diff --git a/src/dsp/lossless.c b/src/dsp/lossless.c
index 0af443ca..13328b2c 100644
--- a/src/dsp/lossless.c
+++ b/src/dsp/lossless.c
@@ -1133,14 +1133,11 @@ static void GetBestGreenToRed(
   best_tx->green_to_red_ = green_to_red;
 }
 
-static float GetPredictionCostCrossColorBlue(
+static void CollectColorBlueTransforms(
     int tile_x_offset, int tile_y_offset, int all_x_max, int all_y_max,
-    int xsize, VP8LMultipliers prev_x, VP8LMultipliers prev_y,
-    int green_to_blue, int red_to_blue, const int accumulated_blue_histo[256],
+    int xsize, int green_to_blue, int red_to_blue, int* histo,
     const uint32_t* const argb) {
   int all_y;
-  int histo[256] = { 0 };
-  float cur_diff;
   for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) {
     int all_x;
     int ix = all_y * xsize + tile_x_offset;
@@ -1148,6 +1145,20 @@ static float GetPredictionCostCrossColorBlue(
       ++histo[TransformColorBlue(green_to_blue, red_to_blue, argb[ix])];
     }
   }
+}
+
+static float GetPredictionCostCrossColorBlue(
+    int tile_x_offset, int tile_y_offset, int all_x_max, int all_y_max,
+    int xsize, VP8LMultipliers prev_x, VP8LMultipliers prev_y,
+    int green_to_blue, int red_to_blue, const int accumulated_blue_histo[256],
+    const uint32_t* const argb) {
+  int histo[256] = { 0 };
+  float cur_diff;
+
+  VP8LCollectColorBlueTransforms(tile_x_offset, tile_y_offset, all_x_max,
+                                 all_y_max, xsize, green_to_blue, red_to_blue,
+                                 histo, argb);
+
   cur_diff = PredictionCostCrossColor(accumulated_blue_histo, histo);
   if ((uint8_t)green_to_blue == prev_x.green_to_blue_) {
     cur_diff -= 3;  // favor keeping the areas locally similar
@@ -1726,6 +1737,8 @@ VP8LConvertFunc VP8LConvertBGRAToRGBA4444;
 VP8LConvertFunc VP8LConvertBGRAToRGB565;
 VP8LConvertFunc VP8LConvertBGRAToBGR;
 
+VP8LCollectColorBlueTransformsFunc VP8LCollectColorBlueTransforms;
+
 VP8LFastLog2SlowFunc VP8LFastLog2Slow;
 VP8LFastLog2SlowFunc VP8LFastSLog2Slow;
 
@@ -1765,6 +1778,8 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInit(void) {
   VP8LConvertBGRAToRGB565 = VP8LConvertBGRAToRGB565_C;
   VP8LConvertBGRAToBGR = VP8LConvertBGRAToBGR_C;
 
+  VP8LCollectColorBlueTransforms = CollectColorBlueTransforms;
+
   VP8LFastLog2Slow = FastLog2Slow;
   VP8LFastSLog2Slow = FastSLog2Slow;
 
diff --git a/src/dsp/lossless.h b/src/dsp/lossless.h
index 44469277..17e024d0 100644
--- a/src/dsp/lossless.h
+++ b/src/dsp/lossless.h
@@ -59,6 +59,12 @@ extern VP8LConvertFunc VP8LConvertBGRAToRGBA4444;
 extern VP8LConvertFunc VP8LConvertBGRAToRGB565;
 extern VP8LConvertFunc VP8LConvertBGRAToBGR;
 
+typedef void (*VP8LCollectColorBlueTransformsFunc)(
+    int tile_x_offset, int tile_y_offset, int all_x_max, int all_y_max,
+    int xsize, int green_to_blue, int red_to_blue, int* histo,
+    const uint32_t* const argb);
+extern VP8LCollectColorBlueTransformsFunc VP8LCollectColorBlueTransforms;
+
 // Expose some C-only fallback functions
 void VP8LTransformColor_C(const VP8LMultipliers* const m,
                           uint32_t* data, int num_pixels);
diff --git a/src/dsp/lossless_mips_dsp_r2.c b/src/dsp/lossless_mips_dsp_r2.c
index d85811b5..914a46a3 100644
--- a/src/dsp/lossless_mips_dsp_r2.c
+++ b/src/dsp/lossless_mips_dsp_r2.c
@@ -367,6 +367,63 @@ static void TransformColor(const VP8LMultipliers* const m, uint32_t* data,
   }
 }
 
