tree eb1d2226c8eb80c974cdf90ebb04cd55893b0ae0
parent fb330c2b6b952b467f794bdf302b9d6ecbcdc8ad
author Michael Ludwig <michaelludwig@google.com> 1603299694 -0400
committer Skia Commit-Bot <skia-commit-bot@chromium.org> 1603301832 +0000

Make small epsilons more rigorous for gpu gaussian blurs

This also has several other bug fixes and refactorings within it that
I realized were possible while updating every where that had checked
sigma > 0 to be sigma > kEffectivelyZeroSigma.

The big things are that SkBlurPriv.h goes away and its functions are
just moved into SkGpuBlurUtils since they were only used by the GPU.
The implementations of those functions are also collected into
SkGpuBlurUtils.cpp.  I removed the GrMatrixConvolution::MakeGaussian,
in favor of SkGpuBlurUtils filling in the kernel itself and then calling
the regular Make. This let me consolidate two different 1D kernel
computing functions, and remove the 1D fallback code from the 2D kernel
calculation because GaussianBlur() can detect that earlier.

The new GM, BlurSigmaSmall, originally drew incorrectly on the GPU
backend because it's small but non-zero sigma would trick the sigma > 0
checks in various places so we'd do a full 2 pass X/Y blur. However,
when the sigma was too small, the kernel was just filled with 0s so the
Y pass would effectively clear everything. While I could have just fixed
that to be a [0, 1, 0] kernel, updating the blur pipeline to compare
against integer radii seems more robust.

Change-Id: I3c41e0235a27615a9056b25e627ffedd995264bd
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/328797
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Brian Salomon <bsalomon@google.com>