[FFmpeg-devel] [PATCH] Vulkan hwcontext and filters

[Anton Khirnov]

Quoting Lynne (2020-01-21 21:52:52)
> Jan 21, 2020, 18:22 by anton at khirnov.net:
> >> Its size too? Didn't know that.
> >> I do think its a good idea to be able to append fields to it, so I've
> >> added a av_vk_frame_alloc() function. I've followed what
> >> av_frame_alloc is called, though I'm not sure if its too close to that
> >> one's name.
> >>
> >
> > My original intent with this API was that calles are allowed to provide
> > their own pools. Not sure if that's still allowed or works though. But
> > if it is, the caller needs to be able to allocate/free AVkFrame.
> >
> 
> They are of course. The first wip of the cuda interop exploited that IIRC.
> But I think the issue is that in order for API users to make their own
> pools they need to know the size in bytes of AVVkFrame since that's
> what av_buffer_create() accepts.
> I could make a function to return that, but I'm wondering if its
> somewhat too big of a mess already.
> I could instead reserve some memory in the struct, a few hundred bytes
> would be enough since every Vulkan object has to be allocated on heap.

The requirement for av_buffer_create() is just formal here, we shouldn't
let it guide our API design (ideally we'd make a standalone refcount
type and stop abusing AVBuffer for it). You can just pass sizeof(void*)
to it. Or make av_vk_frame_alloc() write sizeof(AVVkFrame) into a
supplied pointer, if you want to be really proper about it.

-- 
Anton Khirnov