[FFmpeg-devel] [PATCH 3/3] libavutil/hwcontext_opencl: fix a bug for mapping qsv frame to opencl

Soft Works softworkz at hotmail.com
Thu Dec 30 02:17:31 EET 2021



> -----Original Message-----
> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of Mark
> Thompson
> Sent: Thursday, December 30, 2021 12:23 AM
> To: ffmpeg-devel at ffmpeg.org
> Subject: Re: [FFmpeg-devel] [PATCH 3/3] libavutil/hwcontext_opencl: fix a bug
> for mapping qsv frame to opencl
> 
> On 28/12/2021 19:04, Soft Works wrote:>> -----Original Message-----
> >> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of Mark
> >> Thompson
> >> Sent: Tuesday, December 28, 2021 1:54 PM
> >> To: ffmpeg-devel at ffmpeg.org
> >> Subject: Re: [FFmpeg-devel] [PATCH 3/3] libavutil/hwcontext_opencl: fix a
> bug
> >> for mapping qsv frame to opencl
> >>
> >> On 28/12/2021 01:17, Soft Works wrote:
> >>>> -----Original Message-----
> >>>> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of Mark
> >>>> Thompson
> >>>> Sent: Tuesday, December 28, 2021 12:46 AM
> >>>> To: ffmpeg-devel at ffmpeg.org
> >>>> Subject: Re: [FFmpeg-devel] [PATCH 3/3] libavutil/hwcontext_opencl: fix
> a
> >> bug
> >>>> for mapping qsv frame to opencl
> >>>>
> >>>> On 27/12/2021 20:31, Soft Works wrote:>> -----Original Message-----
> >>>>>> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of Mark
> >>>>>> Thompson
> >>>>>> Sent: Monday, December 27, 2021 7:51 PM
> >>>>>> To: ffmpeg-devel at ffmpeg.org
> >>>>>> Subject: Re: [FFmpeg-devel] [PATCH 3/3] libavutil/hwcontext_opencl:
> fix
> >> a
> >>>> bug
> >>>>>> for mapping qsv frame to opencl
> >>>>>>
> >>>>>> On 16/11/2021 08:16, Wenbin Chen wrote:
> >>>>>>> From: nyanmisaka <nst799610810 at gmail.com>
> >>>>>>>
> >>>>>>> mfxHDLPair was added to qsv, so modify qsv->opencl map function as
> >> well.
> >>>>>>> Now the following commandline works:
> >>>>>>>
> >>>>>>> ffmpeg -v verbose -init_hw_device vaapi=va:/dev/dri/renderD128 \
> >>>>>>> -init_hw_device qsv=qs at va -init_hw_device opencl=ocl at va -
> >> filter_hw_device
> >>>>>> ocl \
> >>>>>>> -hwaccel qsv -hwaccel_output_format qsv -hwaccel_device qs -c:v
> >> h264_qsv
> >>>> \
> >>>>>>> -i input.264 -vf
> >>>> "hwmap=derive_device=opencl,format=opencl,avgblur_opencl,
> >>>>>> \
> >>>>>>> hwmap=derive_device=qsv:reverse=1:extra_hw_frames=32,format=qsv" \
> >>>>>>> -c:v h264_qsv output.264
> >>>>>>>
> >>>>>>> Signed-off-by: nyanmisaka <nst799610810 at gmail.com>
> >>>>>>> Signed-off-by: Wenbin Chen <wenbin.chen at intel.com>
> >>>>>>> ---
> >>>>>>>      libavutil/hwcontext_opencl.c | 3 ++-
> >>>>>>>      1 file changed, 2 insertions(+), 1 deletion(-)
> >>>>>>>
> >>>>>>> diff --git a/libavutil/hwcontext_opencl.c
> >> b/libavutil/hwcontext_opencl.c
> >>>>>>> index 26a3a24593..4b6e74ff6f 100644
> >>>>>>> --- a/libavutil/hwcontext_opencl.c
> >>>>>>> +++ b/libavutil/hwcontext_opencl.c
> >>>>>>> @@ -2249,7 +2249,8 @@ static int
> opencl_map_from_qsv(AVHWFramesContext
> >>>>>> *dst_fc, AVFrame *dst,
> >>>>>>>      #if CONFIG_LIBMFX
> >>>>>>>          if (src->format == AV_PIX_FMT_QSV) {
> >>>>>>>              mfxFrameSurface1 *mfx_surface = (mfxFrameSurface1*)src-
> >>>>> data[3];
> >>>>>>> -        va_surface = *(VASurfaceID*)mfx_surface->Data.MemId;
> >>>>>>> +        mfxHDLPair *pair = (mfxHDLPair*)mfx_surface->Data.MemId;
> >>>>>>> +        va_surface = *(VASurfaceID*)pair->first;
> >>>>>>>          } else
> >>>>>>>      #endif
> >>>>>>>              if (src->format == AV_PIX_FMT_VAAPI) {
> >>>>>>
> >>>>>> Since these frames can be user-supplied, this implies that the user-
> >> facing
> >>>>>> API/ABI for AV_PIX_FMT_QSV has changed.
> >>>>>>
> >>>>>> It looks like this was broken by using HDLPairs when D3D11 was
> >> introduced,
> >>>>>> which silently changed the existing API for DXVA2 and VAAPI as well.
> >>>>>>
> >>>>>> Could someone related to that please document it properly (clearly not
> >> all
> >>>>>> possible valid mfxFrameSurface1s are allowed), and note in APIchanges
> >> when
> >>>>>> the API change happened?
> >>>>>
> >>>>> Hi Mark,
> >>>>>
