gpu_testing.md

GPU Testing

This set of pages documents the setup and operation of the GPU bots and try servers, which verify the correctness of Chrome's graphically accelerated rendering pipeline.

[TOC]

Overview

The GPU bots run a different set of tests than the majority of the Chromium test machines. The GPU bots specifically focus on tests which exercise the graphics processor, and whose results are likely to vary between graphics card vendors.

Most of the tests on the GPU bots are run via the Telemetry framework. Telemetry was originally conceived as a performance testing framework, but has proven valuable for correctness testing as well. Telemetry directs the browser to perform various operations, like page navigation and test execution, from external scripts written in Python. The GPU bots launch the full Chromium browser via Telemetry for the majority of the tests. Using the full browser to execute tests, rather than smaller test harnesses, has yielded several advantages: testing what is shipped, improved reliability, and improved performance.

A subset of the tests, called "pixel tests", grab screen snapshots of the web page in order to validate Chromium's rendering architecture end-to-end. Where necessary, GPU-specific results are maintained for these tests. Some of these tests verify just a few pixels, using handwritten code, in order to use the same validation for all brands of GPUs.

The GPU bots use the Chrome infrastructure team's recipe framework, and specifically the chromium and chromium_trybot recipes, to describe what tests to execute. Compared to the legacy master-side buildbot scripts, recipes make it easy to add new steps to the bots, change the bots' configuration, and run the tests locally in the same way that they are run on the bots. Additionally, the chromium and chromium_trybot recipes make it possible to send try jobs which add new steps to the bots. This single capability is a huge step forward from the previous configuration where new steps were added blindly, and could cause failures on the tryservers. For more details about the configuration of the bots, see the GPU bot details.

The physical hardware for the GPU bots lives in the Swarming pool*. The Swarming infrastructure (new docs, older but currently more complete docs) provides many benefits:

• Increased parallelism for the tests; all steps for a given tryjob or waterfall build run in parallel.
• Simpler scaling: just add more hardware in order to get more capacity. No manual configuration or distribution of hardware needed.
• Easier to run certain tests only on certain operating systems or types of GPUs.
• Easier to add new operating systems or types of GPUs.
• Clearer description of the binary and data dependencies of the tests. If they run successfully locally, they'll run successfully on the bots.

(* All but a few one-off GPU bots are in the swarming pool. The exceptions to the rule are described in the GPU bot details.)

The bots on the chromium.gpu.fyi waterfall are configured to always test top-of-tree ANGLE. This setup is done with a few lines of code in the tools/build workspace; search the code for "angle".

These aspects of the bots are described in more detail below, and in linked pages. There is a presentation which gives a brief overview of this documentation and links back to various portions.

Fleet Status

Please see the GPU Pixel Wrangling instructions for links to dashboards showing the status of various bots in the GPU fleet.

Using the GPU Bots

Most Chromium developers interact with the GPU bots in two ways:

1. Observing the bots on the waterfalls.
2. Sending try jobs to them.

The GPU bots are grouped on the chromium.gpu and chromium.gpu.fyi waterfalls. Their current status can be easily observed there.

To send try jobs, you must first upload your CL to the codereview server. Then, either clicking the "CQ dry run" link or running from the command line:

git cl try

Sends your job to the default set of try servers.

The GPU tests are part of the default set for Chromium CLs, and are run as part of the following tryservers' jobs:

• linux-rel, formerly on the tryserver.chromium.linux waterfall
• mac-rel, formerly on the tryserver.chromium.mac waterfall
• win-rel, formerly on the tryserver.chromium.win waterfall

Scan down through the steps looking for the text "GPU"; that identifies those tests run on the GPU bots. For each test the "trigger" step can be ignored; the step further down for the test of the same name contains the results.

It's usually not necessary to explicitly send try jobs just for verifying GPU tests. If you want to, you must invoke "git cl try" separately for each tryserver master you want to reference, for example:

git cl try -b linux-rel
git cl try -b mac-rel
git cl try -b win7-rel

Alternatively, the Gerrit UI can be used to send a patch set to these try servers.

