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.
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:
tast run $HOST video.DecodeAccelPerfVP8*
tast run $HOST video.DecodeAccelPerf*1080P60FPS
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>]
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:
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_vd use the new VD-based video decoders, instead of the default 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. --gtest_help display the gtest help and exit. --help display this help and exit.
See the video decoder performance tests source code.