Mesa (master): nir: Rearrange and-compounded float compares

Tue Jan 30 23:40:26 UTC 2018

Module: Mesa
Branch: master
Commit: 053be9f0201db183a761a2a4ce813a114a766b72
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=053be9f0201db183a761a2a4ce813a114a766b72

Author: Ian Romanick <ian.d.romanick at intel.com>
Date:   Thu Jan  4 15:21:30 2018 -0800

nir: Rearrange and-compounded float compares

If both comparisons are used as sources for instructions other than the
iand, this transformation is detrimental.  If the non-identical value in
both compares is constant, the fmin or fmax will be constant-folded
away, so the transformation is always a win.

It is interesting to me that on Iron Lake only 81 shaders have
instruction counts changed, but 726 shaders have cycle counts changed.

shader-db results:

Skylake
total instructions in shared programs: 14525728 -> 14521017 (-0.03%)
instructions in affected programs: 1164726 -> 1160015 (-0.40%)
helped: 1692
HURT: 5
helped stats (abs) min: 1 max: 637 x̄: 2.79 x̃: 2
helped stats (rel) min: 0.07% max: 16.36% x̄: 0.81% x̃: 0.33%
HURT stats (abs)   min: 1 max: 12 x̄: 3.20 x̃: 1
HURT stats (rel)   min: 0.38% max: 2.86% x̄: 2.36% x̃: 2.86%
95% mean confidence interval for instructions value: -3.52 -2.03
95% mean confidence interval for instructions %-change: -0.86% -0.74%
Instructions are helped.

total cycles in shared programs: 533115449 -> 532991404 (-0.02%)
cycles in affected programs: 119401803 -> 119277758 (-0.10%)
helped: 1145
HURT: 467
helped stats (abs) min: 1 max: 34644 x̄: 145.92 x̃: 18
helped stats (rel) min: <.01% max: 45.33% x̄: 1.58% x̃: 0.42%
HURT stats (abs)   min: 1 max: 1590 x̄: 92.15 x̃: 15
HURT stats (rel)   min: <.01% max: 13.48% x̄: 1.26% x̃: 0.39%
95% mean confidence interval for cycles value: -122.16 -31.74
95% mean confidence interval for cycles %-change: -0.94% -0.57%
Cycles are helped.

total spills in shared programs: 9597 -> 9534 (-0.66%)
spills in affected programs: 403 -> 340 (-15.63%)
helped: 1
HURT: 1

total fills in shared programs: 13904 -> 13790 (-0.82%)
fills in affected programs: 1627 -> 1513 (-7.01%)
helped: 2
HURT: 1

LOST:   0
GAINED: 2

Broadwell
total instructions in shared programs: 14816966 -> 14812590 (-0.03%)
instructions in affected programs: 1499885 -> 1495509 (-0.29%)
helped: 1672
HURT: 15
helped stats (abs) min: 1 max: 455 x̄: 2.70 x̃: 2
helped stats (rel) min: 0.05% max: 16.36% x̄: 0.81% x̃: 0.33%
HURT stats (abs)   min: 1 max: 21 x̄: 9.20 x̃: 8
HURT stats (rel)   min: 0.08% max: 2.86% x̄: 1.06% x̃: 0.53%
95% mean confidence interval for instructions value: -3.14 -2.05
95% mean confidence interval for instructions %-change: -0.85% -0.73%
Instructions are helped.

total cycles in shared programs: 559353622 -> 559345595 (<.01%)
cycles in affected programs: 139893703 -> 139885676 (<.01%)
helped: 921
HURT: 697
helped stats (abs) min: 1 max: 42424 x̄: 143.45 x̃: 18
helped stats (rel) min: <.01% max: 36.23% x̄: 2.02% x̃: 0.87%
HURT stats (abs)   min: 1 max: 2370 x̄: 178.03 x̃: 38
HURT stats (rel)   min: <.01% max: 17.35% x̄: 0.71% x̃: 0.14%
95% mean confidence interval for cycles value: -59.64 49.72
95% mean confidence interval for cycles %-change: -1.02% -0.66%
Inconclusive result (value mean confidence interval includes 0).

total spills in shared programs: 78902 -> 78861 (-0.05%)
spills in affected programs: 2418 -> 2377 (-1.70%)
helped: 1
HURT: 11

total fills in shared programs: 83782 -> 83678 (-0.12%)
fills in affected programs: 3515 -> 3411 (-2.96%)
helped: 2
HURT: 11

LOST:   0
GAINED: 5

Haswell and Ivy Bridge had similar results. Haswell shown.
total instructions in shared programs: 9033898 -> 9032010 (-0.02%)
instructions in affected programs: 308064 -> 306176 (-0.61%)
helped: 921
HURT: 4
helped stats (abs) min: 1 max: 20 x̄: 2.05 x̃: 1
helped stats (rel) min: 0.17% max: 17.54% x̄: 0.80% x̃: 0.35%
HURT stats (abs)   min: 1 max: 1 x̄: 1.00 x̃: 1
HURT stats (rel)   min: 3.23% max: 3.23% x̄: 3.23% x̃: 3.23%
95% mean confidence interval for instructions value: -2.21 -1.87
95% mean confidence interval for instructions %-change: -0.88% -0.68%
Instructions are helped.

total cycles in shared programs: 84628949 -> 84620520 (<.01%)
cycles in affected programs: 2164913 -> 2156484 (-0.39%)
helped: 518
HURT: 359
helped stats (abs) min: 1 max: 440 x̄: 41.52 x̃: 20
helped stats (rel) min: <.01% max: 17.17% x̄: 1.95% x̃: 1.01%
HURT stats (abs)   min: 1 max: 586 x̄: 36.43 x̃: 8
HURT stats (rel)   min: 0.04% max: 18.65% x̄: 1.47% x̃: 0.40%
95% mean confidence interval for cycles value: -15.17 -4.05
95% mean confidence interval for cycles %-change: -0.77% -0.32%
Cycles are helped.

