Cache blur mask for rects which can not break into nine-patch

With this CL performance improves:
blurrectsnonninepatch   42.4us -> 20.5us        0.48x

BUG=431021,skia:3118

Review URL: https://codereview.chromium.org/729463002

3 files changed, 104 insertions, 2 deletions

diff --git a/src/core/SkMaskCache.cpp b/src/core/SkMaskCache.cpp
index b4b8b4faf3..207ea0e77d 100644
--- a/src/core/SkMaskCache.cpp
+++ b/src/core/SkMaskCache.cpp
@@ -18,6 +18,28 @@ struct MaskValue {
 namespace {
 static unsigned gRRectBlurKeyNamespaceLabel;
 
+static bool copy_cacheddata_to_mask(SkCachedData* data, SkMask* mask) {
+    const size_t size = data->size();
+    SkASSERT(mask->computeTotalImageSize() <= size);
+
+    mask->fImage = SkMask::AllocImage(size);
+    if (mask->fImage) {
+        memcpy(mask->fImage, data->data(), size);
+        return true;
+    }
+    return false;
+}
+
+static SkCachedData* copy_mask_to_cacheddata(const SkMask& mask) {
+    const size_t size = mask.computeTotalImageSize();
+    SkCachedData* data = SkResourceCache::NewCachedData(size);
+    if (data) {
+        memcpy(data->writable_data(), mask.fImage, size);
+        return data;
+    }
+    return NULL;
+}
+
 struct RRectBlurKey : public SkResourceCache::Key {
 public:
     RRectBlurKey(SkScalar sigma, const SkRRect& rrect, SkBlurStyle style, SkBlurQuality quality)
@@ -90,6 +112,20 @@ void SkMaskCache::Add(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
     return CHECK_LOCAL(localCache, add, Add, SkNEW_ARGS(RRectBlurRec, (key, mask, data)));
 }
 
+bool SkMaskCache::FindAndCopy(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
+                              const SkRRect& rrect, SkMask* mask) {
+    SkAutoTUnref<SkCachedData> data(SkMaskCache::FindAndRef(sigma, style, quality, rrect, mask));
+    return data.get() && copy_cacheddata_to_mask(data, mask);
+}
+
+void SkMaskCache::AddAndCopy(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
+                             const SkRRect& rrect, const SkMask& mask) {
+    SkAutoTUnref<SkCachedData> data(copy_mask_to_cacheddata(mask));
+    if (data.get()) {
+        SkMaskCache::Add(sigma, style, quality, rrect, mask, data);
+    }
+}
+
 //////////////////////////////////////////////////////////////////////////////////////////
 
 namespace {
@@ -181,3 +217,17 @@ void SkMaskCache::Add(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
     RectsBlurKey key(sigma, style, quality, rects, count);
     return CHECK_LOCAL(localCache, add, Add, SkNEW_ARGS(RectsBlurRec, (key, mask, data)));
 }
+
+bool SkMaskCache::FindAndCopy(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
+                              const SkRect rects[], int count, SkMask* mask) {
+    SkAutoTUnref<SkCachedData> data(SkMaskCache::FindAndRef(sigma, style, quality, rects, count, mask));
+    return data.get() && copy_cacheddata_to_mask(data, mask);
+}
+
+void SkMaskCache::AddAndCopy(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
+                             const SkRect rects[], int count, const SkMask& mask) {
+    SkAutoTUnref<SkCachedData> data(copy_mask_to_cacheddata(mask));
+    if (data.get()) {
+        SkMaskCache::Add(sigma, style, quality, rects, count, mask, data);
+    }
+}
diff --git a/src/core/SkMaskCache.h b/src/core/SkMaskCache.h
index f98387b206..2cfb971dde 100644
--- a/src/core/SkMaskCache.h
+++ b/src/core/SkMaskCache.h
@@ -39,6 +39,25 @@ public:
     static void Add(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
                     const SkRect rects[], int count, const SkMask& mask, SkCachedData* data,
                     SkResourceCache* localCache = NULL);
+
+    /**
+     * On success, set mask with cached value, allocate memory for mask->fImage,
+     * copy pixels from SkCachedData in the cache to mask->fImage, then return true.
+     *
+     * On failure, return false, no memory allocated for mask->fImage.
+     */
+    static bool FindAndCopy(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
+                            const SkRRect& rrect, SkMask* mask);
+    static bool FindAndCopy(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
+                            const SkRect rects[], int count, SkMask* mask);
+
+    /**
+     * Create a new SkCachedData, copy pixels from mask.fImage to it, then add it into cache.
+     */
+    static void AddAndCopy(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
+                           const SkRRect& rrect, const SkMask& mask);
+    static void AddAndCopy(SkScalar sigma, SkBlurStyle style, SkBlurQuality quality,
+                           const SkRect rects[], int count, const SkMask& mask);
 };
 
 #endif
diff --git a/src/core/SkMaskFilter.cpp b/src/core/SkMaskFilter.cpp
index c9783e76d6..3a4e3a41e9 100644
--- a/src/core/SkMaskFilter.cpp
+++ b/src/core/SkMaskFilter.cpp
@@ -10,6 +10,7 @@
 #include "SkMaskFilter.h"
 #include "SkBlitter.h"
 #include "SkDraw.h"
+#include "SkMaskCache.h"
 #include "SkRasterClip.h"
 #include "SkRRect.h"
 #include "SkTypes.h"
@@ -263,9 +264,41 @@ bool SkMaskFilter::filterPath(const SkPath& devPath, const SkMatrix& matrix,
     }
     SkAutoMaskFreeImage autoSrc(srcM.fImage);
 
-    if (!this->filterMask(&dstM, srcM, matrix, NULL)) {
-        return false;
+    BlurRec rec;
+    if (this->asABlur(&rec) && rectCount) {
+        SkScalar scaledSigma = matrix.mapRadius(rec.fSigma);
+        if (!SkMaskCache::FindAndCopy(scaledSigma, rec.fStyle, rec.fQuality,
+                                      rects, rectCount, &dstM)) {
+            if (!this->filterMask(&dstM, srcM, matrix, NULL)) {
+                return false;
+            }
+            SkMaskCache::AddAndCopy(scaledSigma, rec.fStyle, rec.fQuality, rects, rectCount, dstM);
+        } else {
+            // Compute the correct bounds of dst mask if dst mask is got from cache.
+            SkMask tmpSrc, tmpDst;
+            tmpSrc = srcM;
+            tmpSrc.fImage = NULL;
+            if (!this->filterMask(&tmpDst, tmpSrc, matrix, NULL)) {
+                return false;
+            }
+
+            // Fallback to original calculation if size of bounds is different with
+            // size of the cached mask.
+            if (dstM.fBounds.width() != tmpDst.fBounds.width() ||
+                dstM.fBounds.height() != tmpDst.fBounds.height()) {
+                if (!this->filterMask(&dstM, srcM, matrix, NULL)) {
+                    return false;
+                }
+            } else {
+                dstM.fBounds = tmpDst.fBounds;
+            }
+        }
+    } else {
+        if (!this->filterMask(&dstM, srcM, matrix, NULL)) {
+            return false;
+        }
     }
+
     SkAutoMaskFreeImage autoDst(dstM.fImage);
 
     // if we get here, we need to (possibly) resolve the clip and blitter