Experimental blur code with 32 bit fix.

This uses a new method of blurring that runs the three
passes of the box filter in a single pass. This implementation
currently only does 1x1 pixel at a time, but it should be simple
to expand to 4x4 pixels at a time.

On the  blur_10_normal_high_quality benchmark, the new is 7% faster
than the old code. For the blur_100.50_normal_high_quality
benchmark, the new code is 11% slower.

Bug: skia:
Change-Id: I847270906b0ceac1dfbf43ab5446756689ef660f
Reviewed-on: https://skia-review.googlesource.com/22700
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Herb Derby <herb@google.com>

5 files changed, 399 insertions, 38 deletions

diff --git a/gn/core.gni b/gn/core.gni
index 187a6de902..39a80fef85 100644
--- a/gn/core.gni
+++ b/gn/core.gni
@@ -178,6 +178,8 @@ skia_core_sources = [
   "$_src/core/SkMD5.h",
   "$_src/core/SkMallocPixelRef.cpp",
   "$_src/core/SkMask.cpp",
+  "$_src/core/SkMaskBlurFilter.h",
+  "$_src/core/SkMaskBlurFilter.cpp",
   "$_src/core/SkMaskCache.cpp",
   "$_src/core/SkMaskFilter.cpp",
   "$_src/core/SkMaskGamma.cpp",
diff --git a/src/core/SkMakeUnique.h b/src/core/SkMakeUnique.h
index 188eb05ff4..860ea2e8a7 100644
--- a/src/core/SkMakeUnique.h
+++ b/src/core/SkMakeUnique.h
@@ -18,6 +18,11 @@ std::unique_ptr<T> make_unique(Args&&... args) {
     return std::unique_ptr<T>(new T(std::forward<Args>(args)...));
 }
 
+template<typename T>
+std::unique_ptr<T> make_unique_default(size_t n) {
+    return std::unique_ptr<T>(new typename std::remove_extent<T>::type[n]);
+}
+
 }
 
