add sse2/ssse3 simd optimization

3 files changed, 59 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]];
     }
   }
 }
diff --git a/fmdsp/fmdsp-vramlookup-ssse3.c b/fmdsp/fmdsp-vramlookup-ssse3.c
new file mode 100644
index 0000000..30e7311
--- /dev/null
+++ b/fmdsp/fmdsp-vramlookup-ssse3.c
@@ -0,0 +1,48 @@
+#include "fmdsp/fmdsp.h"
+#include <tmmintrin.h>
+
+void fmdsp_vramlookup_ssse3(uint8_t *vram32, const uint8_t *vram, const uint8_t *palette, int stride) {
+  __m128i z = _mm_setzero_si128();
+  __m128i p[3];
+  {
+    union {
+      __m128i xmm;
+      uint8_t u8[16];
+    } pi[3];
+    for (int i = 0; i < FMDSP_PALETTE_COLORS; i++) {
+      for (int c = 0; c < 3; c++) {
+        pi[c].u8[i] = palette[i*3+c];
+      }
+    }
+    for (int c = 0; c < 3; c++) {
+      p[c] = _mm_load_si128(&pi[c].xmm);
+    }
+  }
+
+  for (int y = 0; y < PC98_H; y++) {
+    for (int x = 0; x < 40; x++) {
+      // 16 pixels
+      __m128i v = _mm_loadu_si128((__m128i *)&vram[y*PC98_W+x*16]);
+
+      __m128i r = _mm_shuffle_epi8(p[0], v);
+      __m128i g = _mm_shuffle_epi8(p[1], v);
+      __m128i b = _mm_shuffle_epi8(p[2], v);
+
+      __m128i gb[2], zr[2];
+      gb[0] = _mm_unpacklo_epi8(b, g);
+      gb[1] = _mm_unpackhi_epi8(b, g);
+      zr[0] = _mm_unpacklo_epi8(r, z);
+      zr[1] = _mm_unpackhi_epi8(r, z);
+      
+      __m128i o[4];
+      o[0] = _mm_unpacklo_epi16(gb[0], zr[0]);
+      o[1] = _mm_unpackhi_epi16(gb[0], zr[0]);
+      o[2] = _mm_unpacklo_epi16(gb[1], zr[1]);
+      o[3] = _mm_unpackhi_epi16(gb[1], zr[1]);
+      for (int i = 0; i < 4; i++) {
+        _mm_storeu_si128((__m128i *)&vram32[(x*4+i)*16], o[i]);
+      }
+    }
+    vram32 += stride;
+  }
+}
diff --git a/fmdsp/fmdsp.h b/fmdsp/fmdsp.h
index a7e4aab..a8c3edc 100644
--- a/fmdsp/fmdsp.h
+++ b/fmdsp/fmdsp.h
@@ -59,6 +59,9 @@ void fmdsp_vramlookup_c(uint8_t *vram32,
                         const uint8_t *vram,
                         const uint8_t *palette,
                         int stride);
+
+void fmdsp_vramlookup_neon(uint8_t *, const uint8_t *, const uint8_t *, int);
+void fmdsp_vramlookup_ssse3(uint8_t *, const uint8_t *, const uint8_t *, int);
 #ifdef __cplusplus
 }
 #endif