On Wed, 01 Jul 2009 11:53:14 -0400 Owen Taylor <otay...@redhat.com> wrote:
> On Tue, 2009-06-30 at 09:18 -0700, Jesse Barnes wrote: > > That depends... If you're getting hit by scheduling latency (i.e. > > your app doesn't actually get scheduled on the CPU until several ms > > after the vblank interrupt fires), then there's no easy way to fix > > it. You could try realtime scheduling and configuring things just > > right so that your waiting apps always get scheduled in time, but > > that's going to be difficult in a general purpose environment. > > > > If the interrupt is arriving too late, you could try one of the > > other hardware mechanisms (fire an interrupt at a scanline near the > > bottom of the screen instead, to give you some extra time or > > something). > > Without looking at the demo source code, I suspect the problem might > be something else. The implementation of sgi_video_sync in Mesa simply > waits for the vblank interval and does nothing else. > > So, if I: > > - Do a bunch of rendering > - Wait for vblank > - glXSwapBuffers > > Then my command buffers will not be flushed to the video card until > after vblank. If the command buffers take time to execute, then the > swap buffers is delayed and you get a tear. So, you need to glFlush() > before waiting for vblank. Yeah, that's a good point. In the case of this demo we're just doing a clear, so there shouldn't be much latency involved in buffer submission for the swapbuffers (though there surely is some, sending a request to the server for the copyregion isn't free). > > There's a more subtle reason why you want to actually glFinish() > rather than glFlush() - if your rendering takes long enough so that > you can't keep up with the refresh rate: > > - Do a bunch of rendering > - glFlush > - Wait for vblank > - glXSwapBuffers > > Will cause the glXSwapBuffers to happen in the middle of a frame, > while > > - Do a bunch of rendering > - glFinish > - Wait for vblank > - glXSwapBuffers > > Will cause it to properly happen at the *second* vblank. Of course glXSwapBuffers *should* just do the right thing and not tear by default (this happens with current radeon & intel drivers afaik). So I'd discourage people from using SGI_video_sync for tear-avoidance in general. If you get tearing with glXSwapBuffers, file a bug. -- Jesse Barnes, Intel Open Source Technology Center ------------------------------------------------------------------------------ _______________________________________________ Mesa3d-dev mailing list Mesa3d-dev@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/mesa3d-dev
