Add SSE2 version of forward cross-color transform

Encoding speed is roughly the same.

Change-Id: I6b976d0eb24e1847714e719762cb8403768da66c

src/dsp/lossless.c

@@ -833,8 +833,7 @@ static WEBP_INLINE uint32_t MultipliersToColorCode(
          m->green_to_red_;
 }
 
-static WEBP_INLINE void TransformColor(const VP8LMultipliers* const m,
+void VP8LTransformColor_C(const VP8LMultipliers* const m, uint32_t* data,
-                                       uint32_t* data,
                           int num_pixels) {
   int i;
   for (i = 0; i < num_pixels; ++i) {
@@ -1078,7 +1077,7 @@ static void CopyTileWithColorTransform(int xsize, int ysize,
   int yscan = GetMin(max_tile_size, ysize - tile_y);
   argb += tile_y * xsize + tile_x;
   while (yscan-- > 0) {
-    TransformColor(&color_transform, argb, xscan);
+    VP8LTransformColor(&color_transform, argb, xscan);
     argb += xsize;
   }
 }
@@ -1465,6 +1464,7 @@ VP8LProcessBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
 VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed;
 VP8LPredictorFunc VP8LPredictors[16];
 
+VP8LTransformColorFunc VP8LTransformColor;
 VP8LTransformColorFunc VP8LTransformColorInverse;
 
 VP8LConvertFunc VP8LConvertBGRAToRGB;
@@ -1482,6 +1482,7 @@ void VP8LDspInit(void) {
   VP8LSubtractGreenFromBlueAndRed = VP8LSubtractGreenFromBlueAndRed_C;
   VP8LAddGreenToBlueAndRed = VP8LAddGreenToBlueAndRed_C;
 
+  VP8LTransformColor = VP8LTransformColor_C;
   VP8LTransformColorInverse = VP8LTransformColorInverse_C;
 
   VP8LConvertBGRAToRGB = VP8LConvertBGRAToRGB_C;

src/dsp/lossless.h

@@ -41,6 +41,7 @@ typedef struct {
 } VP8LMultipliers;
 typedef void (*VP8LTransformColorFunc)(const VP8LMultipliers* const m,
                                        uint32_t* argb_data, int num_pixels);
+extern VP8LTransformColorFunc VP8LTransformColor;
 extern VP8LTransformColorFunc VP8LTransformColorInverse;
 
 typedef void (*VP8LConvertFunc)(const uint32_t* src, int num_pixels,
@@ -52,8 +53,10 @@ extern VP8LConvertFunc VP8LConvertBGRAToRGB565;
 extern VP8LConvertFunc VP8LConvertBGRAToBGR;
 
 // Expose some C-only fallback functions
-extern void VP8LTransformColorInverse_C(
+extern void VP8LTransformColor_C(const VP8LMultipliers* const m,
-    const VP8LMultipliers* const m, uint32_t* data, int num_pixels);
+                                 uint32_t* data, int num_pixels);
+extern void VP8LTransformColorInverse_C(const VP8LMultipliers* const m,
+                                        uint32_t* data, int num_pixels);
 
 extern void VP8LConvertBGRAToRGB_C(const uint32_t* src,
                                    int num_pixels, uint8_t* dst);

src/dsp/lossless_sse2.c

@@ -171,6 +171,45 @@ static WEBP_INLINE __m128i ColorTransformDelta(__m128i color_pred,
   return _mm_srli_epi32(signed_mult, 5);
 }
 
+static WEBP_INLINE void TransformColor(const VP8LMultipliers* const m,
+                                       uint32_t* argb_data,
+                                       int num_pixels) {
+  const __m128i g_to_r = _mm_set1_epi32(m->green_to_red_);       // multipliers
+  const __m128i g_to_b = _mm_set1_epi32(m->green_to_blue_);
+  const __m128i r_to_b = _mm_set1_epi32(m->red_to_blue_);
+
+  int i;
+
+  for (i = 0; i + 4 <= num_pixels; i += 4) {
+    const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
+    const __m128i alpha_green_mask = _mm_set1_epi32(0xff00ff00);  // masks
+    const __m128i red_mask = _mm_set1_epi32(0x00ff0000);
+    const __m128i green_mask = _mm_set1_epi32(0x0000ff00);
+    const __m128i lower_8bit_mask  = _mm_set1_epi32(0x000000ff);
+    const __m128i ag = _mm_and_si128(in, alpha_green_mask);      // alpha, green
+    const __m128i r = _mm_srli_epi32(_mm_and_si128(in, red_mask), 16);
+    const __m128i g = _mm_srli_epi32(_mm_and_si128(in, green_mask), 8);
+    const __m128i b = in;
+
+    const __m128i r_delta = ColorTransformDelta(g_to_r, g);      // red
+    const __m128i r_new =
+        _mm_and_si128(_mm_sub_epi32(r, r_delta), lower_8bit_mask);
+    const __m128i r_new_shifted = _mm_slli_epi32(r_new, 16);
+
+    const __m128i b_delta_1 = ColorTransformDelta(g_to_b, g);    // blue
+    const __m128i b_delta_2 = ColorTransformDelta(r_to_b, r);
+    const __m128i b_delta = _mm_add_epi32(b_delta_1, b_delta_2);
+    const __m128i b_new =
+        _mm_and_si128(_mm_sub_epi32(b, b_delta), lower_8bit_mask);
+
+    const __m128i out = _mm_or_si128(_mm_or_si128(ag, r_new_shifted), b_new);
+    _mm_storeu_si128((__m128i*)&argb_data[i], out);
+  }
+
+  // Fall-back to C-version for left-overs.
+  VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
+}
+
 static WEBP_INLINE void TransformColorInverse(const VP8LMultipliers* const m,
                                               uint32_t* argb_data,
                                               int num_pixels) {
@@ -359,6 +398,7 @@ void VP8LDspInitSSE2(void) {
   VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
   VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
 
+  VP8LTransformColor = TransformColor;
   VP8LTransformColorInverse = TransformColorInverse;
 
   VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA;