From 2b01246f492c2cab48a7c07b80c738dd51043786 Mon Sep 17 00:00:00 2001 From: Emma Anholt Date: Tue, 13 Jun 2023 16:18:09 -0700 Subject: [PATCH] intel/fs: Move the defin[]/defout[] screening up to livein[]/liveout[] setup. This keeps us from having to run the loop to propagate up quite so much. steam shader-db time -1.86356% +/- 0.941498% (n=10). There's a small scheduling effect, since previously the scheduler wasn't considering defin/defout: cycles helped: shaders/closed/steam/amnesia-the-dark-descent/high/241.shader_test FS SIMD16: 11428 -> 11422 (-0.05%) (scheduled: scheduled) cycles helped: shaders/humus-volumetricfogging2/1.shader_test FS SIMD32: 13832 -> 13800 (-0.23%) (scheduled: scheduled) cycles helped: shaders/tesseract/479.shader_test FS SIMD32: 9330 -> 8644 (-7.35%) (scheduled: scheduled) cycles HURT: shaders/robclark-shaders/android/angle/aztec_ruins/36.shader_test FS SIMD32: 7870 -> 7940 (0.89%) (scheduled: scheduled) cycles HURT: shaders/robclark-shaders/gfxbench5/gl_5_high_off/57.shader_test FS SIMD32: 7870 -> 7940 (0.89%) (scheduled: scheduled) cycles HURT: shaders/robclark-shaders/gfxbench5/gl_5_normal_off/54.shader_test FS SIMD32: 7870 -> 7940 (0.89%) (scheduled: scheduled) cycles HURT: shaders/robclark-shaders/android/angle/aztec_ruins/30.shader_test FS SIMD32: 8726 -> 8808 (0.94%) (scheduled: scheduled) cycles HURT: shaders/robclark-shaders/gfxbench5/gl_5_high_off/51.shader_test FS SIMD32: 8726 -> 8808 (0.94%) (scheduled: scheduled) cycles HURT: shaders/robclark-shaders/gfxbench5/gl_5_normal_off/48.shader_test FS SIMD32: 8726 -> 8808 (0.94%) (scheduled: scheduled) cycles HURT: shaders/robclark-shaders/gfxbench5/gl_4_off/129.shader_test TCS SIMD8: 3911 -> 3979 (1.74%) (scheduled: scheduled) cycles HURT: shaders/robclark-shaders/gfxbench5/gl_4_off/109.shader_test TCS SIMD8: 3911 -> 3979 (1.74%) (scheduled: scheduled) total cycles in shared programs: 313096438 -> 313096306 (<.01%) cycles in affected programs: 92200 -> 92068 (-0.14%) Reviewed-by: Ian Romanick Part-of: --- src/intel/compiler/brw_fs_live_variables.cpp | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/src/intel/compiler/brw_fs_live_variables.cpp b/src/intel/compiler/brw_fs_live_variables.cpp index fa74fc2cccd..c8f8f8cd989 100644 --- a/src/intel/compiler/brw_fs_live_variables.cpp +++ b/src/intel/compiler/brw_fs_live_variables.cpp @@ -191,6 +191,7 @@ fs_live_variables::compute_live_variables() for (int i = 0; i < bitset_words; i++) { BITSET_WORD new_liveout = (child_bd->livein[i] & ~bd->liveout[i]); + new_liveout &= bd->defout[i]; /* Screen off uses with no reaching def */ if (new_liveout) { bd->liveout[i] |= new_liveout; cont = true; @@ -209,6 +210,7 @@ fs_live_variables::compute_live_variables() BITSET_WORD new_livein = (bd->use[i] | (bd->liveout[i] & ~bd->def[i])); + new_livein &= bd->defin[i]; /* Screen off uses with no reaching def */ if (new_livein & ~bd->livein[i]) { bd->livein[i] |= new_livein; cont = true; @@ -236,8 +238,8 @@ fs_live_variables::compute_start_end() struct block_data *bd = &block_data[block->num]; for (int w = 0; w < bitset_words; w++) { - BITSET_WORD livedefin = bd->livein[w] & bd->defin[w]; - BITSET_WORD livedefout = bd->liveout[w] & bd->defout[w]; + BITSET_WORD livedefin = bd->livein[w]; + BITSET_WORD livedefout = bd->liveout[w]; BITSET_WORD livedefinout = livedefin | livedefout; while (livedefinout) { unsigned b = u_bit_scan(&livedefinout);