Skip to content

v2021.5.0
Compare
Choose a tag to compare
@mav-intel mav-intel released this 04 Aug 00:41
v2021.5.0
dde640e

New in This Release

  • Added option to build dispatcher as a static library
  • Added ability to build dispatcher under MinGW
  • Fixes for tools and samples

For more information on the preview C++/Python APIs and Samples, see
https://software.intel.com/content/www/us/en/develop/articles/onevpl-preview-examples.html

Issues and Limitations

C++/Python previews

  • This is a preview release made available to gather feedback for API improvements. Compatibility with future releases is not guaranteed.

Dispatcher

  • MFXInit() and MFXInitEx() functions have been deprecated in the 2.0 oneVPL Specification. Use MFXLoad() instead. See the transition guide for more details.
  • If initialized with MFXInit() or MFXInitEx(), support is limited to the 1.x API and will load Intel(R) Media SDK hardware library rather than oneVPL hardware library.
  • If both Intel(R) Media SDK and oneVPL hardware libraries are installed, only oneVPL libraries will be loaded via MFXLoad(). See the oneVPL specification for more details.
  • The Media SDK runtimes enabling legacy GPU hardware only provide API features up to 1.35, even when initialized with MFXLoad. For more details, see Upgrading from Intel® Media SDK to Intel® oneAPI Video Processing Library.
  • The libmfx.dll and libmfx.so.2021.1.11 dispatchers are deprecated. They are provided for backwards compatibility only and will be removed from a future release. They do not include any changes since the 2021.1.1 release.
  • Dispatcher support has been extended to include hardware implementations, including those that implement API versions lower then 2.0. Applications patterned after the 2021.1.1 release samples, rather than the 2.0 specification may fail after re-build as the previous sample code made the assumption that 2.0 APIs would be available. To remedy this please follow the coding patterns in the oneVPL spec

Tools and samples

  • Releases installed prior to the 2.4 API update may cause CMAKE build failures on Linux due to compatibility issues. Prior package versions were aligned with release version, rather than API. As a result find_package will look for a current version i.e. 2.4 but find a larger legacy usage i.e. 2021.2.2 and assume forward compatibility.
  • dpccp-blur sample incorrectly states its usage. The sample example commandline should read:
dpcpp-blur -i in.i420 -w 128 -h 96
To view: ffplay -f rawvideo -pixel_format bgra -video_size 256x192 -pixel_format yuv420p out.raw
  • legacy-vpp sample in hardware mode is not supported on Sky Lake.
  • hello-transcode readme incorrectly implies sample runs on hardware.
  • decvpp_tool does not support gen12 platforms on Windows in Hardware mode.
  • Encode B-frame distance setting is inconsistent between CPU and GPU implementations. For the CPU implementation 0 means no B-frames, while for the GPU implementation 1 means no B-frames.
  • Colorspace flags are required when using sample_* tools, if colorspace is not set tools will fail.
  • Input size should be aligned to 16 bytes for sample_encode.
  • The sample_multi_transcode tool included in this release is only intended for GPU operation to remain more closely aligned with the version of this tool released with Intel(R) Media SDK.
  • The sample_multi_transcode tool has a synchronization issue in 1->N pipelines. When decode is faster than encoders, the issue can cause the free frames consumed faster than the framework released, eventually resulting in tool crash when there is no more memory available.
  • hello-transcode sample only supports the oneVPL CPU implementation in this release.
  • sample_decode may stop responding if the input is switched between multiple resolutions.
  • sample_encode occasionally hangs for (CPU) SVT-HEVC encodes when using VBR in Ubuntu 18.04.
  • sample_encode does not work if the input is scaled or color converted.
  • sample_encode, sample_decode, and sample_vpp are not in sync with Media SDK's version of these tools. This will be corrected in future releases.
  • Windows samples are installed to C:\Program Files (x86)\Intel\oneAPI\vpl\latest\examples by default. This should not be a writable location, so you will not be able to build samples here. To build examples, copy this folder to another location first.
  • When configuring systems to run the OpenVINO™ interop samples in Linux gmmlib version conflicts can occur. Debug shows undefined GmmLib symbols as the cause of a runtime error. This can be mitigated by using the media stack integrated with OpenVINO™ installation. Uninstall other versions of libva, iHD, and libgmm then select Media SDK when installing OpenVINO™ as described in the OpenVINO™ install documentation.
Assets 2
Loading