Three optional tryservers are also available which run additional tests. As of this writing, they ran longer-running tests that can't run against all Chromium CLs due to lack of hardware capacity. They are added as part of the included tryservers for code changes to certain sub-directories.

• linux_optional_gpu_tests_rel on the luci.chromium.try waterfall
• mac_optional_gpu_tests_rel on the luci.chromium.try waterfall
• win_optional_gpu_tests_rel on the luci.chromium.try waterfall

Tryservers for the ANGLE project are also present on the tryserver.chromium.angle waterfall. These are invoked from the Gerrit user interface. They are configured similarly to the tryservers for regular Chromium patches, and run the same tests that are run on the chromium.gpu.fyi waterfall, in the same way (e.g., against ToT ANGLE).

If you find it necessary to try patches against other sub-repositories than Chromium (src/) and ANGLE (src/third_party/angle/), please file a bug with component Internals>GPU>Testing.

Running the GPU Tests Locally

All of the GPU tests running on the bots can be run locally from a Chromium build. Many of the tests are simple executables:

• angle_unittests
• gl_tests
• gl_unittests
• tab_capture_end2end_tests

Some run only on the chromium.gpu.fyi waterfall, either because there isn't enough machine capacity at the moment, or because they're closed-source tests which aren't allowed to run on the regular Chromium waterfalls:

• angle_deqp_gles2_tests
• angle_deqp_gles3_tests
• angle_end2end_tests
• audio_unittests

The remaining GPU tests are run via Telemetry. In order to run them, just build the telemetry_gpu_integration_test target (or telemetry_gpu_integration_test_android_chrome for Android) and then invoke src/content/test/gpu/run_gpu_integration_test.py with the appropriate argument. The tests this script can invoke are in src/content/test/gpu/gpu_tests/. For example:

• run_gpu_integration_test.py context_lost --browser=release
• run_gpu_integration_test.py webgl_conformance --browser=release --webgl-conformance-version=1.0.2
• run_gpu_integration_test.py maps --browser=release
• run_gpu_integration_test.py screenshot_sync --browser=release
• run_gpu_integration_test.py trace_test --browser=release

The pixel tests are a bit special. See the section on running them locally for details.

The --browser=release argument can be changed to --browser=debug if you built in a directory such as out/Debug. If you built in some non-standard directory such as out/my_special_gn_config, you can instead specify --browser=exact --browser-executable=out/my_special_gn_config/chrome.

If you're testing on Android, use --browser=android-chromium instead of --browser=release/debug to invoke it. Additionally, Telemetry will likely complain about being unable to find the browser binary on Android if you build in a non-standard output directory. Thus, out/Release or out/Debug are suggested when testing on Android.

If you are running on a platform that does not support multiple browser instances at a time (Android or ChromeOS), it is also recommended that you pass in --jobs=1. This only has an effect on test suites that have parallel test support, but failure to pass in the argument for those tests on these platforms will result in weird failures due to multiple test processes stepping on each other. On other platforms, you are still free to specify --jobs to get more or less parallelization instead of relying on the default of one test process per logical core.

Note: The tests require some third-party Python packages. Obtaining these packages is handled automatically by vpython3, and the script's shebang should use vpython if running the script directly. If you're used to invoking python3 to run a script, simply use vpython3 instead, e.g. vpython3 run_gpu_integration_test.py ....

You can run a subset of tests with this harness:

• run_gpu_integration_test.py webgl_conformance --browser=release --test-filter=conformance_attribs

The exact command used to invoke the test on the bots can be found in one of two ways:

1. Looking at the json.input of the trigger step under requests[task_slices][command]. The arguments after the last -- are used to actually run the test.
2. Looking at the top of a swarming task.

