[Bug 783521] gl: Add "direct" dmabuf uploader

[GStreamer (GNOME Bugzilla)]

https://bugzilla.gnome.org/show_bug.cgi?id=783521

--- Comment #71 from Nicolas Dufresne (ndufresne) <nicolas at ndufresne.ca> ---
First sorry for the confusion on my previous comment, I worked on this on the
plane and didn't realize your comments.

(In reply to Michael Olbrich from comment #70)
> I noticed, that _direct_dma_buf_upload_transform_caps() is not quite correct.
> It should provide empty caps if the downstream format is anything other than
> RGBA.
> 
> I'm not sure if that makes a difference in practice.
> 
> In general, I think that glupload should always prefer RGBA on the sink pad.
> That
> way we can potentially avoid extra conversions. My patch would be a first
> step in
> that direction.
> However, I'm not sure it it's worth the effort. Are there any real use-cases
> where
> the input format is not fixed anyways?

Yes, note that I didn't ignore your patch, "glupload: Do prepend the preferred
caps" implements what you described, but in the specific uploader instead of
globally. So it's the same but narrowed to direct dmabuf upload. Prepending
RGBA should be glcolorconvert jobs for the cases where glupload isn't doing
color conversion. It's nothing new, videoconvert does that same to encourage
upstream into going passthrough.

And I also when for the subset approach like you also described, even though I
agree that something isn't quite right. It fixes the regression that first data
pointer get uploaded twice (not just that unit test failing). I'll need your
help to confirm it also works for you use case, considering that my Intel GPU
do very little formats in direct dmabuf upload.

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
You are the assignee for the bug.