[Mesa-dev] [PATCH 12/12] RFC: intel/compiler: Don't always require precise lowering of flrp

Ian Romanick idr at freedesktop.org
Sat Aug 25 05:52:17 UTC 2018


From: Ian Romanick <ian.d.romanick at intel.com>

No changes on any other Intel platforms.

Iron Lake
total instructions in shared programs: 7727356 -> 7699517 (-0.36%)
instructions in affected programs: 3152595 -> 3124756 (-0.88%)
helped: 12836
HURT: 48
helped stats (abs) min: 1 max: 30 x̄: 2.18 x̃: 1
helped stats (rel) min: 0.03% max: 10.77% x̄: 1.14% x̃: 0.94%
HURT stats (abs)   min: 1 max: 4 x̄: 1.75 x̃: 1
HURT stats (rel)   min: 0.27% max: 5.26% x̄: 1.94% x̃: 1.26%
95% mean confidence interval for instructions value: -2.20 -2.12
95% mean confidence interval for instructions %-change: -1.14% -1.11%
Instructions are helped.

total cycles in shared programs: 177982922 -> 177855192 (-0.07%)
cycles in affected programs: 67514274 -> 67386544 (-0.19%)
helped: 11752
HURT: 507
helped stats (abs) min: 2 max: 600 x̄: 11.13 x̃: 6
helped stats (rel) min: <.01% max: 5.48% x̄: 0.47% x̃: 0.26%
HURT stats (abs)   min: 2 max: 62 x̄: 6.03 x̃: 4
HURT stats (rel)   min: 0.01% max: 3.08% x̄: 0.19% x̃: 0.07%
95% mean confidence interval for cycles value: -10.77 -10.07
95% mean confidence interval for cycles %-change: -0.45% -0.43%
Cycles are helped.

LOST:   0
GAINED: 13

GM45
total instructions in shared programs: 4755025 -> 4740052 (-0.31%)
instructions in affected programs: 1703728 -> 1688755 (-0.88%)
helped: 6578
HURT: 24
helped stats (abs) min: 1 max: 87 x̄: 2.28 x̃: 1
helped stats (rel) min: 0.03% max: 10.45% x̄: 1.12% x̃: 0.93%
HURT stats (abs)   min: 1 max: 4 x̄: 1.75 x̃: 1
HURT stats (rel)   min: 0.27% max: 5.00% x̄: 1.89% x̃: 1.25%
95% mean confidence interval for instructions value: -2.33 -2.21
95% mean confidence interval for instructions %-change: -1.12% -1.08%
Instructions are helped.

total cycles in shared programs: 121824644 -> 121742888 (-0.07%)
cycles in affected programs: 41846900 -> 41765144 (-0.20%)
helped: 6119
HURT: 175
helped stats (abs) min: 2 max: 600 x̄: 13.54 x̃: 6
helped stats (rel) min: <.01% max: 5.48% x̄: 0.49% x̃: 0.28%
HURT stats (abs)   min: 2 max: 48 x̄: 6.23 x̃: 6
HURT stats (rel)   min: 0.01% max: 3.08% x̄: 0.18% x̃: 0.07%
95% mean confidence interval for cycles value: -13.56 -12.42
95% mean confidence interval for cycles %-change: -0.49% -0.46%
Cycles are helped.

total spills in shared programs: 72 -> 61 (-15.28%)
spills in affected programs: 66 -> 55 (-16.67%)
helped: 1
HURT: 0

total fills in shared programs: 108 -> 94 (-12.96%)
fills in affected programs: 96 -> 82 (-14.58%)
helped: 1
HURT: 0

LOST:   13
GAINED: 13

Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
---
 src/intel/compiler/brw_nir.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c
index bd08e1e1c65..1c773882f0f 100644
--- a/src/intel/compiler/brw_nir.c
+++ b/src/intel/compiler/brw_nir.c
@@ -575,13 +575,10 @@ brw_nir_optimize(nir_shader *nir, const struct brw_compiler *compiler,
       OPT(nir_opt_constant_folding);
 
       if (need_to_lower_flrp) {
-         /* To match the old behavior, set always_precise only for scalar
-          * shader stages.
-          */
          if (OPT(nir_lower_flrp,
                  nir->options->lower_flrp32,
                  nir->options->lower_flrp64,
-                 is_scalar /* always_precise */,
+                 false /* always_precise */,
                  compiler->devinfo->gen >= 6)) {
             OPT(nir_opt_constant_folding);
          }
-- 
2.14.4



More information about the mesa-dev mailing list