OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

nir/algebraic: Apply addition property of equality more conservatively

Browse Source 
This allows a lot more CSE.  Depending on where the addition and the
comparison are scheduled, it may also reduce register pressure by
reducing the live range of the addends.

Across all the platforms, the shaders affected for spills or fills were
all fragment shaders from Dirt Rally.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

Tiger Lake and Ice Lake had similar results. (Tiger Lake shown)
total instructions in shared programs: 21043103 -> 21038804 (-0.02%)
instructions in affected programs: 892878 -> 888579 (-0.48%)
helped: 1549
HURT: 724
helped stats (abs) min: 1 max: 225 x̄: 4.14 x̃: 2
helped stats (rel) min: 0.05% max: 11.18% x̄: 1.04% x̃: 0.78%
HURT stats (abs)   min: 1 max: 71 x̄: 2.93 x̃: 1
HURT stats (rel)   min: 0.07% max: 6.90% x̄: 0.80% x̃: 0.56%
95% mean confidence interval for instructions value: -2.33 -1.45
95% mean confidence interval for instructions %-change: -0.50% -0.40%
Instructions are helped.

total cycles in shared programs: 855054155 -> 855757566 (0.08%)
cycles in affected programs: 58275918 -> 58979329 (1.21%)
helped: 1213
HURT: 1680
helped stats (abs) min: 1 max: 107405 x̄: 1684.00 x̃: 10
helped stats (rel) min: <.01% max: 38.09% x̄: 1.51% x̃: 0.25%
HURT stats (abs)   min: 1 max: 126632 x̄: 1634.59 x̃: 12
HURT stats (rel)   min: <.01% max: 85.91% x̄: 2.75% x̃: 0.49%
95% mean confidence interval for cycles value: -98.06 584.35
95% mean confidence interval for cycles %-change: 0.71% 1.22%
Inconclusive result (value mean confidence interval includes 0).

total spills in shared programs: 9843 -> 9771 (-0.73%)
spills in affected programs: 72 -> 0
helped: 5
HURT: 0

total fills in shared programs: 9600 -> 9451 (-1.55%)
fills in affected programs: 149 -> 0
helped: 5
HURT: 0

LOST:   14
GAINED: 9

Skylake
total instructions in shared programs: 18185074 -> 18183866 (<.01%)
instructions in affected programs: 575180 -> 573972 (-0.21%)
helped: 1286
HURT: 468
helped stats (abs) min: 1 max: 15 x̄: 1.55 x̃: 1
helped stats (rel) min: 0.03% max: 4.08% x̄: 0.67% x̃: 0.65%
HURT stats (abs)   min: 1 max: 8 x̄: 1.69 x̃: 1
HURT stats (rel)   min: 0.13% max: 7.69% x̄: 0.87% x̃: 0.45%
95% mean confidence interval for instructions value: -0.77 -0.60
95% mean confidence interval for instructions %-change: -0.30% -0.22%
Instructions are helped.

total cycles in shared programs: 960518105 -> 960608234 (<.01%)
cycles in affected programs: 42536073 -> 42626202 (0.21%)
helped: 1210
HURT: 1714
helped stats (abs) min: 1 max: 7015 x̄: 123.41 x̃: 10
helped stats (rel) min: <.01% max: 33.76% x̄: 1.32% x̃: 0.26%
HURT stats (abs)   min: 1 max: 14474 x̄: 139.71 x̃: 14
HURT stats (rel)   min: <.01% max: 58.94% x̄: 2.00% x̃: 0.44%
95% mean confidence interval for cycles value: 4.02 57.63
95% mean confidence interval for cycles %-change: 0.43% 0.82%
Cycles are HURT.

