video/out/hwdec/hwdec_vaapi: Remove unnecessary explicit synchronization

[langyuxf]

Implicit synchronization will handle that.
Extra explicit synchronization here would bring some overheads.

Read this before you submit this pull request:
https://github.com/mpv-player/mpv/blob/master/DOCS/contribute.md

Reading this link and following the rules will get your pull request reviewed
and merged faster. Nobody wants lazy pull requests.

[philipl]

The documentation for vaExportSurfaceHandle explicitly states:

This does not perform any synchronisation. If the contents of the surface will be read, vaSyncSurface() must be called before doing so. If the contents of the surface are written, then all operations must be completed externally before using the surface again by via VA-API functions.

[philipl]

We've also empirically observed synchronisation problems (displaying old frames) without the sync call.

[langyuxf]

The documentation for vaExportSurfaceHandle explicitly states:

This does not perform any synchronisation. If the contents of the surface will be read, vaSyncSurface() must be called before doing so. If the contents of the surface are written, then all operations must be completed externally before using the surface again by via VA-API functions.

My understanding is it's the consumer's responsibility to do the sync.
If the consumer is GL, actually it would be implicitly synced in kernel mode driver.
For consumers can't do implicitly sync, vaSyncSurface() is necessary.

vaSyncSurface() for decode and process is not implemented in Mesa until https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20133

[philipl]

That MR is only relevant for AMD. vaSyncSurface has been meaningful on Intel for much longer. Also remember that we have to support importing to Vulkan, which most assuredly will not do implicit sync - and without a sync object to import (ie: a semaphore) there is no way to make it sync - which is the situation in which I observed the incorrect behaviour mentioned previously.

[langyuxf]

Ok, it would be better to do the sync in consumer side(e.g., GL, Vulkan) instead of producer side.

[langyuxf]

As far as I know, if the surface is exported as a dma-buf, it will be implicitly synced when imported and mapped either by GL or Vulkan in Linux KMD.

[philipl]

If you can point us to either the code or developer statements that guarantee this behaviour for both GL and Vulkan and Intel and AMD, then we can consider it. But as I've said, the only evidence I have, is that things look wrong in at least one combination without it.

[langyuxf]

Please take a look at this https://docs.kernel.org/next/driver-api/dma-buf.html#c.dma_buf_export_sync_file.
This is for Vulkan to do explicitly sync.

[github-actions]

Download the artifacts for this pull request:

Windows

• mpv-i686-w64-mingw32
• mpv-x86_64-w64-mingw32

[philipl]

Ok. If you want to update the PR to include the necessary sync parts, then we can consider merging it.

[langyuxf]

Thanks, let me have a try.