[go: nahoru, domu]

Using the Video Decode/Encode Accelerator Unittests Manually

VDAtest (or video_decode_accelerator_unittest) and VEAtest (or video_encode_accelerator_unittest) are unit tests that embeds the Chrome video decoding/encoding stack without requiring the whole browser, meaning they can work in a headless environment. They includes a variety of tests to validate the decoding and encoding stacks with h264, vp8 and vp9.

Running these tests manually can be very useful when bringing up a new codec, or in order to make sure that new code does not break hardware decoding and/or encoding. This document is a walk though the prerequisites for running these programs, as well as their most common options.

Prerequisites

The required kernel drivers should be loaded, and there should exist a /dev/video-dec0 symbolic link pointing to the decoder device node (e.g. /dev/video-dec0/dev/video0). Similarly, a /dev/video-enc0 symbolic link should point to the encoder device node.

The unittests can be built by specifying the video_decode_accelerator_unittest and video_encode_accelerator_unittest targets to ninja. If you are building for an ARM board that is not yet supported by the simplechrome workflow, use arm-generic as the board. It should work across all ARM targets.

For unlisted Intel boards, any other Intel target (preferably with the same chipset) should be usable with libva. AMD targets can use amd64-generic.

Basic VDA usage

The media/test/data folder in Chromium's source tree contains files with encoded video data (test-25fps.h264, test-25fps.vp8 and test-25fps.vp9). Each of these files also has a .md5 counterpart, which contains the md5 checksums of valid thumbnails.

Running the VDAtest can be done as follows:

./video_decode_accelerator_unittest --disable_rendering --single-process-tests --test_video_data=test_video

Where test_video is of the form

filename:width:height:numframes:numfragments:minFPSwithRender:minFPSnoRender:profile

The correct value of test_video for each test file follows:

  • H264: test-25fps.h264:320:240:250:258:35:150:1
  • VP8: test-25fps.vp8:320:240:250:250:35:150:11
  • VP9: test-25fps.vp9:320:240:250:250:35:150:12

So in order to run all h264 tests, one would invoke

./video_decode_accelerator_unittest --disable_rendering --single-process-tests --test_video_data=test-25fps.h264:320:240:250:258:35:150:1

Test filtering options

./video_decode_accelerator_unittest --help will list all valid options.

The list of available tests can be retrieved using the --gtest_list_tests option.

By default, all tests are run, which can be a bit too much, especially when bringing up a new codec. The --gtest_filter option can be used to specify a pattern of test names to run. For instance, to only run the TestDecodeTimeMedian test, one can specify --gtest_filter="*TestDecodeTimeMedian*".

So the complete command line to test vp9 decoding with the TestDecodeTimeMedian test only (a good starting point for bringup) would be

./video_decode_accelerator_unittest --disable_rendering --single-process-tests --test_video_data=test-25fps.vp9:320:240:250:250:35:150:12 --gtest_filter="*TestDecodeTimeMedian*"

Verbosity options

The --vmodule options allows to specify a set of source files that should be more verbose about what they are doing. For basic usage, a useful set of vmodule options could be:

--vmodule=*/media/gpu/*=4

Testing performance

Use the --disable_rendering --rendering_fps=0 --gtest_filter="DecodeVariations/*/0" options to max the decoder output and measure its performance.

Testing parallel decoding

Use --gtest_filter="ResourceExhaustion*/0" to run 3 decoders in parallel, and --gtest_filter="ResourceExhaustion*/1" to run 4 decoders in parallel.

Wrap-up

Using all these options together, we can invoke VDAtest in the following way for a verbose H264 decoding test:

./video_decode_accelerator_unittest --single-process-tests --disable_rendering --gtest_filter="*TestDecodeTimeMedian*" --vmodule=*/media/gpu/*=4 --test_video_data=test-25fps.h264:320:240:250:258:35:150:1

Import mode

There are two modes in which VDA runs, ALLOCATE and IMPORT. In ALLOCATE mode, the video decoder is responsible for allocating the buffers containing the decoded frames itself. In IMPORT mode, the buffers are allocated by the client and provided to the decoder during decoding. ALLOCATE mode is used during playback within Chrome (e.g. HTML5 videos), while IMPORT mode is used by ARC++ when Android applications require accelerated decoding.\ VDAtest runs VDA in ALLOCATE mode by default. Use --test_import to run VDA in IMPORT mode. VDA cannot run in IMPORT mode on platforms too old for ARC++ to be enabled.

(Recommended) Frame validator

Use --frame_validator=check to verify the correctness of frames decoded by VideoDecodeAccelerator in all test cases. This validator is based on the fact that a decoded content is deterministic in H.264, VP8 and VP9. It reads the expected md5 value of each frame from *.frames.md5, for example, test-25fps.h264.frames.md5 for test-25fps.h264.\ VDATest is able to read the memory of a decoded frame only if VDA runs in IMPORT mode. Therefore, if --frame_validator=check is specified, VDATest runs as if --test_import is specified. See Import mode about IMPORT mode.

Dump mode

Use --frame_validator=dump to write down all the decoded frames. The output format will be I420 and the saved file name will be frame_%{frame-num}_%{width}x%{height}_I420.yuv in the specified directory or a directory whose name is the test file + .frames if unspecified. Here, width and height are visible width and height. For instance, they will be test-25fps.h264.frames/frame_%{frame-num}_320x180_I420.yuv.

How to generate md5 values of decoded frames for a new video stream

It is necessary to generate md5 values of decoded frames for new test streams. ffmpeg with -f framemd5 can be used for this purpose. For instance, ffmpeg -i test-25fps.h264 -f framemd5 test-25fps.frames.md5

Basic VEA usage

The VEA works in a similar fashion to the VDA, taking raw YUV files in I420 format as input and producing e.g. a H.264 Annex-B byte stream. Sample raw YUV files can be found at the following locations:

It is recommended to rename these files after downloading them to e.g. crowd1080.yuv and bear-320x180.yuv.

The VEA can then be tested as follows:

./video_encode_accelerator_unittest --single-process-tests --disable_flush --gtest_filter=SimpleEncode/VideoEncodeAcceleratorTest.TestSimpleEncode/0 --test_stream_data=bear-320x180.yuv:320:180:1:bear.mp4:100000:30

for the bear file, and

./video_encode_accelerator_unittest --single-process-tests --disable_flush --gtest_filter=SimpleEncode/VideoEncodeAcceleratorTest.TestSimpleEncode/0 --test_stream_data=crowd1080.yuv:1920:1080:1:crowd.mp4:4000000:30

for the larger crowd file. These commands will put the encoded output into bear.mp4 and crowd.mp4 respectively. They can then be copied on the host and played with mplayer -fps 25.

Source code

The VDAtest's source code can be consulted here: https://cs.chromium.org/chromium/src/media/gpu/video_decode_accelerator_unittest.cc.

V4L2 support: https://cs.chromium.org/chromium/src/media/gpu/v4l2/.

VAAPI support: https://cs.chromium.org/chromium/src/media/gpu/vaapi/.