In both cases, the following can be omitted when running locally since they're only necessary on swarming:

• testing/test_env.py
• testing/scripts/run_gpu_integration_test_as_googletest.py
• --isolated-script-test-output
• --isolated-script-test-perf-output

The Maps test requires you to authenticate to cloud storage in order to access the Web Page Reply archive containing the test. See Cloud Storage Credentials for documentation on setting this up.

Bisecting ChromeOS Failures Locally

Failures that occur on the ChromeOS amd64-generic configuration are easy to reproduce due to the VM being readily available for use, but doing so requires some additional steps to the bisect process. The following are steps that can be followed using two terminals and the Simple Chrome SDK to bisect a ChromeOS failure.

1. Terminal 1: Start the bisect as normal git bisect start git bisect good <good_revision> git bisect bad <bad_revision>
2. Terminal 1: Sync to the revision that git spits out gclient sync -r src@<revision>
3. Terminal 2: Enter the Simple Chrome SDK cros chrome-sdk --board amd64-generic-vm --log-level info --download-vm --clear-sdk-cache
4. Terminal 2: Compile the relevant target (probably the GPU integration tests) autoninja -C out_amd64-generic-vm/Release/ telemetry_gpu_integration_test
5. Terminal 2: Start the VM cros_vm --start
6. Terminal 2: Deploy the Chrome binary to the VM deploy_chrome --build-dir out_amd64-generic-vm/Release/ --device 127.0.0.1:9222 This will require you to accept a prompt twice, once because of a board mismatch and once because the VM still has rootfs verification enabled.
7. Terminal 1: Run your test on the VM. For GPU integration tests, this involves specifying --browser cros-chrome --remote 127.0.0.1 --remote-ssh-port 9222
8. Terminal 2: After determining whether the revision is good or bad, shut down the VM cros_vm --stop
9. Terminal 2: Exit the SKD exit
10. Terminal 1: Let git know whether the revision was good or bad git bisect good/git bisect bad
11. Repeat from step 2 with the new revision git spits out.

The repeated entry/exit from the SDK between revisions is to ensure that the VM image is in sync with the Chromium revision, as it is possible for regressions to be caused by an update to the image itself rather than a Chromium change.

Telemetry Test Suites

The Telemetry-based tests are all technically the same target, telemetry_gpu_integration_test, just run with different runtime arguments. The first positional argument passed determines which suite will run, and additional runtime arguments may cause the step name to change on the bots. Here is a list of all suites and resulting step names as of April 15th 2021:

• context_lost
  • context_lost_passthrough_tests
  • context_lost_tests
  • context_lost_validating_tests
• hardware_accelerated_feature
  • hardware_accelerated_feature_tests
• gpu_process
  • gpu_process_launch_tests
• info_collection
  • info_collection_tests
• maps
  • maps_pixel_passthrough_test
  • maps_pixel_test
  • maps_pixel_validating_test
  • maps_tests
• pixel
  • android_webview_pixel_skia_gold_test
  • egl_pixel_skia_gold_test
  • pixel_skia_gold_passthrough_test
  • pixel_skia_gold_validating_test
  • pixel_tests
  • vulkan_pixel_skia_gold_test
• power
  • power_measurement_test
• screenshot_sync
  • screenshot_sync_passthrough_tests
  • screenshot_sync_tests
  • screenshot_sync_validating_tests
• trace_test
  • trace_test
• webgl_conformance
  • webgl2_conformance_d3d11_passthrough_tests
  • webgl2_conformance_gl_passthrough_tests
  • webgl2_conformance_gles_passthrough_tests
  • webgl2_conformance_tests
  • webgl2_conformance_validating_tests
  • webgl_conformance_d3d11_passthrough_tests
  • webgl_conformance_d3d9_passthrough_tests
  • webgl_conformance_fast_call_tests
  • webgl_conformance_gl_passthrough_tests
  • webgl_conformance_gles_passthrough_tests
  • webgl_conformance_metal_passthrough_tests
  • webgl_conformance_swangle_passthrough_tests
  • webgl_conformance_tests
  • webgl_conformance_validating_tests
  • webgl_conformance_vulkan_passthrough_tests

Running the pixel tests locally

The pixel tests are a special case because they use an external Skia service called Gold to handle image approval and storage. See GPU Pixel Testing With Gold for specifics.

