Mesa (main): nir/algebraic: Rearrange some logic-joined comparisons and reduce

Thu May 20 02:09:22 UTC 2021

Module: Mesa
Branch: main
Commit: 64bcfc3a170bba3081ddfa04c6851158a5691ba4
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=64bcfc3a170bba3081ddfa04c6851158a5691ba4

Author: Ian Romanick <ian.d.romanick at intel.com>
Date:   Mon Aug 17 15:26:07 2020 -0700

nir/algebraic: Rearrange some logic-joined comparisons and reduce

On Skylake and Broadwell, a single big compute shader in Dirt Rally has
spills and fills *REALLY* helped.  That same shader is hurt very
slightly for spills and fills on Ice Lake.

v2: Move the patterns earlier to be nearer other patterns that are
similar.  Mark the replacement fmin and fmax exact.  Both suggested by
Rhys.

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

Tiger Lake
total instructions in shared programs: 21073812 -> 21073041 (<.01%)
instructions in affected programs: 77608 -> 76837 (-0.99%)
helped: 522
HURT: 33
helped stats (abs) min: 1 max: 26 x̄: 1.58 x̃: 1
helped stats (rel) min: 0.22% max: 14.29% x̄: 1.29% x̃: 1.02%
HURT stats (abs)   min: 1 max: 8 x̄: 1.67 x̃: 1
HURT stats (rel)   min: 0.25% max: 3.42% x̄: 1.06% x̃: 0.86%
95% mean confidence interval for instructions value: -1.57 -1.20
95% mean confidence interval for instructions %-change: -1.25% -1.05%
Instructions are helped.

total cycles in shared programs: 856224346 -> 856211096 (<.01%)
cycles in affected programs: 2394231 -> 2380981 (-0.55%)
helped: 603
HURT: 25
helped stats (abs) min: 1 max: 5218 x̄: 59.37 x̃: 28
helped stats (rel) min: 0.06% max: 5.61% x̄: 1.52% x̃: 1.37%
HURT stats (abs)   min: 2 max: 21394 x̄: 901.92 x̃: 10
HURT stats (rel)   min: 0.02% max: 5.90% x̄: 0.95% x̃: 0.59%
95% mean confidence interval for cycles value: -93.61 51.41
95% mean confidence interval for cycles %-change: -1.50% -1.34%
Inconclusive result (value mean confidence interval includes 0).

LOST:   1
GAINED: 1

Ice Lake
total instructions in shared programs: 20025692 -> 20024554 (<.01%)
instructions in affected programs: 104981 -> 103843 (-1.08%)
helped: 738
HURT: 0
helped stats (abs) min: 1 max: 30 x̄: 1.54 x̃: 1
helped stats (rel) min: 0.31% max: 10.53% x̄: 1.20% x̃: 1.06%
95% mean confidence interval for instructions value: -1.66 -1.43
95% mean confidence interval for instructions %-change: -1.26% -1.14%
Instructions are helped.

total cycles in shared programs: 979474407 -> 979422333 (<.01%)
cycles in affected programs: 4136364 -> 4084290 (-1.26%)
helped: 759
HURT: 59
helped stats (abs) min: 2 max: 11010 x̄: 72.78 x̃: 28
helped stats (rel) min: 0.03% max: 6.43% x̄: 1.23% x̃: 1.02%
HURT stats (abs)   min: 1 max: 698 x̄: 53.66 x̃: 8
HURT stats (rel)   min: 0.02% max: 24.05% x̄: 1.64% x̃: 0.33%
95% mean confidence interval for cycles value: -97.08 -30.24
95% mean confidence interval for cycles %-change: -1.14% -0.91%
Cycles are helped.

total spills in shared programs: 10568 -> 10569 (<.01%)
spills in affected programs: 102 -> 103 (0.98%)
helped: 0
HURT: 1

total fills in shared programs: 11347 -> 11349 (0.02%)
fills in affected programs: 277 -> 279 (0.72%)
helped: 0
HURT: 1

