AV1 is an open video compression format from the Alliance for Open Media (AOM) .
The reference software implementation for AV1 encoding and decoding can be found in the libxaom library. Significant efforts have been made to optimize the libxaom implementation on Arm platforms. This includes making use of the Neon and SVE2 vector extensions available on Arm Neoverse platforms. The optimized code is available on
Google Git
.
You will need various development tools to build libxaom including CMake and the GNU compiler.
The instructions assume you are running Ubuntu.
Install the required tools by running:
sudo apt install gcc g++ wget cmake p7zip-full -y
Download the libxaom source code:
git clone https://aomedia.googlesource.com/aom
Change directory to the repository, configure the build, and build the library:
mkdir -p aom/build_aom
cd aom/build_aom
cmake ..
make -j$(nproc)
For additional details refer to the README .
The libxaom library has a comprehensive suite of unit tests, written using the GTest framework.
The build above includes the test_libaom executable.
You can run all unit tests by starting test_libaom with no arguments. However, the number of tests is huge, and it takes a long to run. Instead, you can constrain the number of tests by specifying a filter.
You can try this help argument:
./test_libaom --help
There is also an argument to list the tests:
./test_libaom --gtest_list_tests | less
To run a subset of tests, you can use a filter to run only the Neon Sum of Absolute Difference (SAD) tests.
Here is an example with the filter:
./test_libaom --gtest_filter="*NEON*SAD*"
The output is:
Note: Google Test filter = *NEON*SAD*-:NEON_I8MM.*:NEON_I8MM/*:NEON_I8MM_*:SVE.*:SVE/*:SVE_*:SVE2.*:SVE2/*:SVE2_*
[==========] Running 3650 tests from 17 test suites.
[----------] Global test environment set-up.
[----------] 22 tests from NEON/MaskedSADTest
[ RUN ] NEON/MaskedSADTest.OperationCheck/0
[ OK ] NEON/MaskedSADTest.OperationCheck/0 (152 ms)
[ RUN ] NEON/MaskedSADTest.OperationCheck/1
[ OK ] NEON/MaskedSADTest.OperationCheck/1 (152 ms)
[ RUN ] NEON/MaskedSADTest.OperationCheck/2
[ OK ] NEON/MaskedSADTest.OperationCheck/2 (152 ms)
[ RUN ] NEON/MaskedSADTest.OperationCheck/3
[ OK ] NEON/MaskedSADTest.OperationCheck/3 (152 ms)
[ RUN ] NEON/MaskedSADTest.OperationCheck/4
[ OK ] NEON/MaskedSADTest.OperationCheck/4 (153 ms)
[ RUN ] NEON/MaskedSADTest.OperationCheck/5
[ OK ] NEON/MaskedSADTest.OperationCheck/5 (152 ms)
[ RUN ] NEON/MaskedSADTest.OperationCheck/6
[ OK ] NEON/MaskedSADTest.OperationCheck/6 (152 ms)
< output omitted>
[----------] 90 tests from NEON_DOTPROD/SADSkipx4Test (1433 ms total)
[----------] Global test environment tear-down
[==========] 3650 tests from 17 test suites ran. (56720 ms total)
[ PASSED ] 3650 tests.
YOU HAVE 581 DISABLED TESTS
You can benchmark either the encoding or decoding processes. In this section, you will focus on encoding.
For Performance benchmarking, you can select video encoding either on-demand or live stream.
To start, download some example 8-bit FHD, 8-bit 4K and 10-bit 4K video files:
wget https://ultravideo.fi/video/Bosphorus_1920x1080_120fps_420_8bit_YUV_Y4M.7z
wget https://ultravideo.fi/video/Bosphorus_3840x2160_120fps_420_8bit_YUV_Y4M.7z
wget https://ultravideo.fi/video/Bosphorus_3840x2160_120fps_420_10bit_YUV_Y4M.7z
Next, extract the contents of the 7z files:
7za e Bosphorus_1920x1080_120fps_420_8bit_YUV_Y4M.7z
7za e Bosphorus_3840x2160_120fps_420_8bit_YUV_Y4M.7z
7za e Bosphorus_3840x2160_120fps_420_10bit_YUV_Y4M.7z
For on-demand encoding you can experiment different number of processors and monitor performance.
For example, run with --good and use the --cpu-used argument to vary video quality/compression options from 2 to 6.
The --cpu-used flag is used to trade-off encoding speed for resulting video quality/compression, not to determine how many CPUs to use for parallel encoding. Lower numbers indicate better quality and longer encoding time.
Run standard bit depth and change the CPU count and see the results using:
./aomenc --good --cpu-used=4 --bit-depth=8 -o output.mkv Bosphorus_1920x1080_120fps_420_8bit_YUV.y4m
Try the above command with different --cpu-used values.
./aomenc --good --cpu-used=6 --bit-depth=8 -o output_cpu_used_2.mkv Bosphorus_1920x1080_120fps_420_8bit_YUV.y4m
You can do the same for high bit depth:
./aomenc --good --cpu-used=4 --bit-depth=10 -o output.mkv Bosphorus_3840x2160_120fps_420_10bit.y4m
For live stream encoding you can experiment with different number of processors and monitor performance using the --cpus-used flag, varying the range from 5 to 9.
For standard bit depth with 8 CPUs run:
./aomenc --rt --cpu-used=8 --bit-depth=8 -o output.mkv Bosphorus_1920x1080_120fps_420_8bit_YUV.y4m
For high bit depth run:
./aomenc --rt --cpu-used=8 --bit-depth=10 -o output.mkv Bosphorus_3840x2160_120fps_420_10bit.y4m
The encoding frame rate (which is frames per second) for the video files is output at the end of each run.
Shown below is example output from running the libxaom codec on the 8-bit FHD sample video file with different --cpu-used settings to compare the encoding time.
First, run the codec with --cpu-used=2:
./aomenc --good --cpu-used=2 --bit-depth=8 -o output_cpu_used_2.mkv Bosphorus_1920x1080_120fps_420_8bit_YUV.y4m
The output is similar to:
Pass 1/2 frame 600/601 139432B 1859b/f 55770b/s 62641 ms (9.58 fps)
Pass 2/2 frame 600/600 638429B 8512b/f 255360b/s 1103538 ms (0.54 fps)
Now, run the codec with --cpu-used=6:
./aomenc --good --cpu-used=6 --bit-depth=8 -o output_cpu_used_6.mkv Bosphorus_1920x1080_120fps_420_8bit_YUV.y4m
The output is similar to:
Pass 1/2 frame 600/601 139432B 1859b/f 55770b/s 17500 ms (34.28 fps)
Pass 2/2 frame 600/600 637902B 8505b/f 255150b/s 103754 ms (5.78 fps)
As demonstrated in these examples, --cpu-used=6 is significantly faster than --cpu-used=2, by a factor of 2 to 10.