diff --git a/src/core/SkBlurImageFilter.cpp b/src/core/SkBlurImageFilter.cpp
index b87312f..8c78504 100644
--- a/src/core/SkBlurImageFilter.cpp
+++ b/src/core/SkBlurImageFilter.cpp
@@ -65,6 +65,11 @@
             SkIRect inputBounds, SkIRect dstBounds, const OutputProperties& outProps) const;
     #endif
 
+    sk_sp<SkSpecialImage> cpuFilter(
+            SkSpecialImage *source,
+            SkVector sigma, const sk_sp<SkSpecialImage> &input,
+            SkIRect inputBounds, SkIRect dstBounds) const;
+
     SkSize                      fSigma;
     SkBlurImageFilter::TileMode fTileMode;
 };
@@ -153,6 +158,23 @@
 }
 #endif
 
+static void get_box3_params(SkScalar s, int *kernelSize, int* kernelSize3, int *lowOffset,
+                            int *highOffset) {
+    float pi = SkScalarToFloat(SK_ScalarPI);
+    int d = static_cast<int>(floorf(SkScalarToFloat(s) * 3.0f * sqrtf(2.0f * pi) / 4.0f + 0.5f));
+    *kernelSize = d;
+    if (d % 2 == 1) {
+        *lowOffset = *highOffset = (d - 1) / 2;
+        *kernelSize3 = d;
+    } else {
+        *highOffset = d / 2;
+        *lowOffset = *highOffset - 1;
+        *kernelSize3 = d + 1;
+    }
+}
+
+#if !defined(SK_SUPPORT_LEGACY_BLUR_IMAGE)
+
 // This is defined by the SVG spec:
 // https://drafts.fxtf.org/filter-effects/#feGaussianBlurElement
 static int calculate_window(double sigma) {
@@ -393,10 +415,9 @@
     }
 }
 
-// TODO: Implement CPU backend for different fTileMode.
-static sk_sp<SkSpecialImage> cpu_blur(
+static sk_sp<SkSpecialImage> combined_pass_blur(
         SkVector sigma,
-        SkSpecialImage *source, const sk_sp<SkSpecialImage> &input,
+        SkSpecialImage* source, const sk_sp<SkSpecialImage>& input,
         SkIRect srcBounds, SkIRect dstBounds) {
     SkBitmap inputBM;
 
@@ -512,6 +533,7 @@
                                                           dstBounds.height()),
                                           dst, &source->props());
 }
+#endif
 
 sk_sp<SkSpecialImage> SkBlurImageFilterImpl::onFilterImage(SkSpecialImage* source,
                                                            const Context& ctx,
@@ -563,7 +585,11 @@
         if (sigma.x() < kZeroWindow && sigma.y() < kZeroWindow) {
             result = input->makeSubset(inputBounds);
         } else {
-            result = cpu_blur(sigma, source, input, inputBounds, dstBounds);
+            #if defined(SK_SUPPORT_LEGACY_BLUR_IMAGE)
+                result = this->cpuFilter(source, sigma, input, inputBounds, dstBounds);
+            #else
+                result = combined_pass_blur(sigma, source, input, inputBounds, dstBounds);
+            #endif
         }
     }
 
