nir: Transform expressions of b2f(a) and b2f(b) to a || b

All Gen6+ platforms had pretty similar results. (Skylake shown)
total instructions in shared programs: 14277184 -> 14276964 (<.01%)
instructions in affected programs: 10082 -> 9862 (-2.18%)
helped: 37
HURT: 1
helped stats (abs) min: 1 max: 30 x̄: 5.97 x̃: 4
helped stats (rel) min: 0.14% max: 16.00% x̄: 5.23% x̃: 2.04%
HURT stats (abs)   min: 1 max: 1 x̄: 1.00 x̃: 1
HURT stats (rel)   min: 0.70% max: 0.70% x̄: 0.70% x̃: 0.70%
95% mean confidence interval for instructions value: -7.87 -3.71
95% mean confidence interval for instructions %-change: -6.98% -3.16%
Instructions are helped.

total cycles in shared programs: 532577990 -> 532577062 (<.01%)
cycles in affected programs: 170959 -> 170031 (-0.54%)
helped: 33
HURT: 9
helped stats (abs) min: 2 max: 120 x̄: 30.91 x̃: 30
helped stats (rel) min: 0.02% max: 7.65% x̄: 2.66% x̃: 1.13%
HURT stats (abs)   min: 2 max: 24 x̄: 10.22 x̃: 8
HURT stats (rel)   min: 0.09% max: 1.79% x̄: 0.61% x̃: 0.22%
95% mean confidence interval for cycles value: -31.23 -12.96
95% mean confidence interval for cycles %-change: -2.90% -1.02%
Cycles are helped.

Iron Lake and GM45 had similar results. (Iron Lake shown)
total instructions in shared programs: 7781539 -> 7781301 (<.01%)
instructions in affected programs: 10169 -> 9931 (-2.34%)
helped: 32
HURT: 0
helped stats (abs) min: 2 max: 20 x̄: 7.44 x̃: 6
helped stats (rel) min: 0.47% max: 17.02% x̄: 4.03% x̃: 1.88%
95% mean confidence interval for instructions value: -9.53 -5.34
95% mean confidence interval for instructions %-change: -5.94% -2.12%
Instructions are helped.

total cycles in shared programs: 177878590 -> 177878932 (<.01%)
cycles in affected programs: 78706 -> 79048 (0.43%)
helped: 7
HURT: 21
helped stats (abs) min: 6 max: 34 x̄: 24.57 x̃: 28
helped stats (rel) min: 0.15% max: 8.33% x̄: 4.66% x̃: 6.37%
HURT stats (abs)   min: 2 max: 86 x̄: 24.48 x̃: 22
HURT stats (rel)   min: 0.01% max: 4.28% x̄: 1.21% x̃: 0.70%
95% mean confidence interval for cycles value: 0.30 24.13
95% mean confidence interval for cycles %-change: -1.52% 1.01%
Inconclusive result (%-change mean confidence interval includes 0).

v2: s/fmin/fmax/.  Noticed by Thomas Helland.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 8300d6d01f52c979fb5814acc0295583d216ec1d..5ba469f5f65b90dc05faeb23e30aea30364757a2 100644 (file)
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -166,6 +166,25 @@ optimizations = [
 
    (('fge', ('fneg', ('b2f', a)), 0.0), ('inot', a)),
 
+   (('fne', ('fadd', ('b2f', a), ('b2f', b)), 0.0), ('ior', a, b)),
+   (('fne', ('fmax', ('b2f', a), ('b2f', b)), 0.0), ('ior', a, b)),
+   (('fne', ('bcsel', a, 1.0, ('b2f', b))   , 0.0), ('ior', a, b)),
+   (('fne', ('b2f', a), ('fneg', ('b2f', b))),      ('ior', a, b)),
+
+   # -(b2f(a) + b2f(b)) < 0
+   # 0 < b2f(a) + b2f(b)
+   # 0 != b2f(a) + b2f(b)       b2f must be 0 or 1, so the sum is non-negative
+   # a || b
+   (('flt', ('fneg', ('fadd', ('b2f', a), ('b2f', b))), 0.0), ('ior', a, b)),
+   (('flt', 0.0, ('fadd', ('b2f', a), ('b2f', b))), ('ior', a, b)),
+
+   # Some optimizations (below) convert things like (a < b || c < b) into
+   # (min(a, c) < b).  However, this interfers with the previous optimizations
+   # that try to remove comparisons with negated sums of b2f.  This just
+   # breaks that apart.
+   (('flt', ('fmin', c, ('fneg', ('fadd', ('b2f', a), ('b2f', b)))), 0.0),
+    ('ior', ('flt', c, 0.0), ('ior', a, b))),
+
    (('~flt', ('fadd', a, b), a), ('flt', b, 0.0)),
    (('~fge', ('fadd', a, b), a), ('fge', b, 0.0)),
    (('~feq', ('fadd', a, b), a), ('feq', b, 0.0)),