nir/algebraic: Use value range analysis to eliminate useless unary ops

Sandy Bridge is the big winner because it lies at something of a
crossroads.  It supports a fairly high OpenGL version, and it still has
the old style math box.  The high OpenGL version means a lot more
shaders can run on it.  The old style math box means extra moves are
necessary to resolve source modifiers on operands to complex math
instructions like COS, SQRT, and RCP.

v2: Remove a couple patterns that are now redundant.

All Gen7+ platforms had similar results.  (Ice Lake shown)
total instructions in shared programs: 16282006 -> 16278207 (-0.02%)
instructions in affected programs: 174555 -> 170756 (-2.18%)
helped: 661
HURT: 0
helped stats (abs) min: 1 max: 36 x̄: 5.75 x̃: 3
helped stats (rel) min: 0.06% max: 23.68% x̄: 2.81% x̃: 1.94%
95% mean confidence interval for instructions value: -6.16 -5.34
95% mean confidence interval for instructions %-change: -3.02% -2.60%
Instructions are helped.

total cycles in shared programs: 367168597 -> 367134284 (<.01%)
cycles in affected programs: 1105276 -> 1070963 (-3.10%)
helped: 460
HURT: 150
helped stats (abs) min: 1 max: 568 x̄: 96.60 x̃: 82
helped stats (rel) min: 0.02% max: 32.50% x̄: 7.99% x̃: 4.27%
HURT stats (abs)   min: 1 max: 901 x̄: 67.49 x̃: 39
HURT stats (rel)   min: 0.07% max: 20.00% x̄: 4.90% x̃: 4.22%
95% mean confidence interval for cycles value: -65.68 -46.82
95% mean confidence interval for cycles %-change: -5.59% -4.05%
Cycles are helped.

Sandy Bridge
total instructions in shared programs: 10824272 -> 10802557 (-0.20%)
instructions in affected programs: 1237988 -> 1216273 (-1.75%)
helped: 8199
HURT: 0
helped stats (abs) min: 1 max: 41 x̄: 2.65 x̃: 2
helped stats (rel) min: 0.12% max: 20.00% x̄: 2.04% x̃: 1.73%
95% mean confidence interval for instructions value: -2.70 -2.59
95% mean confidence interval for instructions %-change: -2.07% -2.00%
Instructions are helped.

total cycles in shared programs: 154009894 -> 153843598 (-0.11%)
cycles in affected programs: 10650486 -> 10484190 (-1.56%)
helped: 4973
HURT: 1533
helped stats (abs) min: 1 max: 3904 x̄: 40.20 x̃: 20
helped stats (rel) min: 0.02% max: 41.72% x̄: 2.63% x̃: 1.67%
HURT stats (abs)   min: 1 max: 453 x̄: 21.94 x̃: 8
HURT stats (rel)   min: 0.02% max: 41.91% x̄: 1.54% x̃: 0.58%
95% mean confidence interval for cycles value: -28.02 -23.10
95% mean confidence interval for cycles %-change: -1.74% -1.56%
Cycles are helped.

LOST:   0
GAINED: 2

GM45 and Iron Lake had similar results. (Iron Lake shown)
total instructions in shared programs: 8135196 -> 8134888 (<.01%)
instructions in affected programs: 31920 -> 31612 (-0.96%)
helped: 169
HURT: 0
helped stats (abs) min: 1 max: 12 x̄: 1.82 x̃: 2
helped stats (rel) min: 0.43% max: 3.23% x̄: 1.23% x̃: 1.16%
95% mean confidence interval for instructions value: -2.01 -1.64
95% mean confidence interval for instructions %-change: -1.32% -1.15%
Instructions are helped.

total cycles in shared programs: 188575724 -> 188574092 (<.01%)
cycles in affected programs: 406840 -> 405208 (-0.40%)
helped: 169
HURT: 0
helped stats (abs) min: 4 max: 72 x̄: 9.66 x̃: 10
helped stats (rel) min: 0.07% max: 2.16% x̄: 0.57% x̃: 0.47%
95% mean confidence interval for cycles value: -10.72 -8.59
95% mean confidence interval for cycles %-change: -0.63% -0.50%
Cycles are helped.

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
This commit is contained in:
Ian Romanick 2018-01-23 17:35:51 +08:00
parent 8d14380971
commit 5544b2cbbd
1 changed files with 8 additions and 2 deletions

View File

@ -764,8 +764,6 @@ optimizations = [
(('f2u', ('ftrunc', a)), ('f2u', a)),
(('i2b', ('ineg', a)), ('i2b', a)),
(('i2b', ('iabs', a)), ('i2b', a)),
(('fabs', ('b2f', a)), ('b2f', a)),
(('iabs', ('b2i', a)), ('b2i', a)),
(('inot', ('f2b1', a)), ('feq', a, 0.0)),
# Ironically, mark these as imprecise because removing the conversions may
@ -776,6 +774,14 @@ optimizations = [
(('~f2u32', ('i2f', 'a@32')), a),
(('~f2u32', ('u2f', 'a@32')), a),
(('ffloor', 'a(is_integral)'), a),
(('fceil', 'a(is_integral)'), a),
(('ftrunc', 'a(is_integral)'), a),
(('ffract', 'a(is_integral)'), 0.0),
(('fabs', 'a(is_not_negative)'), a),
(('iabs', 'a(is_not_negative)'), a),
(('fsat', 'a(is_not_positive)'), 0.0),
# Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
# says:
#