2014-12-12 19:13:10 +00:00
|
|
|
#
|
|
|
|
# Copyright (C) 2014 Intel Corporation
|
|
|
|
#
|
|
|
|
# Permission is hereby granted, free of charge, to any person obtaining a
|
|
|
|
# copy of this software and associated documentation files (the "Software"),
|
|
|
|
# to deal in the Software without restriction, including without limitation
|
|
|
|
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
|
|
|
|
# and/or sell copies of the Software, and to permit persons to whom the
|
|
|
|
# Software is furnished to do so, subject to the following conditions:
|
|
|
|
#
|
|
|
|
# The above copyright notice and this permission notice (including the next
|
|
|
|
# paragraph) shall be included in all copies or substantial portions of the
|
|
|
|
# Software.
|
|
|
|
#
|
|
|
|
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
|
|
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
|
|
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
|
|
|
|
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
|
|
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
|
|
|
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
|
|
|
|
# IN THE SOFTWARE.
|
|
|
|
#
|
|
|
|
# Authors:
|
|
|
|
# Jason Ekstrand (jason@jlekstrand.net)
|
|
|
|
|
2018-07-05 14:17:32 +01:00
|
|
|
from __future__ import print_function
|
|
|
|
|
2018-06-27 11:37:38 +01:00
|
|
|
from collections import OrderedDict
|
2014-12-12 19:13:10 +00:00
|
|
|
import nir_algebraic
|
2018-12-06 18:56:33 +00:00
|
|
|
from nir_opcodes import type_sizes
|
2018-02-07 01:27:53 +00:00
|
|
|
import itertools
|
2014-12-12 19:13:10 +00:00
|
|
|
|
|
|
|
# Convenience variables
|
|
|
|
a = 'a'
|
|
|
|
b = 'b'
|
|
|
|
c = 'c'
|
|
|
|
d = 'd'
|
2018-09-06 04:45:19 +01:00
|
|
|
e = 'e'
|
2014-12-12 19:13:10 +00:00
|
|
|
|
|
|
|
# Written in the form (<search>, <replace>) where <search> is an expression
|
|
|
|
# and <replace> is either an expression or a value. An expression is
|
2016-03-17 18:04:49 +00:00
|
|
|
# defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
|
2014-12-12 19:13:10 +00:00
|
|
|
# where each source is either an expression or a value. A value can be
|
2015-01-29 00:42:20 +00:00
|
|
|
# either a numeric constant or a string representing a variable name.
|
|
|
|
#
|
2016-03-17 18:04:49 +00:00
|
|
|
# If the opcode in a search expression is prefixed by a '~' character, this
|
|
|
|
# indicates that the operation is inexact. Such operations will only get
|
|
|
|
# applied to SSA values that do not have the exact bit set. This should be
|
|
|
|
# used by by any optimizations that are not bit-for-bit exact. It should not,
|
|
|
|
# however, be used for backend-requested lowering operations as those need to
|
|
|
|
# happen regardless of precision.
|
|
|
|
#
|
2016-05-07 18:01:24 +01:00
|
|
|
# Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
|
|
|
|
# that the given variable will only match constants and the type indicates that
|
2015-01-29 00:42:20 +00:00
|
|
|
# the given variable will only match values from ALU instructions with the
|
2016-05-07 18:01:24 +01:00
|
|
|
# given output type, and (cond) specifies an additional condition function
|
|
|
|
# (see nir_search_helpers.h).
|
2015-01-29 00:42:20 +00:00
|
|
|
#
|
|
|
|
# For constants, you have to be careful to make sure that it is the right
|
|
|
|
# type because python is unaware of the source and destination types of the
|
2014-12-12 19:13:10 +00:00
|
|
|
# opcodes.
|
2016-04-25 20:23:38 +01:00
|
|
|
#
|
|
|
|
# All expression types can have a bit-size specified. For opcodes, this
|
|
|
|
# looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
|
2018-12-18 21:28:22 +00:00
|
|
|
# type and size. In the search half of the expression this indicates that it
|
|
|
|
# should only match that particular bit-size. In the replace half of the
|
|
|
|
# expression this indicates that the constructed value should have that
|
|
|
|
# bit-size.
|
2014-12-12 19:13:10 +00:00
|
|
|
|
|
|
|
optimizations = [
|
2016-05-07 18:01:24 +01:00
|
|
|
|
2019-05-31 18:54:12 +01:00
|
|
|
(('imul', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitshift'),
|
|
|
|
(('imul', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b)))), '!options->lower_bitshift'),
|
|
|
|
(('ishl', a, '#b@32'), ('imul', a, ('ishl', 1, b)), 'options->lower_bitshift'),
|
|
|
|
|
2019-02-25 19:43:53 +00:00
|
|
|
(('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a, b)), ('imul', a, b)),
|
|
|
|
(('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a, b)), ('imul', a, b)),
|
2019-02-15 07:08:39 +00:00
|
|
|
(('imul_2x32_64', a, b), ('pack_64_2x32_split', ('imul', a, b), ('imul_high', a, b)), 'options->lower_mul_2x32_64'),
|
|
|
|
(('umul_2x32_64', a, b), ('pack_64_2x32_split', ('imul', a, b), ('umul_high', a, b)), 'options->lower_mul_2x32_64'),
|
2016-10-19 16:53:10 +01:00
|
|
|
(('udiv', a, 1), a),
|
|
|
|
(('idiv', a, 1), a),
|
|
|
|
(('umod', a, 1), 0),
|
|
|
|
(('imod', a, 1), 0),
|
2019-05-31 18:54:12 +01:00
|
|
|
(('udiv', a, '#b@32(is_pos_power_of_two)'), ('ushr', a, ('find_lsb', b)), '!options->lower_bitshift'),
|
2016-05-07 18:01:24 +01:00
|
|
|
(('idiv', a, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', b))), 'options->lower_idiv'),
|
|
|
|
(('idiv', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', ('iabs', b))))), 'options->lower_idiv'),
|
|
|
|
(('umod', a, '#b(is_pos_power_of_two)'), ('iand', a, ('isub', b, 1))),
|
|
|
|
|
2014-12-12 19:13:10 +00:00
|
|
|
(('fneg', ('fneg', a)), a),
|
|
|
|
(('ineg', ('ineg', a)), a),
|
|
|
|
(('fabs', ('fabs', a)), ('fabs', a)),
|
|
|
|
(('fabs', ('fneg', a)), ('fabs', a)),
|
2018-11-07 23:32:32 +00:00
|
|
|
(('fabs', ('u2f', a)), ('u2f', a)),
|
2014-12-12 19:13:10 +00:00
|
|
|
(('iabs', ('iabs', a)), ('iabs', a)),
|
|
|
|
(('iabs', ('ineg', a)), ('iabs', a)),
|
2018-12-06 19:02:01 +00:00
|
|
|
(('f2b', ('fneg', a)), ('f2b', a)),
|
|
|
|
(('i2b', ('ineg', a)), ('i2b', a)),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fadd', a, 0.0), a),
|
2014-12-12 19:13:10 +00:00
|
|
|
(('iadd', a, 0), a),
|
2015-08-19 06:38:34 +01:00
|
|
|
(('usadd_4x8', a, 0), a),
|
|
|
|
(('usadd_4x8', a, ~0), ~0),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
|
2015-02-28 19:32:31 +00:00
|
|
|
(('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fadd', ('fneg', a), a), 0.0),
|
2015-02-28 19:32:32 +00:00
|
|
|
(('iadd', ('ineg', a), a), 0),
|
2015-12-15 01:44:23 +00:00
|
|
|
(('iadd', ('ineg', a), ('iadd', a, b)), b),
|
|
|
|
(('iadd', a, ('iadd', ('ineg', a), b)), b),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fadd', ('fneg', a), ('fadd', a, b)), b),
|
|
|
|
(('~fadd', a, ('fadd', ('fneg', a), b)), b),
|
2019-03-05 20:46:11 +00:00
|
|
|
(('fadd', ('fsat', a), ('fsat', ('fneg', a))), ('fsat', ('fabs', a))),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fmul', a, 0.0), 0.0),
|
2014-12-12 19:13:10 +00:00
|
|
|
(('imul', a, 0), 0),
|
2015-08-19 06:38:34 +01:00
|
|
|
(('umul_unorm_4x8', a, 0), 0),
|
|
|
|
(('umul_unorm_4x8', a, ~0), a),
|
2014-12-12 19:13:10 +00:00
|
|
|
(('fmul', a, 1.0), a),
|
|
|
|
(('imul', a, 1), a),
|
|
|
|
(('fmul', a, -1.0), ('fneg', a)),
|
|
|
|
(('imul', a, -1), ('ineg', a)),
|
2018-09-10 18:39:42 +01:00
|
|
|
# If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
|
|
|
|
# If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
|
|
|
|
# If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
|
|
|
|
(('fmul', ('fsign', a), ('fmul', a, a)), ('fmul', ('fabs', a), a)),
|
|
|
|
(('fmul', ('fmul', ('fsign', a), a), a), ('fmul', ('fabs', a), a)),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~ffma', 0.0, a, b), b),
|
|
|
|
(('~ffma', a, b, 0.0), ('fmul', a, b)),
|
2014-12-12 19:13:10 +00:00
|
|
|
(('ffma', 1.0, a, b), ('fadd', a, b)),
|
2019-05-10 18:20:02 +01:00
|
|
|
(('ffma', -1.0, a, b), ('fadd', ('fneg', a), b)),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~flrp', a, b, 0.0), a),
|
|
|
|
(('~flrp', a, b, 1.0), b),
|
|
|
|
(('~flrp', a, a, b), a),
|
|
|
|
(('~flrp', 0.0, a, b), ('fmul', a, b)),
|
2018-08-23 03:15:15 +01:00
|
|
|
|
|
|
|
# flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
|
|
|
|
(('~flrp', a, ('fadd(is_used_once)', a, b), c), ('fadd', ('fmul', b, c), a)),
|
|
|
|
(('~flrp@32', a, ('fadd', a, b), c), ('fadd', ('fmul', b, c), a), 'options->lower_flrp32'),
|
|
|
|
(('~flrp@64', a, ('fadd', a, b), c), ('fadd', ('fmul', b, c), a), 'options->lower_flrp64'),
|
|
|
|
|
|
|
|
(('~flrp@32', ('fadd', a, b), ('fadd', a, c), d), ('fadd', ('flrp', b, c, d), a), 'options->lower_flrp32'),
|
|
|
|
(('~flrp@64', ('fadd', a, b), ('fadd', a, c), d), ('fadd', ('flrp', b, c, d), a), 'options->lower_flrp64'),
|
|
|
|
|
2018-08-23 05:55:55 +01:00
|
|
|
(('~flrp@32', a, ('fmul(is_used_once)', a, b), c), ('fmul', ('flrp', 1.0, b, c), a), 'options->lower_flrp32'),
|
|
|
|
(('~flrp@64', a, ('fmul(is_used_once)', a, b), c), ('fmul', ('flrp', 1.0, b, c), a), 'options->lower_flrp64'),
|
|
|
|
|
|
|
|
(('~flrp', ('fmul(is_used_once)', a, b), ('fmul(is_used_once)', a, c), d), ('fmul', ('flrp', b, c, d), a)),
|
|
|
|
|
2018-12-06 03:36:10 +00:00
|
|
|
(('~flrp', a, b, ('b2f', 'c@1')), ('bcsel', c, b, a), 'options->lower_flrp32'),
|
2017-01-03 10:10:33 +00:00
|
|
|
(('~flrp', a, 0.0, c), ('fadd', ('fmul', ('fneg', a), c), a)),
|
2019-04-12 09:12:27 +01:00
|
|
|
(('ftrunc', a), ('bcsel', ('flt', a, 0.0), ('fneg', ('ffloor', ('fabs', a))), ('ffloor', ('fabs', a))), 'options->lower_ftrunc'),
|
2018-09-01 20:15:27 +01:00
|
|
|
(('ffloor', a), ('fsub', a, ('ffract', a)), 'options->lower_ffloor'),
|
nir/algebraic: Replace a-fract(a) with floor(a)
I noticed this while looking at a shader that was affected by Tim's
"more loop unrolling" series.
In review, Tim Arceri asked:
> Why the hurt on Gen6+ is this something that should be in the late
> optimisations pass?
As far as I can tell, it's just because our scheduler is terrible. In
all the fragment shaders that I looked at (some hurt shaders were from
other stages), only one of the SIMD8 or SIMD16 version would be hurt.
In many of those case, the other SIMD width is improved (e.g.,
shaders/closed/steam/brutal-legend/3990.shader_test).
Often it looks like the scheduler decides to differently schedule a SEND
the occurs somewhere early in the shader. Once that happens, everything
is different.
I looked at one vertex shader that was hurt (from Goat Simulator). In
that case, both the floor and fract are used. The optimization
eliminates the add, and it should allow better scheduling. In the area
of the FRC and RNDD instructions, the scheduler does the right thing.
However, later in the shader a MAD and and ADD get scheduled
differently, and that makes it slightly worse.
In light of this, I tried adding some "is_used_once" mark-up, and that
did not fix all the cycles regressions. It also did a lot more harm
than good on SKL (helped 82 vs. hurt 241).
All Gen6+ platforms had similar results. (Skylake shown)
total instructions in shared programs: 15437001 -> 15435259 (-0.01%)
instructions in affected programs: 213651 -> 211909 (-0.82%)
helped: 988
HURT: 0
helped stats (abs) min: 1 max: 27 x̄: 1.76 x̃: 1
helped stats (rel) min: 0.15% max: 11.54% x̄: 1.14% x̃: 0.59%
95% mean confidence interval for instructions value: -1.89 -1.63
95% mean confidence interval for instructions %-change: -1.23% -1.05%
Instructions are helped.
total cycles in shared programs: 383007378 -> 382997063 (<.01%)
cycles in affected programs: 1650825 -> 1640510 (-0.62%)
helped: 679
HURT: 302
helped stats (abs) min: 1 max: 348 x̄: 23.39 x̃: 14
helped stats (rel) min: 0.04% max: 28.77% x̄: 1.61% x̃: 0.98%
HURT stats (abs) min: 1 max: 250 x̄: 18.43 x̃: 7
HURT stats (rel) min: 0.04% max: 25.86% x̄: 1.41% x̃: 0.53%
95% mean confidence interval for cycles value: -13.05 -7.98
95% mean confidence interval for cycles %-change: -0.86% -0.50%
Cycles are helped.
