[Bug 95472] [i915] Feature request: Add support for fencing for PRIME setups

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=95472

--- Comment #9 from Chris Wilson <chris at chris-wilson.co.uk> ---
Bisecting that lockup would be invaluable to me.

DRI2 render offload goes something like:

GL client renders to back
 -> DRI2SwapBuffers
    -> X/dgpu CopyRegion2, copy back to front + post damage
       -> DamageEvent
          -> compositor copies on igfx damaged front onto framebuffer and flips
             -> DRI2SwapBuffers
                -> X/igfx flips

(DRI3 is a couple of steps simpler, in that the dgpu copy is done by the GL
client and only the X/igfx is involved with handling Present)

The point at which we require the fence is between the copy by the dgpu inside
X and the copy by the igfx inside the compositor. The buffer is passed between
dgpu and igfx using a dma-buf, which also carries along the implicit fences. So
long as the fencing is correct that should be fine.

Q: Is it only the PRIME offload that tears? Does enabling TearFree do anything?
(Just trying to ascertain that it is the fence and not the last SwapBuffers
that is at fault.)

If you want to look at intel-gpu-tools/tests/prime_vgem, tests/prime_busy
should test the basic export/import of fences. I'll patch in some debug
messages tomorrow to show whether we are seeing fences in i915 execbuf.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
You are the assignee for the bug.
You are the QA Contact for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-gfx-bugs/attachments/20160804/43e84bc3/attachment.html>