LOST:   16
GAINED: 42

Broadwell
total instructions in shared programs: 17856880 -> 17852158 (-0.03%)
instructions in affected programs: 564836 -> 560114 (-0.84%)
helped: 1243
HURT: 418
helped stats (abs) min: 1 max: 115 x̄: 4.36 x̃: 1
helped stats (rel) min: 0.03% max: 9.67% x̄: 0.90% x̃: 0.67%
HURT stats (abs)   min: 1 max: 8 x̄: 1.67 x̃: 1
HURT stats (rel)   min: 0.14% max: 7.69% x̄: 0.89% x̃: 0.46%
95% mean confidence interval for instructions value: -3.45 -2.23
95% mean confidence interval for instructions %-change: -0.51% -0.38%
Instructions are helped.

total cycles in shared programs: 1031140321 -> 1029856892 (-0.12%)
cycles in affected programs: 66986946 -> 65703517 (-1.92%)
helped: 1084
HURT: 1653
helped stats (abs) min: 1 max: 415168 x̄: 1835.32 x̃: 10
helped stats (rel) min: <.01% max: 57.16% x̄: 1.19% x̃: 0.28%
HURT stats (abs)   min: 1 max: 43930 x̄: 427.14 x̃: 12
HURT stats (rel)   min: <.01% max: 57.53% x̄: 1.32% x̃: 0.39%
95% mean confidence interval for cycles value: -915.76 -22.07
95% mean confidence interval for cycles %-change: 0.17% 0.47%
Inconclusive result (value mean confidence interval and %-change mean confidence interval disagree).

total spills in shared programs: 20891 -> 20335 (-2.66%)
spills in affected programs: 1567 -> 1011 (-35.48%)
helped: 70
HURT: 0

total fills in shared programs: 27307 -> 25905 (-5.13%)
fills in affected programs: 5381 -> 3979 (-26.05%)
helped: 71
HURT: 0

LOST:   17
GAINED: 20

Haswell
total instructions in shared programs: 16411850 -> 16409414 (-0.01%)
instructions in affected programs: 602666 -> 600230 (-0.40%)
helped: 1152
HURT: 781
helped stats (abs) min: 1 max: 103 x̄: 3.59 x̃: 1
helped stats (rel) min: 0.03% max: 8.61% x̄: 0.85% x̃: 0.65%
HURT stats (abs)   min: 1 max: 41 x̄: 2.18 x̃: 1
HURT stats (rel)   min: 0.12% max: 7.69% x̄: 0.88% x̃: 0.69%
95% mean confidence interval for instructions value: -1.74 -0.78
95% mean confidence interval for instructions %-change: -0.21% -0.10%
Instructions are helped.

total cycles in shared programs: 1035338781 -> 1036977801 (0.16%)
cycles in affected programs: 68961096 -> 70600116 (2.38%)
helped: 1246
HURT: 2206
helped stats (abs) min: 1 max: 392022 x̄: 1040.28 x̃: 14
helped stats (rel) min: <.01% max: 56.44% x̄: 2.32% x̃: 0.38%
HURT stats (abs)   min: 1 max: 68630 x̄: 1330.56 x̃: 18
HURT stats (rel)   min: <.01% max: 69.97% x̄: 3.31% x̃: 0.61%
95% mean confidence interval for cycles value: 90.43 859.17
95% mean confidence interval for cycles %-change: 1.02% 1.54%
Cycles are HURT.

total spills in shared programs: 17805 -> 17457 (-1.95%)
spills in affected programs: 1202 -> 854 (-28.95%)
helped: 34
HURT: 31

total fills in shared programs: 20939 -> 20387 (-2.64%)
fills in affected programs: 2702 -> 2150 (-20.43%)
helped: 34
HURT: 31

LOST:   24
GAINED: 45

Ivy Bridge and earlier Intel GPUs had similar results. (Ivy Bridge shown)
total instructions in shared programs: 15515912 -> 15516757 (<.01%)
instructions in affected programs: 396569 -> 397414 (0.21%)
helped: 578
HURT: 858
helped stats (abs) min: 1 max: 9 x̄: 1.32 x̃: 1
helped stats (rel) min: 0.04% max: 3.70% x̄: 0.65% x̃: 0.65%
HURT stats (abs)   min: 1 max: 11 x̄: 1.87 x̃: 1
HURT stats (rel)   min: 0.08% max: 12.90% x̄: 0.95% x̃: 0.53%
95% mean confidence interval for instructions value: 0.47 0.70
95% mean confidence interval for instructions %-change: 0.24% 0.37%
Instructions are HURT.