@@ -633,6 +659,87 @@
 }
 #endif
 
+// TODO: Implement CPU backend for different fTileMode.
+sk_sp<SkSpecialImage> SkBlurImageFilterImpl::cpuFilter(
+        SkSpecialImage *source,
+        SkVector sigma, const sk_sp<SkSpecialImage> &input,
+        SkIRect inputBounds, SkIRect dstBounds) const
+{
+    int kernelSizeX, kernelSizeX3, lowOffsetX, highOffsetX;
+    int kernelSizeY, kernelSizeY3, lowOffsetY, highOffsetY;
+    get_box3_params(sigma.x(), &kernelSizeX, &kernelSizeX3, &lowOffsetX, &highOffsetX);
+    get_box3_params(sigma.y(), &kernelSizeY, &kernelSizeY3, &lowOffsetY, &highOffsetY);
+
+    SkBitmap inputBM;
+
+    if (!input->getROPixels(&inputBM) && inputBM.colorType() != kN32_SkColorType) {
+        return nullptr;
+    }
+
+    SkImageInfo info = SkImageInfo::Make(dstBounds.width(), dstBounds.height(),
+                                         inputBM.colorType(), inputBM.alphaType());
+
+    SkBitmap tmp, dst;
+    if (!tmp.tryAllocPixels(info) || !dst.tryAllocPixels(info)) {
+        return nullptr;
+    }
+
+    // Get ready to blur.
+    const SkPMColor* s = inputBM.getAddr32(inputBounds.x(), inputBounds.y());
+          SkPMColor* t = tmp.getAddr32(0, 0);
+          SkPMColor* d = dst.getAddr32(0, 0);
+
+    // Shift everything from being relative to the orignal input bounds to the destination bounds.
+    inputBounds.offset(-dstBounds.x(), -dstBounds.y());
+    dstBounds.offset(-dstBounds.x(), -dstBounds.y());
+
+    int w  = dstBounds.width(),
+        h  = dstBounds.height(),
+        sw = inputBM.rowBytesAsPixels();
+
+    SkIRect inputBoundsT = SkIRect::MakeLTRB(inputBounds.top(), inputBounds.left(),
+                                             inputBounds.bottom(), inputBounds.right());
+    SkIRect dstBoundsT = SkIRect::MakeWH(dstBounds.height(), dstBounds.width());
+
+    /**
+     *
+     * In order to make memory accesses cache-friendly, we reorder the passes to
+     * use contiguous memory reads wherever possible.
+     *
+     * For example, the 6 passes of the X-and-Y blur case are rewritten as
+     * follows. Instead of 3 passes in X and 3 passes in Y, we perform
+     * 2 passes in X, 1 pass in X transposed to Y on write, 2 passes in X,
+     * then 1 pass in X transposed to Y on write.
+     *
+     * +----+       +----+       +----+        +---+       +---+       +---+        +----+
+     * + AB + ----> | AB | ----> | AB | -----> | A | ----> | A | ----> | A | -----> | AB |
+     * +----+ blurX +----+ blurX +----+ blurXY | B | blurX | B | blurX | B | blurXY +----+
+     *                                         +---+       +---+       +---+
+     *
+     * In this way, two of the y-blurs become x-blurs applied to transposed
+     * images, and all memory reads are contiguous.
+     */
+    if (kernelSizeX > 0 && kernelSizeY > 0) {
+        SkOpts::box_blur_xx(s, sw,  inputBounds,  t, kernelSizeX,  lowOffsetX,  highOffsetX, w, h);
+        SkOpts::box_blur_xx(t,  w,  dstBounds,    d, kernelSizeX,  highOffsetX, lowOffsetX,  w, h);
+        SkOpts::box_blur_xy(d,  w,  dstBounds,    t, kernelSizeX3, highOffsetX, highOffsetX, w, h);
+        SkOpts::box_blur_xx(t,  h,  dstBoundsT,   d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
+        SkOpts::box_blur_xx(d,  h,  dstBoundsT,   t, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
+        SkOpts::box_blur_xy(t,  h,  dstBoundsT,   d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
+    } else if (kernelSizeX > 0) {
+        SkOpts::box_blur_xx(s, sw,  inputBounds,  d, kernelSizeX,  lowOffsetX,  highOffsetX, w, h);
+        SkOpts::box_blur_xx(d,  w,  dstBounds,    t, kernelSizeX,  highOffsetX, lowOffsetX,  w, h);
+        SkOpts::box_blur_xx(t,  w,  dstBounds,    d, kernelSizeX3, highOffsetX, highOffsetX, w, h);
+    } else if (kernelSizeY > 0) {
+        SkOpts::box_blur_yx(s, sw,  inputBoundsT, d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
+        SkOpts::box_blur_xx(d,  h,  dstBoundsT,   t, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
+        SkOpts::box_blur_xy(t,  h,  dstBoundsT,   d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
+    }
+
+    return SkSpecialImage::MakeFromRaster(SkIRect::MakeSize(dstBounds.size()),
+                                          dst, &source->props());
+}
+
 sk_sp<SkImageFilter> SkBlurImageFilterImpl::onMakeColorSpace(SkColorSpaceXformer* xformer)
 const {
     SkASSERT(1 == this->countInputs());
diff --git a/src/core/SkOpts.cpp b/src/core/SkOpts.cpp
index 8a3d30e..4131546 100644
--- a/src/core/SkOpts.cpp
+++ b/src/core/SkOpts.cpp
@@ -38,6 +38,7 @@
 
