Fix gpu memory accounting for auto-resolving MSAA.

src/gpu/gl/GrGLRenderTarget.cpp

Index: src/gpu/gl/GrGLRenderTarget.cpp
diff --git a/src/gpu/gl/GrGLRenderTarget.cpp b/src/gpu/gl/GrGLRenderTarget.cpp
index 3eb2ae09fea489802fe38faa77523be8c6c17fc9..f90689339b5494dbc2826bee3958aac5beb7eb84 100644
--- a/src/gpu/gl/GrGLRenderTarget.cpp
+++ b/src/gpu/gl/GrGLRenderTarget.cpp
@@ -43,7 +43,10 @@ void GrGLRenderTarget::init(const GrSurfaceDesc& desc, const IDDesc& idDesc) {
     if (fTexFBOID != fRTFBOID) {
         // If we own the resolve buffer then that is one more sample per pixel.
         fColorValuesPerPixel += 1;
-    } 
+    } else if (fTexFBOID != 0) {
+        // For auto-resolving FBOs, the MSAA buffer is free.
+        fColorValuesPerPixel = 1;

[bsalomon, 2015/04/30 17:34:28]

looks ok to me, but can we rename this to something like
fColorSamplesForSizeComputation or something that makes it more clear that this
is really about byte size computation and not logical sample count? Or, we could
just compute the total size here and store that instead of the sample cnt. I
think it didn't work that way in the past because we didn't yet know whether we
had a stencil buffer. However, this class no longer has to account for the
stencil's memory

+    }
 }
 
 size_t GrGLRenderTarget::onGpuMemorySize() const {