GStreamer VA-API Troubleshooting

GStreamer VA-API is not a trivial piece of software. Even though, in my opinion it is a bit over-engineered, the complexity relies on its layered architecture: the user must troubleshoot in which layer is the failure.

So, bear in mind this architecture:

GStreamer VA-API is not a trivial piece of software. Even though, in my opinion it is a bit over-engineered, the complexity relies on its layered architecture: the user must troubleshoot in which layer is the failure.

So, bear in mind this architecture:

libva architecture
libva architecture

And the point of failure could be anywhere.

Drivers

libva is a library designed to load another library called driver or back-end. This driver is responsible to talk with the kernel, windowing platform, memory handling library, or any other piece of software or hardware that actually will do the video processing.

There are many drivers in the wild. As it is an API aiming to stateless video processing, and the industry is moving towards that way to process video, it is expected more drivers would appear in the future.

Nonetheless, not all the drivers have the same level of maturity, and some of them are abandon-ware. For this reason we decided in GStreamer VA-API, some time ago, to add a white list of functional drivers, basically, those developed by Mesa3D and this one from Intel™. If you wish to disable that white-list, you can do it by setting an environment variable:

$ export GST_VAAPI_ALL_DRIVERS=1

Remember, if you set it, you are on your own, since we do not trust on the maturity of that driver yet.

Internal libva↔driver version

Thus, there is an internal API between libva and the driver and it is versioned, meaning that the internal API version of the installed libva library must match with the internal API exposed by the driver. One of the causes that libva could not initialize a driver could be because the internal API version does not match.

Drivers path and driver name

By default there is a path where libva looks for drivers to load. That path is defined at compilation time. Following Debian’s file-system hierarchy standard (FHS) it should be set by distributions in /usr/lib/x86_64-linux-gnu/dri/. But the user can control this path with an environment variable:

$ export LIBVA_DRIVERS_PATH=${HOME}/src/intel-vaapi-driver/src/.libs

The driver path, as a directory, might contain several drivers. libva will try to guess the correct one by querying the instantiated VA display (which could be either KMS/DRM, Wayland, Android or X11). If the user instantiates a VA display different of his running environment, the guess will be erroneous, the library loading will fail.

Although, there is a way for the user to set the driver’s name too. Again, by setting an environment variable:

$ export LIBVA_DRIVER_NAME=iHD

With this setting, libva will try to load iHD_drv_video.so (a new and experimental open source driver from Intel™, targeted for MediaSDK —do not use it yet with GStreamer VAAPI—).

vainfo

vainfo is the diagnostic tool for VA-API. In a couple words, it will iterate on a list of VA displays, in try-and-error strategy, and try to initialize VA. In case of success, vainfo will report the driver signature, and it will query the driver for the available profiles and entry-points.

For example, my skylake board for development will report

$ vainfo
error: can't connect to X server!
libva info: VA-API version 1.1.0
libva info: va_getDriverName() returns 0
libva info: Trying to open /home/vjaquez/gst/master/intel-vaapi-driver/src/.libs/i965_drv_video.so
libva info: Found init function __vaDriverInit_1_1
libva info: va_openDriver() returns 0
vainfo: VA-API version: 1.1 (libva 2.1.1.pre1)
vainfo: Driver version: Intel i965 driver for Intel(R) Skylake - 2.1.1.pre1 (2.1.0-41-g99c3748)
vainfo: Supported profile and entrypoints
      VAProfileMPEG2Simple            : VAEntrypointVLD
      VAProfileMPEG2Simple            : VAEntrypointEncSlice
      VAProfileMPEG2Main              : VAEntrypointVLD
      VAProfileMPEG2Main              : VAEntrypointEncSlice
      VAProfileH264ConstrainedBaseline: VAEntrypointVLD
      VAProfileH264ConstrainedBaseline: VAEntrypointEncSlice
      VAProfileH264ConstrainedBaseline: VAEntrypointEncSliceLP
      VAProfileH264ConstrainedBaseline: VAEntrypointFEI
      VAProfileH264ConstrainedBaseline: VAEntrypointStats
      VAProfileH264Main               : VAEntrypointVLD
      VAProfileH264Main               : VAEntrypointEncSlice
      VAProfileH264Main               : VAEntrypointEncSliceLP
      VAProfileH264Main               : VAEntrypointFEI
      VAProfileH264Main               : VAEntrypointStats
      VAProfileH264High               : VAEntrypointVLD
      VAProfileH264High               : VAEntrypointEncSlice
      VAProfileH264High               : VAEntrypointEncSliceLP
      VAProfileH264High               : VAEntrypointFEI
      VAProfileH264High               : VAEntrypointStats
      VAProfileH264MultiviewHigh      : VAEntrypointVLD
      VAProfileH264MultiviewHigh      : VAEntrypointEncSlice
      VAProfileH264StereoHigh         : VAEntrypointVLD
      VAProfileH264StereoHigh         : VAEntrypointEncSlice
      VAProfileVC1Simple              : VAEntrypointVLD
      VAProfileVC1Main                : VAEntrypointVLD
      VAProfileVC1Advanced            : VAEntrypointVLD
      VAProfileNone                   : VAEntrypointVideoProc
      VAProfileJPEGBaseline           : VAEntrypointVLD
      VAProfileJPEGBaseline           : VAEntrypointEncPicture
      VAProfileVP8Version0_3          : VAEntrypointVLD
      VAProfileVP8Version0_3          : VAEntrypointEncSlice
      VAProfileHEVCMain               : VAEntrypointVLD
      VAProfileHEVCMain               : VAEntrypointEncSlice