LOST:   0
GAINED: 4

Sandy Bridge
total instructions in shared programs: 10544860 -> 10542933 (-0.02%)
instructions in affected programs: 360019 -> 358092 (-0.54%)
helped: 931
HURT: 4
helped stats (abs) min: 1 max: 20 x̄: 2.07 x̃: 1
helped stats (rel) min: 0.11% max: 15.52% x̄: 0.68% x̃: 0.30%
HURT stats (abs)   min: 1 max: 1 x̄: 1.00 x̃: 1
HURT stats (rel)   min: 3.33% max: 3.33% x̄: 3.33% x̃: 3.33%
95% mean confidence interval for instructions value: -2.23 -1.89
95% mean confidence interval for instructions %-change: -0.76% -0.58%
Instructions are helped.

total cycles in shared programs: 146106820 -> 146098397 (<.01%)
cycles in affected programs: 3435047 -> 3426624 (-0.25%)
helped: 572
HURT: 329
helped stats (abs) min: 1 max: 1289 x̄: 32.52 x̃: 15
helped stats (rel) min: <.01% max: 26.29% x̄: 0.97% x̃: 0.33%
HURT stats (abs)   min: 1 max: 1714 x̄: 30.93 x̃: 6
HURT stats (rel)   min: 0.02% max: 41.31% x̄: 1.13% x̃: 0.19%
95% mean confidence interval for cycles value: -16.85 -1.85
95% mean confidence interval for cycles %-change: -0.39% -0.01%
Cycles are helped.

LOST:   1
GAINED: 0

Iron Lake
total instructions in shared programs: 7886925 -> 7886711 (<.01%)
instructions in affected programs: 25763 -> 25549 (-0.83%)
helped: 75
HURT: 6
helped stats (abs) min: 1 max: 13 x̄: 3.33 x̃: 1
helped stats (rel) min: 0.35% max: 17.57% x̄: 1.96% x̃: 0.53%
HURT stats (abs)   min: 1 max: 16 x̄: 6.00 x̃: 1
HURT stats (rel)   min: 2.86% max: 4.79% x̄: 3.49% x̃: 2.86%
95% mean confidence interval for instructions value: -3.69 -1.60
95% mean confidence interval for instructions %-change: -2.54% -0.57%
Instructions are helped.

total cycles in shared programs: 178116888 -> 178115324 (<.01%)
cycles in affected programs: 5858790 -> 5857226 (-0.03%)
helped: 484
HURT: 242
helped stats (abs) min: 2 max: 76 x̄: 5.27 x̃: 6
helped stats (rel) min: 0.01% max: 10.70% x̄: 0.18% x̃: 0.06%
HURT stats (abs)   min: 2 max: 76 x̄: 4.07 x̃: 2
HURT stats (rel)   min: 0.01% max: 3.99% x̄: 0.19% x̃: 0.03%
95% mean confidence interval for cycles value: -2.76 -1.55
95% mean confidence interval for cycles %-change: -0.12% 0.01%
Inconclusive result (%-change mean confidence interval includes 0).

GM45
total instructions in shared programs: 4857870 -> 4857762 (<.01%)
instructions in affected programs: 13994 -> 13886 (-0.77%)
helped: 39
HURT: 5
helped stats (abs) min: 1 max: 13 x̄: 3.28 x̃: 2
helped stats (rel) min: 0.33% max: 17.11% x̄: 1.86% x̃: 0.48%
HURT stats (abs)   min: 1 max: 16 x̄: 4.00 x̃: 1
HURT stats (rel)   min: 2.86% max: 4.71% x̄: 3.23% x̃: 2.86%
95% mean confidence interval for instructions value: -3.86 -1.05
95% mean confidence interval for instructions %-change: -2.61% 0.04%
Inconclusive result (%-change mean confidence interval includes 0).

total cycles in shared programs: 122180744 -> 122179674 (<.01%)
cycles in affected programs: 3686646 -> 3685576 (-0.03%)
helped: 273
HURT: 141
helped stats (abs) min: 2 max: 76 x̄: 5.81 x̃: 6
helped stats (rel) min: 0.01% max: 10.70% x̄: 0.18% x̃: 0.06%
HURT stats (abs)   min: 2 max: 76 x̄: 3.66 x̃: 2
HURT stats (rel)   min: 0.01% max: 3.99% x̄: 0.16% x̃: 0.02%
95% mean confidence interval for cycles value: -3.42 -1.75
95% mean confidence interval for cycles %-change: -0.15% 0.03%
Inconclusive result (%-change mean confidence interval includes 0).

Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias at igalia.com>
Reviewed-by: Elie Tournier <elie.tournier at collabora.com>

---

 src/compiler/nir/nir_opt_algebraic.py | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index bee135c235..9dadb67bd2 100644
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -209,6 +209,14 @@ optimizations = [
    (('~ior', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmin', a, b), c)),
    (('~ior', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmin', b, c))),
    (('~ior', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmax', a, b), c)),
+   (('~iand', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmin', b, c))),
+   (('~iand', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmax', a, b), c)),
+   (('~iand', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmax', b, c))),
+   (('~iand', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmin', a, b), c)),
+   (('~iand', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmin', b, c))),
+   (('~iand', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmax', a, b), c)),
+   (('~iand', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmax', b, c))),
+   (('~iand', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmin', a, b), c)),
 
    # These patterns can result when (a < b || a < c) => (a < min(b, c))
    # transformations occur before constant propagation and loop-unrolling.


