Handle premultiplication correctly for ColorMatrixColorFilters

bug:17405627

Previously, the input content to the color matrix computation was left
premultiplied. Since the color matrix could reduce the alpha channel,
the alpha was re-multiplied, but this was incomplete, and incorrect.

Instead, apply the color matrix in unpremultiplied space.

Change-Id: I87b8e03d2e228e6ded81f7bbfea952605d7a095c

libs/hwui/ProgramCache.cpp

@@ -327,8 +327,10 @@ const char* gFS_Main_ApplyColorOp[3] = {
         // None
         "",
         // Matrix
+        "    fragColor.rgb /= (fragColor.a + 0.0019);\n" // un-premultiply
         "    fragColor *= colorMatrix;\n"
-        "    fragColor += colorMatrixVector;\n",
+        "    fragColor += colorMatrixVector;\n"
+        "    fragColor.rgb *= (fragColor.a + 0.0019);\n", // re-premultiply
         // PorterDuff
         "    fragColor = blendColors(colorBlend, fragColor);\n"
 };