> >>>>> QSV contexts always need to be backed by a child context, which can be
> >>>> DXVA2,
> >>>>> D3D11VA or VAAPI. You can create a QSV context either by deriving from
> >> one
> >>>> of
> >>>>> those contexts or when create a new QSV context, it automatically
> creates
> >>>> an
> >>>>> appropriate child context - either implicitly (auto mode) or
> explicitly,
> >>>> like
> >>>>> the ffmpeg implementation does in most cases.
> >>>>
> >>>> ... or by using the one the user supplies when they create it.
> >>>>
> >>>>> When working with "user-supplied" frames on Linux, you need to create a
> >>>> VAAPI
> >>>>> context with those frames and derive a QSV context from that context.
> >>>>>
> >>>>> There is no way to create or supply QSV frames directly.
> >>>>
> >>>> ???  The ability for the user to set up their own version of these
> things
> >> is
> >>>> literally the whole point of the split alloc/init API.
> >>>>
> >>>>
> >>>> // Some user stuff involving libmfx - has a D3D or VAAPI backing, but
> this
> >>>> code doesn't need to care about it.
> >>>>
> >>>> // It has a session and creates some surfaces to use with MemId filled
> >>>> compatible with ffmpeg.
> >>>> user_session = ...;
> >>>> user_surfaces = ...;
> >>>>
> >>>> // No ffmpeg involved before this, now we want to pass these surfaces
> >> we've
> >>>> got into ffmpeg.
> >>>>
> >>>> // Create a device context using the existing session.
> >>>>
> >>>> mfx_ctx = av_hwdevice_ctx_alloc(MFX);
> >>>>
> >>>> dc = mfx_ctx->data;
> >>>> mfx_dc = dc->hwctx;
> >>>> mfx_dc->session = user_session;
> >>>>
> >>>> av_hwdevice_ctx_init(mfx_ctx);
> >>>>
> >>>> // Create a frames context out of the surfaces we've got.
> >>>>
> >>>> mfx_frames = av_hwframes_ctx_alloc(mfx_ctx);
> >>>>
> >>>> fc = mfx_frames->data;
> >>>> fc.pool = user_surfaces.allocator;
> >>>> fc.width = user_surfaces.width;
> >>>> // etc.
> >>>>
> >>>> mfx_fc = fc->hwctx;
> >>>> mfx_fc.surfaces = user_surfaces.array;
> >>>> mfx_fc.nb_surfaces = user_surfaces.count;
> >>>> mfx_fc.frame_type = user_surfaces.memtype;
> >>>>
> >>>> av_hwframe_ctx_init(frames);
> >>>>
> >>>> // Do stuff with frames.
> >>>
> >>> I wouldn't consider an mfxSession as an entity that could or should be
> >>> shared between implementations. IMO, this is not a valid use case.
> >>> A consumer of the mfx API needs to make certain choices regarding
> >>> the usage of the API, one of which is the way how frames are allocated
> >>> and managed.
> >>> This is not something that is meant to be shared between implementations.
> >>> Even inside ffmpeg, we don't use a single mfx session. We use separate
> >>> sessions for decoding, encoding and filtering that are joined together
> >>> via MFXJoinSession.
> >>> When an (ffmpeg-)API consumer is creating its own MFX session and its
> >>> own frame allocator implementation, it shouldn't be and allowed
> >>> scenario to create an ffmpeg hw context using this session.
> >>
> >> The user is also aware of the thread safety rules.  I was giving the
> example
> >> above entirely serially to avoid that, but if they were using libmfx
> >> elsewhere in parallel with the above then indeed they would need a bit
> more
> >> care (create another session, some sort of locking to ensure
> serialisation)
> >> when passing it to ffmpeg.
> >>
> >>> This shouldn't be considered a public API of ffmpeg because it doesn't
> >>> make sense to share an mfx session like this.
> >>
> >> So, to clarify, your opinion is that none of hwcontext_qsv.h should be
> public
> >> API?  If it were actually removed (not visible in installed headers) then
> I
> >> agree that would fix the problem.
> >
> > I think that exposing the mfx session handle is OK in order to allow an
> > API user to interop in a way that you create (and manage) your own session
> > and your own surface allocation, and use the exposes mfx session handle to
> > join the ffmpeg-created (and managed) session.
> 
> I'm not sure what you gain by making this one direction only.  Setting it is
> currently supported, and hasn't had an API-breaking internal change like the
> MemIds.