And my AMD board with stable packages replies:

$ vainfo
libva info: VA-API version 0.40.0
libva info: va_getDriverName() returns 0
libva info: Trying to open /usr/lib64/dri/radeonsi_drv_video.so
libva info: Found init function __vaDriverInit_0_40
libva info: va_openDriver() returns 0
vainfo: VA-API version: 0.40 (libva )
vainfo: Driver version: mesa gallium vaapi
vainfo: Supported profile and entrypoints
      VAProfileMPEG2Simple            : VAEntrypointVLD
      VAProfileMPEG2Main              : VAEntrypointVLD
      VAProfileVC1Simple              : VAEntrypointVLD
      VAProfileVC1Main                : VAEntrypointVLD
      VAProfileVC1Advanced            : VAEntrypointVLD
      VAProfileH264ConstrainedBaseline: VAEntrypointVLD
      VAProfileH264ConstrainedBaseline: VAEntrypointEncSlice
      VAProfileH264Main               : VAEntrypointVLD
      VAProfileH264Main               : VAEntrypointEncSlice
      VAProfileH264High               : VAEntrypointVLD
      VAProfileH264High               : VAEntrypointEncSlice
      VAProfileNone                   : VAEntrypointVideoProc

Does this mean that VA-API processes video? No. It means that there is an usable VA display which could open a driver correctly and libva can extract symbols from it.

I would like to mention another tool, not official, but I like it a lot, since it extracts almost of the VA information available in the driver: vadumpcaps.c, written by Mark Thompson.

GStreamer VA-API registration

When GStreamer is launched, normally it will register all the available plugins and plugin features (elements, device providers, etc.). All that data is cache and keep until the cache file is deleted or the cache invalidated by some event.

At registration time, GStreamer VA-API will instantiate a DRM-based VA display, which works with no need of a real display (in other words, headless), and will query the driver for the profiles and entry-points tuples, in order to register only the available elements (encoders, decoders. sink, post-processor). If the DRM VA display fails, a list of VA displays will be tried.

In the case that libva could not load any driver, or the driver is not in the white-list, GStreamer VA-API will not register any element. Otherwise gst-inspect-1.0 will show the registered elements:

$ gst-inspect-1.0 vaapi
Plugin Details:
  Name                     vaapi
  Description              VA-API based elements
  Filename                 /usr/lib/x86_64-linux-gnu/gstreamer-1.0/libgstvaapi.so
  Version                  1.12.4
  License                  LGPL
  Source module            gstreamer-vaapi
  Source release date      2017-12-07
  Binary package           gstreamer-vaapi
  Origin URL               http://bugzilla.gnome.org/enter_bug.cgi?product=GStreamer

  vaapijpegdec: VA-API JPEG decoder
  vaapimpeg2dec: VA-API MPEG2 decoder
  vaapih264dec: VA-API H264 decoder
  vaapivc1dec: VA-API VC1 decoder
  vaapivp8dec: VA-API VP8 decoder
  vaapih265dec: VA-API H265 decoder
  vaapipostproc: VA-API video postprocessing
  vaapidecodebin: VA-API Decode Bin
  vaapisink: VA-API sink
  vaapimpeg2enc: VA-API MPEG-2 encoder
  vaapih265enc: VA-API H265 encoder
  vaapijpegenc: VA-API JPEG encoder
  vaapih264enc: VA-API H264 encoder

  13 features:
  +-- 13 elements

Beside the normal behavior, GStreamer VA-API will also invalidate GStreamer’s cache at every boot, or when any of the mentioned environment variables change.

