[FFmpeg-user] avfilter: hwoverlay

[Paul B Mahol]

On 11/26/18, msanders <marksanders at protonmail.com> wrote:
> ‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
> On Monday, 26 de November de 2018 14:00, Micael Silva
> <micaelsilva at gmail.com> wrote:
>
>> On Mon, Nov 26, 2018 at 7:47 AM msanders marksanders at protonmail.com wrote:
>>
>> > Hi,
>> > As now we have many filters compatible with the GPU, and in particular
>> > “scale” filters. So I think we need a filter like “hwoverlay” or
>> > “overlay_cuda/_qsv/etc.”.
>> > With this kind of filter we can do a PiP using only a GPU pipeline:
>> >
>> > -   GPU decoder --> spplit (hw) --> hwscale --> hwoverlay --> GPU
>> > encoder
>> >
>> > In fact, I feel it’s not too complex to do.
>> > What you think?
>> > M. Sanders
>>
>> It's a good idea, although working with hardware accelerators it's never
>> so
>> straightforward. It depends how that API is construct, what methods are
>> allowed, and many bugs and crashes surface. Hardware processing is less
>> flexible in many aspects.
>>
>> Maybe possible, but I think that requiring more specialized people may
>> take
>> more time to realize
>>
> Hi,
>
> The complexity varies.
> - If the implementation is "generic", for example using some function like
> "av_frame_clone()" that works with RAM and GPU frames, then it will be more
> complex... at start. In this case new functions "av_frame_clip()" and
> "av_frame_overlay()" should be required.
> - And if the implementation is "particular", for example implementing
> "overlay_cuda|_qsv", then it will be more simple.
>
> In fact, using the code of "vf_scale_cuda.c" can be easy to translate it to
> a "vf_overlay_cuda.c". The reason it's because the GPU copy functions are
> already implemented. And the algorithm is summarized in: copy the complete
> frame and then overwrite with the small frame. So, only "copy GPU" functions
> are required.
>
> Anyone interested?

Nope.