 #include "SkBlitMask_opts.h"
 #include "SkBlitRow_opts.h"
+#include "SkBlurImageFilter_opts.h"
 #include "SkChecksum_opts.h"
 #include "SkMorphologyImageFilter_opts.h"
 #include "SkSwizzler_opts.h"
@@ -52,6 +53,10 @@
 #define DEFINE_DEFAULT(name) decltype(name) name = SK_OPTS_NS::name
     DEFINE_DEFAULT(create_xfermode);
 
+    DEFINE_DEFAULT(box_blur_xx);
+    DEFINE_DEFAULT(box_blur_xy);
+    DEFINE_DEFAULT(box_blur_yx);
+
     DEFINE_DEFAULT(dilate_x);
     DEFINE_DEFAULT(dilate_y);
     DEFINE_DEFAULT( erode_x);
diff --git a/src/core/SkOpts.h b/src/core/SkOpts.h
index 6e6e168..329a75c 100644
--- a/src/core/SkOpts.h
+++ b/src/core/SkOpts.h
@@ -25,6 +25,9 @@
     // May return nullptr if we haven't specialized the given Mode.
     extern SkXfermode* (*create_xfermode)(SkBlendMode);
 
+    typedef void (*BoxBlur)(const SkPMColor*, int, const SkIRect& srcBounds, SkPMColor*, int, int, int, int, int);
+    extern BoxBlur box_blur_xx, box_blur_xy, box_blur_yx;
+
     typedef void (*Morph)(const SkPMColor*, SkPMColor*, int, int, int, int, int);
     extern Morph dilate_x, dilate_y, erode_x, erode_y;
 