 #endif  // SkMakeUnique_DEFINED
diff --git a/src/core/SkMaskBlurFilter.cpp b/src/core/SkMaskBlurFilter.cpp
new file mode 100644
index 0000000000..0f0286354c
--- /dev/null
+++ b/src/core/SkMaskBlurFilter.cpp
@@ -0,0 +1,267 @@
+/*
+ * Copyright 2017 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "SkMaskBlurFilter.h"
+
+#include <cmath>
+
+#include "SkMakeUnique.h"
+
+static const double kPi = 3.14159265358979323846264338327950288;
+
+static uint64_t weight_from_diameter(uint32_t d) {
+    uint64_t d2 = d * d;
+    uint64_t d3 = d2 * d;
+    if ((d&1) == 0) {
+        // d * d * (d + 1);
+        return d3 + d2;
+    }
+
+    return d3;
+}
+
+static uint32_t filter_window(double sigma) {
+    auto possibleWindow = static_cast<uint32_t>(floor(sigma * 3 * sqrt(2*kPi)/4 + 0.5));
+    return std::max(1u, possibleWindow);
+}
+
+SkMaskBlurFilter::FilterInfo::FilterInfo(double sigma)
+    : fFilterWindow{filter_window(sigma)}
+    , fScaledWeight{(static_cast<uint64_t>(1) << 32) / weight_from_diameter(fFilterWindow)} {}
+
+uint64_t SkMaskBlurFilter::FilterInfo::weight() const {
+    return weight_from_diameter(fFilterWindow);
+
+}
+uint32_t SkMaskBlurFilter::FilterInfo::borderSize() const {
+    if ((fFilterWindow&1) == 0) {
+        return 3 * (fFilterWindow / 2) - 1;
+    }
+    return 3 * (fFilterWindow / 2);
+}
+
+size_t SkMaskBlurFilter::FilterInfo::diameter(uint8_t pass) const {
+    SkASSERT(pass <= 2);
+
+    if ((fFilterWindow&1) == 0) {
+        // Handle even case.
+        switch (pass) {
+            case 0: return fFilterWindow;
+            case 1: return fFilterWindow;
+            case 2: return fFilterWindow+1;
+        }
+    }
+
+    return fFilterWindow;
+}
+
+uint64_t SkMaskBlurFilter::FilterInfo::scaledWeight() const {
+    return fScaledWeight;
+}
+
+SkMaskBlurFilter::SkMaskBlurFilter(double sigmaW, double sigmaH)
+    : fInfoW{sigmaW}, fInfoH{sigmaH}
+    , fBuffer0{skstd::make_unique_default<uint32_t[]>(bufferSize(0))}
+    , fBuffer1{skstd::make_unique_default<uint32_t[]>(bufferSize(1))}
+    , fBuffer2{skstd::make_unique_default<uint32_t[]>(bufferSize(2))} {
+}
+
+SkIPoint SkMaskBlurFilter::blur(const SkMask& src, SkMask* dst) const {
+
+    uint64_t weightW = fInfoW.weight();
+    uint64_t weightH = fInfoH.weight();
+
+    size_t borderW = fInfoW.borderSize();
+    size_t borderH = fInfoH.borderSize();
+
+    size_t srcW = src.fBounds.width();
+    size_t srcH = src.fBounds.height();
+
+    size_t dstW = srcW + 2 * borderW;
+    size_t dstH = srcH + 2 * borderH;
+
+    dst->fBounds.set(0, 0, dstW, dstH);
+    dst->fBounds.offset(src.fBounds.x(), src.fBounds.y());
+    dst->fBounds.offset(-SkTo<int32_t>(borderW), -SkTo<int32_t>(borderH));
+
+    dst->fImage = nullptr;
+    dst->fRowBytes = dstW;
+    dst->fFormat = SkMask::kA8_Format;
+
+    if (src.fImage == nullptr) {
+        return {SkTo<int32_t>(borderW), SkTo<int32_t>(borderH)};
+    }
+
+    dst->fImage = SkMask::AllocImage(dstW * dstH);
+
+    if (weightW > 1 && weightH > 1) {
+        // Blur both directions.
+        size_t tmpW = srcH;
+        size_t tmpH = dstW;
+        auto tmp = skstd::make_unique_default<uint8_t[]>(tmpW * tmpH);
+
+        // Blur horizontally, and transpose.
+        for (size_t y = 0; y < srcH; y++) {
+            auto srcStart = &src.fImage[y * src.fRowBytes];
+            auto tmpStart = &tmp[y];
+            this->blurOneScan(fInfoW,
+                              srcStart, 1, srcStart + srcW,
+                              tmpStart, tmpW, tmpStart + tmpW * tmpH);
+        }
+
+        // Blur vertically (scan in memory order because of the transposition),
+        // and transpose back to the original orientation.
+        for (size_t y = 0; y < tmpH; y++) {
+            auto tmpStart = &tmp[y * tmpW];
+            auto dstStart = &dst->fImage[y];
+            this->blurOneScan(fInfoH,
+                              tmpStart, 1, tmpStart + tmpW,
