ANGLE Restricted Traces

The files in this directory are traces of real applications. We host them internally because they may contain third party IP which we don't want to share publicly.

Accessing the traces

In order to compile and run with these, you must be granted access by Google, then authenticate with CIPD. Googlers, use your @google account.

cipd auth-login

Add the following to ANGLE's .gclient file:

    "custom_vars": {
      "checkout_angle_restricted_traces": True
    },

Then use gclient to pull down binary files from CIPD.

gclient sync -D

This should result in a number of directories created in src/tests/restricted_traces that contain the trace files listed in restricted_traces.json:

$ ls -d src/tests/restricted_traces/*/
src/tests/restricted_traces/aliexpress/
src/tests/restricted_traces/angry_birds_2_1500/
src/tests/restricted_traces/arena_of_valor/
src/tests/restricted_traces/asphalt_8/
src/tests/restricted_traces/brawl_stars/
src/tests/restricted_traces/bus_simulator_indonesia/
src/tests/restricted_traces/candy_crush_500/
src/tests/restricted_traces/clash_of_clans/
src/tests/restricted_traces/clash_royale/
src/tests/restricted_traces/cod_mobile/
...

Building the trace tests

To build for Android, follow the steps in DevSetupAndroid.md (Recommend using the Performance arguments for best performance)

To build for Desktop, follow the steps in DevSetup.md

When that is working, add the following GN arg to your setup:

build_angle_trace_perf_tests = true

To build the angle_perftests:

autoninja -C out/<config> angle_perftests

Run them like so:

out/<config>/angle_perftests --gtest_filter=TracePerfTest*

(Optional) Reducing the trace count

Since the traces are numerous, you can limit build to a subset of them if needed with the angle_restricted_traces GN arg. This arg takes a list of traces as found in restricted_traces.json. For example:

angle_restricted_traces = ["world_of_kings 5", "worms_zone_io 5"]

Capturing and adding new Android traces

Generally we want to use a Debug setup for recording new traces. That allows us to see asserts and errors if the tracer needs to be improved. Add the following GN arg to your Debug setup:

angle_with_capture_by_default = true

After building and installing the APK with the above arg, we're ready to start capturing.

Determine the target app

We first need to identify which application we want to trace. That can generally be done by looking at the web-based Play Store entry for your app. For instance, Angry Birds 2 is listed here: https://play.google.com/store/apps/details?id=com.rovio.baba

If there is no Play Store entry for your app, there are a couple of ways you can determine the app's name.

If you have a userdebug build of Android, you can check logcat when you launch the application. You should see an entry like this:

GraphicsEnvironment: ANGLE Developer option for 'com.rovio.baba' set to: 'default'

If you just have an APK, you can use the following command to find the package name:

$ aapt dump badging angry_birds_2.apk | grep package
package: name='com.rovio.baba' versionCode='24900001' versionName='2.49.1' platformBuildVersionName=''

You can also just guess at the package name, then check your device to see if it is installed. Keep trying combinations until you find it:

$ adb shell pm list packages | grep rovio
package:com.rovio.baba

Track the package name for use in later steps:

export PACKAGE_NAME=com.rovio.baba

Choose a trace name

Next, we need to chose a name for the trace. Choose something simple that identifies the app, then use snake case. This will be the name of the trace files, including the trace directory. Changing this value later is possible, but not recommended.

export LABEL=angry_birds_2

Opt the application into ANGLE

Next, opt the application into using your ANGLE with capture enabled by default:

adb shell settings put global angle_debug_package org.chromium.angle
adb shell settings put global angle_gl_driver_selection_pkgs $PACKAGE_NAME
adb shell settings put global angle_gl_driver_selection_values angle

Set up some Capture/Replay properties

We also need to set some debug properties used by the tracer.

