From: Kenneth Graunke Date: Mon, 10 Jun 2019 21:03:03 +0000 (-0700) Subject: iris: Bypass half-float pack/unpack lowering. X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=a8588f512b41cbba37d9a0f3d31e5275a8adb573;p=mesa.git iris: Bypass half-float pack/unpack lowering. This skips GLSL IR lowering of pack/unpackHalf operations, allowing the NIR optimizer to see them Improves performance in Synmark2's OglCSDof by about 2x, by cutting about 90% of the cycles from one of the compute shaders. shader-db statistics on Skylake: 4 compute shaders went from SIMD8 to SIMD16. total instructions in shared programs: 15598871 -> 15542568 (-0.36%) instructions in affected programs: 143016 -> 86713 (-39.37%) helped: 144 HURT: 0 helped stats (abs) min: 17 max: 4669 x̄: 390.99 x̃: 164 helped stats (rel) min: 7.48% max: 85.28% x̄: 30.17% x̃: 24.22% 95% mean confidence interval for instructions value: -510.50 -271.49 95% mean confidence interval for instructions %-change: -32.70% -27.65% Instructions are helped. total cycles in shared programs: 371973958 -> 368902103 (-0.83%) cycles in affected programs: 5557722 -> 2485867 (-55.27%) helped: 144 HURT: 0 helped stats (abs) min: 106 max: 1026600 x̄: 21332.33 x̃: 1697 helped stats (rel) min: 0.53% max: 88.98% x̄: 36.12% x̃: 34.67% 95% mean confidence interval for cycles value: -41570.02 -1094.64 95% mean confidence interval for cycles %-change: -38.44% -33.80% Cycles are helped. total spills in shared programs: 11936 -> 11903 (-0.28%) spills in affected programs: 110 -> 77 (-30.00%) helped: 3 HURT: 2 total fills in shared programs: 25644 -> 25178 (-1.82%) fills in affected programs: 677 -> 211 (-68.83%) helped: 5 HURT: 0 total loops in shared programs: 4830 -> 4829 (-0.02%) loops in affected programs: 1 -> 0 helped: 1 HURT: 0 --- diff --git a/src/gallium/drivers/iris/iris_screen.c b/src/gallium/drivers/iris/iris_screen.c index b3c4466805e..83f5208e65e 100644 --- a/src/gallium/drivers/iris/iris_screen.c +++ b/src/gallium/drivers/iris/iris_screen.c @@ -151,6 +151,7 @@ iris_get_param(struct pipe_screen *pscreen, enum pipe_cap param) case PIPE_CAP_TGSI_VS_LAYER_VIEWPORT: case PIPE_CAP_TGSI_TES_LAYER_VIEWPORT: case PIPE_CAP_TGSI_FS_FINE_DERIVATIVE: + case PIPE_CAP_TGSI_PACK_HALF_FLOAT: case PIPE_CAP_ACCELERATED: case PIPE_CAP_UMA: case PIPE_CAP_CONDITIONAL_RENDER_INVERTED: