commit b7f28d6ecd07eb97fdd069fb166c1fdd9ce24f27
author senorblanco@chromium.org <senorblanco@chromium.org@2bbb7eff-a529-9590-31e7-b0007b416f81> Thu Jul 21 22:07:53 2011 +0000
committer senorblanco@chromium.org <senorblanco@chromium.org@2bbb7eff-a529-9590-31e7-b0007b416f81> Thu Jul 21 22:07:53 2011 +0000
tree 02b38e24a0020ccd5dffce8fc99081e4a5164b8b
parent 6dc745506e6d6cc0936fed4de24443dc1ecb5a34

Improve GPU blur speed by clearing only when necessary.  This gives a
1.8X speedup on the Blurs sample, and 2.3X on the BigBlur sample.

We don't need to clear while downsampling, since each step reads only the pixels
written in the previous step.  We can avoid destination clears before
convolution by disabling blending.  We also don't need to clear when upsampling,
since the upsample step also only reads pixels written by the convolution.  The
only clears we then need to do are on each side of the srcRect used for
convolution.  Since our srcRect is always offset to (0, 0), we only need to
clear on the right and bottom.

Review URL:  http://codereview.appspot.com/4803048/



git-svn-id: http://skia.googlecode.com/svn/trunk@1937 2bbb7eff-a529-9590-31e7-b0007b416f81

src/gpu/SkGpuDevice.cpp

diff --git a/src/gpu/SkGpuDevice.cpp b/src/gpu/SkGpuDevice.cpp
index c4e0ebe..e54e0de 100644
--- a/src/gpu/SkGpuDevice.cpp
+++ b/src/gpu/SkGpuDevice.cpp
@@ -913,9 +913,7 @@
     GrClip oldClip = context->getClip();
     context->setRenderTarget(dstTexture->asRenderTarget());
     context->setClip(srcRect);
-    // FIXME:  could just clear bounds
     context->clear(NULL, 0);
-    GrMatrix transM;
     GrPaint tempPaint;
     tempPaint.reset();
 
@@ -959,9 +957,6 @@
         context->setRenderTarget(dstTexture->asRenderTarget());
         SkRect dstRect(srcRect);
         scaleRect(&dstRect, 0.5f);
-        // Clear out 1 pixel border for linear filtering.
-        // FIXME:  for now, clear everything
-        context->clear(NULL, 0);
         paint.setTexture(0, srcTexture);
         context->drawRectToRect(paint, dstRect, srcRect);
         srcRect = dstRect;
@@ -972,13 +967,23 @@
     float* kernel = kernelStorage.get();
     buildKernel(sigma, kernel, kernelWidth);
 
+    // Clear out a halfWidth to the right of the srcRect to prevent the
+    // X convolution from reading garbage.
+    SkIRect clearRect = SkIRect::MakeXYWH(
+        srcRect.fRight, srcRect.fTop, halfWidth, srcRect.height());
+    context->clear(&clearRect, 0x0);
+
     context->setRenderTarget(dstTexture->asRenderTarget());
-    context->clear(NULL, 0);
     context->convolveInX(srcTexture, srcRect, kernel, kernelWidth);
     SkTSwap(srcTexture, dstTexture);
 
+    // Clear out a halfWidth below the srcRect to prevent the Y
+    // convolution from reading garbage.
+    clearRect = SkIRect::MakeXYWH(
+        srcRect.fLeft, srcRect.fBottom, srcRect.width(), halfWidth);
+    context->clear(&clearRect, 0x0);
+
     context->setRenderTarget(dstTexture->asRenderTarget());
-    context->clear(NULL, 0);
     context->convolveInY(srcTexture, srcRect, kernel, kernelWidth);
     SkTSwap(srcTexture, dstTexture);
 
@@ -988,9 +993,6 @@
         sampleM.setIDiv(srcTexture->width(), srcTexture->height());
         paint.getTextureSampler(0)->setMatrix(sampleM);
         context->setRenderTarget(dstTexture->asRenderTarget());
-        // Clear out 2 pixel border for bicubic filtering.
-        // FIXME:  for now, clear everything
-        context->clear(NULL, 0);
         paint.setTexture(0, srcTexture);
         SkRect dstRect(srcRect);
         scaleRect(&dstRect, scaleFactor);