On Fri, Aug 9, 2013 at 1:31 PM, Tom Cooksey tom.cooksey@arm.com wrote:

...

...

...

So in the above, after X receives the second DRI2SwapBuffers, it doesn't need to get scheduled again for the next frame to be both rendered by the GPU and issued to the display for scanout.

well, this is really only an issue if you are so loaded that you don't get a chance to schedule for ~16ms.. which is pretty long time.

Yes - it really is 16ms (minus interrupt/workqueue latency) isn't it? Hmmm, that does sound very long. Will try out some experiments and see.

yeah

...

...

If you are triple buffering, it should not end up in the critical path (since the gpu already has the 3rd buffer to start on the next frame). And, well, if you do it all in the kernel you probably need to toss things over to a workqueue anyways.

Just a quick comment on the kernel flip queue issue.

16 ms scheduling latency sounds awful but totally doable with a less than stellar ddx driver going into limbo land and so preventing your single threaded X from doing more useful stuff. Is this really the linux scheduler being stupid?

Ahahhaaa!! Yes!!! Really good point. We generally don't have 2D HW and so rely on pixman to perform all 2D operations which does indeed tie up that thread for fairly long periods of time.

We've had internal discussions about introducing a thread (gulp) in the DDX to off-load drawing operations to. I think we were all a bit scared by that idea though.

thread does sound a bit scary.. it probably could be done if you treat it like a virtual cpu and have WaitMarker or PrepareAccess for sw fallbacks synchronize properly..

I bet you'd be much better off just making non-scanout pixmaps cached and doing cache sync ops when needed for dri2 buffers. Sw fallbacks on uncached buffers probably aren't exactly the hot ticket.

BTW: I wasn't suggesting it was the linux scheduler being stupid, just that there is sometimes lots of contention over the CPU cores and X is just one thread among many wanting to run.

...

At least my impression was that the hw/kernel flip queue is to save power so that you can queue up a few frames and everything goes to sleep for half a second or so (at 24fps or whatever movie your showing). Needing to schedule 5 frames ahead with pageflips under load is just guaranteed to result in really horrible interactivity and so awful user experience

Agreed. There's always a tradeoff between tolerance to variable frame rendering time/system latency (lot of buffers) and UI latency (few buffers).

As a side note, video playback is one use-case for explicit sync objects which implicit/buffer-based sync doesn't handle: Queue up lots of video frames for display, but mark those "display buffer" operations as depending on explicit sync objects which get signalled by the audio clock. Not sure Android actually does that yet though. Anyway, off topic.

w/ dmafence, rather than explicit fences, I suppose you could add some way to queue the buffer to the audio device and have the audio device signal the fence. I suppose it does sound a bit funny for ALSA to have a DMA_BUF_AV_SYNC ioctl for this sort of case?

I don't think there is anything like it in EGL, but there is oml_sync_control extension for more precise control of presentation time. But this is all implemented in userspace and doesn't really work out w/ >double buffering. This is part of the reason for the timing information in vblank events. Of course it doesn't have any tie in to audio subsystem, but in practice this really shouldn't be needed. Audio samples are either rendered at a very predictable rate, or sound like sh** with lots of pops and cut outs.

BR, -R

Cheers,

Tom