[libav-devel] qsv: How about replace software parsers with MFXVideoDECODE_DecodeHeader

Li, Zhong zhong.li at intel.com
Tue Jan 23 06:28:35 CET 2018


> wm4 <nfxjfg at googlemail.com> wrote:
> 
> > >   Appreciated for any comment. If we are agree with that, patches will
> be sent soon (about following next two weeks).
> >
> > I don't know how qsvdec  works in particular (doesn't it pretty much
> > parse all the codec stream in MFX?).

If you check qsvdec.c, av_parser_parse2() is called. That is one thing different from cuvid.

 But if you look how it went for
> > cuvid vs. nvdec, using as much parsing functionality from libavcodec
> > as possible is superior.
> >
> > Cuvid/nvdec are nvidia's decoder wrappers as they are named in FFmpeg.
> > (They don't necessarily match with the naming nvidia itself uses.)
> > Cuvid is a "full stream" decoder, which lets nvidia's software parse
> > the full stream. nvdec is a hwaccel, using nvidia's low level API on
> > top of libavcodec's own decoders, and which gets only slices
> 
> Sorry, accidentally hit sent. Here's the rest:
> 
> ... nvdec is a hwaccel, using nvidia's low level API on top of libavcodec's own
> decoders, and which gets only slices plus some metadata. nvdec is actually
> pretty much like the dxva2/d3d11va API.
> 
> When I tried cuvid vs. nvdec myself, I found nvdec superior, because it
> integrated so much better into libavcodec. It supported parsing all the
> metadata libavcodec did, it avoided some weird nvidia (or cuvid
> wrapper) issues, and it was faster, because it used libavcodec's native data
> flow instead of having to go over weird FIFOs. Also, nvdec is essentially
> feature complete (except missing deinterlacing), while cuvid still misses
> exporting some metadata. With cuvid we'd be busy adding support to export
> new types of metadata every time, while nvdec simply benefits from
> development of the native libavcodec decoders.
> 
> I get that Intel devs probably prefer qsvdec over dxva2/vaapi/etc.
> because they develop libmfx, but for libavcodec the hwaccel approach will
> always be superior.
> 
> (I.e. basically what Mark said.)

Yes, basically speaking, cuvid vs nvdec is similar as qsv vs dxva2/vaapi. 
For linux, I am open to vaapi and planning do some contribution on this. 
But I think it should be another topic. 
Using software parser (i.e. av_parser_parse2()) for qsvdec seems can't bring too much advantage compared with using MSDK API, though it is stable for some codecs since it is existed a while.



More information about the libav-devel mailing list