Implement SSE2-based implementations of the morphology filters (dilate &

src/effects/SkMorphologyImageFilter.cpp

Index: src/effects/SkMorphologyImageFilter.cpp
diff --git a/src/effects/SkMorphologyImageFilter.cpp b/src/effects/SkMorphologyImageFilter.cpp
index df015cac7356fb2cf4d9b1c67612cf9c155fb7aa..8adeb38a2bd6a6f9da685f0197725347d9357327 100644
--- a/src/effects/SkMorphologyImageFilter.cpp
+++ b/src/effects/SkMorphologyImageFilter.cpp
@@ -10,6 +10,7 @@
 #include "SkColorPriv.h"
 #include "SkFlattenableBuffers.h"
 #include "SkRect.h"
+#include "SkMorphology_opts.h"
 #if SK_SUPPORT_GPU
 #include "GrContext.h"
 #include "GrTexture.h"
@@ -38,11 +39,15 @@ void SkMorphologyImageFilter::flatten(SkFlattenableWriteBuffer& buffer) const {
     buffer.writeInt(fRadius.fHeight);
 }
 
+template<bool inX>

[reed1, 2013/10/30 20:14:28]

very nice code sharing!

bike shed : seeing erode<true> and erode<false> is not self-documenting. A
little enum like

enum MorphDirection {
  kX, kY
};

would be clearer, but perhaps that can be for another day.

[Stephen White, 2013/10/30 20:48:06]

Done.

 static void erode(const SkPMColor* src, SkPMColor* dst,
                   int radius, int width, int height,
-                  int srcStrideX, int srcStrideY,
-                  int dstStrideX, int dstStrideY)
+                  int srcStride, int dstStride)
 {
+    const int srcStrideX = inX ? 1 : srcStride;
+    const int dstStrideX = inX ? 1 : dstStride;
+    const int srcStrideY = inX ? srcStride : 1;
+    const int dstStrideY = inX ? dstStride : 1;
     radius = SkMin32(radius, width - 1);
     const SkPMColor* upperSrc = src + radius * srcStrideX;
     for (int x = 0; x < width; ++x) {
@@ -74,23 +79,35 @@ static void erode(const SkPMColor* src, SkPMColor* dst,
 
 static void erodeX(const SkBitmap& src, SkBitmap* dst, int radiusX, const SkIRect& bounds)
 {
-    erode(src.getAddr32(bounds.left(), bounds.top()), dst->getAddr32(0, 0),
-          radiusX, bounds.width(), bounds.height(),
-          1, src.rowBytesAsPixels(), 1, dst->rowBytesAsPixels());
+    SkMorphologyProc erodeXProc = SkMorphologyGetPlatformProc(SkMorphologyErodeX_Type);
+    if (!erodeXProc) {
+        erodeXProc = erode<true>;
+    }
+    erodeXProc(src.getAddr32(bounds.left(), bounds.top()), dst->getAddr32(0, 0),
+               radiusX, bounds.width(), bounds.height(),
+               src.rowBytesAsPixels(), dst->rowBytesAsPixels());
 }
 
 static void erodeY(const SkBitmap& src, SkBitmap* dst, int radiusY, const SkIRect& bounds)
 {
-    erode(src.getAddr32(bounds.left(), bounds.top()), dst->getAddr32(0, 0),
-          radiusY, bounds.height(), bounds.width(),
-          src.rowBytesAsPixels(), 1, dst->rowBytesAsPixels(), 1);
+    SkMorphologyProc erodeYProc = SkMorphologyGetPlatformProc(SkMorphologyErodeY_Type);
+    if (!erodeYProc) {
+        erodeYProc = erode<false>;
+    }
+    erodeYProc(src.getAddr32(bounds.left(), bounds.top()), dst->getAddr32(0, 0),
+               radiusY, bounds.height(), bounds.width(),
+               src.rowBytesAsPixels(), dst->rowBytesAsPixels());
 }
 
+template<bool inX>
 static void dilate(const SkPMColor* src, SkPMColor* dst,
                    int radius, int width, int height,
-                   int srcStrideX, int srcStrideY,
-                   int dstStrideX, int dstStrideY)
+                   int srcStride, int dstStride)
 {
+    const int srcStrideX = inX ? 1 : srcStride;
+    const int dstStrideX = inX ? 1 : dstStride;
+    const int srcStrideY = inX ? srcStride : 1;
+    const int dstStrideY = inX ? dstStride : 1;
     radius = SkMin32(radius, width - 1);
     const SkPMColor* upperSrc = src + radius * srcStrideX;
     for (int x = 0; x < width; ++x) {
@@ -122,16 +139,24 @@ static void dilate(const SkPMColor* src, SkPMColor* dst,
 
 static void dilateX(const SkBitmap& src, SkBitmap* dst, int radiusX, const SkIRect& bounds)
 {
-    dilate(src.getAddr32(bounds.left(), bounds.top()), dst->getAddr32(0, 0),
-           radiusX, bounds.width(), bounds.height(),
-           1, src.rowBytesAsPixels(), 1, dst->rowBytesAsPixels());
+    SkMorphologyProc dilateXProc = SkMorphologyGetPlatformProc(SkMorphologyDilateX_Type);
+    if (!dilateXProc) {
+        dilateXProc = dilate<true>;
+    }
+    dilateXProc(src.getAddr32(bounds.left(), bounds.top()), dst->getAddr32(0, 0),
+                radiusX, bounds.width(), bounds.height(),
+                src.rowBytesAsPixels(), dst->rowBytesAsPixels());
 }
 
 static void dilateY(const SkBitmap& src, SkBitmap* dst, int radiusY, const SkIRect& bounds)
 {
-    dilate(src.getAddr32(bounds.left(), bounds.top()), dst->getAddr32(0, 0),
-           radiusY, bounds.height(), bounds.width(),
-           src.rowBytesAsPixels(), 1, dst->rowBytesAsPixels(), 1);
+    SkMorphologyProc dilateYProc = SkMorphologyGetPlatformProc(SkMorphologyDilateY_Type);
+    if (!dilateYProc) {
+        dilateYProc = dilate<false>;
+    }
+    dilateYProc(src.getAddr32(bounds.left(), bounds.top()), dst->getAddr32(0, 0),
+                radiusY, bounds.height(), bounds.width(),
+                src.rowBytesAsPixels(), dst->rowBytesAsPixels());
 }
 
 bool SkErodeImageFilter::onFilterImage(Proxy* proxy,