total cycles in shared programs: 584395455 -> 584466352 (0.01%)
cycles in affected programs: 20346570 -> 20417467 (0.35%)
helped: 1192
HURT: 1896
helped stats (abs) min: 1 max: 4108 x̄: 123.27 x̃: 14
helped stats (rel) min: <.01% max: 37.20% x̄: 2.27% x̃: 0.46%
HURT stats (abs)   min: 1 max: 3698 x̄: 114.89 x̃: 19
HURT stats (rel)   min: <.01% max: 70.28% x̄: 3.02% x̃: 0.71%
95% mean confidence interval for cycles value: 10.75 35.16
95% mean confidence interval for cycles %-change: 0.73% 1.23%
Cycles are HURT.

LOST:   20
GAINED: 12
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9374>
This commit is contained in:
Ian Romanick
2018-01-10 18:50:58 -08:00
committed by Marge Bot
parent 206495cac4
commit 33031bdab6
1 changed files with 10 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/compiler/nir/nir_opt_algebraic.py
View File

@@ -2161,20 +2161,20 @@ late_optimizations = [
# a == -b true true false false
# (a+b) != 0 true true true true
# a != -b false false true true
(('flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))),
(('flt', ('fneg', ('fadd', a, b)), 0.0), ('flt', ('fneg', a), b)),
(('~fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))),
(('~fge', ('fneg', ('fadd', a, b)), 0.0), ('fge', ('fneg', a), b)),
(('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))),
(('~fneu', ('fadd', a, b), 0.0), ('fneu', a, ('fneg', b))),
(('flt', ('fadd(is_used_once)', a, b), 0.0), ('flt', a, ('fneg', b))),
(('flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a, b)), 0.0), ('flt', ('fneg', a), b)),
(('~fge', ('fadd(is_used_once)', a, b), 0.0), ('fge', a, ('fneg', b))),
(('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a, b)), 0.0), ('fge', ('fneg', a), b)),
(('~feq', ('fadd(is_used_once)', a, b), 0.0), ('feq', a, ('fneg', b))),
(('~fneu', ('fadd(is_used_once)', a, b), 0.0), ('fneu', a, ('fneg', b))),
# If either source must be finite, then the original (a+b) cannot produce
# NaN due to Inf-Inf. The patterns and the replacements produce the same
# result if b is NaN. Therefore, the replacements are exact.
(('fge', ('fadd', 'a(is_finite)', b), 0.0), ('fge', a, ('fneg', b))),
(('fge', ('fneg', ('fadd', 'a(is_finite)', b)), 0.0), ('fge', ('fneg', a), b)),
(('feq', ('fadd', 'a(is_finite)', b), 0.0), ('feq', a, ('fneg', b))),
(('fneu', ('fadd', 'a(is_finite)', b), 0.0), ('fneu', a, ('fneg', b))),
(('fge', ('fadd(is_used_once)', 'a(is_finite)', b), 0.0), ('fge', a, ('fneg', b))),
(('fge', ('fneg(is_used_once)', ('fadd(is_used_once)', 'a(is_finite)', b)), 0.0), ('fge', ('fneg', a), b)),
(('feq', ('fadd(is_used_once)', 'a(is_finite)', b), 0.0), ('feq', a, ('fneg', b))),
(('fneu', ('fadd(is_used_once)', 'a(is_finite)', b), 0.0), ('fneu', a, ('fneg', b))),
# nir_lower_to_source_mods will collapse this, but its existence during the
# optimization loop can prevent other optimizations.