Iron Lake and GM45 had similar results. (GM45 shown)
total instructions in shared programs: 5043616 -> 5043010 (-0.01%)
instructions in affected programs: 119691 -> 119085 (-0.51%)
helped: 432
HURT: 0
helped stats (abs) min: 1 max: 27 x̄: 1.40 x̃: 1
helped stats (rel) min: 0.10% max: 8.11% x̄: 0.66% x̃: 0.39%
95% mean confidence interval for instructions value: -1.58 -1.23
95% mean confidence interval for instructions %-change: -0.72% -0.59%
Instructions are helped.
total cycles in shared programs: 128139812 -> 128135762 (<.01%)
cycles in affected programs: 3829724 -> 3825674 (-0.11%)
helped: 602
HURT: 0
helped stats (abs) min: 2 max: 486 x̄: 6.73 x̃: 6
helped stats (rel) min: 0.02% max: 4.85% x̄: 0.19% x̃: 0.10%
95% mean confidence interval for cycles value: -8.40 -5.05
95% mean confidence interval for cycles %-change: -0.22% -0.16%
Cycles are helped.
Reviewed-by: Elie Tournier <tournier.elie@gmail.com>
2019-02-23 00:47:06 +00:00
|
|
|
(('fadd', a, ('fneg', ('ffract', a))), ('ffloor', a), '!options->lower_ffloor'),
|
2015-09-14 16:13:19 +01:00
|
|
|
(('ffract', a), ('fsub', a, ('ffloor', a)), 'options->lower_ffract'),
|
2018-11-12 17:49:32 +00:00
|
|
|
(('fceil', a), ('fneg', ('ffloor', ('fneg', a))), 'options->lower_fceil'),
|
2019-03-13 00:43:38 +00:00
|
|
|
(('~fadd', ('fmul', a, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b, ('b2f', c))), ('bcsel', c, b, a), 'options->lower_flrp32'),
|
|
|
|
(('~fadd@32', ('fmul', a, ('fadd', 1.0, ('fneg', c ) )), ('fmul', b, c )), ('flrp', a, b, c), '!options->lower_flrp32'),
|
|
|
|
(('~fadd@64', ('fmul', a, ('fadd', 1.0, ('fneg', c ) )), ('fmul', b, c )), ('flrp', a, b, c), '!options->lower_flrp64'),
|
|
|
|
# These are the same as the previous three rules, but it depends on
|
2018-03-28 06:57:07 +01:00
|
|
|
# 1-fsat(x) <=> fsat(1-x). See below.
|
2019-03-13 00:43:38 +00:00
|
|
|
(('~fadd@32', ('fmul', a, ('fsat', ('fadd', 1.0, ('fneg', c )))), ('fmul', b, ('fsat', c))), ('flrp', a, b, ('fsat', c)), '!options->lower_flrp32'),
|
|
|
|
(('~fadd@64', ('fmul', a, ('fsat', ('fadd', 1.0, ('fneg', c )))), ('fmul', b, ('fsat', c))), ('flrp', a, b, ('fsat', c)), '!options->lower_flrp64'),
|
|
|
|
|
2018-12-06 03:36:10 +00:00
|
|
|
(('~fadd', a, ('fmul', ('b2f', 'c@1'), ('fadd', b, ('fneg', a)))), ('bcsel', c, b, a), 'options->lower_flrp32'),
|
2016-04-26 08:35:30 +01:00
|
|
|
(('~fadd@32', a, ('fmul', c , ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp32'),
|
|
|
|
(('~fadd@64', a, ('fmul', c , ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp64'),
|
2015-01-30 21:49:48 +00:00
|
|
|
(('ffma', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma'),
|
2016-05-06 00:58:44 +01:00
|
|
|
(('~fadd', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma'),
|
nir/algebraic: Optimize common array indexing sequence
Some shaders include code that looks like:
uniform int i;
uniform vec4 bones[...];
foo(bones[i * 3], bones[i * 3 + 1], bones[i * 3 + 2]);
CSE would do some work on this:
x = i * 3
foo(bones[x], bones[x + 1], bones[x + 2]);
The compiler may then add '<< 4 + base' to the index calculations.
This results in expressions like
x = i * 3
foo(bones[x << 4], bones[(x + 1) << 4], bones[(x + 2) << 4]);
Just rearranging the math to produce (i * 48) + 16 saves an
instruction, and it allows CSE to do more work.
x = i * 48;
foo(bones[x], bones[x + 16], bones[x + 32]);
So, ~6 instructions becomes ~3.
Some individual shader-db results look pretty bad. However, I have a
really, really hard time believing the change in estimated cycles in,
for example, 3dmmes-taiji/51.shader_test after looking that change in
the generated code.
G45
total instructions in shared programs: 4020840 -> 4010070 (-0.27%)
instructions in affected programs: 177460 -> 166690 (-6.07%)
helped: 894
HURT: 0
total cycles in shared programs: 98829000 -> 98784990 (-0.04%)
cycles in affected programs: 3936648 -> 3892638 (-1.12%)
helped: 894
HURT: 0
Ironlake
total instructions in shared programs: 6418887 -> 6408117 (-0.17%)
instructions in affected programs: 177460 -> 166690 (-6.07%)
helped: 894
HURT: 0
total cycles in shared programs: 143504542 -> 143460532 (-0.03%)
cycles in affected programs: 3936648 -> 3892638 (-1.12%)
helped: 894
HURT: 0
Sandy Bridge
total instructions in shared programs: 8357887 -> 8339251 (-0.22%)
instructions in affected programs: 432715 -> 414079 (-4.31%)
helped: 2795
HURT: 0
total cycles in shared programs: 118284184 -> 118207412 (-0.06%)
cycles in affected programs: 6114626 -> 6037854 (-1.26%)
helped: 2478
HURT: 317
Ivy Bridge
total instructions in shared programs: 7669390 -> 7653822 (-0.20%)
instructions in affected programs: 388234 -> 372666 (-4.01%)
helped: 2795
HURT: 0
total cycles in shared programs: 68381982 -> 68263684 (-0.17%)
cycles in affected programs: 1972658 -> 1854360 (-6.00%)
helped: 2458
HURT: 307
Haswell
total instructions in shared programs: 7082636 -> 7067068 (-0.22%)
instructions in affected programs: 388234 -> 372666 (-4.01%)
helped: 2795
HURT: 0
total cycles in shared programs: 68282020 -> 68164158 (-0.17%)
cycles in affected programs: 1891820 -> 1773958 (-6.23%)
helped: 2459
HURT: 261
Broadwell
total instructions in shared programs: 9002466 -> 8985875 (-0.18%)
instructions in affected programs: 658784 -> 642193 (-2.52%)
helped: 2795
HURT: 5
total cycles in shared programs: 78503092 -> 78450404 (-0.07%)
cycles in affected programs: 2873304 -> 2820616 (-1.83%)
helped: 2275
HURT: 415
Skylake
total instructions in shared programs: 9156978 -> 9140387 (-0.18%)
instructions in affected programs: 682625 -> 666034 (-2.43%)
helped: 2795
HURT: 5
total cycles in shared programs: 75591392 -> 75550574 (-0.05%)
cycles in affected programs: 3192120 -> 3151302 (-1.28%)
helped: 2271
HURT: 425
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2016-07-20 01:47:38 +01:00
|
|
|
|
2018-09-06 04:45:19 +01:00
|
|
|
(('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b, c)), '#d'), '#e'),
|
|
|
|
('bcsel', a, ('fmul', ('fadd', ('fmul', b, c), d), e), ('fmul', d, e))),
|
|
|
|
|
2018-02-15 22:49:55 +00:00
|
|
|
(('fdot4', ('vec4', a, b, c, 1.0), d), ('fdph', ('vec3', a, b, c), d)),
|
|
|
|
(('fdot4', ('vec4', a, 0.0, 0.0, 0.0), b), ('fmul', a, b)),
|
|
|
|
(('fdot4', ('vec4', a, b, 0.0, 0.0), c), ('fdot2', ('vec2', a, b), c)),
|
|
|
|
(('fdot4', ('vec4', a, b, c, 0.0), d), ('fdot3', ('vec3', a, b, c), d)),
|
|
|
|
|
|
|
|
(('fdot3', ('vec3', a, 0.0, 0.0), b), ('fmul', a, b)),
|
|
|
|
(('fdot3', ('vec3', a, b, 0.0), c), ('fdot2', ('vec2', a, b), c)),
|
|
|
|
|
2018-03-28 06:57:07 +01:00
|
|
|
# If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
|
|
|
|
# If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
|
|
|
|
# If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
|
|
|
|
(('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))),
|
|
|
|
|
nir/algebraic: Optimize common array indexing sequence
Some shaders include code that looks like:
uniform int i;
uniform vec4 bones[...];
foo(bones[i * 3], bones[i * 3 + 1], bones[i * 3 + 2]);
CSE would do some work on this:
x = i * 3
foo(bones[x], bones[x + 1], bones[x + 2]);
The compiler may then add '<< 4 + base' to the index calculations.
This results in expressions like
x = i * 3
foo(bones[x << 4], bones[(x + 1) << 4], bones[(x + 2) << 4]);
Just rearranging the math to produce (i * 48) + 16 saves an
instruction, and it allows CSE to do more work.
x = i * 48;
foo(bones[x], bones[x + 16], bones[x + 32]);
So, ~6 instructions becomes ~3.
Some individual shader-db results look pretty bad. However, I have a
really, really hard time believing the change in estimated cycles in,
for example, 3dmmes-taiji/51.shader_test after looking that change in
the generated code.
G45
total instructions in shared programs: 4020840 -> 4010070 (-0.27%)
instructions in affected programs: 177460 -> 166690 (-6.07%)
helped: 894
HURT: 0
total cycles in shared programs: 98829000 -> 98784990 (-0.04%)
cycles in affected programs: 3936648 -> 3892638 (-1.12%)
helped: 894
HURT: 0
Ironlake
total instructions in shared programs: 6418887 -> 6408117 (-0.17%)
instructions in affected programs: 177460 -> 166690 (-6.07%)
helped: 894
HURT: 0
total cycles in shared programs: 143504542 -> 143460532 (-0.03%)
cycles in affected programs: 3936648 -> 3892638 (-1.12%)
helped: 894
HURT: 0
Sandy Bridge
total instructions in shared programs: 8357887 -> 8339251 (-0.22%)
instructions in affected programs: 432715 -> 414079 (-4.31%)
helped: 2795
HURT: 0
total cycles in shared programs: 118284184 -> 118207412 (-0.06%)
cycles in affected programs: 6114626 -> 6037854 (-1.26%)
helped: 2478
HURT: 317
Ivy Bridge
total instructions in shared programs: 7669390 -> 7653822 (-0.20%)
instructions in affected programs: 388234 -> 372666 (-4.01%)
helped: 2795
HURT: 0
total cycles in shared programs: 68381982 -> 68263684 (-0.17%)
cycles in affected programs: 1972658 -> 1854360 (-6.00%)
helped: 2458
HURT: 307
Haswell
total instructions in shared programs: 7082636 -> 7067068 (-0.22%)
instructions in affected programs: 388234 -> 372666 (-4.01%)
helped: 2795
HURT: 0
total cycles in shared programs: 68282020 -> 68164158 (-0.17%)
cycles in affected programs: 1891820 -> 1773958 (-6.23%)
helped: 2459
HURT: 261
Broadwell
total instructions in shared programs: 9002466 -> 8985875 (-0.18%)
instructions in affected programs: 658784 -> 642193 (-2.52%)
helped: 2795
HURT: 5
total cycles in shared programs: 78503092 -> 78450404 (-0.07%)
cycles in affected programs: 2873304 -> 2820616 (-1.83%)
helped: 2275
HURT: 415
Skylake
total instructions in shared programs: 9156978 -> 9140387 (-0.18%)
instructions in affected programs: 682625 -> 666034 (-2.43%)
helped: 2795
HURT: 5
total cycles in shared programs: 75591392 -> 75550574 (-0.05%)
cycles in affected programs: 3192120 -> 3151302 (-1.28%)
helped: 2271
HURT: 425
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2016-07-20 01:47:38 +01:00
|
|
|
# (a * #b + #c) << #d
|
|
|
|
# ((a * #b) << #d) + (#c << #d)
|
|
|
|
# (a * (#b << #d)) + (#c << #d)
|
|
|
|
(('ishl', ('iadd', ('imul', a, '#b'), '#c'), '#d'),
|
|
|
|
('iadd', ('imul', a, ('ishl', b, d)), ('ishl', c, d))),
|
|
|
|
|
|
|
|
# (a * #b) << #c
|
|
|
|
# a * (#b << #c)
|
|
|
|
(('ishl', ('imul', a, '#b'), '#c'), ('imul', a, ('ishl', b, c))),
|
|
|
|
|
2015-01-22 07:36:01 +00:00
|
|
|
# Comparison simplifications
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~inot', ('flt', a, b)), ('fge', a, b)),