+                              dstStart, dst->fRowBytes, dstStart + dst->fRowBytes * dstH);
+        }
+    } else if (weightW > 1) {
+        // Blur only horizontally.
+
+        for (size_t y = 0; y < srcH; y++) {
+            auto srcStart = &src.fImage[y * src.fRowBytes];
+            auto dstStart = &dst->fImage[y * dst->fRowBytes];
+            this->blurOneScan(fInfoW,
+                              srcStart, 1, srcStart + srcW,
+                              dstStart, 1, dstStart + dstW);
+        }
+    } else if (weightH > 1) {
+        // Blur only vertically.
+
+        for (size_t x = 0; x < srcW; x++) {
+            auto srcStart = &src.fImage[x];
+            auto srcEnd   = &src.fImage[src.fRowBytes * srcH];
+            auto dstStart = &dst->fImage[x];
+            auto dstEnd   = &dst->fImage[dst->fRowBytes * dstH];
+            this->blurOneScan(fInfoH,
+                              srcStart, src.fRowBytes, srcEnd,
+                              dstStart, dst->fRowBytes, dstEnd);
+        }
+    } else {
+        // Copy to dst. No Blur.
+
+        for (size_t y = 0; y < srcH; y++) {
+            std::memcpy(&dst->fImage[y * dst->fRowBytes], &src.fImage[y * src.fRowBytes], dstW);
+        }
+    }
+
+    return {SkTo<int32_t>(borderW), SkTo<int32_t>(borderH)};
+}
+
+size_t SkMaskBlurFilter::bufferSize(uint8_t bufferPass) const {
+    return std::max(fInfoW.diameter(bufferPass), fInfoH.diameter(bufferPass)) - 1;
+}
+
+// Blur one horizontal scan into the dst.
+void SkMaskBlurFilter::blurOneScan(
+    FilterInfo info,
+    const uint8_t* src, size_t srcStride, const uint8_t* srcEnd,
+          uint8_t* dst, size_t dstStride,       uint8_t* dstEnd) const {
+
+    auto buffer0Begin = &fBuffer0[0];
+    auto buffer1Begin = &fBuffer1[0];
+    auto buffer2Begin = &fBuffer2[0];
+
+    auto buffer0Cursor = buffer0Begin;
+    auto buffer1Cursor = buffer1Begin;
+    auto buffer2Cursor = buffer2Begin;
+
+    auto buffer0End = &fBuffer0[0] + info.diameter(0) - 1;
+    auto buffer1End = &fBuffer1[0] + info.diameter(1) - 1;
+    auto buffer2End = &fBuffer2[0] + info.diameter(2) - 1;
+
+    std::memset(&fBuffer0[0], 0, (buffer0End - buffer0Begin) * sizeof(fBuffer0[0]));
+    std::memset(&fBuffer1[0], 0, (buffer1End - buffer1Begin) * sizeof(fBuffer1[0]));
+    std::memset(&fBuffer2[0], 0, (buffer2End - buffer2Begin) * sizeof(fBuffer2[0]));
+
+    uint32_t sum0 = 0;
+    uint32_t sum1 = 0;
+    uint32_t sum2 = 0;
+
+    const uint64_t half = static_cast<uint64_t>(1) << 31;
+
+    // Consume the source generating pixels.
+    for (auto srcCursor = src; srcCursor < srcEnd; dst += dstStride, srcCursor += srcStride) {
+        uint32_t s = *srcCursor;
+        sum0 += s;
+        sum1 += sum0;
+        sum2 += sum1;
+
+        *dst = SkTo<uint8_t>((info.scaledWeight() * sum2 + half) >> 32);
+
+        sum2 -= *buffer2Cursor;
+        *buffer2Cursor = sum1;
+        buffer2Cursor = (buffer2Cursor + 1) < buffer2End ? buffer2Cursor + 1 : &fBuffer2[0];
+
+        sum1 -= *buffer1Cursor;
+        *buffer1Cursor = sum0;
+        buffer1Cursor = (buffer1Cursor + 1) < buffer1End ? buffer1Cursor + 1 : &fBuffer1[0];
+
+        sum0 -= *buffer0Cursor;
+        *buffer0Cursor = s;
+        buffer0Cursor = (buffer0Cursor + 1) < buffer0End ? buffer0Cursor + 1 : &fBuffer0[0];
+    }
+
+    // This handles the case when both ends of the box are not between [src, srcEnd), and both
+    // are zero at that point.
+    for (auto i = 0; i < static_cast<ptrdiff_t>(2 * info.borderSize()) - (srcEnd - src); i++) {
+        uint32_t s = 0;
+        sum0 += s;
+        sum1 += sum0;
+        sum2 += sum1;
+
+        *dst = SkTo<uint8_t>((info.scaledWeight() * sum2 + half) >> 32);
+
+        sum2 -= *buffer2Cursor;
+        *buffer2Cursor = sum1;
+        buffer2Cursor = (buffer2Cursor + 1) < buffer2End ? buffer2Cursor + 1 : &fBuffer2[0];
+
+        sum1 -= *buffer1Cursor;
+        *buffer1Cursor = sum0;
