This is a collection of patches I've written and collected that aim to reduce the pain when dropping brw_do_channel_expressions() and brw_do_vector_splitting() two passes that combined take about 14% of the time it takes to run shader-db. This series includes Ken's move comparisions series with a couple of small changes.
I have a few more changes that make good progress at removing regressions but they are all basically poor mans attempts at tree height lowering and we would be much better off implementing a proper pass to do this. Jenkins is reporting that fbo-generatemipmap-formats is crashing due to this series on g965/m64, ilk/m64, g45/m64 but I can't reproduce that with INTEL_DEVID_OVERRIDE so I assume its telling me lies. Each opt patch has its own shader-db results. Shader-db results for the whole series on BDW: total instructions in shared programs: 13070792 -> 13044633 (-0.20%) instructions in affected programs: 2718825 -> 2692666 (-0.96%) helped: 8499 HURT: 269 total cycles in shared programs: 256597034 -> 256406494 (-0.07%) cycles in affected programs: 168145092 -> 167954552 (-0.11%) helped: 8794 HURT: 3706 total loops in shared programs: 2950 -> 2942 (-0.27%) loops in affected programs: 8 -> 0 helped: 7 HURT: 0 total spills in shared programs: 15107 -> 14745 (-2.40%) spills in affected programs: 9339 -> 8977 (-3.88%) helped: 353 HURT: 0 total fills in shared programs: 20213 -> 19850 (-1.80%) fills in affected programs: 12863 -> 12500 (-2.82%) helped: 353 HURT: 0 LOST: 2 GAINED: 23 _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