Conclusion

A simple task list to review when GStreamer VA-API is not working at all is this:

#. Check your LIBVA_* environment variables
#. Verify that vainfo returns sensible information
#. Invalidate GStreamer’s cache (or just delete the file)
#. Check the output of gst-inspect-1.0 vaapi

And, if you decide to file a bug in bugzilla, please do not forget to attach the output of vainfo and the logs if the developer asks for them.

GStreamer VA-API 1.14: what’s new?

As you may already know, there is a new release of GStreamer, 1.14. In this blog post we will talk about the new features and improvements of GStreamer VA-API module, though you have a more comprehensive list of changes in the release notes.

Most of the topics explained along this blog post are already mentioned in the release notes, but a bit more detailed.

DMABuf usage

We have improved DMA-buf’s usage, mostly at downstream.

In the case of upstream, we just got rid a nasty hack which detected when to instantiate and use a buffer pool in sink pad with a dma-buf based allocator. This functionality has been already broken for a while, and that code was the wrong way to enabled it. The sharing of a dma-buf based buffer pool to upstream is going to be re-enabled after bug 792034 is merged.

For downstream, we have added the handling of memory:DMABuf caps feature. The purpose of this caps feature is to negotiate a media when the buffers are not map-able onto user space, because of digital rights or platform restrictions.

For example, currently intel-vaapi-driver doesn’t allow the mapping of its produced dma-buf descriptors. But, as we cannot know if a back-end produces or not map-able dma-buf descriptors, gstreamer-vaapi, when the allocator is instantiated, creates a dummy buffer and tries to map it, if it fails, memory:DMABuf caps feature is negotiated, otherwise, normal video caps are used.

VA-API usage

First of all, GStreamer VA-API has support now for libva-2.0, this means VA-API 1.10. We had to guard some deprecated symbols and the new ones. Nowadays most of distributions have upgraded to libva-2.0.

We have improved the initialization of the VA display internal structure (GstVaapiDisplay). Previously, if a X based display was instantiated, immediately it tried to grab the screen resolution. Obviously, this broke the usage of headless systems. We just delay the screen resolution check to when the VA display truly requires that information.

New API were added into VA, particularly for log handling. Now it is possible to redirect the log messages into a callback. Thus, we use it to redirect VA-API message into the GStreamer log mechanisms, uncluttering the console’s output.

Also, we have blacklisted, in autoconf and meson, libva version 0.99.0, because that version is used internally by the closed-source version of Intel MediaSDK, which is incompatible with official libva. By the way, there is a new open-source version of MediaSDK, but we will talk about it in a future blog post.

Application VA Display sharing

Normally, the object GstVaapiDisplay is shared among the pipeline through the GstContext mechanism. But this class is defined internally and it is not exposed to users since release 1.6. This posed a problem when an application wanted to create its own VA Display and share it with an embedded pipeline. The solution is a new context application message: gst.vaapi.app.Display, defined as a GstStructure with two fields: va-display with the application’s vaDisplay, and x11-display with the application’s X11 native display. In the future, a Wayland’s native handler will be processed too. Please note that this context message is only processed by vaapisink.

One precondition for this solution was the removal of the VA display cache mechanism, a lingered request from users, which, of course, we did.

Interoperability with appsink and similar

A hardware accelerated driver, as the Intel one, may have custom offsets and strides for specific memory regions. We use the GstVideoMeta to set this custom values. The problem comes when downstream does not handle this meta, for example, appsink. Then, the user expect the “normal” values for those variable, but in the case of GStreamer VA-API with a hardware based driver, when the user displays the frame, it is shown corrupted.

In order to fix this, we have to make a memory copy, from our custom VA-API images to an allocated system memory. Of course there is a big CPU penalty, but it is better than delivering wrong video frames. If the user wants a better performance, then they should seek for a different approach.

