Use texture barriers to read directly from the RT

src/gpu/GrGpu.h

Index: src/gpu/GrGpu.h
diff --git a/src/gpu/GrGpu.h b/src/gpu/GrGpu.h
index 82f4931d6c49f6e0cd45d09f67d7e5d03eb7a7d3..3cf7fbb907095e652a0dbd027f1804822e47d123 100644
--- a/src/gpu/GrGpu.h
+++ b/src/gpu/GrGpu.h
@@ -304,6 +304,10 @@ public:
                              const SkIRect& srcRect,
                              const SkIPoint& dstPoint) = 0;
 
+    // Inserts a boundary between render passes that touch the same pixels. Some blending extensions
+    // require this in order to guarantee coherent results.
+    virtual void blendBarrier() = 0;

[Mark Kilgard, 2015/04/02 23:02:38]

seems like you could just provide a no-op implementation {} rather than a base
virtual method

having a no-op default blendBarrier makes sense to me

long term, I don't expect hardware to need the barrier to do anything

+
     struct DrawArgs {
         typedef GrDrawTarget::DrawInfo DrawInfo;
         DrawArgs(const GrPrimitiveProcessor* primProc,