[FFmpeg-devel] [PATCH] avfilter/setparams: add FF_FILTER_FLAG_HWFRAME_AWARE

[Mark Thompson]

On 29/09/2020 19:49, Pavel Koshevoy wrote:
> On Tue, Sep 29, 2020 at 12:14 PM Mark Thompson <sw at jkqxz.net> wrote:
> 
>> On 29/09/2020 18:14, Pavel Koshevoy wrote:
>>> On Tue, Sep 29, 2020 at 10:09 AM Mark Thompson <sw at jkqxz.net> wrote:
>>>
>>
>> <snip>
>>
>>>> - Mark
>>>>
>>>>
>>> It's pretty much this use case, except I'm not using ffmpeg cli but the
>>> avfilter api to configure the filter chain, and I'm working with
>>> AV_PIX_FMT_CUDA frames.
>>> Perhaps I am mis-using the api, but my patch was sufficient for my needs:
>>>
>>> ```
>>>       bool
>>>       VideoFilterChain::setup_filter_links(int num_threads)
>>>       {
>>>           graph_ = avfilter_graph_alloc();
>>>           graph_->nb_threads = num_threads;
>>>           int err = avfilter_graph_parse2(graph_, filters_.c_str(), &in_,
>>> &out_);
>>>           UL_FAIL_IF_AVERROR(err);
>>>
>>>           if (hw_frames_.ref_)
>>>           {
>>>               AVHWFramesContext * hw_frames_ctx =
>>> hw_frames_.get<AVHWFramesContext>();
>>>               AVBufferRef * device_ref = hw_frames_ctx->device_ref;
>>>
>>>               for (int i = 0; i < graph_->nb_filters; i++)
>>>               {
>>>                   AVFilterContext * filter_ctx = graph_->filters[i];
>>>                   UL_ASSERT(!filter_ctx->hw_device_ctx);
>>>                   filter_ctx->hw_device_ctx = av_buffer_ref(device_ref);
>>>
>>>                   bool found_hwdownload = strcmp(filter_ctx->filter->name,
>>> "hwdownload") == 0;
>>>                   if (found_hwdownload)
>>>                   {
>>>                       break;
>>>                   }
>>>
>>>                   for (int j = 0; j < filter_ctx->nb_outputs; j++)
>>>                   {
>>>                       AVFilterLink * link = filter_ctx->outputs[j];
>>>                       UL_ASSERT(!link->hw_frames_ctx);
>>>                       link->hw_frames_ctx =
>> av_buffer_ref(hw_frames_.ref_);
>>
>>
>> <
>> http://git.videolan.org/?p=ffmpeg.git;a=blob;f=libavfilter/avfilter.h;h=99297ae798aa325ac37836a3a90d9a3f8e1e7a95;hb=HEAD#l497
>>>
>>
>> Don't write to internal fields.
>>
>> I'm not sure exactly what you're trying to do by writing the internal
>> context fields in this section, but I suspect that if you just remove it
>> entirely then the expected context propagation will happen and it will work.
>>
>>
> Okay, if I fix my API mis-use and context propagation happens
> automagically  (idk where it would get the context in the 1st place,
> yet)...

If you are putting hardware frames into a filter graph then it comes from buffersrc.  If you are using hwupload inside the filter graph then it gets made there from the device you provide as AVFilterContext.hw_device_ctx.

>         won't that still leave me with setparams that is not
> FF_FILTER_FLAG_HWFRAME_AWARE and avfilter_config_links would still fail?

Why would it fail?

- Mark