|
|
|
|
(('~inot', ('fge', a, b)), ('flt', a, b)),
|
|
|
|
(('~inot', ('feq', a, b)), ('fne', a, b)),
|
|
|
|
(('~inot', ('fne', a, b)), ('feq', a, b)),
|
2015-01-22 07:36:01 +00:00
|
|
|
(('inot', ('ilt', a, b)), ('ige', a, b)),
|
2018-05-30 11:32:24 +01:00
|
|
|
(('inot', ('ult', a, b)), ('uge', a, b)),
|
2015-01-22 07:36:01 +00:00
|
|
|
(('inot', ('ige', a, b)), ('ilt', a, b)),
|
2018-05-30 11:32:24 +01:00
|
|
|
(('inot', ('uge', a, b)), ('ult', a, b)),
|
2015-05-06 06:54:59 +01:00
|
|
|
(('inot', ('ieq', a, b)), ('ine', a, b)),
|
|
|
|
(('inot', ('ine', a, b)), ('ieq', a, b)),
|
2016-03-02 23:18:34 +00:00
|
|
|
|
2019-03-06 22:07:18 +00:00
|
|
|
# This helps some shaders because, after some optimizations, they end up
|
|
|
|
# with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
|
|
|
|
# matching would be handled by CSE.
|
|
|
|
(('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)),
|
|
|
|
(('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)),
|
|
|
|
(('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)),
|
|
|
|
(('fne', ('fneg', a), ('fneg', b)), ('fne', b, a)),
|
|
|
|
(('flt', ('fneg', a), -1.0), ('flt', 1.0, a)),
|
|
|
|
(('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)),
|
|
|
|
(('fge', ('fneg', a), -1.0), ('fge', 1.0, a)),
|
|
|
|
(('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)),
|
|
|
|
(('fne', ('fneg', a), -1.0), ('fne', 1.0, a)),
|
|
|
|
(('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)),
|
|
|
|
|
2019-03-05 22:54:35 +00:00
|
|
|
(('flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)),
|
|
|
|
(('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)),
|
|
|
|
(('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)),
|
|
|
|
(('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)),
|
|
|
|
(('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)),
|
|
|
|
(('fne', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fne', a, b)),
|
|
|
|
|
|
|
|
(('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)),
|
|
|
|
(('flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)),
|
|
|
|
(('fge', 0.0, ('fsat(is_used_once)', a)), ('fge', 0.0, a)),
|
|
|
|
(('flt', 0.0, ('fsat(is_used_once)', a)), ('flt', 0.0, a)),
|
|
|
|
|
2016-03-02 23:18:34 +00:00
|
|
|
# 0.0 >= b2f(a)
|
|
|
|
# b2f(a) <= 0.0
|
|
|
|
# b2f(a) == 0.0 because b2f(a) can only be 0 or 1
|
|
|
|
# inot(a)
|
2018-12-06 03:36:10 +00:00
|
|
|
(('fge', 0.0, ('b2f', 'a@1')), ('inot', a)),
|
|
|
|
|
|
|
|
(('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a)),
|
|
|
|
|
|
|
|
(('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
|
|
|
|
(('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
|
|
|
|
(('fne', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a, b)),
|
|
|
|
(('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a, b)),
|
|
|
|
(('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
|
|
|
|
(('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
|
|
|
|
(('fne', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a, b)),
|
|
|
|
(('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a, b)),
|
|
|
|
(('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a, b)),
|
|
|
|
(('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a, b)),
|
|
|
|
(('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
|
|
|
|
(('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
|
|
|
|
(('feq', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a, b))),
|
|
|
|
(('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
|
|
|
|
(('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
|
|
|
|
(('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
|
|
|
|
(('feq', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a, b))),
|
|
|
|
(('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a, b)),
|
|
|
|
(('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a, b)),
|
|
|
|
(('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a, b)),
|
2018-08-02 01:18:07 +01:00
|
|
|
|
|
|
|
# -(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
|
2018-12-06 03:36:10 +00:00
|
|
|
(('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a, b)),
|
|
|
|
(('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a, b)),
|
2018-08-02 01:18:07 +01:00
|
|
|
|
2018-08-02 03:27:01 +01:00
|
|
|
# -(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)
|
2018-12-06 03:36:10 +00:00
|
|
|
(('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a, b))),
|
|
|
|
(('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
|
2018-08-02 03:27:01 +01:00
|
|
|
|
2018-12-18 06:29:26 +00:00
|
|
|
(('flt', a, ('fneg', a)), ('flt', a, 0.0)),
|
|
|
|
(('fge', a, ('fneg', a)), ('fge', a, 0.0)),
|
|
|
|
|
2018-08-02 01:18:07 +01:00
|
|
|
# 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.
|
2018-12-06 03:36:10 +00:00
|
|
|
(('flt', ('fmin', c, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
|
2018-08-02 01:18:07 +01:00
|
|
|
('ior', ('flt', c, 0.0), ('ior', a, b))),
|
|
|
|
|
2018-02-13 02:58:53 +00:00
|
|
|
(('~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)),
|
|
|
|
(('~fne', ('fadd', a, b), a), ('fne', b, 0.0)),
|
|
|
|
|
|
|
|
# Cannot remove the addition from ilt or ige due to overflow.
|
|
|
|
(('ieq', ('iadd', a, b), a), ('ieq', b, 0)),
|
|
|
|
(('ine', ('iadd', a, b), a), ('ine', b, 0)),
|
|
|
|
|
2018-01-31 19:11:02 +00:00
|
|
|
# fmin(-b2f(a), b) >= 0.0
|
|
|
|
# -b2f(a) >= 0.0 && b >= 0.0
|
|
|
|
# -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
|
|
|
|
# b2f(a) == 0.0 && b >= 0.0
|
|
|
|
# a == False && b >= 0.0
|
|
|
|
# !a && b >= 0.0
|
|
|
|
#
|
|
|
|
# The fge in the second replacement is not a typo. I leave the proof that
|
|
|
|
# "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
|
|
|
|
# reader.
|
2018-12-06 03:36:10 +00:00
|
|
|
(('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
|
|
|
|
(('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
|
2018-01-31 19:11:02 +00:00
|
|
|
|
2018-12-06 03:36:10 +00:00
|
|
|
(('feq', ('b2f', 'a@1'), 0.0), ('inot', a)),
|
|
|
|
(('fne', ('b2f', 'a@1'), 0.0), a),
|
|
|
|
(('ieq', ('b2i', 'a@1'), 0), ('inot', a)),
|
|
|
|
(('ine', ('b2i', 'a@1'), 0), a),
|
2016-03-02 03:05:14 +00:00
|
|
|
|
2018-11-07 23:32:32 +00:00
|
|
|
(('fne', ('u2f', a), 0.0), ('ine', a, 0)),
|
|
|
|
(('feq', ('u2f', a), 0.0), ('ieq', a, 0)),
|
|
|
|
(('fge', ('u2f', a), 0.0), True),
|
|
|
|
(('fge', 0.0, ('u2f', a)), ('uge', 0, a)), # ieq instead?
|
|
|
|
(('flt', ('u2f', a), 0.0), False),
|
|
|
|
(('flt', 0.0, ('u2f', a)), ('ult', 0, a)), # ine instead?
|
|
|
|
(('fne', ('i2f', a), 0.0), ('ine', a, 0)),
|
|
|
|
(('feq', ('i2f', a), 0.0), ('ieq', a, 0)),
|
|
|
|
(('fge', ('i2f', a), 0.0), ('ige', a, 0)),
|
|
|
|
(('fge', 0.0, ('i2f', a)), ('ige', 0, a)),
|
|
|
|
(('flt', ('i2f', a), 0.0), ('ilt', a, 0)),
|
|
|
|
(('flt', 0.0, ('i2f', a)), ('ilt', 0, a)),
|
2018-07-03 20:39:54 +01:00
|
|
|
|
2016-03-02 23:36:14 +00:00
|
|
|
# 0.0 < fabs(a)
|
|
|
|
# fabs(a) > 0.0
|
|
|
|
# fabs(a) != 0.0 because fabs(a) must be >= 0
|
|
|
|
# a != 0.0
|
2018-08-01 18:13:55 +01:00
|
|
|
(('~flt', 0.0, ('fabs', a)), ('fne', a, 0.0)),
|
2016-03-02 23:36:14 +00:00
|
|
|
|
2018-07-27 02:26:18 +01:00
|
|
|
# -fabs(a) < 0.0
|
|
|
|
# fabs(a) > 0.0
|
|
|
|
(('~flt', ('fneg', ('fabs', a)), 0.0), ('fne', a, 0.0)),
|
|
|
|
|
2018-08-02 03:46:26 +01:00
|
|
|
# 0.0 >= fabs(a)
|
|
|
|
# 0.0 == fabs(a) because fabs(a) must be >= 0
|
|
|
|
# 0.0 == a
|
|
|
|
(('fge', 0.0, ('fabs', a)), ('feq', a, 0.0)),
|
|
|
|
|
|
|
|
# -fabs(a) >= 0.0
|
|
|
|
# 0.0 >= fabs(a)
|
|
|
|
(('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
|
|
|
|
|
2018-12-06 03:36:10 +00:00
|
|
|
(('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a, b))),
|
|
|
|
(('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a, b)))),
|
|
|
|
(('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
|
|
|
|
(('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a, b)))),
|
2016-03-08 19:11:00 +00:00
|
|
|
|
2018-02-03 01:39:54 +00:00
|
|
|
# fmin(b2f(a), b)
|
|
|
|
# bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
|
|
|
|
# bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
|
|
|
|
# bcsel(a, fmin(1.0, b), fmin(0.0, b))
|
|
|
|
#
|
|
|
|
# Since b is a constant, constant folding will eliminate the fmin and the
|
|
|
|
# fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
|
2018-12-06 03:36:10 +00:00
|
|
|
(('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a, ('fmin', b, 1.0), ('fmin', b, 0.0))),
|
2018-02-03 01:39:54 +00:00
|
|
|
|
2018-03-14 23:25:07 +00:00
|
|
|
(('flt', ('fadd(is_used_once)', a, ('fneg', b)), 0.0), ('flt', a, b)),
|
2017-01-12 06:01:03 +00:00
|
|
|
|
2014-12-12 19:13:10 +00:00
|
|
|
(('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
|
2018-02-26 22:49:47 +00:00
|
|
|
(('~bcsel', ('flt', b, a), b, a), ('fmin', a, b)),
|
|
|
|
(('~bcsel', ('flt', a, b), b, a), ('fmax', a, b)),
|
2018-01-22 09:29:48 +00:00
|
|
|
(('~bcsel', ('fge', a, b), b, a), ('fmin', a, b)),
|
|
|
|
(('~bcsel', ('fge', b, a), b, a), ('fmax', a, b)),
|
2018-12-04 00:30:44 +00:00
|
|
|
(('bcsel', ('i2b', a), b, c), ('bcsel', ('ine', a, 0), b, c)),
|
2016-08-17 23:03:01 +01:00
|
|
|
(('bcsel', ('inot', a), b, c), ('bcsel', a, c, b)),
|
2015-02-20 09:20:34 +00:00
|
|
|
(('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)),
|
2018-08-01 17:58:19 +01:00
|
|
|
(('bcsel', a, b, ('bcsel', a, c, d)), ('bcsel', a, b, d)),
|
2018-08-02 00:51:35 +01:00
|
|
|
(('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
|
|
|
|
(('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
|
|
|
|
(('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
|
|
|
|
(('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
|
2019-03-30 03:51:20 +00:00
|
|
|
(('bcsel', a, True, b), ('ior', a, b)),
|
2019-03-29 22:12:47 +00:00
|
|
|
(('bcsel', a, a, b), ('ior', a, b)),
|
|
|
|
(('bcsel', a, b, False), ('iand', a, b)),
|
|
|
|
(('bcsel', a, b, a), ('iand', a, b)),
|
2015-05-06 04:20:30 +01:00
|
|
|
(('fmin', a, a), a),
|
|
|
|
(('fmax', a, a), a),
|
|
|
|
(('imin', a, a), a),
|
|
|
|
(('imax', a, a), a),
|
|
|
|
(('umin', a, a), a),
|
|
|
|
(('umax', a, a), a),
|
2018-08-15 10:29:42 +01:00
|
|
|
(('fmax', ('fmax', a, b), b), ('fmax', a, b)),
|
|
|
|
(('umax', ('umax', a, b), b), ('umax', a, b)),
|
|
|
|
(('imax', ('imax', a, b), b), ('imax', a, b)),
|
|
|
|
(('fmin', ('fmin', a, b), b), ('fmin', a, b)),
|
|
|
|
(('umin', ('umin', a, b), b), ('umin', a, b)),
|
|
|
|
(('imin', ('imin', a, b), b), ('imin', a, b)),
|
2018-07-23 07:57:07 +01:00
|
|
|
(('fmax', a, ('fneg', a)), ('fabs', a)),
|
|
|
|
(('imax', a, ('ineg', a)), ('iabs', a)),
|
2017-01-18 11:23:22 +00:00
|
|
|
(('fmin', a, ('fneg', a)), ('fneg', ('fabs', a))),
|
|
|
|
(('imin', a, ('ineg', a)), ('ineg', ('iabs', a))),
|
|
|
|
(('fmin', a, ('fneg', ('fabs', a))), ('fneg', ('fabs', a))),
|
|
|
|
(('imin', a, ('ineg', ('iabs', a))), ('ineg', ('iabs', a))),
|
|
|
|
(('fmin', a, ('fabs', a)), a),
|
|
|
|
(('imin', a, ('iabs', a)), a),
|
|
|
|
(('fmax', a, ('fneg', ('fabs', a))), a),
|
|
|
|
(('imax', a, ('ineg', ('iabs', a))), a),
|
|
|
|
(('fmax', a, ('fabs', a)), ('fabs', a)),
|
|
|
|
(('imax', a, ('iabs', a)), ('iabs', a)),
|
|
|
|
(('fmax', a, ('fneg', a)), ('fabs', a)),
|
|
|
|
(('imax', a, ('ineg', a)), ('iabs', a)),
|
2019-06-04 20:54:12 +01:00
|
|
|
(('~fmax', ('fabs', a), 0.0), ('fabs', a)),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fmin', ('fmax', a, 0.0), 1.0), ('fsat', a), '!options->lower_fsat'),
|
|
|
|
(('~fmax', ('fmin', a, 1.0), 0.0), ('fsat', a), '!options->lower_fsat'),
|
2019-03-05 20:46:11 +00:00
|
|
|
(('~fmin', ('fmax', a, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_negate && !options->lower_fsat'),
|
|
|
|
(('~fmax', ('fmin', a, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_negate && !options->lower_fsat'),
|
2018-09-09 21:40:02 +01:00
|
|
|
(('fsat', ('fsign', a)), ('b2f', ('flt', 0.0, a))),
|
2019-03-01 22:41:59 +00:00
|
|
|
(('fsat', ('b2f', a)), ('b2f', a)),
|
2015-01-30 21:53:39 +00:00
|
|
|
(('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'),
|
2015-01-30 22:10:58 +00:00
|
|
|
(('fsat', ('fsat', a)), ('fsat', a)),
|
2018-10-23 22:30:41 +01:00
|
|
|
(('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a, b))), ('fsat', ('fadd', ('fneg', a), ('fneg', b))), '!options->lower_negate && !options->lower_fsat'),
|
|
|
|
(('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fneg', a), b)), '!options->lower_negate && !options->lower_fsat'),
|
|
|
|
(('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fabs', a), ('fabs', b))), '!options->lower_fsat'),
|
2016-04-09 02:31:16 +01:00
|
|
|
(('fmin', ('fmax', ('fmin', ('fmax', a, b), c), b), c), ('fmin', ('fmax', a, b), c)),
|
|
|
|
(('imin', ('imax', ('imin', ('imax', a, b), c), b), c), ('imin', ('imax', a, b), c)),
|
|
|
|
(('umin', ('umax', ('umin', ('umax', a, b), c), b), c), ('umin', ('umax', a, b), c)),