+static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue,
+                                              uint8_t red_to_blue,
+                                              uint32_t argb) {
+  const uint32_t green = argb >> 8;
+  const uint32_t red = argb >> 16;
+  uint8_t new_blue = argb;
+  new_blue -= ColorTransformDelta(green_to_blue, green);
+  new_blue -= ColorTransformDelta(red_to_blue, red);
+  return (new_blue & 0xff);
+}
+
+static void CollectColorBlueTransforms(
+    int tile_x_offset, int tile_y_offset, int all_x_max, int all_y_max,
+    int xsize, int green_to_blue, int red_to_blue, int* histo,
+    const uint32_t* const argb) {
+  const int rtb = (red_to_blue << 16) | (red_to_blue & 0xffff);
+  const int gtb = (green_to_blue << 16) | (green_to_blue & 0xffff);
+  const uint32_t mask = 0xff00ffu;
+  int ix = tile_y_offset * xsize + tile_x_offset;
+  int all_y;
+  for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) {
+    uint32_t* p_argb = (uint32_t*)&argb[ix];
+    const int loop_cnt = all_x_max - tile_x_offset;
+    int all_x;
+    ix += xsize;
+    for (all_x = 0; all_x < (loop_cnt >> 1); ++all_x) {
+      int temp0, temp1, temp2, temp3, temp4, temp5, temp6;
+      __asm__ volatile (
+        "lw           %[temp0],  0(%[p_argb])             \n\t"
+        "lw           %[temp1],  4(%[p_argb])             \n\t"
+        "precr.qb.ph  %[temp2],  %[temp0],  %[temp1]      \n\t"
+        "ins          %[temp1],  %[temp0],  16,    16     \n\t"
+        "shra.ph      %[temp2],  %[temp2],  8             \n\t"
+        "shra.ph      %[temp3],  %[temp1],  8             \n\t"
+        "mul.ph       %[temp5],  %[temp2],  %[rtb]        \n\t"
+        "mul.ph       %[temp6],  %[temp3],  %[gtb]        \n\t"
+        "and          %[temp4],  %[temp1],  %[mask]       \n\t"
+        "addiu        %[p_argb], %[p_argb], 8             \n\t"
+        "shra.ph      %[temp5],  %[temp5],  5             \n\t"
+        "shra.ph      %[temp6],  %[temp6],  5             \n\t"
+        "subu.qb      %[temp2],  %[temp4],  %[temp5]      \n\t"
+        "subu.qb      %[temp2],  %[temp2],  %[temp6]      \n\t"
+        : [p_argb]"+&r"(p_argb), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),
+          [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4),
+          [temp5]"=&r"(temp5), [temp6]"=&r"(temp6)
+        : [rtb]"r"(rtb), [gtb]"r"(gtb), [mask]"r"(mask)
+        : "memory", "hi", "lo"
+      );
+      ++histo[(uint8_t)(temp2 >> 16)];
+      ++histo[(uint8_t)temp2];
+    }
+    if (loop_cnt & 1) {
+      ++histo[TransformColorBlue(green_to_blue, red_to_blue, *p_argb)];
+    }
+  }
+}
+
 #endif  // WEBP_USE_MIPS_DSP_R2
 
 //------------------------------------------------------------------------------
@@ -388,6 +445,7 @@ void VP8LDspInitMIPSdspR2(void) {
   VP8LPredictors[13] = Predictor13;
   VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
   VP8LTransformColor = TransformColor;
+  VP8LCollectColorBlueTransforms = CollectColorBlueTransforms;
 #endif  // WEBP_USE_MIPS_DSP_R2
 }