Here's a series of patches to implement GL_KHR_flush_control and GLX_ARB_flush_control. The idea of the extension is quite simple and just makes it so that the GL doesn't implicitly flush the context when the context is unbound. This should be a speed improvement for applications that are using multiple contexts and have to switch between them often.
The actual work of the extension is just to stop calling _mesa_flush in _mesa_make_current when the attribute is specified. However the patch is a bit more fiddly because it involves a new attribute for glXCreateContextAttribs which needs modifications on quite a few layers. Sadly the X server needs patching as well because the server side of GLX seems to verify the attributes passed to glXCreateContextAttribs as well even for direct contexts. I'll post the patches for the X server and glproto shortly. I've written an example of using the extension here: https://github.com/bpeel/multi-context-test The example renders a grid of rectangles to three different windows with three different contexts. The rendering of each frame of each window is interleaved so that the context is switched after every row of rectangles. If I set the release behaviour to flush then the FPS is 80 FPS whereas if I set it to none then it is 192 FPS. Hopefully that should be convincing enough that I haven't missed some flush somewhere. There is already a Piglit test proposed by Fredrik Höglund which I think would be enough to test the extension if we can fix it up a bit. http://lists.freedesktop.org/archives/piglit/2014-August/012271.html I've run the patch series against all of the existing Piglit tests on Ivybridge and there are no regressions. _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