|
2016-11-30 01:33:30 +00:00
|
|
|
(('fmax', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a, b))),
|
|
|
|
(('fmin', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a, b))),
|
2016-04-09 02:31:16 +01:00
|
|
|
(('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)),
|
2018-01-04 21:30:49 +00:00
|
|
|
(('~ior', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))),
|
|
|
|
(('~ior', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)),
|
|
|
|
(('~ior', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))),
|
|
|
|
(('~ior', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)),
|
|
|
|
(('~ior', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmax', b, c))),
|
|
|
|
(('~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)),
|
2018-01-04 23:21:30 +00:00
|
|
|
(('~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)),
|
2018-01-04 21:30:49 +00:00
|
|
|
|
2018-01-05 21:20:46 +00:00
|
|
|
(('ior', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imax', b, c))),
|
|
|
|
(('ior', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imin', a, b), c)),
|
|
|
|
(('ior', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imin', b, c))),
|
|
|
|
(('ior', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imax', a, b), c)),
|
|
|
|
(('ior', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umax', b, c))),
|
|
|
|
(('ior', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umin', a, b), c)),
|
|
|
|
(('ior', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umin', b, c))),
|
|
|
|
(('ior', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umax', a, b), c)),
|
|
|
|
(('iand', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imin', b, c))),
|
|
|
|
(('iand', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imax', a, b), c)),
|
|
|
|
(('iand', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imax', b, c))),
|
|
|
|
(('iand', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imin', a, b), c)),
|
|
|
|
(('iand', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umin', b, c))),
|
|
|
|
(('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)),
|
|
|
|
|
2019-02-12 00:47:34 +00:00
|
|
|
# Common pattern like 'if (i == 0 || i == 1 || ...)'
|
|
|
|
(('ior', ('ieq', a, 0), ('ieq', a, 1)), ('uge', 1, a)),
|
|
|
|
(('ior', ('uge', 1, a), ('ieq', a, 2)), ('uge', 2, a)),
|
|
|
|
(('ior', ('uge', 2, a), ('ieq', a, 3)), ('uge', 3, a)),
|
|
|
|
|
2018-02-22 02:30:20 +00:00
|
|
|
# The (i2f32, ...) part is an open-coded fsign. When that is combined with
|
|
|
|
# the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
|
|
|
|
# so emit an open-coded version of that.
|
|
|
|
(('bcsel@32', ('feq', a, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
|
|
|
|
('ior', 0x3f800000, ('iand', a, 0x80000000))),
|
|
|
|
|
2019-03-30 03:51:20 +00:00
|
|
|
(('ior', a, ('ieq', a, False)), True),
|
2018-10-19 18:43:43 +01:00
|
|
|
(('ior', a, ('inot', a)), -1),
|
2018-07-16 06:19:29 +01:00
|
|
|
|
2018-12-18 05:34:11 +00:00
|
|
|
(('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a, b)),
|
|
|
|
(('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a, b))),
|
|
|
|
|
2018-05-30 09:48:31 +01:00
|
|
|
(('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
|
|
|
|
|
2018-01-05 21:29:26 +00:00
|
|
|
# These patterns can result when (a < b || a < c) => (a < min(b, c))
|
|
|
|
# transformations occur before constant propagation and loop-unrolling.
|
|
|
|
(('~flt', a, ('fmax', b, a)), ('flt', a, b)),
|
|
|
|
(('~flt', ('fmin', a, b), a), ('flt', b, a)),
|
|
|
|
(('~fge', a, ('fmin', b, a)), True),
|
|
|
|
(('~fge', ('fmax', a, b), a), True),
|
|
|
|
(('~flt', a, ('fmin', b, a)), False),
|
|
|
|
(('~flt', ('fmax', a, b), a), False),
|
nir: Add some missing "optimization undo" patterns
d8d18516b0a and 03fb13f6467 added some patterns to undo conversions like
(('ior', ('flt', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c)))
If further optimization cause some of the operands to either be the same
or be constants, undoing the transformation can lead to further savings.
I don't know why these patterns were not added in those patches. I did
not check to see which specific patterns actually helped. I just added
all of them for symmetry. This prevents some loop unrolling regressions
Plane Shift caused by Samuel's "nir: implement the GLSL equivalent of if
simplication in nir_opt_if" patch.
Skylake and Broadwell had similar results. (Skylake shown)
total instructions in shared programs: 14369768 -> 14369557 (<.01%)
instructions in affected programs: 44076 -> 43865 (-0.48%)
helped: 141
HURT: 0
helped stats (abs) min: 1 max: 5 x̄: 1.50 x̃: 1
helped stats (rel) min: 0.07% max: 1.52% x̄: 0.66% x̃: 0.60%
95% mean confidence interval for instructions value: -1.67 -1.32
95% mean confidence interval for instructions %-change: -0.72% -0.59%
Instructions are helped.
total cycles in shared programs: 532430629 -> 532425772 (<.01%)
cycles in affected programs: 1170832 -> 1165975 (-0.41%)
helped: 101
HURT: 5
helped stats (abs) min: 1 max: 160 x̄: 48.54 x̃: 32
helped stats (rel) min: <.01% max: 8.49% x̄: 2.76% x̃: 2.03%
HURT stats (abs) min: 2 max: 22 x̄: 9.20 x̃: 4
HURT stats (rel) min: <.01% max: 0.05% x̄: 0.02% x̃: <.01%
95% mean confidence interval for cycles value: -53.64 -38.00
95% mean confidence interval for cycles %-change: -3.06% -2.20%
Cycles are helped.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
2018-05-31 02:06:57 +01:00
|
|
|
(('~fge', a, ('fmax', b, a)), ('fge', a, b)),
|
|
|
|
(('~fge', ('fmin', a, b), a), ('fge', b, a)),
|
2018-01-05 21:29:26 +00:00
|
|
|
|
2018-01-05 21:20:46 +00:00
|
|
|
(('ilt', a, ('imax', b, a)), ('ilt', a, b)),
|
|
|
|
(('ilt', ('imin', a, b), a), ('ilt', b, a)),
|
|
|
|
(('ige', a, ('imin', b, a)), True),
|
|
|
|
(('ige', ('imax', a, b), a), True),
|
|
|
|
(('ult', a, ('umax', b, a)), ('ult', a, b)),
|
|
|
|
(('ult', ('umin', a, b), a), ('ult', b, a)),
|
|
|
|
(('uge', a, ('umin', b, a)), True),
|
|
|
|
(('uge', ('umax', a, b), a), True),
|
nir: Add some missing "optimization undo" patterns
d8d18516b0a and 03fb13f6467 added some patterns to undo conversions like
(('ior', ('flt', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c)))
If further optimization cause some of the operands to either be the same
or be constants, undoing the transformation can lead to further savings.
I don't know why these patterns were not added in those patches. I did
not check to see which specific patterns actually helped. I just added
all of them for symmetry. This prevents some loop unrolling regressions
Plane Shift caused by Samuel's "nir: implement the GLSL equivalent of if
simplication in nir_opt_if" patch.
Skylake and Broadwell had similar results. (Skylake shown)
total instructions in shared programs: 14369768 -> 14369557 (<.01%)
instructions in affected programs: 44076 -> 43865 (-0.48%)
helped: 141
HURT: 0
helped stats (abs) min: 1 max: 5 x̄: 1.50 x̃: 1
helped stats (rel) min: 0.07% max: 1.52% x̄: 0.66% x̃: 0.60%
95% mean confidence interval for instructions value: -1.67 -1.32
95% mean confidence interval for instructions %-change: -0.72% -0.59%
Instructions are helped.
total cycles in shared programs: 532430629 -> 532425772 (<.01%)
cycles in affected programs: 1170832 -> 1165975 (-0.41%)
helped: 101
HURT: 5
helped stats (abs) min: 1 max: 160 x̄: 48.54 x̃: 32
helped stats (rel) min: <.01% max: 8.49% x̄: 2.76% x̃: 2.03%
HURT stats (abs) min: 2 max: 22 x̄: 9.20 x̃: 4
HURT stats (rel) min: <.01% max: 0.05% x̄: 0.02% x̃: <.01%
95% mean confidence interval for cycles value: -53.64 -38.00
95% mean confidence interval for cycles %-change: -3.06% -2.20%
Cycles are helped.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
2018-05-31 02:06:57 +01:00
|
|
|
(('ilt', a, ('imin', b, a)), False),
|
|
|
|
(('ilt', ('imax', a, b), a), False),
|
|
|
|
(('ige', a, ('imax', b, a)), ('ige', a, b)),
|
|
|
|
(('ige', ('imin', a, b), a), ('ige', b, a)),
|
|
|
|
(('ult', a, ('umin', b, a)), False),
|
|
|
|
(('ult', ('umax', a, b), a), False),
|
|
|
|
(('uge', a, ('umax', b, a)), ('uge', a, b)),
|
|
|
|
(('uge', ('umin', a, b), a), ('uge', b, a)),
|
2019-01-25 11:48:44 +00:00
|
|
|
(('ult', a, ('iand', b, a)), False),
|
|
|
|
(('ult', ('ior', a, b), a), False),
|
|
|
|
(('uge', a, ('iand', b, a)), True),
|
|
|
|
(('uge', ('ior', a, b), a), True),
|
2018-01-05 21:20:46 +00:00
|
|
|
|
nir: Add some missing "optimization undo" patterns
d8d18516b0a and 03fb13f6467 added some patterns to undo conversions like
(('ior', ('flt', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c)))
If further optimization cause some of the operands to either be the same
or be constants, undoing the transformation can lead to further savings.
I don't know why these patterns were not added in those patches. I did
not check to see which specific patterns actually helped. I just added
all of them for symmetry. This prevents some loop unrolling regressions
Plane Shift caused by Samuel's "nir: implement the GLSL equivalent of if
simplication in nir_opt_if" patch.
Skylake and Broadwell had similar results. (Skylake shown)
total instructions in shared programs: 14369768 -> 14369557 (<.01%)
instructions in affected programs: 44076 -> 43865 (-0.48%)
helped: 141
HURT: 0
helped stats (abs) min: 1 max: 5 x̄: 1.50 x̃: 1
helped stats (rel) min: 0.07% max: 1.52% x̄: 0.66% x̃: 0.60%
95% mean confidence interval for instructions value: -1.67 -1.32
95% mean confidence interval for instructions %-change: -0.72% -0.59%
Instructions are helped.
total cycles in shared programs: 532430629 -> 532425772 (<.01%)
cycles in affected programs: 1170832 -> 1165975 (-0.41%)
helped: 101
HURT: 5
helped stats (abs) min: 1 max: 160 x̄: 48.54 x̃: 32
helped stats (rel) min: <.01% max: 8.49% x̄: 2.76% x̃: 2.03%
HURT stats (abs) min: 2 max: 22 x̄: 9.20 x̃: 4
HURT stats (rel) min: <.01% max: 0.05% x̄: 0.02% x̃: <.01%
95% mean confidence interval for cycles value: -53.64 -38.00
95% mean confidence interval for cycles %-change: -3.06% -2.20%
Cycles are helped.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
2018-05-31 02:06:57 +01:00
|
|
|
(('ilt', '#a', ('imax', '#b', c)), ('ior', ('ilt', a, b), ('ilt', a, c))),
|
|
|
|
(('ilt', ('imin', '#a', b), '#c'), ('ior', ('ilt', a, c), ('ilt', b, c))),
|
|
|
|
(('ige', '#a', ('imin', '#b', c)), ('ior', ('ige', a, b), ('ige', a, c))),
|
|
|
|
(('ige', ('imax', '#a', b), '#c'), ('ior', ('ige', a, c), ('ige', b, c))),
|
|
|
|
(('ult', '#a', ('umax', '#b', c)), ('ior', ('ult', a, b), ('ult', a, c))),
|
|
|
|
(('ult', ('umin', '#a', b), '#c'), ('ior', ('ult', a, c), ('ult', b, c))),
|
|
|
|
(('uge', '#a', ('umin', '#b', c)), ('ior', ('uge', a, b), ('uge', a, c))),
|
|
|
|
(('uge', ('umax', '#a', b), '#c'), ('ior', ('uge', a, c), ('uge', b, c))),
|
2018-01-05 21:20:46 +00:00
|
|
|
(('ilt', '#a', ('imin', '#b', c)), ('iand', ('ilt', a, b), ('ilt', a, c))),
|
|
|
|
(('ilt', ('imax', '#a', b), '#c'), ('iand', ('ilt', a, c), ('ilt', b, c))),
|
|
|
|
(('ige', '#a', ('imax', '#b', c)), ('iand', ('ige', a, b), ('ige', a, c))),
|
|
|
|
(('ige', ('imin', '#a', b), '#c'), ('iand', ('ige', a, c), ('ige', b, c))),
|
|
|
|
(('ult', '#a', ('umin', '#b', c)), ('iand', ('ult', a, b), ('ult', a, c))),
|
|
|
|
(('ult', ('umax', '#a', b), '#c'), ('iand', ('ult', a, c), ('ult', b, c))),
|
|
|
|
(('uge', '#a', ('umax', '#b', c)), ('iand', ('uge', a, b), ('uge', a, c))),
|
|
|
|
(('uge', ('umin', '#a', b), '#c'), ('iand', ('uge', a, c), ('uge', b, c))),
|
|
|
|
|
2019-03-10 15:06:56 +00:00
|
|
|
# Thanks to sign extension, the ishr(a, b) is negative if and only if a is
|
|
|
|
# negative.
|
|
|
|
(('bcsel', ('ilt', a, 0), ('ineg', ('ishr', a, b)), ('ishr', a, b)),
|
|
|
|
('iabs', ('ishr', a, b))),
|
|
|
|
(('iabs', ('ishr', ('iabs', a), b)), ('ishr', ('iabs', a), b)),
|
|
|
|
|
2016-03-02 21:47:56 +00:00
|
|
|
(('fabs', ('slt', a, b)), ('slt', a, b)),
|
|
|
|
(('fabs', ('sge', a, b)), ('sge', a, b)),
|
|
|
|
(('fabs', ('seq', a, b)), ('seq', a, b)),
|
|
|
|
(('fabs', ('sne', a, b)), ('sne', a, b)),
|
2015-03-31 16:25:19 +01:00
|
|
|
(('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'),
|
|
|
|
(('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'),
|
|
|
|
(('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'),
|
|
|
|
(('sne', a, b), ('b2f', ('fne', a, b)), 'options->lower_scmp'),
|
2015-08-06 12:36:04 +01:00
|
|
|
(('fne', ('fneg', a), a), ('fne', a, 0.0)),
|
2015-08-06 12:36:05 +01:00
|
|
|
(('feq', ('fneg', a), a), ('feq', a, 0.0)),
|
2015-03-18 06:30:32 +00:00
|
|
|
# Emulating booleans
|
2018-12-06 03:36:10 +00:00
|
|
|
(('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))),
|
|
|
|
(('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
|
|
|
|
(('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a, b))),
|
2018-12-06 03:32:11 +00:00
|
|
|
(('iand', 'a@bool32', 1.0), ('b2f', a)),
|
2016-08-09 01:05:18 +01:00
|
|
|
# True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
|
2018-11-07 19:43:40 +00:00
|
|
|
(('ineg', ('b2i32', 'a@32')), a),
|
2018-12-06 03:36:10 +00:00
|
|
|
(('flt', ('fneg', ('b2f', 'a@1')), 0), a), # Generated by TGSI KILL_IF.
|
|
|
|
(('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a), # Generated by TGSI KILL_IF.
