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

src/effects/SkMorphologyImageFilter.cpp

 /*
  * Copyright 2012 The Android Open Source Project
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkMorphologyImageFilter.h"
 #include "SkBitmap.h"
 #include "SkColorPriv.h"
 #include "SkFlattenableBuffers.h"
 #include "SkRect.h"
+#include "SkMorphology_opts.h"
 #if SK_SUPPORT_GPU
 #include "GrContext.h"
 #include "GrTexture.h"
 #include "GrTBackendEffectFactory.h"
 #include "gl/GrGLEffect.h"
 #include "effects/Gr1DKernelEffect.h"
 #include "SkImageFilterUtils.h"
 #endif
 
 SkMorphologyImageFilter::SkMorphologyImageFilter(SkFlattenableReadBuffer& buffer)
   : INHERITED(buffer) {
     fRadius.fWidth = buffer.readInt();
     fRadius.fHeight = buffer.readInt();
     buffer.validate((fRadius.fWidth >= 0) &&
                     (fRadius.fHeight >= 0));
 }
 
 SkMorphologyImageFilter::SkMorphologyImageFilter(int radiusX, int radiusY, SkImageFilter* input, const CropRect* cropRect)
     : INHERITED(input, cropRect), fRadius(SkISize::Make(radiusX, radiusY)) {
 }
 
 
 void SkMorphologyImageFilter::flatten(SkFlattenableWriteBuffer& buffer) const {
     this->INHERITED::flatten(buffer);
     buffer.writeInt(fRadius.fWidth);
     buffer.writeInt(fRadius.fHeight);
 }
 
+enum MorphDirection {
+    kX, kY
+};
+
+template<MorphDirection direction>
 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 = direction == kX ? 1 : srcStride;
+    const int dstStrideX = direction == kX ? 1 : dstStride;
+    const int srcStrideY = direction == kX ? srcStride : 1;
+    const int dstStrideY = direction == kX ? dstStride : 1;
     radius = SkMin32(radius, width - 1);
     const SkPMColor* upperSrc = src + radius * srcStrideX;
     for (int x = 0; x < width; ++x) {
         const SkPMColor* lp = src;
         const SkPMColor* up = upperSrc;
         SkPMColor* dptr = dst;
         for (int y = 0; y < height; ++y) {
             int minB = 255, minG = 255, minR = 255, minA = 255;
             for (const SkPMColor* p = lp; p <= up; p += srcStrideX) {
                 int b = SkGetPackedB32(*p);
@@ skipping 11 matching lines @@
             up += srcStrideY;
         }
         if (x >= radius) src += srcStrideX;
         if (x + radius < width - 1) upperSrc += srcStrideX;
         dst += dstStrideX;
     }
 }
 
 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(kErodeX_SkMorphologyProcType);
+    if (!erodeXProc) {
+        erodeXProc = erode<kX>;
+    }
+    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(kErodeY_SkMorphologyProcType);
+    if (!erodeYProc) {
+        erodeYProc = erode<kY>;
+    }
+    erodeYProc(src.getAddr32(bounds.left(), bounds.top()), dst->getAddr32(0, 0),
+               radiusY, bounds.height(), bounds.width(),
+               src.rowBytesAsPixels(), dst->rowBytesAsPixels());
 }
 
+template<MorphDirection direction>
 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 = direction == kX ? 1 : srcStride;
+    const int dstStrideX = direction == kX ? 1 : dstStride;
+    const int srcStrideY = direction == kX ? srcStride : 1;
+    const int dstStrideY = direction == kX ? dstStride : 1;
     radius = SkMin32(radius, width - 1);
     const SkPMColor* upperSrc = src + radius * srcStrideX;
     for (int x = 0; x < width; ++x) {
         const SkPMColor* lp = src;
         const SkPMColor* up = upperSrc;
         SkPMColor* dptr = dst;
         for (int y = 0; y < height; ++y) {
             int maxB = 0, maxG = 0, maxR = 0, maxA = 0;
             for (const SkPMColor* p = lp; p <= up; p += srcStrideX) {
                 int b = SkGetPackedB32(*p);
@@ skipping 11 matching lines @@
             up += srcStrideY;
         }
         if (x >= radius) src += srcStrideX;
         if (x + radius < width - 1) upperSrc += srcStrideX;
         dst += dstStrideX;
     }
 }
 
 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(kDilateX_SkMorphologyProcType);
+    if (!dilateXProc) {
+        dilateXProc = dilate<kX>;
+    }
+    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(kDilateY_SkMorphologyProcType);
+    if (!dilateYProc) {
+        dilateYProc = dilate<kY>;
+    }
+    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,
                                        const SkBitmap& source, const SkMatrix& ctm,
                                        SkBitmap* dst, SkIPoint* offset) {
     SkBitmap src = source;
     if (getInput(0) && !getInput(0)->filterImage(proxy, source, ctm, &src, offset)) {
         return false;
     }
 
@@ skipping 440 matching lines @@
 
     if (!apply_morphology(input, bounds, GrMorphologyEffect::kErode_MorphologyType, radius(), result)) {
         return false;
     }
     offset->fX += bounds.left();
     offset->fY += bounds.top();
     return true;
 }
 
 #endif