GpuScheduler can unschedule a command buffer until the GPU has made progress up to a fence.

GpuScheduler can unschedule a command buffer until the GPU has made progress up to a fence.

Other command buffers on the channel can still be processed but if any more calls are issued on the unscheduled command buffer, the entire channel is unscheduled. This prevents reordering of GL commands wrt flushes.

It'll be used by this: 
http://codereview.chromium.org/8060045/
Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=107154

[apatrick_chromium, 2011-10-25 00:02:34]

Oh, forgot to mention. This should do something in the case where GL_NV_fence is
not available. Suggestions? I could fall back to just unscheduling the
GpuScheduler as normal or, if this is unlikely to be unsupported in practice, I
could just fail GPU process initialization if it's not supported. ANGLE always
supports it and currently it's the only time I would want to use DeferToFence.
See referenced CL.

[jbauman, 2011-10-25 00:16:02]

On 2011/10/25 00:02:34, apatrick_chromium wrote:
> Oh, forgot to mention. This should do something in the case where GL_NV_fence
is
> not available. Suggestions? I could fall back to just unscheduling the
> GpuScheduler as normal or, if this is unlikely to be unsupported in practice,
I
> could just fail GPU process initialization if it's not supported. ANGLE always
> supports it and currently it's the only time I would want to use DeferToFence.
> See referenced CL.

How would you reschedule it if you just unscheduled it like normal?

Actually, I suppose you could just do a glFinish in that case. If that's never
actually called in practice, I suppose the performance won't matter.

[apatrick_chromium, 2011-10-25 00:44:24]

glFinish it is then.

http://codereview.chromium.org/8387008/diff/5/gpu/command_buffer/service/gpu_...
File gpu/command_buffer/service/gpu_scheduler.cc (right):

http://codereview.chromium.org/8387008/diff/5/gpu/command_buffer/service/gpu_...
gpu/command_buffer/service/gpu_scheduler.cc:62: glGenFencesNV = NULL;
glFinish code path tested scientifically. Poor performance confirmed. I deleted
this line already.

[apatrick_chromium, 2011-10-25 00:54:45]

http://codereview.chromium.org/8387008/diff/5001/gpu/command_buffer/service/g...
File gpu/command_buffer/service/gpu_scheduler.cc (right):

http://codereview.chromium.org/8387008/diff/5001/gpu/command_buffer/service/g...
gpu/command_buffer/service/gpu_scheduler.cc:189: glFlush();
I might as well flush here in the no GL_NV_fence case as well so I moved it out
of the "if". It could make the glFinish return faster.

[jbauman, 2011-10-25 01:37:25]

lgtm