|
2015-02-07 01:12:59 +00:00
|
|
|
# Comparison with the same args. Note that these are not done for
|
|
|
|
# the float versions because NaN always returns false on float
|
|
|
|
# inequalities.
|
|
|
|
(('ilt', a, a), False),
|
|
|
|
(('ige', a, a), True),
|
|
|
|
(('ieq', a, a), True),
|
|
|
|
(('ine', a, a), False),
|
|
|
|
(('ult', a, a), False),
|
|
|
|
(('uge', a, a), True),
|
2015-01-19 22:57:38 +00:00
|
|
|
# Logical and bit operations
|
|
|
|
(('fand', a, 0.0), 0.0),
|
|
|
|
(('iand', a, a), a),
|
2015-07-31 17:12:48 +01:00
|
|
|
(('iand', a, ~0), a),
|
2015-01-19 22:57:38 +00:00
|
|
|
(('iand', a, 0), 0),
|
|
|
|
(('ior', a, a), a),
|
|
|
|
(('ior', a, 0), a),
|
2016-07-06 23:00:45 +01:00
|
|
|
(('ior', a, True), True),
|
2015-01-19 22:57:38 +00:00
|
|
|
(('fxor', a, a), 0.0),
|
|
|
|
(('ixor', a, a), 0),
|
2016-02-02 00:34:46 +00:00
|
|
|
(('ixor', a, 0), a),
|
2015-01-19 22:57:38 +00:00
|
|
|
(('inot', ('inot', a)), a),
|
2018-07-23 08:35:02 +01:00
|
|
|
(('ior', ('iand', a, b), b), b),
|
|
|
|
(('ior', ('ior', a, b), b), ('ior', a, b)),
|
|
|
|
(('iand', ('ior', a, b), b), b),
|
|
|
|
(('iand', ('iand', a, b), b), ('iand', a, b)),
|
2015-01-19 22:57:38 +00:00
|
|
|
# DeMorgan's Laws
|
|
|
|
(('iand', ('inot', a), ('inot', b)), ('inot', ('ior', a, b))),
|
|
|
|
(('ior', ('inot', a), ('inot', b)), ('inot', ('iand', a, b))),
|
2015-01-22 07:25:56 +00:00
|
|
|
# Shift optimizations
|
|
|
|
(('ishl', 0, a), 0),
|
|
|
|
(('ishl', a, 0), a),
|
|
|
|
(('ishr', 0, a), 0),
|
|
|
|
(('ishr', a, 0), a),
|
|
|
|
(('ushr', 0, a), 0),
|
2015-04-14 16:40:22 +01:00
|
|
|
(('ushr', a, 0), a),
|
2017-08-01 02:26:49 +01:00
|
|
|
(('iand', 0xff, ('ushr@32', a, 24)), ('ushr', a, 24)),
|
|
|
|
(('iand', 0xffff, ('ushr@32', a, 16)), ('ushr', a, 16)),
|
2015-01-22 07:47:06 +00:00
|
|
|
# Exponential/logarithmic identities
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a
|
|
|
|
(('~flog2', ('fexp2', a)), a), # lg2(2^a) = a
|
2015-01-28 18:32:57 +00:00
|
|
|
(('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
|
|
|
|
(('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2', c), d))),
|
|
|
|
('~fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
|
2018-02-05 14:08:03 +00:00
|
|
|
(('~fexp2', ('fmul', ('flog2', a), 2.0)), ('fmul', a, a)),
|
|
|
|
(('~fexp2', ('fmul', ('flog2', a), 4.0)), ('fmul', ('fmul', a, a), ('fmul', a, a))),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fpow', a, 1.0), a),
|
|
|
|
(('~fpow', a, 2.0), ('fmul', a, a)),
|
|
|
|
(('~fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))),
|
|
|
|
(('~fpow', 2.0, a), ('fexp2', a)),
|
|
|
|
(('~fpow', ('fpow', a, 2.2), 0.454545), a),
|
|
|
|
(('~fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)),
|
|
|
|
(('~fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))),
|
|
|
|
(('~frcp', ('fexp2', a)), ('fexp2', ('fneg', a))),
|
|
|
|
(('~frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))),
|
|
|
|
(('~flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))),
|
|
|
|
(('~flog2', ('frcp', a)), ('fneg', ('flog2', a))),
|
|
|
|
(('~flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))),
|
|
|
|
(('~flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))),
|
2018-02-05 15:07:45 +00:00
|
|
|
(('~fmul', ('fexp2(is_used_once)', a), ('fexp2(is_used_once)', b)), ('fexp2', ('fadd', a, b))),
|
2018-10-25 03:48:49 +01:00
|
|
|
(('bcsel', ('flt', a, 0.0), 0.0, ('fsqrt', a)), ('fsqrt', ('fmax', a, 0.0))),
|
2015-01-22 07:52:17 +00:00
|
|
|
# Division and reciprocal
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fdiv', 1.0, a), ('frcp', a)),
|
nir: Add a lower_fdiv option, turn fdiv into fmul/frcp.
The nir_opt_algebraic rule
(('fadd', ('flog2', a), ('fneg', ('flog2', b))), ('flog2', ('fdiv', a, b))),
can produce new fdiv operations, which need to be lowered on i965,
as we don't actually implement fdiv. (Normally, we handle this in
GLSL IR's lower_instructions pass, but in the above case we introduce
an fdiv after that point. So, make NIR do it for us.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Cc: mesa-stable@lists.freedesktop.org
2016-01-05 13:09:46 +00:00
|
|
|
(('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~frcp', ('frcp', a)), a),
|
|
|
|
(('~frcp', ('fsqrt', a)), ('frsq', a)),
|
2015-01-28 18:39:29 +00:00
|
|
|
(('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
|
2015-01-29 00:55:03 +00:00
|
|
|
# Boolean simplifications
|
2018-12-04 00:30:44 +00:00
|
|
|
(('i2b32(is_used_by_if)', a), ('ine32', a, 0)),
|
|
|
|
(('i2b1(is_used_by_if)', a), ('ine', a, 0)),
|
2019-03-30 03:51:20 +00:00
|
|
|
(('ieq', a, True), a),
|
|
|
|
(('ine(is_not_used_by_if)', a, True), ('inot', a)),
|
|
|
|
(('ine', a, False), a),
|
|
|
|
(('ieq(is_not_used_by_if)', a, False), ('inot', 'a')),
|
2016-08-17 23:03:01 +01:00
|
|
|
(('bcsel', a, True, False), a),
|
|
|
|
(('bcsel', a, False, True), ('inot', a)),
|
2016-08-09 09:44:38 +01:00
|
|
|
(('bcsel@32', a, 1.0, 0.0), ('b2f', a)),
|
|
|
|
(('bcsel@32', a, 0.0, 1.0), ('b2f', ('inot', a))),
|
|
|
|
(('bcsel@32', a, -1.0, -0.0), ('fneg', ('b2f', a))),
|
|
|
|
(('bcsel@32', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a)))),
|
2015-01-29 00:53:51 +00:00
|
|
|
(('bcsel', True, b, c), b),
|
|
|
|
(('bcsel', False, b, c), c),
|
2018-11-09 00:08:37 +00:00
|
|
|
(('bcsel', a, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a, b, c))),
|
2015-01-19 22:57:38 +00:00
|
|
|
|
2015-01-29 23:50:18 +00:00
|
|
|
(('bcsel', a, b, b), b),
|
|
|
|
(('fcsel', a, b, b), b),
|
|
|
|
|
2018-10-19 22:58:36 +01:00
|
|
|
# D3D Boolean emulation
|
|
|
|
(('bcsel', a, -1, 0), ('ineg', ('b2i', 'a@1'))),
|
|
|
|
(('bcsel', a, 0, -1), ('ineg', ('b2i', ('inot', a)))),
|
|
|
|
(('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
|
|
|
|
('ineg', ('b2i', ('iand', a, b)))),
|
|
|
|
(('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
|
|
|
|
('ineg', ('b2i', ('ior', a, b)))),
|
|
|
|
(('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a)),
|
|
|
|
(('ieq', ('ineg', ('b2i', 'a@1')), -1), a),
|
|
|
|
(('ine', ('ineg', ('b2i', 'a@1')), 0), a),
|
|
|
|
(('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a)),
|
2018-12-18 21:28:22 +00:00
|
|
|
(('iand', ('ineg', ('b2i', a)), 1.0), ('b2f', a)),
|
2018-10-19 22:58:36 +01:00
|
|
|
|
2019-02-22 21:05:07 +00:00
|
|
|
# SM5 32-bit shifts are defined to use the 5 least significant bits
|
|
|
|
(('ishl', 'a@32', ('iand', 31, b)), ('ishl', a, b)),
|
|
|
|
(('ishr', 'a@32', ('iand', 31, b)), ('ishr', a, b)),
|
|
|
|
(('ushr', 'a@32', ('iand', 31, b)), ('ushr', a, b)),
|
|
|
|
|
2015-03-26 17:09:42 +00:00
|
|
|
# Conversions
|
2018-11-07 19:43:40 +00:00
|
|
|
(('i2b32', ('b2i', 'a@32')), a),
|
2018-11-07 23:32:32 +00:00
|
|
|
(('f2i', ('ftrunc', a)), ('f2i', a)),
|
|
|
|
(('f2u', ('ftrunc', a)), ('f2u', a)),
|
2016-03-02 02:59:57 +00:00
|
|
|
(('i2b', ('ineg', a)), ('i2b', a)),
|
|
|
|
(('i2b', ('iabs', a)), ('i2b', a)),
|
2016-03-02 21:46:50 +00:00
|
|
|
(('fabs', ('b2f', a)), ('b2f', a)),
|
|
|
|
(('iabs', ('b2i', a)), ('b2i', a)),
|
2018-12-06 03:36:10 +00:00
|
|
|
(('inot', ('f2b1', a)), ('feq', a, 0.0)),
|
2015-03-26 17:09:42 +00:00
|
|
|
|
2018-07-03 21:57:06 +01:00
|
|
|
# Ironically, mark these as imprecise because removing the conversions may
|
|
|
|
# preserve more precision than doing the conversions (e.g.,
|
|
|
|
# uint(float(0x81818181u)) == 0x81818200).
|
2018-11-07 23:32:32 +00:00
|
|
|
(('~f2i32', ('i2f', 'a@32')), a),
|
|
|
|
(('~f2i32', ('u2f', 'a@32')), a),
|
|
|
|
(('~f2u32', ('i2f', 'a@32')), a),
|
|
|
|
(('~f2u32', ('u2f', 'a@32')), a),
|
2018-07-03 21:57:06 +01:00
|
|
|
|
2019-02-13 03:59:39 +00:00
|
|
|
# Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
|
|
|
|
# says:
|
|
|
|
#
|
|
|
|
# It is undefined to convert a negative floating-point value to an
|
|
|
|
# uint.
|
|
|
|
#
|
|
|
|
# Assuming that (uint)some_float behaves like (uint)(int)some_float allows
|
|
|
|
# some optimizations in the i965 backend to proceed.
|
|
|
|
(('ige', ('f2u', a), b), ('ige', ('f2i', a), b)),
|
|
|
|
(('ige', b, ('f2u', a)), ('ige', b, ('f2i', a))),
|
|
|
|
(('ilt', ('f2u', a), b), ('ilt', ('f2i', a), b)),
|
|
|
|
(('ilt', b, ('f2u', a)), ('ilt', b, ('f2i', a))),
|
|
|
|
|
2019-03-05 20:46:11 +00:00
|
|
|
(('~fmin', ('fabs', a), 1.0), ('fsat', ('fabs', a)), '!options->lower_fsat'),
|
|
|
|
|
2018-10-25 04:13:45 +01:00
|
|
|
# The result of the multiply must be in [-1, 0], so the result of the ffma
|
|
|
|
# must be in [0, 1].
|
|
|
|
(('flt', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), False),
|
|
|
|
(('flt', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), False),
|
|
|
|
(('fmax', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0)),
|
|
|
|
(('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0)),
|
|
|
|
|
2017-02-15 06:22:36 +00:00
|
|
|
# Packing and then unpacking does nothing
|
|
|
|
(('unpack_64_2x32_split_x', ('pack_64_2x32_split', a, b)), a),
|
|
|
|
(('unpack_64_2x32_split_y', ('pack_64_2x32_split', a, b)), b),
|
|
|
|
(('pack_64_2x32_split', ('unpack_64_2x32_split_x', a),
|
|
|
|
('unpack_64_2x32_split_y', a)), a),
|
|
|
|
|
nir: Recognize open-coded extract_u8.
Two shaders that appear in Unigine benchmarks (Heaven and Valley) unpack
three bytes from an integer and convert each into a float:
float((val >> 16u) & 0xffu)
float((val >> 8u) & 0xffu)
float((val >> 0u) & 0xffu)
Instead of shifting, masking, and type converting like this:
shr(8) g15<1>UD g25<8,8,1>UD 0x00000010UD
and(8) g16<1>UD g15<8,8,1>UD 0x000000ffUD
mov(8) g17<1>F g16<8,8,1>UD
shr(8) g18<1>UD g25<8,8,1>UD 0x00000008UD
and(8) g19<1>UD g18<8,8,1>UD 0x000000ffUD
mov(8) g20<1>F g19<8,8,1>UD
and(8) g21<1>UD g25<8,8,1>UD 0x000000ffUD
mov(8) g22<1>F g21<8,8,1>UD
i965 can simply extract a byte and convert to float in a single
instruction:
mov(8) g17<1>F g25.2<32,8,4>UB
mov(8) g20<1>F g25.1<32,8,4>UB
mov(8) g22<1>F g25.0<32,8,4>UB
This patch implements the first step: recognizing byte extraction. A
later patch will optimize out the conversion to float.
instructions in affected programs: 28568 -> 27450 (-3.91%)
helped: 7
cycles in affected programs: 210076 -> 203144 (-3.30%)
helped: 7
This patch decreases the number of instructions in the two Unigine
programs by:
#1721: 4520 -> 4374 instructions (-3.23%)
#1706: 3752 -> 3582 instructions (-4.53%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-21 17:09:48 +00:00
|
|
|
# Byte extraction
|
2019-02-28 04:15:32 +00:00
|
|
|
(('ushr', 'a@16', 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
|
2018-08-14 21:10:22 +01:00
|
|
|
(('ushr', 'a@32', 24), ('extract_u8', a, 3), '!options->lower_extract_byte'),
|
2019-02-28 04:12:46 +00:00
|
|
|
(('ushr', 'a@64', 56), ('extract_u8', a, 7), '!options->lower_extract_byte'),
|
2019-02-28 04:15:32 +00:00
|
|
|
(('ishr', 'a@16', 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
|
2018-08-14 20:37:39 +01:00
|
|
|
(('ishr', 'a@32', 24), ('extract_i8', a, 3), '!options->lower_extract_byte'),
|
2019-02-28 04:12:46 +00:00
|
|
|
(('ishr', 'a@64', 56), ('extract_i8', a, 7), '!options->lower_extract_byte'),
|
2019-02-28 03:52:12 +00:00
|
|
|
(('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte')
|
|
|
|
]
|
|
|
|
|
2019-02-28 04:08:38 +00:00
|
|
|
# After the ('extract_u8', a, 0) pattern, above, triggers, there will be
|
|
|
|
# patterns like those below.
|
2019-02-28 03:52:12 +00:00
|
|
|
for op in ('ushr', 'ishr'):
|
|
|
|
optimizations.extend([(('extract_u8', (op, 'a@16', 8), 0), ('extract_u8', a, 1))])
|
|
|
|
optimizations.extend([(('extract_u8', (op, 'a@32', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 4)])
|
|
|
|
optimizations.extend([(('extract_u8', (op, 'a@64', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 8)])
|
|
|
|
|
|
|
|
optimizations.extend([(('extract_u8', ('extract_u16', a, 1), 0), ('extract_u8', a, 2))])
|
nir: Recognize open-coded extract_u8.
