From 0c2b3a7fc00575cc2ef3ab296aef36262f05a4b3 Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Wed, 24 Apr 2019 23:49:30 -0700 Subject: [PATCH] nir/algebraic: Rearrange 1-((1-a) * (1-b)) into flrp-friendly form MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit No shader-db changes Ice Lake, Iron Lake, or GM45 as these platforms lack a LRP instruction. v2: Convert the pattern directly to flrp. There were negligible improvements on Gen4 and Gen5, and Gen11 was actually hurt. I believe the problem is this optimization conflicts with the (1-x)*y => ffma(-x, y, y) optimization on Gen11. Skylake total instructions in shared programs: 15046487 -> 15041996 (-0.03%) instructions in affected programs: 194681 -> 190190 (-2.31%) helped: 880 HURT: 20 helped stats (abs) min: 1 max: 19 x̄: 5.13 x̃: 4 helped stats (rel) min: 0.19% max: 36.36% x̄: 4.85% x̃: 3.33% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.11% max: 1.06% x̄: 0.28% x̃: 0.17% 95% mean confidence interval for instructions value: -5.25 -4.73 95% mean confidence interval for instructions %-change: -5.11% -4.36% Instructions are helped. total cycles in shared programs: 354340839 -> 354303333 (-0.01%) cycles in affected programs: 1753622 -> 1716116 (-2.14%) helped: 786 HURT: 182 helped stats (abs) min: 1 max: 1842 x̄: 56.52 x̃: 22 helped stats (rel) min: 0.03% max: 43.17% x̄: 3.90% x̃: 2.84% HURT stats (abs) min: 1 max: 440 x̄: 37.99 x̃: 9 HURT stats (rel) min: 0.03% max: 29.37% x̄: 1.96% x̃: 0.32% 95% mean confidence interval for cycles value: -45.90 -31.59 95% mean confidence interval for cycles %-change: -3.09% -2.50% Cycles are helped. All Gen6-Gen8 platforms had similar results. (Broadwell shown) total instructions in shared programs: 15055907 -> 15051466 (-0.03%) instructions in affected programs: 196370 -> 191929 (-2.26%) helped: 871 HURT: 26 helped stats (abs) min: 1 max: 19 x̄: 5.13 x̃: 4 helped stats (rel) min: 0.19% max: 36.36% x̄: 4.76% x̃: 3.27% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.11% max: 1.06% x̄: 0.24% x̃: 0.12% 95% mean confidence interval for instructions value: -5.21 -4.69 95% mean confidence interval for instructions %-change: -4.99% -4.24% Instructions are helped. total cycles in shared programs: 387729170 -> 387699745 (<.01%) cycles in affected programs: 1816409 -> 1786984 (-1.62%) helped: 788 HURT: 172 helped stats (abs) min: 1 max: 662 x̄: 47.29 x̃: 22 helped stats (rel) min: 0.03% max: 31.26% x̄: 3.55% x̃: 2.76% HURT stats (abs) min: 1 max: 404 x̄: 45.59 x̃: 14 HURT stats (rel) min: 0.03% max: 22.92% x̄: 1.53% x̃: 0.43% 95% mean confidence interval for cycles value: -35.69 -25.61 95% mean confidence interval for cycles %-change: -2.88% -2.40% Cycles are helped. total fills in shared programs: 34712 -> 34710 (<.01%) fills in affected programs: 7 -> 5 (-28.57%) helped: 1 HURT: 0 LOST: 0 GAINED: 2 Reviewed-by: Matt Turner --- src/compiler/nir/nir_opt_algebraic.py | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py index 09eaeaf3f23..8fcf0a1cbff 100644 --- a/src/compiler/nir/nir_opt_algebraic.py +++ b/src/compiler/nir/nir_opt_algebraic.py @@ -186,6 +186,16 @@ optimizations = [ # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))), + # 1 - ((1 - a) * (1 - b)) + # 1 - (1 - a - b + a*b) + # 1 - 1 + a + b - a*b + # a + b - a*b + # a + b*(1 - a) + # b*(1 - a) + 1*a + # flrp(b, 1, a) + (('~fadd@32', 1.0, ('fneg', ('fmul', ('fadd', 1.0, ('fneg', a)), ('fadd', 1.0, ('fneg', b))))), + ('flrp', b, 1.0, a), '!options->lower_flrp32'), + # (a * #b + #c) << #d # ((a * #b) << #d) + (#c << #d) # (a * (#b << #d)) + (#c << #d) -- 2.30.2