add sse2/ssse3 simd optimization

1 files changed, 8 insertions, 5 deletions

diff --git a/fmdsp/fmdsp-vramlookup-c.c b/fmdsp/fmdsp-vramlookup-c.c
index f900c8d..3d06f71 100644
--- a/fmdsp/fmdsp-vramlookup-c.c
+++ b/fmdsp/fmdsp-vramlookup-c.c
@@ -1,14 +1,17 @@
 #include "fmdsp/fmdsp.h"
 
 void fmdsp_vramlookup_c(uint8_t *vram32, const uint8_t *vram, const uint8_t *palette, int stride) {
+  uint32_t palette32[FMDSP_PALETTE_COLORS];
+  for (int i = 0; i < FMDSP_PALETTE_COLORS; i++) {
+    uint8_t r = palette[i*3+0];
+    uint8_t g = palette[i*3+1];
+    uint8_t b = palette[i*3+2];
+    palette32[i] = (((uint32_t)r)<<16) | (((uint32_t)g)<<8) | ((uint32_t)b);
+  }
   for (int y = 0; y < PC98_H; y++) {
     for (int x = 0; x < PC98_W; x++) {
-      uint8_t r = palette[vram[y*PC98_W+x]*3+0];
-      uint8_t g = palette[vram[y*PC98_W+x]*3+1];
-      uint8_t b = palette[vram[y*PC98_W+x]*3+2];
-      uint32_t data = (((uint32_t)r)<<16) | (((uint32_t)g)<<8) | ((uint32_t)b);
       uint32_t *row = (uint32_t *)(vram32 + y*stride);
-      row[x] = data;
+      row[x] = palette32[vram[y*PC98_W+x]];
     }
   }
 }