+        buffer1Cursor = (buffer1Cursor + 1) < buffer1End ? buffer1Cursor + 1 : &fBuffer1[0];
+
+        sum0 -= *buffer0Cursor;
+        *buffer0Cursor = s;
+        buffer0Cursor = (buffer0Cursor + 1) < buffer0End ? buffer0Cursor + 1 : &fBuffer0[0];
+        dst += dstStride;
+    }
+
+    // Starting from the right, fill in the rest of the buffer.
+    std::memset(&fBuffer0[0], 0, (buffer0End - &fBuffer0[0]) * sizeof(fBuffer0[0]));
+    std::memset(&fBuffer1[0], 0, (buffer1End - &fBuffer1[0]) * sizeof(fBuffer1[0]));
+    std::memset(&fBuffer2[0], 0, (buffer2End - &fBuffer2[0]) * sizeof(fBuffer2[0]));
+
+    sum0 = sum1 = sum2 = 0;
+
+    uint8_t* dstCursor = dstEnd;
+    const uint8_t* srcCursor = srcEnd;
+    do {
+        dstCursor -= dstStride;
+        srcCursor -= srcStride;
+        uint32_t s = *srcCursor;
+        sum0 += s;
+        sum1 += sum0;
+        sum2 += sum1;
+
+        *dstCursor = SkTo<uint8_t>((info.scaledWeight() * sum2 + half) >> 32);
+
+        sum2 -= *buffer2Cursor;
+        *buffer2Cursor = sum1;
+        buffer2Cursor = (buffer2Cursor + 1) < buffer2End ? buffer2Cursor + 1 : &fBuffer2[0];
+
+        sum1 -= *buffer1Cursor;
+        *buffer1Cursor = sum0;
+        buffer1Cursor = (buffer1Cursor + 1) < buffer1End ? buffer1Cursor + 1 : &fBuffer1[0];
+
+        sum0 -= *buffer0Cursor;
+        *buffer0Cursor = s;
+        buffer0Cursor = (buffer0Cursor + 1) < buffer0End ? buffer0Cursor + 1 : &fBuffer0[0];
+    } while (dstCursor > dst);
+}
diff --git a/src/core/SkMaskBlurFilter.h b/src/core/SkMaskBlurFilter.h
new file mode 100644
index 0000000000..9becadca39
--- /dev/null
+++ b/src/core/SkMaskBlurFilter.h
@@ -0,0 +1,67 @@
+/*
+ * Copyright 2017 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkBlurMaskFilter_DEFINED
+#define SkBlurMaskFilter_DEFINED
+
+#include <algorithm>
+#include <memory>
+
+#include "SkMask.h"
+#include "SkTypes.h"
+
+// Implement a single channel Gaussian blur. The specifics for implementation are taken from:
+// https://drafts.fxtf.org/filters/#feGaussianBlurElement
+class SkMaskBlurFilter {
+public:
+    // Given a filter specified by sigma, generate various quantities.
+    class FilterInfo {
+    public:
+        explicit FilterInfo(double sigma);
+
+        // The final weight to divide by given a box size calculated from sigma accumulated for
+        // all three passes. For example, if the box size is 5, then the final weight for all
+        // three passes is 5^3 or 125.
+        uint64_t weight() const;
+
+        // The distance between the first value of the dst and the first value of the src.
+        uint32_t borderSize() const;
+
+        // The size of the box filter.
+        size_t   diameter(uint8_t) const;
+
+        // A factor used to simulate division using multiplication and shift.
+        uint64_t scaledWeight() const;
+
+    private:
+        const uint32_t fFilterWindow;
+        const uint64_t fScaledWeight;
+    };
+
+    // Create an object suitable for filtering an SkMask using a filter with width sigmaW and
+    // height sigmaH.
+    SkMaskBlurFilter(double sigmaW, double sigmaH);
+
+    // Given a src SkMask, generate dst SkMask returning the border width and height.
+    SkIPoint blur(const SkMask& src, SkMask* dst) const;
+
+private:
+    size_t bufferSize(uint8_t bufferPass) const;
+
+    void blurOneScan(FilterInfo gen,
+                     const uint8_t* src, size_t srcStride, const uint8_t* srcEnd,
+                           uint8_t* dst, size_t dstStride,       uint8_t* dstEnd) const;
+
+
+    const FilterInfo            fInfoW,
+                                fInfoH;
+    std::unique_ptr<uint32_t[]> fBuffer0,
+                                fBuffer1,
+                                fBuffer2;
+};
+
+#endif  // SkBlurMaskFilter_DEFINED
diff --git a/src/effects/SkBlurMask.cpp b/src/effects/SkBlurMask.cpp
index eee16313f6..fe59ab8269 100644
--- a/src/effects/SkBlurMask.cpp
+++ b/src/effects/SkBlurMask.cpp
@@ -7,6 +7,7 @@
 
