diff --git a/README.chromium b/README.chromium
index 440dac6a1..7858de66b 100644
--- a/README.chromium
+++ b/README.chromium
@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 282
+Version: 283
 License: BSD
 License File: LICENSE
 
diff --git a/include/libyuv/version.h b/include/libyuv/version.h
index cba66fc71..b5043f59e 100644
--- a/include/libyuv/version.h
+++ b/include/libyuv/version.h
@@ -11,7 +11,7 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 282
+#define LIBYUV_VERSION 283
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_
 
diff --git a/source/planar_functions.cc b/source/planar_functions.cc
index 6383552eb..bdd1dfed4 100644
--- a/source/planar_functions.cc
+++ b/source/planar_functions.cc
@@ -1709,26 +1709,50 @@ int ARGBBlur(const uint8* src_argb, int src_stride_argb,
              uint8* dst_argb, int dst_stride_argb,
              int32* dst_cumsum, int dst_stride32_cumsum,
              int width, int height, int radius) {
+  void (*ComputeCumulativeSumRow)(const uint8* row, int32* cumsum,
+      int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
   void (*CumulativeSumToAverage)(const int32* topleft, const int32* botleft,
       int width, int area, uint8* dst, int count) = CumulativeSumToAverage_C;
 #if defined(HAS_CUMULATIVESUMTOAVERAGE_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
     CumulativeSumToAverage = CumulativeSumToAverage_SSE2;
   }
 #endif
-
   ARGBComputeCumulativeSum(src_argb, src_stride_argb,
                            dst_cumsum, dst_stride32_cumsum,
-                           width, height);
+                           width, radius);
+
+  src_argb = src_argb + radius * src_stride_argb;
+  int32* cumsum_bot_row = &dst_cumsum[(radius - 1) * dst_stride32_cumsum];
+
+  int32* max_cumsum_bot_row =
+      &dst_cumsum[(radius * 2 + 2) * dst_stride32_cumsum];
+  int32* cumsum_top_row = &dst_cumsum[0];
 
   for (int y = 0; y < height; ++y) {
     int top_y = ((y - radius - 1) >= 0) ? (y - radius - 1) : 0;
     int bot_y = ((y + radius) < height) ? (y + radius) : (height - 1);
-    int32* cumsum_top_row = &dst_cumsum[top_y * dst_stride32_cumsum];
-    int32* cumsum_bot_row = &dst_cumsum[bot_y * dst_stride32_cumsum];
+    int area = radius * (bot_y - top_y);
+
+    if (top_y) {
+      cumsum_top_row += dst_stride32_cumsum;
+      if (cumsum_top_row >= max_cumsum_bot_row) {
+        cumsum_top_row = dst_cumsum;
+      }
+    }
+    if ((y + radius) < height) {
+      int32* prev_cumsum_bot_row = cumsum_bot_row;
+      cumsum_bot_row += dst_stride32_cumsum;
+      if (cumsum_bot_row >= max_cumsum_bot_row) {
+        cumsum_bot_row = dst_cumsum;
+      }
+      ComputeCumulativeSumRow(src_argb, cumsum_bot_row, prev_cumsum_bot_row,
+                              width);
+      src_argb += src_stride_argb;
+    }
 
     // Left clipped.
-    int area = radius * (bot_y - top_y);
     int boxwidth = radius * 4;
     int x;
     for (x = 0; x < radius + 1; ++x) {
diff --git a/source/row_posix.cc b/source/row_posix.cc
index a1d499abc..0fdb0923b 100644
--- a/source/row_posix.cc
+++ b/source/row_posix.cc
@@ -2989,8 +2989,8 @@ void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
   "10:                                         \n"
     "movd      (%0),%%xmm2                     \n"
     "lea       0x4(%0),%0                      \n"
-    "punpcklbw %%xmm4,%%xmm2                   \n"
-    "punpcklwd %%xmm4,%%xmm2                   \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
     "paddd     %%xmm2,%%xmm0                   \n"
     "movdqu    (%1,%2,1),%%xmm2                \n"
     "paddd     %%xmm0,%%xmm2                   \n"
diff --git a/source/row_win.cc b/source/row_win.cc
index 2fd6f5c02..6d8bcecca 100644
--- a/source/row_win.cc
+++ b/source/row_win.cc
@@ -3187,8 +3187,8 @@ void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
   l1:
     movd       xmm2, dword ptr [eax]  // 1 argb pixel 4 bytes.
     lea        eax, [eax + 4]
-    punpcklbw  xmm2, xmm4
-    punpcklwd  xmm2, xmm4
+    punpcklbw  xmm2, xmm1
+    punpcklwd  xmm2, xmm1
     paddd      xmm0, xmm2
     movdqu     xmm2, [edx + esi]
     paddd      xmm2, xmm0