The video decoder performance tests are a set of tests used to measure the performance of various video decoder implementations. These tests run directly on top of the video decoder implementation, and don't require the full Chrome browser stack. They are build on top of the GoogleTest framework.
[TOC]
The Tast framework provides an easy way to run the video decoder performance tests from a ChromeOS chroot. Test data is automatically deployed to the device being tested. To run all video decoder performance tests use:
tast run $HOST video.DecodeAccelPerf*
Wildcards can be used to run specific sets of tests:
- Run all VP8 performance tests:
tast run $HOST video.DecodeAccelPerfVP8*
- Run all 1080p 60fps performance tests:
tast run $HOST video.DecodeAccelPerf*1080P60FPS
Check the tast video folder for a list of all available tests. See the Tast quickstart guide for more information about the Tast framework.
To run the video decoder performance tests manually the video_decode_accelerator_perf_tests target needs to be built and deployed to the device being tested. Running the video decoder performance tests can be done by executing:
./video_decode_accelerator_perf_tests [<video path>] [<video metadata path>]
e.g.: ./video_decode_accelerator_perf_tests test-25fps.h264
Test videos: Test videos are present for multiple codecs in the media/test/data folder in Chromium's source tree (e.g. test-25fps.vp8). If no video is specified test-25fps.h264 will be used.
Video Metadata: These videos also have an accompanying metadata .json file that needs to be deployed alongside the test video. They can also be found in the media/test/data folder (e.g. test-25fps.h264.json). If no metadata file is specified <video path>.json will be used. The video metadata file contains info about the video such as its codec profile, dimensions and number of frames.
Note: Various CPU scaling and thermal throttling mechanisms might be active on the device (e.g. scaling_governor, intel_pstate). As these can influence test results it's advised to disable them.
To measure decoder performance two different test scenarios are used.
Uncapped decoder performance: In this scenario the specified test video is decoded from start to finish as fast as possible. This test scenario provides an estimate of the decoder's maximum performance (e.g. the maximum FPS).
Capped decoder performance: This scenario simulates a more realistic environment by decoding a video from start to finish at its actual frame rate, while simulating real-time rendering. Frames that are not decoded by the time they should be rendered will be dropped.
Various performance metrics are collected by these tests:
- FPS: The average number of frames the decoder was able to decode per second.
- Frames Dropped: The number of frames that were dropped during playback, only relevant for the capped performance test.
- Dropped frame percentage: The percentage of frames dropped, only relevant for the capped performance test.
- Frame delivery time: The time between subsequent frame deliveries. The average frame delivery time and 25/50/75 percentiles are calculated.
- Frame decode time: The time between scheduling a frame to be decoded and getting the decoded frame. This metric provides a measure of the decoder's latency. The average decode time and 25/50/75 percentiles are calculated.
All performance metrics are written to perf_metrics/<test_name>.json.
Multiple command line arguments can be given to the command:
-v enable verbose mode, e.g. -v=2.
--vmodule enable verbose mode for the specified module,
e.g. --vmodule=*media/gpu*=2.
--output_folder overwrite the output folder used to store
performance metrics, if not specified results
will be stored in the current working directory.
--use-legacy use the legacy VDA-based video decoders.
--use_vd_vda use the new VD-based video decoders with a
wrapper that translates to the VDA interface,
used to test interaction with older components
expecting the VDA interface.
--linear_output use linear buffers as the final output of the
decoder which may require the use of an image
processor internally. This flag only works in
conjunction with --use_vd_vda.
Disabled by default.
--use-gl specify which GPU backend to use, possible values
include desktop (GLX), egl (GLES w/ ANGLE), and
swiftshader (software rendering)
--ozone-platform specify which Ozone platform to use, possible values
depend on build configuration but normally include
x11, drm, wayland, and headless
--disable_vaapi_lock disable the global VA-API lock if applicable,
i.e., only on devices that use the VA-API with a libva
backend that's known to be thread-safe and only in
portions of the Chrome stack that should be able to
deal with the absence of the lock
(not the VaapiVideoDecodeAccelerator).
--gtest_help display the gtest help and exit.
--help display this help and exit.
See the video decoder performance tests source code.