 
 #include "SkBlurMask.h"
+#include "SkMaskBlurFilter.h"
 #include "SkMath.h"
 #include "SkTemplates.h"
 #include "SkEndian.h"
@@ -406,13 +407,7 @@ static int boxBlurInterp(const uint8_t* src, int src_y_stride, uint8_t* dst,
     return new_width;
 }
 
-static void get_adjusted_radii(SkScalar passRadius, int *loRadius, int *hiRadius)
-{
-    *loRadius = *hiRadius = SkScalarCeilToInt(passRadius);
-    if (SkIntToScalar(*hiRadius) - passRadius > 0.5f) {
-        *loRadius = *hiRadius - 1;
-    }
-}
+
 
 #include "SkColorPriv.h"
 
@@ -487,6 +482,17 @@ bool SkBlurMask::BoxBlur(SkMask* dst, const SkMask& src,
         return false;
     }
 
+    SkIPoint border;
+
+#ifdef SK_SUPPORT_LEGACY_MASK_BLUR
+
+    auto get_adjusted_radii = [](SkScalar passRadius, int *loRadius, int *hiRadius) {
+        *loRadius = *hiRadius = SkScalarCeilToInt(passRadius);
+        if (SkIntToScalar(*hiRadius) - passRadius > 0.5f) {
+            *loRadius = *hiRadius - 1;
+        }
+    };
+
     // Force high quality off for small radii (performance)
     if (!force_quality && sigma <= SkIntToScalar(2)) {
         quality = kLow_SkBlurQuality;
@@ -496,12 +502,12 @@ bool SkBlurMask::BoxBlur(SkMask* dst, const SkMask& src,
     if (kHigh_SkBlurQuality == quality) {
         // For the high quality path the 3 pass box blur kernel width is
         // 6*rad+1 while the full Gaussian width is 6*sigma.
-        passRadius = sigma - (1/6.0f);
+        passRadius = sigma - (1 / 6.0f);
     } else {
         // For the low quality path we only attempt to cover 3*sigma of the
         // Gaussian blur area (1.5*sigma on each side). The single pass box
         // blur's kernel size is 2*rad+1.
-        passRadius = 1.5f*sigma - 0.5f;
+        passRadius = 1.5f * sigma - 0.5f;
     }
 
     // highQuality: use three box blur passes as a cheap way
@@ -522,9 +528,8 @@ bool SkBlurMask::BoxBlur(SkMask* dst, const SkMask& src,
     int padx = passCount * rx;
     int pady = passCount * ry;
 
-    if (margin) {
-        margin->set(padx, pady);
-    }
+    border = {padx, pady};
+
     dst->fBounds.set(src.fBounds.fLeft - padx, src.fBounds.fTop - pady,
                      src.fBounds.fRight + padx, src.fBounds.fBottom + pady);
 
@@ -538,15 +543,15 @@ bool SkBlurMask::BoxBlur(SkMask* dst, const SkMask& src,
             return false;   // too big to allocate, abort
         }
 
-        int             sw = src.fBounds.width();
-        int             sh = src.fBounds.height();
-        const uint8_t*  sp = src.fImage;
-        uint8_t*        dp = SkMask::AllocImage(dstSize);
+        int sw = src.fBounds.width();
+        int sh = src.fBounds.height();
+        const uint8_t* sp = src.fImage;
+        uint8_t* dp = SkMask::AllocImage(dstSize);
         SkAutoTCallVProc<uint8_t, SkMask_FreeImage> autoCall(dp);
 
         // build the blurry destination
-        SkAutoTMalloc<uint8_t>  tmpBuffer(dstSize);
-        uint8_t*                tp = tmpBuffer.get();
+        SkAutoTMalloc<uint8_t> tmpBuffer(dstSize);
+        uint8_t* tp = tmpBuffer.get();
         int w = sw, h = sh;
 