Two shaders that appear in Unigine benchmarks (Heaven and Valley) unpack
three bytes from an integer and convert each into a float:
float((val >> 16u) & 0xffu)
float((val >> 8u) & 0xffu)
float((val >> 0u) & 0xffu)
Instead of shifting, masking, and type converting like this:
shr(8) g15<1>UD g25<8,8,1>UD 0x00000010UD
and(8) g16<1>UD g15<8,8,1>UD 0x000000ffUD
mov(8) g17<1>F g16<8,8,1>UD
shr(8) g18<1>UD g25<8,8,1>UD 0x00000008UD
and(8) g19<1>UD g18<8,8,1>UD 0x000000ffUD
mov(8) g20<1>F g19<8,8,1>UD
and(8) g21<1>UD g25<8,8,1>UD 0x000000ffUD
mov(8) g22<1>F g21<8,8,1>UD
i965 can simply extract a byte and convert to float in a single
instruction:
mov(8) g17<1>F g25.2<32,8,4>UB
mov(8) g20<1>F g25.1<32,8,4>UB
mov(8) g22<1>F g25.0<32,8,4>UB
This patch implements the first step: recognizing byte extraction. A
later patch will optimize out the conversion to float.
instructions in affected programs: 28568 -> 27450 (-3.91%)
helped: 7
cycles in affected programs: 210076 -> 203144 (-3.30%)
helped: 7
This patch decreases the number of instructions in the two Unigine
programs by:
#1721: 4520 -> 4374 instructions (-3.23%)
#1706: 3752 -> 3582 instructions (-4.53%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-21 17:09:48 +00:00
|
|
|
|
2019-02-28 04:08:38 +00:00
|
|
|
# After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
|
|
|
|
# patterns like those below.
|
2019-02-28 03:52:12 +00:00
|
|
|
for op in ('extract_u8', 'extract_i8'):
|
2019-02-28 04:15:32 +00:00
|
|
|
optimizations.extend([((op, ('ishl', 'a@16', 8), 1), (op, a, 0))])
|
2019-02-28 03:52:12 +00:00
|
|
|
optimizations.extend([((op, ('ishl', 'a@32', 24 - 8 * i), 3), (op, a, i)) for i in range(2, -1, -1)])
|
2019-02-28 04:12:46 +00:00
|
|
|
optimizations.extend([((op, ('ishl', 'a@64', 56 - 8 * i), 7), (op, a, i)) for i in range(6, -1, -1)])
|
2019-02-28 03:52:12 +00:00
|
|
|
|
|
|
|
optimizations.extend([
|
2016-02-03 22:28:31 +00:00
|
|
|
# Word extraction
|
2018-08-14 20:37:39 +01:00
|
|
|
(('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
|
2018-08-14 21:10:22 +01:00
|
|
|
(('ushr', 'a@32', 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
|
2018-08-14 20:37:39 +01:00
|
|
|
(('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a, 0), '!options->lower_extract_word'),
|
|
|
|
(('ishr', 'a@32', 16), ('extract_i16', a, 1), '!options->lower_extract_word'),
|
2016-02-03 22:28:31 +00:00
|
|
|
(('iand', 0xffff, a), ('extract_u16', a, 0), '!options->lower_extract_word'),
|
|
|
|
|
2015-01-28 18:52:53 +00:00
|
|
|
# Subtracts
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fsub', a, ('fsub', 0.0, b)), ('fadd', a, b)),
|
2015-02-20 08:00:27 +00:00
|
|
|
(('isub', a, ('isub', 0, b)), ('iadd', a, b)),
|
2015-08-19 06:38:34 +01:00
|
|
|
(('ussub_4x8', a, 0), a),
|
|
|
|
(('ussub_4x8', a, ~0), 0),
|
2015-04-04 13:13:44 +01:00
|
|
|
(('fsub', a, b), ('fadd', a, ('fneg', b)), 'options->lower_sub'),
|
|
|
|
(('isub', a, b), ('iadd', a, ('ineg', b)), 'options->lower_sub'),
|
2015-01-28 18:52:53 +00:00
|
|
|
(('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'),
|
|
|
|
(('ineg', a), ('isub', 0, a), 'options->lower_negate'),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~fadd', a, ('fsub', 0.0, b)), ('fsub', a, b)),
|
2015-01-28 18:52:53 +00:00
|
|
|
(('iadd', a, ('isub', 0, b)), ('isub', a, b)),
|
|
|
|
(('fabs', ('fsub', 0.0, a)), ('fabs', a)),
|
|
|
|
(('iabs', ('isub', 0, a)), ('iabs', a)),
|
2016-01-06 23:30:38 +00:00
|
|
|
|
2016-02-22 23:16:49 +00:00
|
|
|
# Propagate negation up multiplication chains
|
2018-10-23 22:30:41 +01:00
|
|
|
(('fmul(is_used_by_non_fsat)', ('fneg', a), b), ('fneg', ('fmul', a, b))),
|
2016-03-31 03:18:16 +01:00
|
|
|
(('imul', ('ineg', a), b), ('ineg', ('imul', a, b))),
|
2016-02-22 23:16:49 +00:00
|
|
|
|
2017-01-12 02:10:55 +00:00
|
|
|
# Propagate constants up multiplication chains
|
|
|
|
(('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a, c), b)),
|
|
|
|
(('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a, c), b)),
|
|
|
|
(('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a, c), b)),
|
|
|
|
(('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a, c), b)),
|
|
|
|
|
2016-04-07 06:34:39 +01:00
|
|
|
# Reassociate constants in add/mul chains so they can be folded together.
|
2017-01-12 02:10:55 +00:00
|
|
|
# For now, we mostly only handle cases where the constants are separated by
|
2016-04-07 06:34:39 +01:00
|
|
|
# a single non-constant. We could do better eventually.
|
2019-02-15 22:52:20 +00:00
|
|
|
(('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a, c), b)),
|
|
|
|
(('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a, c), b)),
|
|
|
|
(('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a, c), b)),
|
|
|
|
(('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a, ('fneg', c)), ('fneg', b))),
|
|
|
|
(('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a, c), b)),
|
2016-04-07 06:34:39 +01:00
|
|
|
|
2017-06-30 23:48:19 +01:00
|
|
|
# By definition...
|
|
|
|
(('bcsel', ('ige', ('find_lsb', a), 0), ('find_lsb', a), -1), ('find_lsb', a)),
|
|
|
|
(('bcsel', ('ige', ('ifind_msb', a), 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
|
|
|
|
(('bcsel', ('ige', ('ufind_msb', a), 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
|
|
|
|
|
|
|
|
(('bcsel', ('ine', a, 0), ('find_lsb', a), -1), ('find_lsb', a)),
|
|
|
|
(('bcsel', ('ine', a, 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
|
|
|
|
(('bcsel', ('ine', a, 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
|
|
|
|
|
|
|
|
(('bcsel', ('ine', a, -1), ('ifind_msb', a), -1), ('ifind_msb', a)),
|
|
|
|
|
2016-01-06 23:30:38 +00:00
|
|
|
# Misc. lowering
|
2019-06-03 21:18:55 +01:00
|
|
|
(('fmod@16', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod'),
|
|
|
|
(('fmod@32', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod'),
|
|
|
|
(('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod'),
|
2016-04-26 04:59:06 +01:00
|
|
|
(('uadd_carry@32', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'),
|
|
|
|
(('usub_borrow@32', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'),
|
2016-01-13 19:08:37 +00:00
|
|
|
|
|
|
|
(('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
|
2019-05-02 14:28:59 +01:00
|
|
|
('bcsel', ('ult', 31, 'bits'), 'insert',
|
2016-01-13 19:08:37 +00:00
|
|
|
('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
|
|
|
|
'options->lower_bitfield_insert'),
|
2018-07-12 14:02:27 +01:00
|
|
|
(('ihadd', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
|
|
|
|
(('uhadd', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
|
|
|
|
(('irhadd', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
|
|
|
|
(('urhadd', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
|
|
|
|
(('uadd_sat', a, b), ('bcsel', ('ult', ('iadd', a, b), a), -1, ('iadd', a, b)), 'options->lower_add_sat'),
|
|
|
|
(('usub_sat', a, b), ('bcsel', ('ult', a, b), 0, ('isub', a, b)), 'options->lower_add_sat'),
|
2016-01-13 19:09:11 +00:00
|
|
|
|
2018-05-02 22:13:23 +01:00
|
|
|
# Alternative lowering that doesn't rely on bfi.
|
|
|
|
(('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
|
2019-01-26 08:12:46 +00:00
|
|
|
('ior',
|
|
|
|
('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
|
|
|
|
('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
|
2018-05-02 22:13:23 +01:00
|
|
|
'options->lower_bitfield_insert_to_shifts'),
|
|
|
|
|
2016-01-13 19:09:11 +00:00
|
|
|
(('ibitfield_extract', 'value', 'offset', 'bits'),
|
2019-05-02 14:28:59 +01:00
|
|
|
('bcsel', ('ult', 31, 'bits'), 'value',
|
2016-01-13 19:09:11 +00:00
|
|
|
('ibfe', 'value', 'offset', 'bits')),
|
|
|
|
'options->lower_bitfield_extract'),
|
|
|
|
|
|
|
|
(('ubitfield_extract', 'value', 'offset', 'bits'),
|
|
|
|
('bcsel', ('ult', 31, 'bits'), 'value',
|
|
|
|
('ubfe', 'value', 'offset', 'bits')),
|
|
|
|
'options->lower_bitfield_extract'),
|
2016-01-21 17:09:29 +00:00
|
|
|
|
2019-01-25 12:56:49 +00:00
|
|
|
# Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
|
|
|
|
(('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
|
|
|
|
(('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
|
|
|
|
(('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
|
|
|
|
(('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
|
|
|
|
(('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
|
|
|
|
(('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
|
|
|
|
|
2018-05-02 21:02:21 +01:00
|
|
|
(('ibitfield_extract', 'value', 'offset', 'bits'),
|
|
|
|
('bcsel', ('ieq', 0, 'bits'),
|
|
|
|
0,
|
|
|
|
('ishr',
|
|
|
|
('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
|
|
|
|
('isub', 32, 'bits'))),
|
|
|
|
'options->lower_bitfield_extract_to_shifts'),
|
|
|
|
|
|
|
|
(('ubitfield_extract', 'value', 'offset', 'bits'),
|
|
|
|
('iand',
|
|
|
|
('ushr', 'value', 'offset'),
|
|
|
|
('bcsel', ('ieq', 'bits', 32),
|
|
|
|
0xffffffff,
|
2019-01-26 08:12:46 +00:00
|
|
|
('isub', ('ishl', 1, 'bits'), 1))),
|
2018-05-02 21:02:21 +01:00
|
|
|
'options->lower_bitfield_extract_to_shifts'),
|
|
|
|
|
2018-05-04 21:33:47 +01:00
|
|
|
(('ifind_msb', 'value'),
|
|
|
|
('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
|
|
|
|
'options->lower_ifind_msb'),
|
|
|
|
|
2018-05-04 22:02:55 +01:00
|
|
|
(('find_lsb', 'value'),
|
|
|
|
('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
|
|
|
|
'options->lower_find_lsb'),
|
|
|
|
|
nir: Shift count for shift opcodes is always 32-bits
Previously both sources were unsized. This caused problems when the
thing being shifted was 64-bit but the shift count was 32-bit. The
expectation in NIR is that all unsized sources (and destination) will
ultimately have the same size.
The changes in nir_opt_algebraic.py are to prevent errors like:
Failed to parse transformation:
03:12:25 (('extract_i8', 'a', 'b'), ('ishr', ('ishl', 'a', ('imul', ('isub', 3, 'b'), 8)), 24), 'options->lower_extract_byte')
03:12:25 Traceback (most recent call last):
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 610, in __init__
03:12:25 xform = SearchAndReplace(xform)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 495, in __init__
03:12:25 BitSizeValidator(varset).validate(self.search, self.replace)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 311, in validate
03:12:25 validate_dst_class = self._validate_bit_class_up(replace)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 414, in _validate_bit_class_up
03:12:25 src_class = self._validate_bit_class_up(val.sources[i])
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 420, in _validate_bit_class_up
03:12:25 assert src_class == src_type_bits
03:12:25 AssertionError
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Suggested-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Cc: Jason Ekstrand <jason@jlekstrand.net>
2016-10-27 11:02:58 +01:00
|
|
|
(('extract_i8', a, 'b@32'),
|
2016-01-21 17:09:29 +00:00
|
|
|
('ishr', ('ishl', a, ('imul', ('isub', 3, b), 8)), 24),
|
|
|
|
'options->lower_extract_byte'),
|
|
|
|
|
nir: Shift count for shift opcodes is always 32-bits
Previously both sources were unsized. This caused problems when the
thing being shifted was 64-bit but the shift count was 32-bit. The
expectation in NIR is that all unsized sources (and destination) will
ultimately have the same size.
The changes in nir_opt_algebraic.py are to prevent errors like:
Failed to parse transformation:
03:12:25 (('extract_i8', 'a', 'b'), ('ishr', ('ishl', 'a', ('imul', ('isub', 3, 'b'), 8)), 24), 'options->lower_extract_byte')
03:12:25 Traceback (most recent call last):
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 610, in __init__
03:12:25 xform = SearchAndReplace(xform)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 495, in __init__
03:12:25 BitSizeValidator(varset).validate(self.search, self.replace)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 311, in validate
03:12:25 validate_dst_class = self._validate_bit_class_up(replace)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 414, in _validate_bit_class_up
03:12:25 src_class = self._validate_bit_class_up(val.sources[i])
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 420, in _validate_bit_class_up
03:12:25 assert src_class == src_type_bits
03:12:25 AssertionError
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Suggested-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Cc: Jason Ekstrand <jason@jlekstrand.net>
2016-10-27 11:02:58 +01:00
|
|
|
(('extract_u8', a, 'b@32'),
|
2016-01-21 17:09:29 +00:00
|
|
|
('iand', ('ushr', a, ('imul', b, 8)), 0xff),
|
|
|
|
'options->lower_extract_byte'),
|
|
|
|
|
nir: Shift count for shift opcodes is always 32-bits
Previously both sources were unsized. This caused problems when the
thing being shifted was 64-bit but the shift count was 32-bit. The
expectation in NIR is that all unsized sources (and destination) will
ultimately have the same size.
