On Mon, Feb 16, 2015 at 10:40 AM, Francisco Jerez <curroje...@riseup.net> wrote: > My intuition is that the huge performance improvement Matt observed by > disabling the third scheduling heuristic is more likely to have been > caused by a decrease in the amount of cache thrashing caused by the fact > that he was running less channels concurrently rather than by the > scheduling heuristic itself. Matt, did you rule out that possibility?
Actually, the fragment shader used in the demo [1] doesn't can't register allocate in SIMD16 at all. From my Dec/Jan status report: > As an aside, I did an experiment to see what would happen if we used > SCHEDULE_PRE and spilled anyway (I added a couple of extra instructions > to increase register pressure over the threshold). I changed the window > size to 2048x2048 and rendered 1000 frames. > > SCHEDULE_PRE with no spills: 17.5 seconds to render 1000 fames > SCHEDULE_PRE with 4 spills (8 sends): 17.5 seconds > SCHEDULE_PRE_NON_LIFO with no spills: 28 seconds The demo does no texturing (no send instructions at all if no spilling), so it's not affected by caching as far as I understand. [1] https://github.com/ashima/webgl-noise _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