It's a logical consequence of the MSDK API that there can't be "two
owners" of an mfx session and that an mfx session is not meant to be shared
between implementations where one implementation tries to "match" the frame 
allocation logic of the other one.
Even decoding, processing and encoding inside ffmpeg are having their own mfx 
sessions.

But there also never existed a documented way how it could/should be used. 

You said that one could have inferred such way from MSDK samples, but looking 
at the MSDK samples for VAAPI, there are already two different ones using 
different structures for mfxMemId.

https://github.com/Intel-Media-SDK/MediaSDK/blob/510d19dcace1d8c57567fdd40b557155ab11ab8e/tutorials/common/common_vaapi.cpp#L199-L207

https://github.com/Intel-Media-SDK/MediaSDK/blob/510d19dcace1d8c57567fdd40b557155ab11ab8e/_studio/shared/include/libmfx_allocator_vaapi.h#L33-L39

So, I wouldn't say that it's obvious or straightforward how it would be 
implemented "normally".


> >>> Besides that, is there any ffmpeg documentation indicating that
> >>> the memId field of mfxFrameSurface1 could be casted to VASurfaceId?
> >>
> >> It is user-visible API, and it matched the behaviour of the libmfx
> examples
> >> for D3D9 (identical pointer to IDirect3DSurface9) and VAAPI (a compatible
> >> pointer to a structure with the VASurfaceID as the first member) so doing
> the
> >> obvious thing worked.
> >
> > Also I think it's OK to expose the array of mfxSurface1 pointers allowing
> > to use them in a opaque way (without making use of and any assumptions
> > about the mfxMemId field).
> >
> > This would allow an API user to interoperate with a joined session, e.g.
> > taking frames from a joined ffmpeg session as input to a user-created
> > session doing custom VPP operations where the output frames are from (and
> > managed by) the user-created session.
> >
> > On the other side, when and API user wants to access and interop with
> frames
> > directly, then the proper way would be to use derive-from or derive-to
> > (VAAPI, D3D9, D3D11) to get an appropriate frames context to work with.
> >
> > That's how I would see it. What do you think?
> I don't much like the idea of a special-case for read-only structures here,
> unlike all other hwcontexts.
> 
> Any users will already have been broken by the API change.  Is there any
> reason to believe that it will change again?  If not, we could document the
> new behaviour and say that's the new API.

I don't know... Maybe openVPL will start to support Vulkan as a child context
and maybe mfxHandlePair won't suffice in that case..?

It just appears an odd case to me to "support" a case where the ownership
of an mfx session is shared and not clearly separated. 

> If we do expect it to change further, then I guess that what you say is
> probably the least-bad option.  (It would want some extra code to disallow
> external pools so that users can't get into the bad case again, too.)

I don't think it's bad at all, but the most reasonable way as it implies
constraints that are implied by the libmfx API.
Though, I don't want to destroy sombody's use case, but I think when there 
would be anybody who uses the qsv context like that, he would already know 
what he's doing (and that it's more a kind of a hack than regular 
implementation).

Maybe the easiest solution is to leave it as-is and put a note somewhere 
about that change.
(while in reality, nobody will read that note in the first place. one would
notice that it doesn’t work like before and start researching why..)

Kind regards,
softworkz


> - Mark
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".


More information about the ffmpeg-devel mailing list