Ensure frame capture is enabled. This might be redundant, but ensure the property isn't set to zero, which disables frame capture.

adb shell setprop debug.angle.capture.enabled 1

Empty the start and end frames. Again, this might be redundant, but it is less confusing.

adb shell setprop debug.angle.capture.frame_start '""'
adb shell setprop debug.angle.capture.frame_end '""'

Set the label to be used in the trace files

adb shell setprop debug.angle.capture.label $LABEL

Set a trigger value to be used by the tracer. This should be set to the number of frames you want to capture. We typically use 10 to get an idea of how a scene is running, but some workloads require more. Use your discretion here:

adb shell setprop debug.angle.capture.trigger 10

Create output location

We need to write out the trace file in a location accessible by the app. We use the app‘s data storage on sdcard, but create a subfolder to isolate ANGLE’s files:

adb shell mkdir -p /sdcard/Android/data/$PACKAGE_NAME/angle_capture

Start the target app

From here, you can start the application. You should see logcat entries like the following, indicating that we've succesfully turned on capturing:

ANGLE   : INFO: Capture trigger detected, disabling capture start/end frame.
ANGLE   : INFO: Limiting binary format support count to zero while FrameCapture enabled
ANGLE   : INFO: Limiting image unit count to 8 while FrameCapture enabled
ANGLE   : INFO: Setting uniform buffer offset alignment to 256 while FrameCapture enabled
ANGLE   : INFO: Disabling GL_EXT_map_buffer_range and GL_OES_mapbuffer during capture, which are not supported on some native drivers
ANGLE   : INFO: Disabling GL_CHROMIUM_bind_uniform_location during capture, which is not supported on native drivers
ANGLE   : INFO: Disabling GL_NV_shader_noperspective_interpolation during capture, which is not supported on some native drivers
ANGLE   : INFO: Limiting draw buffer count to 4 while FrameCapture enabled

Trigger the capture

When you have reached the content in your application that you want to record, set the trigger value to zero:

adb shell setprop debug.angle.capture.trigger 0

In logcat we'll see another entry corresponding to this:

ANGLE   : INFO: Capture triggered after frame 30440 for 10 frames

The app may pause briefly when the capture is completing. You can check its progress by looking at the file system:

adb shell ls -la /sdcard/Android/data/$PACKAGE_NAME/angle_capture

Allow the app to run until the *angledata.gz file is non-zero and no longer growing. The app should continue rendering after that:

$ adb shell ls -s -w 1 /sdcard/Android/data/$PACKAGE_NAME/angle_capture
30528 angry_birds_2_capture_context1.angledata.gz
    8 angry_birds_2_capture_context1.cpp
    4 angry_birds_2_capture_context1_files.txt
  768 angry_birds_2_capture_context1_frame001.cpp
  100 angry_birds_2_capture_context1_frame002.cpp
  100 angry_birds_2_capture_context1_frame003.cpp
  100 angry_birds_2_capture_context1_frame004.cpp
  100 angry_birds_2_capture_context1_frame005.cpp
  104 angry_birds_2_capture_context1_frame006.cpp
  100 angry_birds_2_capture_context1_frame007.cpp
  100 angry_birds_2_capture_context1_frame008.cpp
  100 angry_birds_2_capture_context1_frame009.cpp
  100 angry_birds_2_capture_context1_frame010.cpp
  120 angry_birds_2_capture_context1_frame011.cpp
    8 angry_birds_2_capture_context1.h

Note, you may see multiple contexts captured in the output. When this happens, look at the size of the files. The larger files should be the context you care about it. You should move or delete the other context files.

Pull the trace files

Next, we want to pull those files over to the host and run some scripts.

cd $CHROMIUM_SRC/third_party/angle/src/tests/restricted_traces
mkdir -p $LABEL
adb pull /sdcard/Android/data/$PACKAGE_NAME/angle_capture/. $LABEL/

Add the new trace to the JSON list

The list of traces is tracked in restricted_traces.json. Manually add your new trace to this list. Use version “1” for the trace version.

On Linux, you can also use a tool called jq to update the list. This ensures we get them in alphabetical order with no duplicates. It can also be done by hand if you are unable to install it, for some reason.

sudo apt-get install jq

Then run the following command:

export VERSION=1
jq ".traces = (.traces + [\"$LABEL $VERSION\"] | unique)" restricted_traces.json | sponge restricted_traces.json

Run code auto-generation

The gen_restricted_traces script auto-generates entries in our checkout dependencies to sync restricted trace data on checkout. To trigger code generation run the following from the angle root folder:

python ./scripts/run_code_generation.py

After this you should be able to git diff and see changes in the following files:

  • DEPS
  • scripts/code_generation_hashes/restricted_traces.json
  • src/tests/restricted_traces/restricted_traces.json (this is the file you originally modified)

Note the absence of the traces themselves listed above. They are automatically ignored by .gitignore since they won't be checked in directly to the repo.

Upload your trace to CIPD

Once you feel good about your trace, you can upload it to our collection of traces. This can only be done by Googlers with write access to the trace CIPD prefix. If you need write access contact someone listed in the OWNERS file.