The changes in nir_opt_algebraic.py are to prevent errors like:
Failed to parse transformation:
03:12:25 (('extract_i8', 'a', 'b'), ('ishr', ('ishl', 'a', ('imul', ('isub', 3, 'b'), 8)), 24), 'options->lower_extract_byte')
03:12:25 Traceback (most recent call last):
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 610, in __init__
03:12:25 xform = SearchAndReplace(xform)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 495, in __init__
03:12:25 BitSizeValidator(varset).validate(self.search, self.replace)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 311, in validate
03:12:25 validate_dst_class = self._validate_bit_class_up(replace)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 414, in _validate_bit_class_up
03:12:25 src_class = self._validate_bit_class_up(val.sources[i])
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 420, in _validate_bit_class_up
03:12:25 assert src_class == src_type_bits
03:12:25 AssertionError
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Suggested-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Cc: Jason Ekstrand <jason@jlekstrand.net>
2016-10-27 11:02:58 +01:00
|
|
|
(('extract_i16', a, 'b@32'),
|
2016-01-21 17:09:29 +00:00
|
|
|
('ishr', ('ishl', a, ('imul', ('isub', 1, b), 16)), 16),
|
|
|
|
'options->lower_extract_word'),
|
|
|
|
|
nir: Shift count for shift opcodes is always 32-bits
Previously both sources were unsized. This caused problems when the
thing being shifted was 64-bit but the shift count was 32-bit. The
expectation in NIR is that all unsized sources (and destination) will
ultimately have the same size.
The changes in nir_opt_algebraic.py are to prevent errors like:
Failed to parse transformation:
03:12:25 (('extract_i8', 'a', 'b'), ('ishr', ('ishl', 'a', ('imul', ('isub', 3, 'b'), 8)), 24), 'options->lower_extract_byte')
03:12:25 Traceback (most recent call last):
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 610, in __init__
03:12:25 xform = SearchAndReplace(xform)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 495, in __init__
03:12:25 BitSizeValidator(varset).validate(self.search, self.replace)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 311, in validate
03:12:25 validate_dst_class = self._validate_bit_class_up(replace)
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 414, in _validate_bit_class_up
03:12:25 src_class = self._validate_bit_class_up(val.sources[i])
03:12:25 File "/home/jenkins/workspace/Leeroy_2/repos/mesa/src/compiler/nir/nir_algebraic.py", line 420, in _validate_bit_class_up
03:12:25 assert src_class == src_type_bits
03:12:25 AssertionError
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Suggested-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Cc: Jason Ekstrand <jason@jlekstrand.net>
2016-10-27 11:02:58 +01:00
|
|
|
(('extract_u16', a, 'b@32'),
|
2016-01-21 17:09:29 +00:00
|
|
|
('iand', ('ushr', a, ('imul', b, 16)), 0xffff),
|
|
|
|
'options->lower_extract_word'),
|
2016-01-25 19:05:52 +00:00
|
|
|
|
|
|
|
(('pack_unorm_2x16', 'v'),
|
|
|
|
('pack_uvec2_to_uint',
|
2017-03-08 03:54:37 +00:00
|
|
|
('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
|
2016-01-25 19:05:52 +00:00
|
|
|
'options->lower_pack_unorm_2x16'),
|
|
|
|
|
|
|
|
(('pack_unorm_4x8', 'v'),
|
|
|
|
('pack_uvec4_to_uint',
|
2017-03-08 03:54:37 +00:00
|
|
|
('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
|
2016-01-25 19:05:52 +00:00
|
|
|
'options->lower_pack_unorm_4x8'),
|
|
|
|
|
|
|
|
(('pack_snorm_2x16', 'v'),
|
|
|
|
('pack_uvec2_to_uint',
|
2017-03-08 03:54:37 +00:00
|
|
|
('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
|
2016-01-25 19:05:52 +00:00
|
|
|
'options->lower_pack_snorm_2x16'),
|
|
|
|
|
|
|
|
(('pack_snorm_4x8', 'v'),
|
|
|
|
('pack_uvec4_to_uint',
|
2017-03-08 03:54:37 +00:00
|
|
|
('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
|
2016-01-25 19:05:52 +00:00
|
|
|
'options->lower_pack_snorm_4x8'),
|
2016-01-25 19:07:02 +00:00
|
|
|
|
|
|
|
(('unpack_unorm_2x16', 'v'),
|
2017-03-08 03:54:37 +00:00
|
|
|
('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
|
2018-11-07 23:32:32 +00:00
|
|
|
('extract_u16', 'v', 1))),
|
2016-01-25 19:07:02 +00:00
|
|
|
65535.0),
|
|
|
|
'options->lower_unpack_unorm_2x16'),
|
|
|
|
|
|
|
|
(('unpack_unorm_4x8', 'v'),
|
2017-03-08 03:54:37 +00:00
|
|
|
('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
|
2018-11-07 23:32:32 +00:00
|
|
|
('extract_u8', 'v', 1),
|
|
|
|
('extract_u8', 'v', 2),
|
|
|
|
('extract_u8', 'v', 3))),
|
2016-01-25 19:07:02 +00:00
|
|
|
255.0),
|
|
|
|
'options->lower_unpack_unorm_4x8'),
|
|
|
|
|
|
|
|
(('unpack_snorm_2x16', 'v'),
|
2018-11-07 23:32:32 +00:00
|
|
|
('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
|
|
|
|
('extract_i16', 'v', 1))),
|
2016-01-25 19:07:02 +00:00
|
|
|
32767.0))),
|
|
|
|
'options->lower_unpack_snorm_2x16'),
|
|
|
|
|
|
|
|
(('unpack_snorm_4x8', 'v'),
|
2018-11-07 23:32:32 +00:00
|
|
|
('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
|
|
|
|
('extract_i8', 'v', 1),
|
|
|
|
('extract_i8', 'v', 2),
|
|
|
|
('extract_i8', 'v', 3))),
|
2016-01-25 19:07:02 +00:00
|
|
|
127.0))),
|
|
|
|
'options->lower_unpack_snorm_4x8'),
|
2019-02-06 21:12:25 +00:00
|
|
|
|
|
|
|
(('isign', a), ('imin', ('imax', a, -1), 1), 'options->lower_isign'),
|
2019-04-16 21:49:41 +01:00
|
|
|
(('fsign', a), ('fsub', ('b2f', ('flt', 0.0, a)), ('b2f', ('flt', a, 0.0))), 'options->lower_fsign'),
|
2019-02-28 03:52:12 +00:00
|
|
|
])
|
2014-12-12 19:13:10 +00:00
|
|
|
|
2018-07-12 14:02:27 +01:00
|
|
|
# bit_size dependent lowerings
|
|
|
|
for bit_size in [8, 16, 32, 64]:
|
|
|
|
# convenience constants
|
|
|
|
intmax = (1 << (bit_size - 1)) - 1
|
|
|
|
intmin = 1 << (bit_size - 1)
|
|
|
|
|
|
|
|
optimizations += [
|
|
|
|
(('iadd_sat@' + str(bit_size), a, b),
|
|
|
|
('bcsel', ('ige', b, 1), ('bcsel', ('ilt', ('iadd', a, b), a), intmax, ('iadd', a, b)),
|
|
|
|
('bcsel', ('ilt', a, ('iadd', a, b)), intmin, ('iadd', a, b))), 'options->lower_add_sat'),
|
|
|
|
(('isub_sat@' + str(bit_size), a, b),
|
|
|
|
('bcsel', ('ilt', b, 0), ('bcsel', ('ilt', ('isub', a, b), a), intmax, ('isub', a, b)),
|
|
|
|
('bcsel', ('ilt', a, ('isub', a, b)), intmin, ('isub', a, b))), 'options->lower_add_sat'),
|
|
|
|
]
|
|
|
|
|
2018-06-27 11:37:38 +01:00
|
|
|
invert = OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
|
2018-02-07 01:27:53 +00:00
|
|
|
|
2018-08-09 09:27:19 +01:00
|
|
|
for left, right in itertools.combinations_with_replacement(invert.keys(), 2):
|
2018-02-07 01:27:53 +00:00
|
|
|
optimizations.append((('inot', ('ior(is_used_once)', (left, a, b), (right, c, d))),
|
|
|
|
('iand', (invert[left], a, b), (invert[right], c, d))))
|
|
|
|
optimizations.append((('inot', ('iand(is_used_once)', (left, a, b), (right, c, d))),
|
|
|
|
('ior', (invert[left], a, b), (invert[right], c, d))))
|
|
|
|
|
2018-12-06 18:56:33 +00:00
|
|
|
# Optimize x2bN(b2x(x)) -> x
|
|
|
|
for size in type_sizes('bool'):
|
|
|
|
aN = 'a@' + str(size)
|
|
|
|
f2bN = 'f2b' + str(size)
|
|
|
|
i2bN = 'i2b' + str(size)
|
|
|
|
optimizations.append(((f2bN, ('b2f', aN)), a))
|
|
|
|
optimizations.append(((i2bN, ('b2i', aN)), a))
|
|
|
|
|
2018-10-11 04:36:52 +01:00
|
|
|
# Optimize x2yN(b2x(x)) -> b2y
|
|
|
|
for x, y in itertools.product(['f', 'u', 'i'], ['f', 'u', 'i']):
|
|
|
|
if x != 'f' and y != 'f' and x != y:
|
|
|
|
continue
|
|
|
|
|
|
|
|
b2x = 'b2f' if x == 'f' else 'b2i'
|
|
|
|
b2y = 'b2f' if y == 'f' else 'b2i'
|
2018-11-07 22:25:31 +00:00
|
|
|
x2yN = '{}2{}'.format(x, y)
|
|
|
|
optimizations.append(((x2yN, (b2x, a)), (b2y, a)))
|
2018-10-11 04:36:52 +01:00
|
|
|
|
2019-04-23 00:03:04 +01:00
|
|
|
# Optimize away x2xN(a@N)
|
|
|
|
for t in ['int', 'uint', 'float']:
|
|
|
|
for N in type_sizes(t):
|
|
|
|
x2xN = '{0}2{0}{1}'.format(t[0], N)
|
|
|
|
aN = 'a@{0}'.format(N)
|
|
|
|
optimizations.append(((x2xN, aN), a))
|
|
|
|
|
|
|
|
# Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
|
|
|
|
# In particular, we can optimize away everything except upcast of downcast and
|
|
|
|
# upcasts where the type differs from the other cast
|
|
|
|
for N, M in itertools.product(type_sizes('uint'), type_sizes('uint')):
|
|
|
|
if N < M:
|
|
|
|
# The outer cast is a down-cast. It doesn't matter what the size of the
|
|
|
|
# argument of the inner cast is because we'll never been in the upcast
|
|
|
|
# of downcast case. Regardless of types, we'll always end up with y2yN
|
|
|
|
# in the end.
|
|
|
|
for x, y in itertools.product(['i', 'u'], ['i', 'u']):
|
|
|
|
x2xN = '{0}2{0}{1}'.format(x, N)
|
|
|
|
y2yM = '{0}2{0}{1}'.format(y, M)
|
|
|
|
y2yN = '{0}2{0}{1}'.format(y, N)
|
|
|
|
optimizations.append(((x2xN, (y2yM, a)), (y2yN, a)))
|
|
|
|
elif N > M:
|
|
|
|
# If the outer cast is an up-cast, we have to be more careful about the
|
|
|
|
# size of the argument of the inner cast and with types. In this case,
|
|
|
|
# the type is always the type of type up-cast which is given by the
|
|
|
|
# outer cast.
|
|
|
|
for P in type_sizes('uint'):
|
|
|
|
# We can't optimize away up-cast of down-cast.
|
|
|
|
if M < P:
|
|
|
|
continue
|
|
|
|
|
|
|
|
# Because we're doing down-cast of down-cast, the types always have
|
|
|
|
# to match between the two casts
|
|
|
|
for x in ['i', 'u']:
|
|
|
|
x2xN = '{0}2{0}{1}'.format(x, N)
|
|
|
|
x2xM = '{0}2{0}{1}'.format(x, M)
|
|
|
|
aP = 'a@{0}'.format(P)
|
|
|
|
optimizations.append(((x2xN, (x2xM, aP)), (x2xN, a)))
|
|
|
|
else:
|
|
|
|
# The N == M case is handled by other optimizations
|
|
|
|
pass
|
|
|
|
|
2016-04-27 19:18:04 +01:00
|
|
|
def fexp2i(exp, bits):
|
|
|
|
# We assume that exp is already in the right range.
|
2018-05-24 08:54:45 +01:00
|
|
|
if bits == 16:
|
|
|
|
return ('i2i16', ('ishl', ('iadd', exp, 15), 10))
|
|
|
|
elif bits == 32:
|
2016-04-27 19:18:04 +01:00
|
|
|
return ('ishl', ('iadd', exp, 127), 23)
|
|
|
|
elif bits == 64:
|
2017-02-15 06:15:16 +00:00
|
|
|
return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20))
|
2016-04-27 19:18:04 +01:00
|
|
|
else:
|
|
|
|
assert False
|
|
|
|
|
|
|
|
def ldexp(f, exp, bits):
|
2016-03-25 19:12:38 +00:00
|
|
|
# First, we clamp exp to a reasonable range. The maximum possible range
|
|
|
|
# for a normal exponent is [-126, 127] and, throwing in denormals, you get
|
|
|
|
# a maximum range of [-149, 127]. This means that we can potentially have
|
|
|
|
# a swing of +-276. If you start with FLT_MAX, you actually have to do
|
|
|
|
# ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
|
|
|
|
# spec, on the other hand, only requires that we handle an exponent value
|
|
|
|
# in the range [-126, 128]. This implementation is *mostly* correct; it
|
|
|
|
# handles a range on exp of [-252, 254] which allows you to create any
|
|
|
|
# value (including denorms if the hardware supports it) and to adjust the
|
|
|
|
# exponent of any normal value to anything you want.
|
2018-05-24 08:54:45 +01:00
|
|
|
if bits == 16:
|
|
|
|
exp = ('imin', ('imax', exp, -28), 30)
|
|
|
|
elif bits == 32:
|
2016-04-27 19:18:04 +01:00
|
|
|
exp = ('imin', ('imax', exp, -252), 254)
|
|
|
|
elif bits == 64:
|
|
|
|
exp = ('imin', ('imax', exp, -2044), 2046)
|
|
|
|
else:
