.cpp b/src/resources/dye/dyepalette_replaceaoglcolor.cpp
index 2b2d7a54c..8ed7af643 100644
--- a/src/resources/dye/dyepalette_replaceaoglcolor.cpp
+++ b/src/resources/dye/dyepalette_replaceaoglcolor.cpp
@@ -33,14 +33,6 @@
 #ifdef __x86_64__
 // avx2
 #include "immintrin.h"
-
-#if SDL_BYTEORDER == SDL_BIG_ENDIAN
-#define buildHex(a, b, c, d) \
-    (d) * 16777216U + (c) * 65536U + (b) * 256U + CAST_U32(a)
-#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
-#define buildHex(a, b, c, d) \
-    (a) * 16777216U + (b) * 65536U + (c) * 256U + CAST_U32(d)
-#endif  // SDL_BYTEORDER == SDL_BIG_ENDIAN
 #endif  // __x86_64__
 
 #include "debug.h"
@@ -185,8 +177,8 @@ void DyePalette::replaceAOGLColorSimd(uint32_t *restrict pixels,
             ++ it;
             const DyeColor &col2 = *it;
 
-            __m256i newMask = _mm256_set1_epi32(buildHex(col2.value[3], col2.value[2], col2.value[1], col2.value[0]));
-            __m256i cmpMask = _mm256_set1_epi32(buildHex(col.value[3], col.value[2], col.value[1],  col.value[0]));
+            __m256i newMask = _mm256_set1_epi32(col2.valueAOgl);
+            __m256i cmpMask = _mm256_set1_epi32(col.valueAOgl);
             __m256i cmpRes = _mm256_cmpeq_epi32(base, cmpMask);
             __m256i srcAnd = _mm256_andnot_si256(cmpRes, base);
             __m256i dstAnd = _mm256_and_si256(cmpRes, newMask);
diff --git a/src/resources/dye/dyepalette_replacescolor.cpp b/src/resources/dye/dyepalette_replacescolor.cpp
index 22f4e6ac4..02412b186 100644
--- a/src/resources/dye/dyepalette_replacescolor.cpp
+++ b/src/resources/dye/dyepalette_replacescolor.cpp
@@ -29,14 +29,6 @@
 #ifdef __x86_64__
 // avx2
 #include "immintrin.h"
-
-#if SDL_BYTEORDER == SDL_BIG_ENDIAN
-#define buildHex(a, b, c, d) \
-    (d) * 16777216U + (c) * 65536U + (b) * 256U + CAST_U32(a)
-#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
-#define buildHex(a, b, c, d) \
-    (a) * 16777216U + (b) * 65536U + (c) * 256U + CAST_U32(d)
-#endif  // SDL_BYTEORDER == SDL_BIG_ENDIAN
 #endif  // __x86_64__
 
 #include "debug.h"
@@ -178,9 +170,9 @@ void DyePalette::replaceSColorSimd(uint32_t *restrict pixels,
 
             __m256i base2 = _mm256_and_si256(mask, base);
             //print256("base2  ", base2);
-            __m256i newMask = _mm256_set1_epi32(buildHex(col2.value[0], col2.value[1], col2.value[2], 0));
+            __m256i newMask = _mm256_set1_epi32(col2.valueS);
             //print256("newMask  ", newMask);
-            __m256i cmpMask = _mm256_set1_epi32(buildHex(col.value[0], col.value[1], col.value[2], 0));
+            __m256i cmpMask = _mm256_set1_epi32(col.valueS);
             //print256("cmpMask  ", cmpMask);
             __m256i cmpRes = _mm256_cmpeq_epi32(base2, cmpMask);
             //print256("cmpRes  ", cmpRes);
diff --git a/src/resources/dye/dyepalette_replacesoglcolor.cpp b/src/resources/dye/dyepalette_replacesoglcolor.cpp
index a023d8f79..899286240 100644
--- a/src/resources/dye/dyepalette_replacesoglcolor.cpp
+++ b/src/resources/dye/dyepalette_replacesoglcolor.cpp
@@ -31,14 +31,6 @@
 #ifdef __x86_64__
 // avx2
 #include "immintrin.h"
-
-#if SDL_BYTEORDER == SDL_BIG_ENDIAN
-#define buildHex(a, b, c, d) \
-    (d) * 16777216U + (c) * 65536U + (b) * 256U + CAST_U32(a)
-#else  // SDL_BYTEORDER == SDL_BIG_ENDIAN
-#define buildHex(a, b, c, d) \
-    (a) * 16777216U + (b) * 65536U + (c) * 256U + CAST_U32(d)
-#endif  // SDL_BYTEORDER == SDL_BIG_ENDIAN
 #endif  // __x86_64__
 
 #include "debug.h"
@@ -186,9 +178,9 @@ void DyePalette::replaceSOGLColorSimd(uint32_t *restrict pixels,
 
             __m256i base2 = _mm256_and_si256(mask, base);
             //print256("base2  ", base2);
-            __m256i newMask = _mm256_set1_epi32(buildHex(0, col2.value[2], col2.value[1], col2.value[0]));
+            __m256i newMask = _mm256_set1_epi32(col2.valueSOgl);
             //print256("newMask  ", newMask);
-            __m256i cmpMask = _mm256_set1_epi32(buildHex(0, col.value[2], col.value[1], col.value[0]));
+            __m256i cmpMask = _mm256_set1_epi32(col.valueSOgl);
             //print256("cmpMask  ", cmpMask);
             __m256i cmpRes = _mm256_cmpeq_epi32(base2, cmpMask);
             //print256("cmpRes  ", cmpRes);