TL;DR is that the pixel tests use a binary called goldctl to download and upload data when running pixel tests.

Normally, goldctl uploads images and image metadata to the Gold server when used. This is not desirable when running locally for a couple reasons:

1. Uploading requires the user to be whitelisted on the server, and whitelisting everyone who wants to run the tests locally is not a viable solution.
2. Images produced during local runs are usually slightly different from those that are produced on the bots due to hardware/software differences. Thus, most images uploaded to Gold from local runs would likely only ever actually be used by tests run on the machine that initially generated those images, which just adds noise to the list of approved images.

Additionally, the tests normally rely on the Gold server for viewing images produced by a test run. This does not work if the data is not actually uploaded.

The pixel tests contain logic to automatically determine whether they are running on a workstation or not, as well as to determine what git revision is being tested. This should mean that the pixel tests will automatically work when run locally. However, if the local run detection code fails for some reason, you can manually pass some flags to force the same behavior:

In order to get around the local run issues, simply pass the --local-pixel-tests flag to the tests. This will disable uploading, but otherwise go through the same steps as a test normally would. Each test will also print out file:// URLs to the produced image, the closest image for the test known to Gold, and the diff between the two.

Because the image produced by the test locally is likely slightly different from any of the approved images in Gold, local test runs are likely to fail during the comparison step. In order to cut down on the amount of noise, you can also pass the --no-skia-gold-failure flag to not fail the test on a failed image comparison. When using --no-skia-gold-failure, you'll also need to pass the --passthrough flag in order to actually see the link output.

Example usage: run_gpu_integration_test.py pixel --no-skia-gold-failure --local-pixel-tests --passthrough

If, for some reason, the local run code is unable to determine what the git revision is, simply pass --git-revision aabbccdd. Note that aabbccdd must be replaced with an actual Chromium src revision (typically whatever revision origin/main is currently synced to) in order for the tests to work. This can be done automatically using: run_gpu_integration_test.py pixel --no-skia-gold-failure --local-pixel-tests --passthrough --git-revision `git rev-parse origin/main`

Running Binaries from the Bots Locally

Any binary run remotely on a bot can also be run locally, assuming the local machine loosely matches the architecture and OS of the bot.

The easiest way to do this is to find the ID of the swarming task and use "swarming.py reproduce" to re-run it:

• ./src/tools/luci-go/swarming reproduce -S https://chromium-swarm.appspot.com [task ID]

The task ID can be found in the stdio for the "trigger" step for the test. For example, look at a recent build from the Mac Release (Intel) bot, and look at the gl_unittests step. You will see something like:

Triggered task: gl_unittests on Intel GPU on Mac/Mac-10.12.6/[TRUNCATED_ISOLATE_HASH]/Mac Release (Intel)/83664
To collect results, use:
  swarming.py collect -S https://chromium-swarm.appspot.com --json /var/folders/[PATH_TO_TEMP_FILE].json
Or visit:
  https://chromium-swarm.appspot.com/user/task/[TASK_ID]

There is a difference between the isolate's hash and Swarming's task ID. Make sure you use the task ID and not the isolate's hash.

As of this writing, there seems to be a bug when attempting to re-run the Telemetry based GPU tests in this way. For the time being, this can be worked around by instead downloading the contents of the isolate. To do so, look into the "Reproducing the task locally" section on a swarming task, which contains something like:

Download inputs files into directory foo:
# (if needed, use "\${platform}" as-is) cipd install "infra/tools/luci/cas/\${platform}" -root bar
# (if needed) ./bar/cas login
./bar/cas download -cas-instance projects/chromium-swarm/instances/default_instance -digest 68ae1d6b22673b0ab7b4427ca1fc2a4761c9ee53474105b9076a23a67e97a18a/647 -dir foo

Before attempting to download an isolate, you must ensure you have permission to access the isolate server. Full instructions can be found here. For most cases, you can simply run:

• ./src/tools/luci-go/isolate login

The above link requires that you log in with your @google.com credentials. It's not known at the present time whether this works with @chromium.org accounts. Email kbr@ if you try this and find it doesn't work.