LOST:   2
GAINED: 2

Skylake
total instructions in shared programs: 18190419 -> 18188523 (-0.01%)
instructions in affected programs: 102502 -> 100606 (-1.85%)
helped: 791
HURT: 0
helped stats (abs) min: 1 max: 676 x̄: 2.40 x̃: 1
helped stats (rel) min: 0.34% max: 20.23% x̄: 1.41% x̃: 1.23%
95% mean confidence interval for instructions value: -4.07 -0.72
95% mean confidence interval for instructions %-change: -1.47% -1.34%
Instructions are helped.

total cycles in shared programs: 960737969 -> 960498951 (-0.02%)
cycles in affected programs: 4435351 -> 4196333 (-5.39%)
helped: 804
HURT: 67
helped stats (abs) min: 1 max: 198540 x̄: 300.54 x̃: 24
helped stats (rel) min: 0.03% max: 25.41% x̄: 1.21% x̃: 0.92%
HURT stats (abs)   min: 2 max: 680 x̄: 39.06 x̃: 6
HURT stats (rel)   min: 0.05% max: 23.98% x̄: 1.12% x̃: 0.19%
95% mean confidence interval for cycles value: -722.03 173.20
95% mean confidence interval for cycles %-change: -1.15% -0.91%
Inconclusive result (value mean confidence interval includes 0).

total spills in shared programs: 9757 -> 9722 (-0.36%)
spills in affected programs: 138 -> 103 (-25.36%)
helped: 1
HURT: 0

total fills in shared programs: 9861 -> 9576 (-2.89%)
fills in affected programs: 564 -> 279 (-50.53%)
helped: 1
HURT: 0

LOST:   5
GAINED: 2

Broadwell
total instructions in shared programs: 17853870 -> 17852414 (<.01%)
instructions in affected programs: 101276 -> 99820 (-1.44%)
helped: 777
HURT: 0
helped stats (abs) min: 1 max: 264 x̄: 1.87 x̃: 1
helped stats (rel) min: 0.34% max: 8.44% x̄: 1.37% x̃: 1.23%
95% mean confidence interval for instructions value: -2.54 -1.21
95% mean confidence interval for instructions %-change: -1.42% -1.32%
Instructions are helped.

total cycles in shared programs: 1029846029 -> 1029725458 (-0.01%)
cycles in affected programs: 4435791 -> 4315220 (-2.72%)
helped: 813
HURT: 43
helped stats (abs) min: 2 max: 68560 x̄: 149.95 x̃: 24
helped stats (rel) min: 0.02% max: 73.73% x̄: 1.43% x̃: 0.92%
HURT stats (abs)   min: 2 max: 726 x̄: 31.12 x̃: 13
HURT stats (rel)   min: 0.01% max: 8.43% x̄: 0.62% x̃: 0.31%
95% mean confidence interval for cycles value: -299.58 17.87
95% mean confidence interval for cycles %-change: -1.63% -1.02%
Inconclusive result (value mean confidence interval includes 0).

total spills in shared programs: 20333 -> 20307 (-0.13%)
spills in affected programs: 151 -> 125 (-17.22%)
helped: 1
HURT: 0

total fills in shared programs: 25899 -> 25775 (-0.48%)
fills in affected programs: 573 -> 449 (-21.64%)
helped: 1
HURT: 0

LOST:   5
GAINED: 0

Sandy Bridge, Ivy Bridge, and Haswell had similar results. (Haswell shown)
total instructions in shared programs: 16417658 -> 16416320 (<.01%)
instructions in affected programs: 96495 -> 95157 (-1.39%)
helped: 774
HURT: 0
helped stats (abs) min: 1 max: 18 x̄: 1.73 x̃: 1
helped stats (rel) min: 0.33% max: 9.80% x̄: 1.52% x̃: 1.20%
95% mean confidence interval for instructions value: -1.83 -1.63
95% mean confidence interval for instructions %-change: -1.59% -1.46%
Instructions are helped.

