GitLab

The work-around code forced us to take a slower path, which is unnecessary now that filter
is precise for opaque inputs.

remove unneeded file from mac utils

* changes:
  pull clipper fix from skia/trunk rev.404

* changes:
  skip S32A_D565_Blend_neon case for now, as it has an artifact when src-alpha==0

* changes:
  moto neon opts for bitmap tiling

* changes:
  Fix http://b/2154737 by setting the size of memory buffer used for jpeg decode in Skia.

During record, our clip is wide-open (in the bug's case, 40K pixels tall). If we call saveLayer
with no bounds (as webkit does), we try to allocate an offscreen the size of the document, which
can be arbitrarily big. This cause us to run out of memory. In reality, I think there is not
benefit to calling through to saveLayer during record. On playback, we will of course call saveLayer,
but then we are on a real device (screen) with finite bounds, and the offscreen allocation is
reasonable.

It is computing out-of-range indices. When I debug it, I'll renable it with the fix.

* changes:
  use correct rowbytes, instead of assuming it equals bounds.width()

bitmap cracking when we're scaled but not filtered.

* changes:
  add SkSize.h add golden-master (gm) test app

add golden-master (gm) test app

refresh misc. skia files

know we're going to run a non-zero final proc for the endcap

of gradients' ability to pre-dither their cache.

- edge case fix in qsort compare proc (overflowing subtract)
- special case 1x1 bitmaps to draw as a color (much faster)

- correctly respect dither for index bitmaps
- fix copyTo to respect isOpaque()
- sanity check in antipath (need to investigate root cause)
- warning fix in bitmap sampler