Allow FPs to elevate default precision for the entire fragment program

Currently, GrConfigConversionEffect is able to round-trip on many
mobile GPUs because it uses highp for all intermediate variables
(including the texture fetch result). Separating the texture sample
into a different processor breaks that.

This is a blunt instrument, not to be used lightly.

This reverts commit dffe9827b18444d1426859d9035f9f0087201f44.

Bug: skia:
Change-Id: I940af3256c47e6672a008d516db9e55669672ca3
Reviewed-on: https://skia-review.googlesource.com/11345
Reviewed-by: Brian Salomon <bsalomon@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
diff --git a/bench/GLVertexAttributesBench.cpp b/bench/GLVertexAttributesBench.cpp
index 4783d55..7efd984 100644
--- a/bench/GLVertexAttributesBench.cpp
+++ b/bench/GLVertexAttributesBench.cpp
@@ -117,7 +117,7 @@
     // setup fragment shader
     GrShaderVar oFragColor("o_FragColor", kVec4f_GrSLType, GrShaderVar::kOut_TypeModifier);
     SkString fshaderTxt(version);
-    GrGLSLAppendDefaultFloatPrecisionDeclaration(kDefault_GrSLPrecision, *shaderCaps, &fshaderTxt);
+    GrGLSLAppendDefaultFloatPrecisionDeclaration(kMedium_GrSLPrecision, *shaderCaps, &fshaderTxt);
 
     const char* fsOutName;
     if (shaderCaps->mustDeclareFragmentShaderOutput()) {