         if (outerWeight == 255) {
@@ -555,33 +560,40 @@ bool SkBlurMask::BoxBlur(SkMask* dst, const SkMask& src,
             if (kHigh_SkBlurQuality == quality) {
                 // Do three X blurs, with a transpose on the final one.
                 w = boxBlur<false>(sp, src.fRowBytes, tp, loRadius, hiRadius, w, h);
-                w = boxBlur<false>(tp, w,             dp, hiRadius, loRadius, w, h);
-                w = boxBlur<true>(dp, w,             tp, hiRadius, hiRadius, w, h);
+                w = boxBlur<false>(tp, w, dp, hiRadius, loRadius, w, h);
+                w = boxBlur<true>(dp, w, tp, hiRadius, hiRadius, w, h);
                 // Do three Y blurs, with a transpose on the final one.
-                h = boxBlur<false>(tp, h,             dp, loRadius, hiRadius, h, w);
-                h = boxBlur<false>(dp, h,             tp, hiRadius, loRadius, h, w);
-                h = boxBlur<true>(tp, h,             dp, hiRadius, hiRadius, h, w);
+                h = boxBlur<false>(tp, h, dp, loRadius, hiRadius, h, w);
+                h = boxBlur<false>(dp, h, tp, hiRadius, loRadius, h, w);
+                h = boxBlur<true>(tp, h, dp, hiRadius, hiRadius, h, w);
             } else {
                 w = boxBlur<true>(sp, src.fRowBytes, tp, rx, rx, w, h);
-                h = boxBlur<true>(tp, h,             dp, ry, ry, h, w);
+                h = boxBlur<true>(tp, h, dp, ry, ry, h, w);
             }
         } else {
             if (kHigh_SkBlurQuality == quality) {
                 // Do three X blurs, with a transpose on the final one.
                 w = boxBlurInterp<false>(sp, src.fRowBytes, tp, rx, w, h, outerWeight);
-                w = boxBlurInterp<false>(tp, w,             dp, rx, w, h, outerWeight);
-                w = boxBlurInterp<true>(dp, w,             tp, rx, w, h, outerWeight);
+                w = boxBlurInterp<false>(tp, w, dp, rx, w, h, outerWeight);
+                w = boxBlurInterp<true>(dp, w, tp, rx, w, h, outerWeight);
                 // Do three Y blurs, with a transpose on the final one.
-                h = boxBlurInterp<false>(tp, h,             dp, ry, h, w, outerWeight);
-                h = boxBlurInterp<false>(dp, h,             tp, ry, h, w, outerWeight);
-                h = boxBlurInterp<true>(tp, h,             dp, ry, h, w, outerWeight);
+                h = boxBlurInterp<false>(tp, h, dp, ry, h, w, outerWeight);
+                h = boxBlurInterp<false>(dp, h, tp, ry, h, w, outerWeight);
+                h = boxBlurInterp<true>(tp, h, dp, ry, h, w, outerWeight);
             } else {
                 w = boxBlurInterp<true>(sp, src.fRowBytes, tp, rx, w, h, outerWeight);
-                h = boxBlurInterp<true>(tp, h,             dp, ry, h, w, outerWeight);
+                h = boxBlurInterp<true>(tp, h, dp, ry, h, w, outerWeight);
             }
         }
 
-        dst->fImage = dp;
+        dst->fImage = autoCall.release();
+    }
+#else
+    SkMaskBlurFilter blurFilter{sigma, sigma};
+    border = blurFilter.blur(src, dst);
+#endif  // SK_SUPPORT_LEGACY_MASK_BLUR
+
+    if (src.fImage != nullptr) {
         // if need be, alloc the "real" dst (same size as src) and copy/merge
         // the blur into it (applying the src)
         if (style == kInner_SkBlurStyle) {
@@ -590,17 +602,21 @@ bool SkBlurMask::BoxBlur(SkMask* dst, const SkMask& src,
             if (0 == srcSize) {
                 return false;   // too big to allocate, abort
             }
+            auto blur = dst->fImage;
             dst->fImage = SkMask::AllocImage(srcSize);
+            auto blurStart = &blur[border.x() + border.y() * dst->fRowBytes];
             merge_src_with_blur(dst->fImage, src.fRowBytes,
-                                sp, src.fRowBytes,
-                                dp + passCount * (rx + ry * dst->fRowBytes),
-                                dst->fRowBytes, sw, sh);
-            SkMask::FreeImage(dp);
+                                src.fImage, src.fRowBytes,
+                                blurStart,
+                                dst->fRowBytes,
+                                src.fBounds.width(), src.fBounds.height());
+            SkMask::FreeImage(blur);
         } else if (style != kNormal_SkBlurStyle) {
-            clamp_with_orig(dp + passCount * (rx + ry * dst->fRowBytes),
-                            dst->fRowBytes, sp, src.fRowBytes, sw, sh, style);
+            auto dstStart = &dst->fImage[border.x() + border.y() * dst->fRowBytes];
+            clamp_with_orig(dstStart,
+                            dst->fRowBytes, src.fImage, src.fRowBytes,
+                            src.fBounds.width(), src.fBounds.height(), style);
         }
-        (void)autoCall.release();
     }
 
     if (style == kInner_SkBlurStyle) {
@@ -608,6 +624,10 @@ bool SkBlurMask::BoxBlur(SkMask* dst, const SkMask& src,
         dst->fRowBytes = src.fRowBytes;
     }
 
+    if (margin != nullptr) {
+        *margin = border;
+    }
+
     return true;
 }