This removes the 0.5 bias that was added before doing the transformation. A similar thing is done in AFFINE_FILTER_NAME in SkBitmapProcState_matrix.h

George, can you get this upstreamed?

I can try

The issue is more or less as the previous patch describes, that when linear filtering is applied we need to make sure we subtract the offset of half a texel. However, we need to ensure that this only happens when we are actually doing a form of linear filtering. This only occurs in shadeSpan_linear_vertical_lerp - a strictly vertical gradient in the 32bpp case. The other cases don't try to emulate any sort of linear filtering so shouldn't have the half-texel offset. The fix ends up just being a one-liner.

This code no longer exists in recent versions of Skia, so I am going to mark this as invalid. If a similar issue turns up in the new Skia code, we can file a new bug.