|
|
|
|
assert False
|
2016-03-25 19:12:38 +00:00
|
|
|
|
|
|
|
# Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
|
|
|
|
# (We use ishr which isn't the same for -1, but the -1 case still works
|
|
|
|
# since we use exp-exp/2 as the second exponent.) While the spec
|
|
|
|
# technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
|
|
|
|
# work with denormals and doesn't allow for the full swing in exponents
|
|
|
|
# that you can get with normalized values. Instead, we create two powers
|
|
|
|
# of two and multiply by them each in turn. That way the effective range
|
|
|
|
# of our exponent is doubled.
|
2016-04-27 19:18:04 +01:00
|
|
|
pow2_1 = fexp2i(('ishr', exp, 1), bits)
|
|
|
|
pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits)
|
2016-03-25 19:12:38 +00:00
|
|
|
return ('fmul', ('fmul', f, pow2_1), pow2_2)
|
|
|
|
|
2016-04-27 19:18:04 +01:00
|
|
|
optimizations += [
|
2018-05-24 08:54:45 +01:00
|
|
|
(('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
|
2018-02-27 08:19:21 +00:00
|
|
|
(('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
|
|
|
|
(('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
|
2016-04-27 19:18:04 +01:00
|
|
|
]
|
2016-03-25 19:12:38 +00:00
|
|
|
|
2016-02-02 00:44:18 +00:00
|
|
|
# Unreal Engine 4 demo applications open-codes bitfieldReverse()
|
|
|
|
def bitfield_reverse(u):
|
|
|
|
step1 = ('ior', ('ishl', u, 16), ('ushr', u, 16))
|
|
|
|
step2 = ('ior', ('ishl', ('iand', step1, 0x00ff00ff), 8), ('ushr', ('iand', step1, 0xff00ff00), 8))
|
|
|
|
step3 = ('ior', ('ishl', ('iand', step2, 0x0f0f0f0f), 4), ('ushr', ('iand', step2, 0xf0f0f0f0), 4))
|
|
|
|
step4 = ('ior', ('ishl', ('iand', step3, 0x33333333), 2), ('ushr', ('iand', step3, 0xcccccccc), 2))
|
|
|
|
step5 = ('ior', ('ishl', ('iand', step4, 0x55555555), 1), ('ushr', ('iand', step4, 0xaaaaaaaa), 1))
|
|
|
|
|
|
|
|
return step5
|
|
|
|
|
2016-04-26 04:59:06 +01:00
|
|
|
optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
|
2016-02-02 00:44:18 +00:00
|
|
|
|
2016-12-20 05:29:51 +00:00
|
|
|
# For any float comparison operation, "cmp", if you have "a == a && a cmp b"
|
|
|
|
# then the "a == a" is redundant because it's equivalent to "a is not NaN"
|
|
|
|
# and, if a is a NaN then the second comparison will fail anyway.
|
|
|
|
for op in ['flt', 'fge', 'feq']:
|
|
|
|
optimizations += [
|
|
|
|
(('iand', ('feq', a, a), (op, a, b)), (op, a, b)),
|
|
|
|
(('iand', ('feq', a, a), (op, b, a)), (op, b, a)),
|
|
|
|
]
|
2016-02-02 00:44:18 +00:00
|
|
|
|
2015-01-29 00:53:51 +00:00
|
|
|
# Add optimizations to handle the case where the result of a ternary is
|
|
|
|
# compared to a constant. This way we can take things like
|
|
|
|
#
|
|
|
|
# (a ? 0 : 1) > 0
|
|
|
|
#
|
|
|
|
# and turn it into
|
|
|
|
#
|
|
|
|
# a ? (0 > 0) : (1 > 0)
|
|
|
|
#
|
|
|
|
# which constant folding will eat for lunch. The resulting ternary will
|
|
|
|
# further get cleaned up by the boolean reductions above and we will be
|
|
|
|
# left with just the original variable "a".
|
|
|
|
for op in ['flt', 'fge', 'feq', 'fne',
|
|
|
|
'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
|
|
|
|
optimizations += [
|
|
|
|
((op, ('bcsel', 'a', '#b', '#c'), '#d'),
|
|
|
|
('bcsel', 'a', (op, 'b', 'd'), (op, 'c', 'd'))),
|
|
|
|
((op, '#d', ('bcsel', a, '#b', '#c')),
|
|
|
|
('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))),
|
|
|
|
]
|
|
|
|
|
2016-03-02 23:39:09 +00:00
|
|
|
|
|
|
|
# For example, this converts things like
|
|
|
|
#
|
|
|
|
# 1 + mix(0, a - 1, condition)
|
|
|
|
#
|
|
|
|
# into
|
|
|
|
#
|
|
|
|
# mix(1, (a-1)+1, condition)
|
|
|
|
#
|
|
|
|
# Other optimizations will rearrange the constants.
|
|
|
|
for op in ['fadd', 'fmul', 'iadd', 'imul']:
|
|
|
|
optimizations += [
|
|
|
|
((op, ('bcsel(is_used_once)', a, '#b', c), '#d'), ('bcsel', a, (op, b, d), (op, c, d)))
|
|
|
|
]
|
|
|
|
|
2019-03-27 22:07:20 +00:00
|
|
|
# For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
|
|
|
|
# states:
|
|
|
|
#
|
|
|
|
# If neither layout qualifier is specified, derivatives in compute shaders
|
|
|
|
# return zero, which is consistent with the handling of built-in texture
|
|
|
|
# functions like texture() in GLSL 4.50 compute shaders.
|
|
|
|
for op in ['fddx', 'fddx_fine', 'fddx_coarse',
|
|
|
|
'fddy', 'fddy_fine', 'fddy_coarse']:
|
|
|
|
optimizations += [
|
|
|
|
((op, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
|
|
|
|
]
|
|
|
|
|
2019-05-12 23:09:38 +01:00
|
|
|
# Some optimizations for ir3-specific instructions.
|
|
|
|
optimizations += [
|
|
|
|
# 'al * bl': If either 'al' or 'bl' is zero, return zero.
|
|
|
|
(('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
|
|
|
|
# '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
|
2019-06-10 20:38:39 +01:00
|
|
|
(('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
|
|
|
|
(('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
|
2019-05-12 23:09:38 +01:00
|
|
|
]
|
|
|
|
|
2017-01-13 06:25:11 +00:00
|
|
|
# This section contains "late" optimizations that should be run before
|
|
|
|
# creating ffmas and calling regular optimizations for the final time.
|
|
|
|
# Optimizations should go here if they help code generation and conflict
|
|
|
|
# with the regular optimizations.
|
|
|
|
before_ffma_optimizations = [
|
|
|
|
# Propagate constants down multiplication chains
|
|
|
|
(('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a, c), b)),
|
|
|
|
(('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a, c), b)),
|
|
|
|
(('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a, c), b)),
|
|
|
|
(('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a, c), b)),
|
|
|
|
|
|
|
|
(('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
|
|
|
|
(('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
|
|
|
|
(('~fadd', ('fneg', a), a), 0.0),
|
|
|
|
(('iadd', ('ineg', a), a), 0),
|
|
|
|
(('iadd', ('ineg', a), ('iadd', a, b)), b),
|
|
|
|
(('iadd', a, ('iadd', ('ineg', a), b)), b),
|
|
|
|
(('~fadd', ('fneg', a), ('fadd', a, b)), b),
|
|
|
|
(('~fadd', a, ('fadd', ('fneg', a), b)), b),
|
2018-08-23 03:15:15 +01:00
|
|
|
|
|
|
|
(('~flrp@32', ('fadd(is_used_once)', a, -1.0), ('fadd(is_used_once)', a, 1.0), d), ('fadd', ('flrp', -1.0, 1.0, d), a)),
|
|
|
|
(('~flrp@32', ('fadd(is_used_once)', a, 1.0), ('fadd(is_used_once)', a, -1.0), d), ('fadd', ('flrp', 1.0, -1.0, d), a)),
|
|
|
|
(('~flrp@32', ('fadd(is_used_once)', a, '#b'), ('fadd(is_used_once)', a, '#c'), d), ('fadd', ('fmul', d, ('fadd', c, ('fneg', b))), ('fadd', a, b))),
|
2017-01-13 06:25:11 +00:00
|
|
|
]
|
|
|
|
|
2015-03-24 00:11:49 +00:00
|
|
|
# This section contains "late" optimizations that should be run after the
|
|
|
|
# regular optimizations have finished. Optimizations should go here if
|
|
|
|
# they help code generation but do not necessarily produce code that is
|
|
|
|
# more easily optimizable.
|
|
|
|
late_optimizations = [
|
2016-03-17 18:31:48 +00:00
|
|
|
# Most of these optimizations aren't quite safe when you get infinity or
|
|
|
|
# Nan involved but the first one should be fine.
|
2018-01-09 23:32:47 +00:00
|
|
|
(('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)),
|
2016-03-17 18:31:48 +00:00
|
|
|
(('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))),
|
|
|
|
(('~fne', ('fadd', a, b), 0.0), ('fne', a, ('fneg', b))),
|
|
|
|
|
2019-03-05 20:08:29 +00:00
|
|
|
# nir_lower_to_source_mods will collapse this, but its existence during the
|
|
|
|
# optimization loop can prevent other optimizations.
|
|
|
|
(('fneg', ('fneg', a)), a),
|
|
|
|
|
2019-03-05 22:54:35 +00:00
|
|
|
# These are duplicated from the main optimizations table. The late
|
|
|
|
# patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
|
|
|
|
# new patterns like these. The patterns that compare with zero are removed
|
|
|
|
# because they are unlikely to be created in by anything in
|
|
|
|
# late_optimizations.
|
|
|
|
(('flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)),
|
|
|
|
(('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)),
|
|
|
|
(('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)),
|
|
|
|
(('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)),
|
|
|
|
(('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)),
|
|
|
|
(('fne', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fne', a, b)),
|
|
|
|
|
|
|
|
(('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)),
|
|
|
|
(('flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)),
|
|
|
|
|
2018-01-11 22:14:25 +00:00
|
|
|
(('~fge', ('fmin(is_used_once)', ('fadd(is_used_once)', a, b), ('fadd', c, d)), 0.0), ('iand', ('fge', a, ('fneg', b)), ('fge', c, ('fneg', d)))),
|
|
|
|
|
2019-03-06 22:07:18 +00:00
|
|
|
(('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)),
|
|
|
|
(('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)),
|
|
|
|
(('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)),
|
|
|
|
(('fne', ('fneg', a), ('fneg', b)), ('fne', b, a)),
|
|
|
|
(('flt', ('fneg', a), -1.0), ('flt', 1.0, a)),
|
|
|
|
(('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)),
|
|
|
|
(('fge', ('fneg', a), -1.0), ('fge', 1.0, a)),
|
|
|
|
(('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)),
|
|
|
|
(('fne', ('fneg', a), -1.0), ('fne', 1.0, a)),
|
|
|
|
(('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)),
|
|
|
|
|
|
|
|
(('ior', a, a), a),
|
|
|
|
(('iand', a, a), a),
|
|
|
|
|
2018-03-28 06:57:07 +01:00
|
|
|
(('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))),
|
|
|
|
|
2015-09-10 18:51:46 +01:00
|
|
|
(('fdot2', a, b), ('fdot_replicated2', a, b), 'options->fdot_replicates'),
|
|
|
|
(('fdot3', a, b), ('fdot_replicated3', a, b), 'options->fdot_replicates'),
|
|
|
|
(('fdot4', a, b), ('fdot_replicated4', a, b), 'options->fdot_replicates'),
|
2015-09-23 00:54:27 +01:00
|
|
|
(('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'),
|
2017-01-03 11:54:48 +00:00
|
|
|
|
2018-08-23 03:15:15 +01:00
|
|
|
(('~flrp@32', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
|
|
|
|
(('~flrp@64', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
|
|
|
|
|
nir/algebraic: Reassociate open-coded flrp(1, b, c)
In a previous verion of this patch, Jason commented,
"Re-associating based on whether or not something has a constant
value of 1.0 seems a bit sneaky. I think it's well within the rules
but it seems like something that could bite you."
That is possibly true. The reassociation will generate different
results if fabs(b) >= 2**24 and fabs(c) < 0.5. The delta increases as
fabs(c) approaches 0.
However, i965 has done this same reassociation indirectly for years.
We would previously allow nir_op_flrp on all pre-Gen11 hardware even
though Gen4 and Gen5 do not have a LRP instruction. Optimizations in
nir_opt_algebraic would convert expressions like a+c(b-a) into flrp(a,
b, c). On Gen7+, the hardware performs the same arithmetic as
a(1-c)+bc. Gen6 seems to implement LRP as a+c(b-a). On Gen4 and
Gen5, we would lower LRP to a sequence of instructions that implement
a(1-c)+bc. The lowering happens after all constant folding, so we
would litterally generate a 1+(-1) instruction sequence in this
scenario: one instruction to load either 1 or -1 in a register, and
another instruction to add either -1 or 1 to it.
This patch just cuts out the middle man. Do the reassociation that
we've always done, but do it explicitly at a time when we can benefit
from other optimizations.
A few cases that were hurt by "nir: Lower flrp(±1, b, c) and flrp(a,
±1, c) differently" are restored by this patch. This includes a few
shaders in ET:QW.
I tried a similar thing for open-coded flrp(-1, b, c), and it hurt
instructions on 35 shaders for ILK without helping any. The helped /
hurt cycles was about even.
No changes on any other Intel platforms.
Iron Lake and GM45 had similar results. (Iron Lake shown)
total instructions in shared programs: 8172020 -> 8164367 (-0.09%)
instructions in affected programs: 1089851 -> 1082198 (-0.70%)
helped: 3285
HURT: 64
helped stats (abs) min: 1 max: 6 x̄: 2.35 x̃: 2
helped stats (rel) min: 0.13% max: 12.00% x̄: 1.15% x̃: 0.83%
HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
HURT stats (rel) min: 0.24% max: 0.64% x̄: 0.39% x̃: 0.38%
95% mean confidence interval for instructions value: -2.32 -2.25
95% mean confidence interval for instructions %-change: -1.16% -1.09%
Instructions are helped.
total cycles in shared programs: 188758338 -> 188719974 (-0.02%)
cycles in affected programs: 20004922 -> 19966558 (-0.19%)
helped: 3012
HURT: 477
helped stats (abs) min: 2 max: 142 x̄: 13.41 x̃: 12
helped stats (rel) min: 0.01% max: 6.37% x̄: 0.52% x̃: 0.24%
HURT stats (abs) min: 2 max: 328 x̄: 4.27 x̃: 4
HURT stats (rel) min: <.01% max: 1.55% x̄: 0.14% x̃: 0.11%
95% mean confidence interval for cycles value: -11.38 -10.62
95% mean confidence interval for cycles %-change: -0.46% -0.41%
Cycles are helped.
Reviewed-by: Matt Turner <mattst88@gmail.com>
2018-08-19 20:42:05 +01:00
|
|
|
(('~fadd@32', 1.0, ('fmul(is_used_once)', c , ('fadd', b, -1.0 ))), ('fadd', ('fadd', 1.0, ('fneg', c)), ('fmul', b, c)), 'options->lower_flrp32'),
|
|
|
|
(('~fadd@64', 1.0, ('fmul(is_used_once)', c , ('fadd', b, -1.0 ))), ('fadd', ('fadd', 1.0, ('fneg', c)), ('fmul', b, c)), 'options->lower_flrp64'),
|
|
|
|
|
2017-01-12 10:51:56 +00:00
|
|
|
# we do these late so that we don't get in the way of creating ffmas
|
|
|
|
(('fmin', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmin', a, b))),
|
|
|
|
(('fmax', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmax', a, b))),
|
2018-12-03 22:41:07 +00:00
|
|
|
|
2019-03-30 03:51:20 +00:00
|
|
|
(('bcsel', a, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a, b)))),
|
2015-03-24 00:11:49 +00:00
|
|
|
]
|
|
|
|
|
2018-07-05 14:17:32 +01:00
|
|
|
print(nir_algebraic.AlgebraicPass("nir_opt_algebraic", optimizations).render())
|
|
|
|
print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_before_ffma",
|
|
|
|
before_ffma_optimizations).render())
|
|
|
|
print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_late",
|
|
|
|
late_optimizations).render())
|