We are failing some webgl_canvas tests which seem to read texture data from Canvas2D. I wonder if something changed in Canvas2D - the Skia change seems is the only thing that seems suspect:

https://chromium.googlesource.com/chromium/src/+/de3f4bede90e857cb62f9f0160e80659561d9324

Maybe the following CL which seems to deal with readback of SkImage pixels?

https://skia.googlesource.com/skia.git/+/f4d443fd1bd8cffbcd7a07a48f3488ae871414e9%5E%21/#F0

Will see if this repros locally.

 Issue 705482  has been merged into this issue.

Taking and upgrading to P0. Working on triaging this.

Seems difficult to reproduce. The same tests don't fail run one-by-one. The following command line does seem to reproduce some failures, though I didn't let the whole shard complete:

./content/test/gpu/run_gpu_integration_test.py webgl_conformance --webgl-conformance-version=2.0.1 --total-shards=15 --shard-index=6 --read-abbreviated-json-results-from content/test/data/gpu/webgl2_conformance_tests_output.json --browser=exact --browser-executable=/Users/kbr/tmp/webgl2-regression/full-build-mac/Google\ Chrome.app/Contents/MacOS/Google\ Chrome 

This is with an official build downloaded at r459600.

Same run is good at r459528.

Last known good build on one machine:
https://build.chromium.org/p/chromium.gpu.fyi/builders/Mac%20Retina%20Release%20%28NVIDIA%29/builds/1160

First known bad build:
https://build.chromium.org/p/chromium.gpu.fyi/builders/Mac%20Retina%20Release%20%28NVIDIA%29/builds/1161

Regression range is r459528..r459556.

Hang on, I'm seeing the same failures at r459528 now...

Confirming that r459491 is a good revision.

OK, r459528 looks good. Must have been a mistake. Continuing manual bisect.

Confirmed that r459543 is the first bad revision:
https://chromium.googlesource.com/chromium/src/+/de3f4bede90e857cb62f9f0160e80659561d9324

by repeatedly downloading the per-revision binaries and running:

./content/test/gpu/run_gpu_integration_test.py webgl_conformance --webgl-conformance-version=2.0.1 --total-shards=15 --shard-index=6 --read-abbreviated-json-results-from content/test/data/gpu/webgl2_conformance_tests_output.json --browser=exact --browser-executable=/Users/kbr/tmp/webgl2-regression/full-build-mac/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --filter-tests-after-sharding

Brian, could you please take care of stopping the Skia auto-roller and rolling back to the previous good revision? Afterward I'll work with you more. Thanks.

Note: with the above command line, as soon as the harness starts running test ~65, it fails WebglConformance_conformance2_textures_webgl_canvas_tex_2d_r16f_red_float . Unfortunately that single test, when run, doesn't fail -- the failure has to do with the tests that run prior to it.

#1
>Maybe the following CL which seems to deal with readback of SkImage pixels?

cc brianosman@chromium.org to check

https://skia-review.googlesource.com/c/10109/2/src/image/SkImage_Gpu.cpp#73

Change to roll back Skia: https://codereview.chromium.org/2777043003/#

Brian Osman is on vacation this week. I'll see if I can figure out if that could have caused this. Because it is affected by earlier tests I suspect this may be an OpenGL state interaction between Chrome and Skia.

Forutnately I can reproduce this with a local build at ToT using Ken's repro steps. I'm going to try manually bisecting the Skia roll to figure out which CL is to blame.

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/706a5ddf48c845431a8de6aabc31d5d0aed3fe45

commit 706a5ddf48c845431a8de6aabc31d5d0aed3fe45
Author: bsalomon <bsalomon@google.com>
Date: Mon Mar 27 21:15:04 2017

Roll Skia back to before WebGL conformance test failure

BUG= 705156 

Review-Url: https://codereview.chromium.org/2777043003
Cr-Commit-Position: refs/heads/master@{#459879}

[modify] https://crrev.com/706a5ddf48c845431a8de6aabc31d5d0aed3fe45/DEPS

The regression begins with:

https://skia.googlesource.com/skia/+/f4d443fd1bd8cffbcd7a07a48f3488ae871414e9
author	Brian Osman <brianosman@google.com>	Fri Mar 24 11:22:31 2017 -0400
committer	Skia Commit-Bot <skia-commit-bot@chromium.org>	Fri Mar 24 17:22:15 2017 +0000
Fix SkImage_Gpu::getROPixels

Reverting that in Skia so that we can restart the autoroller. The optional_gpu_tests have been added to the autoroller to prevent future regressions.

Thanks Brian for your quick action identifying the cause. Please close this out once Skia's been rolled forward with that revert.

Brian's probably gone for today, but his revert went through and the associated bots have rolled green:

https://build.chromium.org/p/chromium.gpu.fyi/builders/Mac%20Retina%20Release%20%28NVIDIA%29/builds/1220
https://build.chromium.org/p/chromium.gpu.fyi/builders/Mac%20Release%20%28Intel%29/builds/1556
https://build.chromium.org/p/chromium.gpu.fyi/builders/Mac%20Retina%20Release%20%28AMD%29/builds/1203

Closing as Fixed. Unfortunately there was no associated Skia bug so I can't link this one to it.

Skia bug for further investigation has been filed here: https://bugs.chromium.org/p/skia/issues/detail?id=6433

The original change is still reverted in Skia and Skia is now rolling into Chrome again.