Thanks Microsoft: RadeonSI Enables Async VA-API For Faster Video Acceleration

Last week I wrote about how Microsoft landed a VA-API improvement in Mesa to support faster Video Acceleration API encoding with FFmpeg. That code was initially only wired up for the Microsoft D3D12 driver within Mesa for WSL use-cases, but now AMD has taken advantage of the new capability for RadeonSI Gallium3D usage with their Radeon GPUs.

Beyond the exciting performance optimizations landing this week in Mesa 22.3, the RadeonSI Gallium3D driver is now supporting the "PIPE_VIDEO_CAP_ENC_SUPPORTS_ASYNC_OPERATION" for enabling the async_depth VA-API support with FFmpeg.

This option allows increasing the maximum processing parallelism for VA-API drivers supporting vaSyncBuffer. Thanks to the Microsoft work for getting this working with Mesa's video acceleration code, it can mean faster video encoding performance and better GPU utilization with the increased parallelism.

Thanks to the heavy lifting by Microsoft, the AMD change is just two lines of code for enabling the video capability with Raven graphics hardware and newer. Basically, AMD Radeon graphics with the Video Core Next "VCN" block can make use of this async functionality.

All of this recent work will be found in Mesa 22.3 due out as stable around the end of November or more than likely in December.