total cycles in shared programs: 1037104346 -> 1037080579 (<.01%)
cycles in affected programs: 3787747 -> 3763980 (-0.63%)
helped: 791
HURT: 53
helped stats (abs) min: 1 max: 5411 x̄: 65.87 x̃: 32
helped stats (rel) min: 0.02% max: 21.17% x̄: 1.44% x̃: 1.18%
HURT stats (abs)   min: 2 max: 14160 x̄: 534.72 x̃: 18
HURT stats (rel)   min: 0.02% max: 15.37% x̄: 5.70% x̃: 0.54%
95% mean confidence interval for cycles value: -69.39 13.07
95% mean confidence interval for cycles %-change: -1.19% -0.80%
Inconclusive result (value mean confidence interval includes 0).

LOST:   12
GAINED: 2

GM45 and Iron Lake had similar results. (Iron Lake shown)
total instructions in shared programs: 8132855 -> 8132703 (<.01%)
instructions in affected programs: 8782 -> 8630 (-1.73%)
helped: 38
HURT: 0
helped stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4
helped stats (rel) min: 1.66% max: 3.23% x̄: 1.77% x̃: 1.72%
95% mean confidence interval for instructions value: -4.00 -4.00
95% mean confidence interval for instructions %-change: -1.88% -1.65%
Instructions are helped.

total cycles in shared programs: 238300850 -> 238298568 (<.01%)
cycles in affected programs: 257202 -> 254920 (-0.89%)
helped: 62
HURT: 2
helped stats (abs) min: 4 max: 58 x̄: 36.90 x̃: 50
helped stats (rel) min: 0.15% max: 1.55% x̄: 0.87% x̃: 1.12%
HURT stats (abs)   min: 2 max: 4 x̄: 3.00 x̃: 3
HURT stats (rel)   min: 0.12% max: 0.22% x̄: 0.17% x̃: 0.17%
95% mean confidence interval for cycles value: -41.34 -29.98
95% mean confidence interval for cycles %-change: -0.95% -0.73%
Cycles are helped.

Fossil-db results:

All Intel platforms had similar results. (Ice Lake shown)
Instructions in all programs: 145296888 -> 145296346 (-0.0%)
SENDs in all programs: 6863696 -> 6863696 (+0.0%)
Loops in all programs: 38334 -> 38334 (+0.0%)
Cycles in all programs: 8800262303 -> 8800258950 (-0.0%)
Spills in all programs: 216880 -> 216880 (+0.0%)
Fills in all programs: 334248 -> 334248 (+0.0%)

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10012>

---

 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 748fae79be9..0c441b8fb8c 100644
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -684,6 +684,16 @@ optimizations.extend([
    (('iand', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umax', a, b), c)),
    (('iand', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umax', b, c))),
    (('iand', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umin', a, b), c)),
+
+   # A number of shaders contain a pattern like a.x < 0.0 || a.x > 1.0 || a.y
+   # < 0.0, || a.y > 1.0 || ...  These patterns rearrange and replace in a
+   # single step.  Doing just the replacement can lead to an infinite loop as
+   # the pattern is repeatedly applied to the result of the previous
+   # application of the pattern.
+   (('ior', ('ior(is_used_once)', ('flt(is_used_once)', a, c), d), ('flt', b, c)), ('ior', ('flt', ('!fmin', a, b), c), d)),
+   (('ior', ('ior(is_used_once)', ('flt', a, c), d), ('flt(is_used_once)', b, c)), ('ior', ('flt', ('!fmin', a, b), c), d)),
+   (('ior', ('ior(is_used_once)', ('flt(is_used_once)', a, b), d), ('flt', a, c)), ('ior', ('flt', a, ('!fmax', b, c)), d)),
+   (('ior', ('ior(is_used_once)', ('flt', a, b), d), ('flt(is_used_once)', a, c)), ('ior', ('flt', a, ('!fmax', b, c)), d)),
 ])
 
 # Float sizes

