To clarify, multiple transform feedback tests seem to be flaking:

WebglConformance_deqp_functional_gles3_transformfeedback_basic_types_interleaved_points
https://ci.chromium.org/buildbot/chromium.gpu.fyi/Win10%20FYI%20Release%20%28NVIDIA%29/413

WebglConformance_deqp_functional_gles3_transformfeedback_basic_types_interleaved_triangles
https://ci.chromium.org/buildbot/tryserver.chromium.angle/win_angle_rel_ng/9666

WebglConformance_deqp_functional_gles3_transformfeedback_basic_types_interleaved_lines
https://ci.chromium.org/buildbot/chromium.gpu.fyi/Win10%20FYI%20Release%20%28NVIDIA%29/412

Corentin, as ANGLE wrangler this week could you please add suppressions for ANGLE's OpenGL backend on Windows for these? All of the transform feedback tests are unfortunately already marked failing in src/content/test/gpu/gpu_tests/webgl2_conformance_expectations.py for Win/Quadro P400/D3D11, so if you could duplicate all of those expectations and put them in place for the OpenGL backend, marking them as Flaky, that would be ideal. Thanks.

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/ca8b8d58636b65c32d3a2046d8a3026358adcdb4

commit ca8b8d58636b65c32d3a2046d8a3026358adcdb4
Author: Corentin Wallez <cwallez@chromium.org>
Date: Fri Mar 16 19:54:52 2018

WebGL2 CTS: Mark interleaved XFB flaky on Win GL NVIDIA

TBR=kbr@chromium.org
BUG=822733

Cq-Include-Trybots: luci.chromium.try:linux_optional_gpu_tests_rel;master.tryserver.chromium.android:android_optional_gpu_tests_rel;master.tryserver.chromium.mac:mac_optional_gpu_tests_rel;master.tryserver.chromium.win:win_optional_gpu_tests_rel
Change-Id: I8d8b0b6c411b8974b130e40776591897cd41760e
Reviewed-on: https://chromium-review.googlesource.com/966884
Reviewed-by: Kenneth Russell <kbr@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Kenneth Russell <kbr@chromium.org>
Cr-Commit-Position: refs/heads/master@{#543793}
[modify] https://crrev.com/ca8b8d58636b65c32d3a2046d8a3026358adcdb4/content/test/gpu/gpu_tests/webgl2_conformance_expectations.py

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/00c7afafacf44740027abc7328940ab2f79b44cb

commit 00c7afafacf44740027abc7328940ab2f79b44cb
Author: Jamie Madill <jmadill@chromium.org>
Date: Mon Mar 19 16:24:22 2018

Expand flaky XFB tests for Win/ANGLE/NVIDIA.

The whole set of tests seems to be flaky on Windows.

Bug: 822733
Tbr: kbr@chromium.org
Cq-Include-Trybots: luci.chromium.try:linux_optional_gpu_tests_rel;master.tryserver.chromium.android:android_optional_gpu_tests_rel;master.tryserver.chromium.mac:mac_optional_gpu_tests_rel;master.tryserver.chromium.win:win_optional_gpu_tests_rel
Change-Id: Ia99bdec79e9a3c363c10d469b134e5e076c07218
Reviewed-on: https://chromium-review.googlesource.com/968761
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Cr-Commit-Position: refs/heads/master@{#544057}
[modify] https://crrev.com/00c7afafacf44740027abc7328940ab2f79b44cb/content/test/gpu/gpu_tests/webgl2_conformance_expectations.py

Olli, can you take a look at these failing tests? It looks very likely to be an Nvidia Windows driver bug. It is easy to repro, just open this link and reload the page a few times:

https://www.khronos.org/registry/webgl/sdk/tests/deqp/functional/gles3/transformfeedback/random_separate_triangles.html?webglVersion=2&quiet=0 

To be sure you're running the same configuration as us, you can pass Chrome the command line arguments --use-gl=angle --use-cmd-decoder=passthrough --use=angle=d3d11. I can also reproduce the problem in Firefox as well.

I can reproduce flakiness with Chrome on Windows 10 with an NVIDIA GPU, and can't reproduce it on Intel.

However, I can't reproduce this with native dEQP tests on top of ANGLE, which makes it hard to verify that this is indeed a driver bug or debug the issue. Are you certain that this is not a Chromium issue? What might be different between Chromium and native ANGLE, maybe framebuffer configuration or other driver calls done by the system?

Well, unfortunately there are a huge number of differences in the software stacks here. Chrome's command buffer multiplexes among many different clients and jdarpinian@ and sunnyps@ have been tracking down strange behavior where we switch contexts more often than we think we should.

However, if the flakiness happens just with this one test, rather than requiring multiple different tests to run, then that seems it should be easier to track down on the driver side. Especially since the flakiness doesn't happen on Intel GPUs.

Could you perhaps try to cut down the JS version of the dEQP test to the smallest variant which reproduces the flakiness, and then send it to the driver team?

I am certain this is not a Chromium issue because I can reproduce it in Firefox.

I've been trying to make the test case more minimal now. I started based on the "basic_types_separate_triangles.html" test, and only a small subset of the cases in it are required to get it to repro. I seem to have better luck reproducing it with floats rather than ints, so it may be specific to floats or smooth interpolation. It seems like the TF output itself is always fine, but the rendered image sometimes gets corrupted. Still continuing investigation.

Did not get farther than that on the triage - flaky issues are always tricky. Leaving this as available.