Debugging a Specific Subset of Tests on a Specific GPU Bot

When a test exhibits flake on the bots, it can be convenient to run it repeatedly with local code modifications on the bot where it is exhibiting flake. One way of doing this is via swarming (see the below section). However, a lower-overhead alternative that also works in the case where you are looking to run on a bot for which you cannot locally build is to locally alter the configuration of the bot in question to specify that it should run only the tests desired, repeating as many times as desired. Instructions for doing this are as follows (see the example CL for a concrete instantiation of these instructions):

1. In testsuite_exceptions.pyl, find the section for the test suite in question (creating it if it doesn't exist).
2. Add modifications for the bot in question and specify arguments such that your desired tests are run for the desired number of iterations.
3. Run testing/buildbot/generate_buildbot_json.py and verify that the JSON file for the bot in question was modified as you would expect.
4. Upload and run tryjobs on that specific bot via "Choose Tryjobs."
5. Examine the test results. (You can verify that the tests run were as you expected by examining the test results for individual shards of the run of the test suite in question.)
6. Add logging/code modifications/etc as desired and go back to step 4, repeating the process until you've uncovered the underlying issue.
7. Remove the the changes to testsuite_exceptions.pyl and the JSON file if turning the CL into one intended for submission!

Here is an example CL that does this.

Running Locally Built Binaries on the GPU Bots

See the Swarming documentation for instructions on how to upload your binaries to the isolate server and trigger execution on Swarming.

Be sure to use the correct swarming dimensions for your desired GPU e.g. "1002:6613" instead of "AMD Radeon R7 240 (1002:6613)" which is how it appears on swarming task page. You can query bots in the chromium.tests.gpu pool to find the correct dimensions:

• tools\luci-go\swarming bots -S chromium-swarm.appspot.com -d pool=chromium.tests.gpu

Moving Test Binaries from Machine to Machine

To create a zip archive of your personal Chromium build plus all of the Telemetry-based GPU tests' dependencies, which you can then move to another machine for testing:

1. Build Chrome (into out/Release in this example).
2. vpython3 tools/mb/mb.py zip out/Release/ telemetry_gpu_integration_test out/telemetry_gpu_integration_test.zip

Then copy telemetry_gpu_integration_test.zip to another machine. Unzip it, and cd into the resulting directory. Invoke content/test/gpu/run_gpu_integration_test.py as above.

This workflow has been tested successfully on Windows with a statically-linked Release build of Chrome.

Note: on one macOS machine, this command failed because of a broken strip-json-comments symlink in src/third_party/catapult/common/node_runner/node_runner/node_modules/.bin. Deleting that symlink allowed it to proceed.

Note also: on the same macOS machine, with a component build, this command failed to zip up a working Chromium binary. The browser failed to start with the following error:

[0626/180440.571670:FATAL:chrome_main_delegate.cc(1057)] Check failed: service_manifest_data_pack_.

In a pinch, this command could be used to bundle up everything, but the "out" directory could be deleted from the resulting zip archive, and the Chromium binaries moved over to the target machine. Then the command line arguments --browser=exact --browser-executable=[path] can be used to launch that specific browser.

See the user guide for mb, the meta-build system, for more details.

Adding New Tests to the GPU Bots

The goal of the GPU bots is to avoid regressions in Chrome's rendering stack. To that end, let's add as many tests as possible that will help catch regressions in the product. If you see a crazy bug in Chrome's rendering which would be easy to catch with a pixel test running in Chrome and hard to catch in any of the other test harnesses, please, invest the time to add a test!