Please kindly go over the trace content with ANGLE code owners before running below commands. You can share your trace through Google Drive for content iterations. We cannot delete trace files once they are up on the CIPD. Doing additional rounds of content check can help us save CIPD resources.

./sync_restricted_traces_to_cipd.py

Upload your CL

Ensure your current working directory is up-to-date, and upload:

git cl upload

You're now ready to run your new trace on CI!

Upgrading existing traces

With tracer updates sometimes we want to re-run tracing to upgrade the trace file format or to take advantage of new tracer improvements. The retrace_restricted_traces script allows us to re-run tracing using SwiftShader on a desktop machine. As of writing we require re-tracing on a Windows machine because of size limitations with a Linux app window.

Prep work: Back up existing traces

This will save the original traces in a temporary folder if you need to revert to the prior trace format:

py ./src/tests/restricted_traces/retrace_restricted_traces.py backup "*"

Note: on Linux, remove the command py prefix to the Python scripts.

This will save the traces to ./retrace-backups. At any time you can revert the trace files by running:

py ./src/tests/restricted_traces/retrace_restricted_traces.py restore "*"

Part 1: Sanity Check with T-Rex

First we'll retrace a single app to verify the workflow is intact. Please ensure you replace the specified variables with paths that work on your configuration and checkout:

Step 1/3: Capture T-Rex with Validation

export TRACE_GN_PATH=out/Debug
export TRACE_NAME=trex_200
py ./src/tests/restricted_traces/retrace_restricted_traces.py upgrade $TRACE_GN_PATH retrace-wip -f $TRACE_NAME --validation --limit 3

The --validation flag will turn on additional validation checks in the trace. The --limit 3 flag forces a maximum of 3 frames of tracing so the test will run more quickly. The trace will end up in the retrace-wip folder.

Step 2/3: Validate T-Rex

The command below will update your copy of the trace, rebuild, the run the test suite with validation enabled:

py ./src/tests/restricted_traces/retrace_restricted_traces.py validate $TRACE_GN_PATH retrace-wip $TRACE_NAME

If the trace failed validation, see the section below on diagnosing tracer errors. Otherwise proceed with the steps below.

Step 3/3: Restore the Canonical T-Rex Trace

py ./src/tests/restricted_traces/retrace_restricted_traces.py restore $TRACE_NAME

Part 2: Do a limited trace upgrade with validation enabled

Step 1/3: Upgrade all traces with a limit of 3 frames

py ./src/tests/restricted_traces/retrace_restricted_traces.py upgrade $TRACE_GN_PATH retrace-wip --validation --limit 3  --no-overwrite

If this process gets interrupted, re-run the upgrade command. The --no-overwrite argument will ensure it will complete eventually.

If any traces failed to upgrade, see the section below on diagnosing tracer errors. Otherwise proceed with the steps below.

Step 2/3: Validate all upgraded traces

py ./src/tests/restricted_traces/retrace_restricted_traces.py validate $TRACE_GN_PATH retrace-wip "*"

If any traces failed validation, see the section below on diagnosing tracer errors. Otherwise proceed with the steps below.

Step 3/3: Restore all traces

py ./src/tests/restricted_traces/retrace_restricted_traces.py restore "*"

Part 3: Do the full trace upgrade

rm -rf retrace-wip
py ./src/tests/restricted_traces/retrace_restricted_traces.py upgrade $TRACE_GN_PATH retrace-wip --no-overwrite

If this process gets interrupted, re-run the upgrade command. The --no-overwrite argument will ensure it will complete eventually.

If any traces failed to upgrade, see the section below on diagnosing tracer errors. Otherwise proceed with the steps below.

Part 4: Test the upgraded traces under an experimental prefix (slow)

Because there still may be trace errors undetected by validation, we first upload the traces to a temporary CIPD path for testing. After a successful run on the CQ, we will then upload them to the main ANGLE prefix.

To enable the experimental prefix, edit restricted_traces.json to use a version number beginning with ‘x’. For example:

  "traces": [
    "aliexpress x1",
    "among_us x1",
    "angry_birds_2_1500 x1",
    "arena_of_valor x1",
    "asphalt_8 x1",
    "avakin_life x1",
... and so on ...

