[Bug 755072] vaapi: expose memory:DMABuf capsfeature

[GStreamer (GNOME Bugzilla)]

https://bugzilla.gnome.org/show_bug.cgi?id=755072

--- Comment #17 from Nicolas Dufresne (stormer) <nicolas.dufresne at collabora.co.uk> ---
(In reply to Julien Isorce from comment #16)
> You would need patch for gst-plugins-base:
> https://bug759358.bugzilla-attachments.gnome.org/attachment.cgi?id=317223
> and for gstglupload:
> https://github.com/CapOM/gst-plugins-bad/commit/
> f274539bc47f300624dee52d304f65cafcd459bb

Assuming there will be documentation in the final version, that patch looks
fine to me. Make sure to keep it separate.

> 
> Nicolas: mapping of dmabuf created by exporting a VASurface fails on desktop
> with: "gst_fd_mem_map: 0x7f39880051c0: fd 26: mmap failed: Permission
> denied". See l195 here
> https://github.com/CapOM/gstreamer-vaapi/commit/
> ffceae0c0cd099967f6bc3b1a27451151550e70a#diff-
> bec9172c05be682d331c790d8a50ba95R195.
> 
> I did not expected it but it was actually good to test switch case where map
> fail and then switch to feature. See regression test list in that branch.

That answers a question I asked a long time ago. I was never sure if the Intel
driver would let us map or not the DMAbuf. Note, it could be laziness in the
driver (as it's often the case for GFX, they implement the strict minimum). It
could also be a very complex process, I don't know. My next question is, can we
trust the format ? It's important, as this is for sharing between drivers. The
DMAbuf objects are just a memory buffers, everything else need to be
communicated to the other driver by userspace.

> 
> The most important thing is to agree on how the following pipelines behave:
> https://github.com/CapOM/gstreamer-vaapi/commit/
> c450b7c7693b1e3417c96dbbc906c394416a6852. I am planning to convert them to
> some automatic tests. (gst-integration-testsuites ? gst-devtools/validate ?)

I'd say create a DMABuf/Intel validation suite (for gst-validate). One that we
can run on such HW. I don't know what is available within our test both at the
moment, they are all virtual machines. We could start creating regression for
VAAPI too, and maybe we can Approach Intel to get a dedicated HW sponsored.

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