There are a couple of different ways to add new tests to the bots:

1. Adding a new test to one of the existing harnesses.
2. Adding an entire new test step to the bots.

Adding a new test to one of the existing test harnesses

Adding new tests to the GTest-based harnesses is straightforward and essentially requires no explanation.

As of this writing it isn't as easy as desired to add a new test to one of the Telemetry based harnesses. See Issue 352807. Let's collectively work to address that issue. It would be great to reduce the number of steps on the GPU bots, or at least to avoid significantly increasing the number of steps on the bots. The WebGL conformance tests should probably remain a separate step, but some of the smaller Telemetry based tests (context_lost_tests, memory_test, etc.) should probably be combined into a single step.

If you are adding a new test to one of the existing tests (e.g., pixel_test), all you need to do is make sure that your new test runs correctly via isolates. See the documentation from the GPU bot details on adding new isolated tests for the gn args and authentication needed to upload isolates to the isolate server. Most likely the new test will be Telemetry based, and included in the telemetry_gpu_test_run isolate.

Adding new steps to the GPU Bots

The tests that are run by the GPU bots are described by a couple of JSON files in the Chromium workspace:

• chromium.gpu.json
• chromium.gpu.fyi.json

These files are autogenerated by the following script:

• generate_buildbot_json.py

This script is documented in testing/buildbot/README.md. The JSON files are parsed by the chromium and chromium_trybot recipes, and describe two basic types of tests:

• GTests: those which use the Googletest and Chromium's base/test/launcher/ frameworks.
• Isolated scripts: tests whose initial entry point is a Python script which follows a simple convention of command line argument parsing.

The majority of the GPU tests are however:

• Telemetry based tests: an isolated script test which is built on the Telemetry framework and which launches the entire browser.

A prerequisite of adding a new test to the bots is that that test run via isolates. Once that is done, modify test_suites.pyl to add the test to the appropriate set of bots. Be careful when adding large new test steps to all of the bots, because the GPU bots are a limited resource and do not currently have the capacity to absorb large new test suites. It is safer to get new tests running on the chromium.gpu.fyi waterfall first, and expand from there to the chromium.gpu waterfall (which will also make them run against every Chromium CL by virtue of the linux-rel, mac-rel, win7-rel and android-marshmallow-arm64-rel tryservers' mirroring of the bots on this waterfall – so be careful!).

Tryjobs which add new test steps to the chromium.gpu.json file will run those new steps during the tryjob, which helps ensure that the new test won't break once it starts running on the waterfall.

Tryjobs which modify chromium.gpu.fyi.json can be sent to the win_optional_gpu_tests_rel, mac_optional_gpu_tests_rel and linux_optional_gpu_tests_rel tryservers to help ensure that they won't break the FYI bots.

Debugging Pixel Test Failures on the GPU Bots

If pixel tests fail on the bots, the build step will contain either one or more links titled gold_triage_link for <test name> or a single link titled Too many artifacts produced to link individually, click for links, which itself will contain links. In either case, these links will direct to Gold pages showing the image produced by the image and the approved image that most closely matches it.

Note that for the tests which programmatically check colors in certain regions of the image (tests with expected_colors fields in pixel_test_pages), there likely won't be a closest approved image since those tests only upload data to Gold in the event of a failure.

Updating and Adding New Pixel Tests to the GPU Bots

If your CL adds a new pixel test or modifies existing ones, it's likely that you will have to approve new images. Simply run your CL through the CQ and follow the steps outline here under the "Check if any pixel test failures are actual failures or need to be rebaselined." step.

If you are adding a new pixel test, it is beneficial to set the grace_period_end argument in the test's definition. This will allow the test to run for a period without actually failing on the waterfall bots, giving you some time to triage any additional images that show up on them. This helps prevent new tests from making the bots red because they're producing slightly different but valid images from the ones triaged while the CL was in review. Example:

from datetime import date

...

