On Mon, Feb 9, 2015 at 7:05 AM, Eero Tamminen <eero.t.tammi...@intel.com> wrote: > Hi, > > On 02/05/2015 06:21 AM, Matt Turner wrote: >> >> total instructions in shared programs: 5895414 -> 5747578 (-2.51%) >> instructions in affected programs: 3618111 -> 3470275 (-4.09%) >> helped: 20492 >> HURT: 4449 >> GAINED: 54 >> LOST: 146 >> >> and with NIR, that already emits MADs always: >> >> total instructions in shared programs: 7992936 -> 7772474 (-2.76%) >> instructions in affected programs: 3738730 -> 3518268 (-5.90%) >> helped: 22082 >> HURT: 3445 >> GAINED: 70 >> LOST: 78 >> >> There are some particularly exciting individual improvements too: >> >> kerbal-space-program/667: FS SIMD8: 1024 -> 561 (-45.21%) >> kerbal-space-program/676: FS SIMD8: 1021 -> 558 (-45.35%) > > > From how old Kerbal version these shaders are from?
They were added to shader-db in November, so that would be my guess. > I just noticed that recent Kerbal outputs large amount of register spill & > alloc failure messages with INTEL_DEBUG=perf, whereas output I've saved from > a last year Kerbal version, had much less of them. I don't have anything to offer about those changes, but Kerbal uses large arrays of immediate values that aren't marked as const and that they variably index. Ken and I talked about it, and making those uniforms would make out code significantly better. _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