Then run:

py ./src/tests/restricted_traces/retrace_restricted_traces.py restore -o retrace-wip "*"
py ./src/tests/restricted_traces/sync_restricted_traces_to_cipd.py
py ./scripts/run_code_generation.py

The restore command will copy the new traces from the retrace-wip directory into the trace folder before we call the sync script.

After these commands complete succesfully, create and upload a CL as normal. Run CQ +1 Dry-Run. If you find a test regression, see the section below on diagnosing tracer errors. Otherwise proceed with the steps below.

Part 5: Upload the verified traces to CIPD under the stable prefix

Now that you've validated the traces on the CQ, update restricted_traces.json to remove the ‘x’ prefix and incrementing the version of the traces (skipping versions if you prefer) and then run:

py ./src/tests/restricted_traces/sync_restricted_traces_to_cipd.py
py ./scripts/run_code_generation.py

Then create and upload a CL as normal. Congratulations, you've finished the trace upgrade!

Diagnosing and fixing tracer errors

Debugging a crash or GLES error

Ensure you're building ANGLE in Debug. Then look in the retrace script output to find the exact command line and environment variables the script uses to produce the failure. For example:

INFO:root:ANGLE_CAPTURE_LABEL=trex_200 ANGLE_CAPTURE_OUT_DIR=C:\src\angle\retrace-wip\trex_200 ANGLE_CAPTURE_FRAME_START=2 ANGLE_CAPTURE_FRAME_END=4 ANGLE_CAPTURE_VALIDATION=1 ANGLE_FEATURE_OVERRIDES_ENABLED=allocateNonZeroMemory:forceInitShaderVariables ANGLE_CAPTURE_TRIM_ENABLED=1 out/Debug\angle_perftests.exe --gtest_filter=TracePerfTest.Run/vulkan_swiftshader_trex_200 --max-steps-performed 3 --retrace-mode --enable-all-trace-tests

Once you can reproduce the issue you can use a debugger or other standard debugging processes to find the root cause and a fix.

Debugging a serialization difference

If you encouter a serialization mismatch in the retrace, you can find the complete serialization output by looking in the retrace script output. ANGLE saves the complete serialization file contents on any mismatch. You can inspect and diff these files in a text editor to help diagnose what objects are faulty.

If the mismatch is with a Buffer or Texture object content, you can manually edit the frame_capture_utils.cpp file to force some or all of the objects to serialize their entire contents. This can help show what kind of pixel or data differences might be causing the issue. For example, change this line:

json->addBlob("data", dataPtr->data(), dataPtr->size());

to

json->addBlobWithMax("data", dataPtr->data(), dataPtr->size(), 1000000);

Note: in the future, we might make this option exposed via an envioronment variable, or even allow serialization of entire data blocks in text-encoded form that could be decoded to separate files.

If you still can't determine what code might be causing the state difference, we can insert finer-grained serialization checkpoints to “bisect” where the coding mismatch is happening. It is not possible to force checkpoints after every GLES call, because serialization and validation is so prohibitively expensive. ANGLE instead has feature in the tracer that allows us to precisely control where the tracer inserts and validates the checkpoints, by using a boolean expression language.

The retrace script command --validation-expr allows us to specify a C-like expression that determines when to add serialization checkpoints. For example, we can specify this validation expression:

((frame == 2) && (call < 1189) && (call > 1100) && ((call % 5) == 0))

Using this expression will insert a serialization checkpoint in the second frame, on every 5th captured call, and when the captured call count is between 1101 and 1188. Here the call keyword denotes the call counter, which resets to 1 every frame, and increments by 1 with every captured GLES API call. The frame keyword denotes the frame counter, which starts at 1 and increments by 1 every captured frame. The expression syntax supports all common C boolean operators.

By finding a starting and ending frame range, and narrowing this range through experimentation, you can pinpoint the exact call that triggers the serialization mismatch, and then diagnose and fix the root cause. In some cases you can use RenderDoc or other frame debugging tools to inspect resource states before/after the bad call once you have found it.

See also: http://crrev.com/c/3136094

Debugging a pixel test failure without a serialization mismatch

Sometimes you manage to complete validation and upload, just to find a golden image pixel difference that manifests in some trace configurations. These problems can be harder to root cause. For instance, some configurations may render undefined pixels that are in practice well-defined on most GLES implementations.

The pixel differences can also be a product of mismatched state even if the trace validation says all states are matched. Because ANGLE's GLES state serialization is incomplete, it can help to check the state serialization logic and add missing features as necessary.