PixelTestPage(
  'foo_pixel_test.html',
  ...
  grace_period_end=date(2020, 1, 1)
)

You should typically set the grace period to end 1-2 days after the the CL will land.

Once your CL passes the CQ, you should be mostly good to go, although you should keep an eye on the waterfall bots for a short period after your CL lands in case any configurations not covered by the CQ need to have images approved, as well. All untriaged images for your test can be found by substituting your test name into:

https://chrome-gpu-gold.skia.org/search?query=name%3D<test name>

NOTE If you have a grace period active for your test, then Gold will be told to ignore results for the test. This is so that it does not comment on unrelated CLs about untriaged images if your test is noisy. Images will still be uploaded to Gold and can be triaged, but will not show up on the main page's untriaged image list, and you will need to enable the "Ignored" toggle at the top of the page when looking at the triage page specific to your test.

Stamping out Flakiness

It's critically important to aggressively investigate and eliminate the root cause of any flakiness seen on the GPU bots. The bots have been known to run reliably for days at a time, and any flaky failures that are tolerated on the bots translate directly into instability of the browser experienced by customers. Critical bugs in subsystems like WebGL, affecting high-profile products like Google Maps, have escaped notice in the past because the bots were unreliable. After much re-work, the GPU bots are now among the most reliable automated test machines in the Chromium project. Let's keep them that way.

Flakiness affecting the GPU tests can come in from highly unexpected sources. Here are some examples:

• Intermittent pixel_test failures on Linux where the captured pixels were black, caused by the Display Power Management System (DPMS) kicking in. Disabled the X server's built-in screen saver on the GPU bots in response.
• GNOME dbus-related deadlocks causing intermittent timeouts (Issue 309093 and related bugs).
• Windows Audio system changes causing intermittent assertion failures in the browser (Issue 310838).
• Enabling assertion failures in the C++ standard library on Linux causing random assertion failures (Issue 328249).
• V8 bugs causing random crashes of the Maps pixel test (V8 issues 3022, 3174).
• TLS changes causing random browser process crashes (Issue 264406).
• Isolated test execution flakiness caused by failures to reliably clean up temporary directories (Issue 340415).
• The Telemetry-based WebGL conformance suite caught a bug in the memory allocator on Android not caught by any other bot (Issue 347919).
• context_lost test failures caused by the compositor's retry logic (Issue 356453).
• Multiple bugs in Chromium's support for lost contexts causing flakiness of the context_lost tests (Issue 365904).
• Maps test timeouts caused by Content Security Policy changes in Blink (Issue 395914).
• Weak pointer assertion failures in various webgl_conformance_tests caused by changes to the media pipeline (Issue 399417).
• A change to a default WebSocket timeout in Telemetry causing intermittent failures to run all WebGL conformance tests on the Mac bots (Issue 403981).
• Chrome leaking suspended sub-processes on Windows, apparently a preexisting race condition that suddenly showed up (Issue 424024).
• Changes to Chrome's cross-context synchronization primitives causing the wrong tiles to be rendered (Issue 584381).
• A bug in V8's handling of array literals causing flaky failures of texture-related WebGL 2.0 tests (Issue 606021).
• Assertion failures in sync point management related to lost contexts that exposed a real correctness bug (Issue 606112).
• A bug in glibc's sem_post/sem_wait primitives breaking V8's parallel garbage collection (Issue 609249).
• A change to Blink's memory purging primitive which caused intermittent timeouts of WebGL conformance tests on all platforms (Issue 840988).
• Screen DPI being inconsistent across seemingly identical Linux machines, causing the Maps pixel test to flakily produce incorrectly sized images (Issue 1091410).

If you notice flaky test failures either on the GPU waterfalls or try servers, please file bugs right away with the component Internals>GPU>Testing and include links to the failing builds and copies of the logs, since the logs expire after a few days. GPU pixel wranglers should give the highest priority to eliminating flakiness on the tree.