Resurrection of GstGLUploadTextureMeta for EGL renders

I know, GstGLUploadTextureMeta must die, right? I am convinced of it. But, Clutter video sink uses it, an it has a vast number of users, so we still have to support it.

Last release we had remove the support for EGL/Wayland in the last minute because we found a terrible bug just before the release. GLX support has always been there.

With Daniel van Vugt efforts, we resurrected the support for that meta in EGL. Though I expect the replacement of Clutter sink with glimagesink someday, soon.

vaapisink demoted in Wayland

vaapisink was demoted to marginal rank on Wayland because COGL cannot display YUV surfaces.

This means, by default, vaapisink won’t be auto-plugged when playing in Wayland.

The reason is because Mutter (aka GNOME) cannot display the frames processed by vaapisink in Wayland. Nonetheless, please note that in Weston, it works just fine.

Decoders

We have improved a little bit upstream renegotiation: if the new stream is compatible with the previous one, there is no need to reset the internal parser, with the exception of changes in codec-data.

low-latency property in H.264

A new property has added only to H.264 decoder: low-latency. Its purpose is for live streams that do not conform the H.264 specification (sadly there are many in the wild) and they need to twitch the spec implementation. This property force to push the frames in the decoded picture buffer as soon as possible.

base-only property in H.264

This is the result of the Google Summer of Code 2017, by Orestis Floros. When this property is enabled, all the MVC (Multiview Video Coding) or SVC (Scalable Video Coding) frames are dropped. This is useful if you want to reduce the processing time or if your VA-API driver does not support those kind of streams.

Encoders

In this release we have put a lot of effort in encoders.

Processing Regions of Interest

It is possible, for certain back-ends and profiles (for example, H.264 and H.265 encoders with Intel driver), to specify a set of regions of interest per frame, with a delta-qp per region. This mean that we would ask more quality in those regions.

In order to process regions of interest, upstream must add to the video frame, a list of GstVideoRegionOfInterestMeta. This list then is traversed by the encoder and it requests them if the VA-API profile, in the driver, supports it.

The common use-case for this feature is if you want to higher definition in regions with faces or text messages in the picture, for example.

New encoding properties

  • quality-level: For all the available encoders. This is number between 1 to 8, where a lower number means higher quality (and slower processing).

  • aud: This is for H.264 encoder only and it is available for certain drivers and platforms. When it is enabled, an AU delimiter is inserted for each encoded frame. This is useful for network streaming, and more particularly for Apple clients.

  • mbbrc: For H.264 only. Controls (auto/on/off) the macro-block bit-rate.

  • temporal-levels: For H.264 only. It specifies the number of temporal levels to include a the hierarchical frame prediction.

  • prediction-type: For H.264 only. It selects the reference picture selection mode.

    The frames are encoded as different layers. A frame in a particular layer will use pictures in lower or same layer as references. This means decoder can drop frames in upper layer but still decode lower layer frames.

    • hierarchical-p: P frames, except in top layer, are reference frames. Base layer frames are I or B.

    • hierarchical-b: B frames , except in top most layer, are reference frames. All the base layer frames are I or P.

  • refs: Added for H.265 (it was already supported for H.264). It specifies the number of reference pictures.

  • qp-ip and qp-ib: For H.264 and H.265 encoders. They handle the QP (quality parameters) difference between the I and P frames, the the I and B frames respectively.

Set media profile via downstream caps

H.264 and H.265 encoders now can configure the desired media profile through the downstream caps.

Contributors

Many thanks to all the contributors and bug reporters.

     1  Daniel van Vugt
    46  Hyunjun Ko
     1  Jan Schmidt
     3  Julien Isorce
     1  Matt Staples
     2  Matteo Valdina
     2  Matthew Waters
     1  Michael Tretter
     4  Nicolas Dufresne
     9  Orestis Floros
     1  Philippe Normand
     4  Sebastian Dröge
    24  Sreerenj Balachandran
     1  Thibault Saunier
    13  Tim-Philipp Müller
     1  Tomas Rataj
     2  U. Artie Eoff
     1  VaL Doroshchuk
   172  Víctor Manuel Jáquez Leal
     3  XuGuangxin
     2  Yi A Wang