diff --git a/src/opts/SkBlurImageFilter_opts.h b/src/opts/SkBlurImageFilter_opts.h
new file mode 100644
index 0000000..08c2121
--- /dev/null
+++ b/src/opts/SkBlurImageFilter_opts.h
@@ -0,0 +1,322 @@
+/*
+ * Copyright 2015 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkBlurImageFilter_opts_DEFINED
+#define SkBlurImageFilter_opts_DEFINED
+
+#include "SkColorData.h"
+#include "SkRect.h"
+
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    #include <immintrin.h>
+#endif
+
+namespace SK_OPTS_NS {
+
+enum class BlurDirection { kX, kY };
+
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
+        // RGBA -> R000 G000 B000 A000
+        static inline __m128i expand(SkPMColor p) {
+            return _mm_cvtepu8_epi32(_mm_cvtsi32_si128(p));
+        };
+        // 000R 000G 000B 000A -> RGBA
+        static inline SkPMColor repack(__m128i p) {
+            const char _ = ~0;  // Don't care what ends up in these bytes.  This zeros them.
+            p = _mm_shuffle_epi8(p, _mm_setr_epi8(3,7,11,15, _,_,_,_, _,_,_,_, _,_,_,_));
+            return _mm_cvtsi128_si32(p);
+        };
+        #define mullo_epi32 _mm_mullo_epi32
+
+    #else
+        static inline __m128i expand(int p) {
+            auto result = _mm_cvtsi32_si128(p);
+            result = _mm_unpacklo_epi8 (result, _mm_setzero_si128());
+            result = _mm_unpacklo_epi16(result, _mm_setzero_si128());
+            return result;
+        };
+        static inline SkPMColor repack(__m128i p) {
+            p = _mm_srli_epi32(p, 24);  // R000 G000 B000 A000
+            p = _mm_packs_epi32(p, p);  // R0G0 B0A0 xxxx xxxx
+            p = _mm_packus_epi16(p, p); // RGBA xxxx xxxx xxxx
+            return _mm_cvtsi128_si32(p);
+        };
+
+        // _mm_mullo_epi32 is not available, so use the standard trick to emulate it.
+        static inline __m128i mullo_epi32(__m128i a, __m128i b) {
+            __m128i p02 = _mm_mul_epu32(a, b),
+                    p13 = _mm_mul_epu32(_mm_srli_si128(a, 4),
+                                        _mm_srli_si128(b, 4));
+            return _mm_unpacklo_epi32(_mm_shuffle_epi32(p02, _MM_SHUFFLE(0,0,2,0)),
+                                      _mm_shuffle_epi32(p13, _MM_SHUFFLE(0,0,2,0)));
+        };
+    #endif
+
+    #define INIT_SCALE const __m128i scale = _mm_set1_epi32((1 << 24) / kernelSize);
+    #define INIT_HALF const __m128i half = _mm_set1_epi32(1 << 23);
+    #define INIT_SUMS __m128i sum = _mm_setzero_si128();
+    #define INCREMENT_SUMS(c) sum = _mm_add_epi32(sum, expand(c))
+    #define DECREMENT_SUMS(c) sum = _mm_sub_epi32(sum, expand(c))
+    #define STORE_SUMS *dptr = repack(_mm_add_epi32(mullo_epi32(sum, scale), half));
+    #define DOUBLE_ROW_OPTIMIZATION /*none*/
+
+#elif defined(SK_ARM_HAS_NEON)
+
+    // val = (sum * scale * 2 + 0x8000) >> 16
+    #define STORE_SUMS_DOUBLE \
+        uint16x8_t resultPixels = vreinterpretq_u16_s16(vqrdmulhq_s16( \
+            vreinterpretq_s16_u16(sum), vreinterpretq_s16_u16(scale))); \
+        if (dstDirection == BlurDirection::kX) { \
+            uint32x2_t px2 = vreinterpret_u32_u8(vmovn_u16(resultPixels)); \
+            vst1_lane_u32(dptr +     0, px2, 0); \
+            vst1_lane_u32(dptr + width, px2, 1); \
+        } else { \
+            vst1_u8((uint8_t*)dptr, vmovn_u16(resultPixels)); \
+        }
+
+    #define INCREMENT_SUMS_DOUBLE(p) sum = vaddw_u8(sum, load_2_pixels(p))
+    #define DECREMENT_SUMS_DOUBLE(p) sum = vsubw_u8(sum, load_2_pixels(p))
+
+    // Fast path for kernel sizes between 2 and 127, working on two rows at a time.
+    template<BlurDirection srcDirection, BlurDirection dstDirection>
+    static int box_blur_double(const SkPMColor** src, int srcStride, const SkIRect& srcBounds,
+                               SkPMColor** dst, int kernelSize,
+                               int leftOffset, int rightOffset, int width, int height) {
+        // Load 2 pixels from adjacent rows.
+        auto load_2_pixels = [&](const SkPMColor* s) {
+            if (srcDirection == BlurDirection::kX) {
+                // 10% faster by adding these 2 prefetches
+                SK_PREFETCH(s + 16);
+                SK_PREFETCH(s + 16 + srcStride);
+                auto one = vld1_lane_u32(s +         0, vdup_n_u32(0), 0),
+                     two = vld1_lane_u32(s + srcStride,           one, 1);
+                return vreinterpret_u8_u32(two);
+            } else {
+                return vld1_u8((uint8_t*)s);
+            }
+        };
+        int left = srcBounds.left();
+        int right = srcBounds.right();
+        int top = srcBounds.top();
+        int bottom = srcBounds.bottom();
+        int incrementStart = SkMax32(left - rightOffset - 1, left - right);
+        int incrementEnd = SkMax32(right - rightOffset - 1, 0);
+        int decrementStart = SkMin32(left + leftOffset, width);
+        int decrementEnd = SkMin32(right + leftOffset, width);
+        const int srcStrideX = srcDirection == BlurDirection::kX ? 1 : srcStride;
+        const int dstStrideX = dstDirection == BlurDirection::kX ? 1 : height;
+        const int srcStrideY = srcDirection == BlurDirection::kX ? srcStride : 1;
+        const int dstStrideY = dstDirection == BlurDirection::kX ? width : 1;
+        const uint16x8_t scale = vdupq_n_u16((1 << 15) / kernelSize);
+
+        for (; bottom - top >= 2; top += 2) {
+            uint16x8_t sum = vdupq_n_u16(0);
+            const SkPMColor* lptr = *src;
+            const SkPMColor* rptr = *src;
+            SkPMColor* dptr = *dst;
+            int x;
+            for (x = incrementStart; x < 0; ++x) {
+                INCREMENT_SUMS_DOUBLE(rptr);
+                rptr += srcStrideX;
+            }
+            // Clear to zero when sampling to the left our domain. "sum" is zero here because we
+            // initialized it above, and the preceeding loop has no effect in this case.
+            for (x = 0; x < incrementStart; ++x) {
+                STORE_SUMS_DOUBLE
+                dptr += dstStrideX;
+            }
+            for (; x < decrementStart && x < incrementEnd; ++x) {
+                STORE_SUMS_DOUBLE
+                dptr += dstStrideX;
+                INCREMENT_SUMS_DOUBLE(rptr);
+                rptr += srcStrideX;
+            }
+            for (x = decrementStart; x < incrementEnd; ++x) {
+                STORE_SUMS_DOUBLE
+                dptr += dstStrideX;
+                INCREMENT_SUMS_DOUBLE(rptr);
+                rptr += srcStrideX;
+                DECREMENT_SUMS_DOUBLE(lptr);
+                lptr += srcStrideX;
+            }
+            for (x = incrementEnd; x < decrementStart; ++x) {
+                STORE_SUMS_DOUBLE
+                dptr += dstStrideX;
+            }
+            for (; x < decrementEnd; ++x) {
+                STORE_SUMS_DOUBLE
+                dptr += dstStrideX;
+                DECREMENT_SUMS_DOUBLE(lptr);
+                lptr += srcStrideX;
+            }
+            // Clear to zero when sampling to the right of our domain. "sum" is
+            // zero here because we added on then subtracted off all of the pixels, leaving zero.
+            for (; x < width; ++x) {
+                STORE_SUMS_DOUBLE
+                dptr += dstStrideX;
+            }
+            *src += srcStrideY * 2;
+            *dst += dstStrideY * 2;
+        }
+        return top;
+    }
+
+    // RGBA -> R0G0 B0A0
+    static inline uint16x4_t expand(SkPMColor p) {
+        return vget_low_u16(vmovl_u8(vreinterpret_u8_u32(vdup_n_u32(p))));
+    };
+
+    #define INIT_SCALE const uint32x4_t scale = vdupq_n_u32((1 << 24) / kernelSize);
+    #define INIT_HALF const uint32x4_t half = vdupq_n_u32(1 << 23);
+    #define INIT_SUMS uint32x4_t sum = vdupq_n_u32(0);
+    #define INCREMENT_SUMS(c) sum = vaddw_u16(sum, expand(c));
+    #define DECREMENT_SUMS(c) sum = vsubw_u16(sum, expand(c));
+
+    #define STORE_SUMS \
+        uint32x4_t result = vmlaq_u32(half, sum, scale); \
+        uint16x4_t result16 = vqshrn_n_u32(result, 16); \
+        uint8x8_t result8 = vqshrn_n_u16(vcombine_u16(result16, result16), 8); \
+        vst1_lane_u32(dptr, vreinterpret_u32_u8(result8), 0);
+
+    #define DOUBLE_ROW_OPTIMIZATION \
+        if (1 < kernelSize && kernelSize < 128) { \
+            top = box_blur_double<srcDirection, dstDirection>(&src, srcStride, srcBounds, &dst, \
+                                                              kernelSize, \
+                                                              leftOffset, rightOffset, \
+                                                              width, height); \
+        }
+
+#else  // Neither NEON nor >=SSE2.
+
+    #define INIT_SCALE uint32_t scale = (1 << 24) / kernelSize;
+    #define INIT_HALF  uint32_t half = 1 << 23;
+    #define INIT_SUMS int sumA = 0, sumR = 0, sumG = 0, sumB = 0;
+    #define INCREMENT_SUMS(c) \
+        sumA += SkGetPackedA32(c); \
+        sumR += SkGetPackedR32(c); \
+        sumG += SkGetPackedG32(c); \
+        sumB += SkGetPackedB32(c)
+    #define DECREMENT_SUMS(c) \
+        sumA -= SkGetPackedA32(c); \
+        sumR -= SkGetPackedR32(c); \
+        sumG -= SkGetPackedG32(c); \
+        sumB -= SkGetPackedB32(c)
+    #define STORE_SUMS \
+        *dptr = SkPackARGB32((sumA * scale + half) >> 24, \
+                             (sumR * scale + half) >> 24, \
+                             (sumG * scale + half) >> 24, \
+                             (sumB * scale + half) >> 24);
+    #define DOUBLE_ROW_OPTIMIZATION
+
+#endif
+
+template<BlurDirection srcDirection, BlurDirection dstDirection>
+/*not static*/ inline
+void box_blur(const SkPMColor* src, int srcStride, const SkIRect& srcBounds, SkPMColor* dst,
+              int kernelSize, int leftOffset, int rightOffset, int width, int height) {
+    int left = srcBounds.left();
+    int right = srcBounds.right();
+    int top = srcBounds.top();
+    int bottom = srcBounds.bottom();
+    int incrementStart = SkMax32(left - rightOffset - 1, left - right);
+    int incrementEnd = SkMax32(right - rightOffset - 1, 0);
+    int decrementStart = SkMin32(left + leftOffset, width);
+    int decrementEnd = SkMin32(right + leftOffset, width);
+    int srcStrideX = srcDirection == BlurDirection::kX ? 1 : srcStride;
+    int dstStrideX = dstDirection == BlurDirection::kX ? 1 : height;
+    int srcStrideY = srcDirection == BlurDirection::kX ? srcStride : 1;
+    int dstStrideY = dstDirection == BlurDirection::kX ? width : 1;
+    INIT_SCALE
+    INIT_HALF
+
+    // Clear to zero when sampling above our domain.
+    for (int y = 0; y < top; y++) {
+        SkColor* dptr = dst;
+        for (int x = 0; x < width; ++x) {
+            *dptr = 0;
+            dptr += dstStrideX;
+        }
+        dst += dstStrideY;
+    }
+
+    DOUBLE_ROW_OPTIMIZATION
+
+    for (int y = top; y < bottom; ++y) {
+        INIT_SUMS
+        const SkPMColor* lptr = src;
+        const SkPMColor* rptr = src;
+        SkColor* dptr = dst;
+        int x;
+        for (x = incrementStart; x < 0; ++x) {
+            INCREMENT_SUMS(*rptr);
+            rptr += srcStrideX;
+            if (srcDirection == BlurDirection::kY) {
+                SK_PREFETCH(rptr);
+            }
+        }
+        // Clear to zero when sampling to the left of our domain.
+        for (x = 0; x < incrementStart; ++x) {
+            *dptr = 0;
+            dptr += dstStrideX;
+        }
+        for (; x < decrementStart && x < incrementEnd; ++x) {
+            STORE_SUMS
+            dptr += dstStrideX;
+            INCREMENT_SUMS(*rptr);
+            rptr += srcStrideX;
+            if (srcDirection == BlurDirection::kY) {
+                SK_PREFETCH(rptr);
+            }
+        }
+        for (x = decrementStart; x < incrementEnd; ++x) {
+            STORE_SUMS
+            dptr += dstStrideX;
+            INCREMENT_SUMS(*rptr);
+            rptr += srcStrideX;
+            if (srcDirection == BlurDirection::kY) {
+                SK_PREFETCH(rptr);
+            }
+            DECREMENT_SUMS(*lptr);
+            lptr += srcStrideX;
+        }
+        for (x = incrementEnd; x < decrementStart; ++x) {
+            STORE_SUMS
+            dptr += dstStrideX;
+        }
+        for (; x < decrementEnd; ++x) {
+            STORE_SUMS
+            dptr += dstStrideX;
+            DECREMENT_SUMS(*lptr);
+            lptr += srcStrideX;
+        }
+        // Clear to zero when sampling to the right of our domain.
+        for (; x < width; ++x) {
+            *dptr = 0;
+            dptr += dstStrideX;
+        }
+        src += srcStrideY;
+        dst += dstStrideY;
+    }
+    // Clear to zero when sampling below our domain.
+    for (int y = bottom; y < height; ++y) {
+        SkColor* dptr = dst;
+        for (int x = 0; x < width; ++x) {
+            *dptr = 0;
+            dptr += dstStrideX;
+        }
+        dst += dstStrideY;
+    }
+}
+
+static auto box_blur_xx = &box_blur<BlurDirection::kX, BlurDirection::kX>,
+            box_blur_xy = &box_blur<BlurDirection::kX, BlurDirection::kY>,
+            box_blur_yx = &box_blur<BlurDirection::kY, BlurDirection::kX>;
+
+}  // namespace SK_OPTS_NS
+
+#endif
diff --git a/src/opts/SkOpts_sse41.cpp b/src/opts/SkOpts_sse41.cpp
index b382799..6fd7c3c 100644
--- a/src/opts/SkOpts_sse41.cpp
+++ b/src/opts/SkOpts_sse41.cpp
@@ -8,10 +8,14 @@
 #include "SkOpts.h"
 
 #define SK_OPTS_NS sse41
+#include "SkBlurImageFilter_opts.h"
 #include "SkBlitRow_opts.h"
 
 namespace SkOpts {
     void Init_sse41() {
+        box_blur_xx          = sse41::box_blur_xx;
+        box_blur_xy          = sse41::box_blur_xy;
+        box_blur_yx          = sse41::box_blur_yx;
         blit_row_s32a_opaque = sse41::blit_row_s32a_opaque;
     }
 }