KonstantinSeurer
/
mesa
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
mesa/src/compiler/nir/nir_opt_algebraic.py

1221 lines
61 KiB
Python
Raw Normal View History

nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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)
python: Use the print function In Python 2, `print` was a statement, but it became a function in Python 3. Using print functions everywhere makes the script compatible with Python versions >= 2.6, including Python 3. Signed-off-by: Mathieu Bridon <bochecha@daitauha.fr> Acked-by: Eric Engestrom <eric.engestrom@intel.com> Acked-by: Dylan Baker <dylan@pnwbakers.com>
2018-07-05 14:17:32 +01:00
from __future__ import print_function
python: Stabilize some script outputs In Python, dictionaries and sets are unordered, and as a result their is no guarantee that running this script twice will produce the same output. Using ordered dicts and explicitly sorting items makes the build more reproducible, and will make it possible to verify that we're not breaking anything when we move the build scripts to Python 3. Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
2018-06-27 11:37:38 +01:00
from collections import OrderedDict
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
import nir_algebraic
nir/algebraic: Generalize an optimization This just makes it nicely scale across bit sizes. Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-12-06 18:56:33 +00:00
from nir_opcodes import type_sizes
nir: Use De Morgan's Law on logic compounded comparisons The replacement of the comparison operators must happen during this step. If it does not, the next pass of nir_opt_algebraic will reapply De Morgan's Law in the "opposite direction" before performing dead code elimination. The resulting infinite loop will eventually get OOM killed. Haswell, Broadwell, and Skylake had similar results. (Broadwell shown) total instructions in shared programs: 14808185 -> 14808036 (<.01%) instructions in affected programs: 13758 -> 13609 (-1.08%) helped: 39 HURT: 0 helped stats (abs) min: 1 max: 10 x̄: 3.82 x̃: 3 helped stats (rel) min: 0.44% max: 1.55% x̄: 0.98% x̃: 1.01% 95% mean confidence interval for instructions value: -4.67 -2.97 95% mean confidence interval for instructions %-change: -1.09% -0.88% Instructions are helped. total cycles in shared programs: 559438333 -> 559435832 (<.01%) cycles in affected programs: 199160 -> 196659 (-1.26%) helped: 42 HURT: 3 helped stats (abs) min: 2 max: 184 x̄: 61.50 x̃: 51 helped stats (rel) min: 0.02% max: 6.94% x̄: 1.41% x̃: 1.40% HURT stats (abs) min: 2 max: 40 x̄: 27.33 x̃: 40 HURT stats (rel) min: 0.05% max: 0.74% x̄: 0.51% x̃: 0.74% 95% mean confidence interval for cycles value: -71.47 -39.69 95% mean confidence interval for cycles %-change: -1.64% -0.93% Cycles are helped. Sandy Bridge and Ivy Bridge had similar results. (Ivy Bridge shown) total instructions in shared programs: 11811776 -> 11811553 (<.01%) instructions in affected programs: 15201 -> 14978 (-1.47%) helped: 39 HURT: 0 helped stats (abs) min: 1 max: 20 x̄: 5.72 x̃: 6 helped stats (rel) min: 0.44% max: 2.53% x̄: 1.30% x̃: 1.26% 95% mean confidence interval for instructions value: -7.21 -4.23 95% mean confidence interval for instructions %-change: -1.48% -1.12% Instructions are helped. total cycles in shared programs: 257617270 -> 257614589 (<.01%) cycles in affected programs: 212107 -> 209426 (-1.26%) helped: 45 HURT: 0 helped stats (abs) min: 2 max: 180 x̄: 59.58 x̃: 54 helped stats (rel) min: 0.02% max: 6.02% x̄: 1.30% x̃: 1.32% 95% mean confidence interval for cycles value: -74.02 -45.14 95% mean confidence interval for cycles %-change: -1.59% -1.01% Cycles are helped. Iron Lake total instructions in shared programs: 7886648 -> 7886515 (<.01%) instructions in affected programs: 14106 -> 13973 (-0.94%) helped: 29 HURT: 0 helped stats (abs) min: 1 max: 10 x̄: 4.59 x̃: 4 helped stats (rel) min: 0.35% max: 1.83% x̄: 0.90% x̃: 0.81% 95% mean confidence interval for instructions value: -5.65 -3.52 95% mean confidence interval for instructions %-change: -1.03% -0.76% Instructions are helped. total cycles in shared programs: 178100812 -> 178100396 (<.01%) cycles in affected programs: 67970 -> 67554 (-0.61%) helped: 29 HURT: 0 helped stats (abs) min: 2 max: 40 x̄: 14.34 x̃: 12 helped stats (rel) min: 0.15% max: 1.69% x̄: 0.58% x̃: 0.54% 95% mean confidence interval for cycles value: -18.30 -10.39 95% mean confidence interval for cycles %-change: -0.71% -0.45% Cycles are helped. GM45 total instructions in shared programs: 4857939 -> 4857872 (<.01%) instructions in affected programs: 7426 -> 7359 (-0.90%) helped: 15 HURT: 0 helped stats (abs) min: 1 max: 10 x̄: 4.47 x̃: 4 helped stats (rel) min: 0.33% max: 1.80% x̄: 0.87% x̃: 0.77% 95% mean confidence interval for instructions value: -6.06 -2.87 95% mean confidence interval for instructions %-change: -1.06% -0.67% Instructions are helped. total cycles in shared programs: 122167930 -> 122167654 (<.01%) cycles in affected programs: 43118 -> 42842 (-0.64%) helped: 15 HURT: 0 helped stats (abs) min: 4 max: 40 x̄: 18.40 x̃: 16 helped stats (rel) min: 0.15% max: 1.69% x̄: 0.62% x̃: 0.54% 95% mean confidence interval for cycles value: -25.03 -11.77 95% mean confidence interval for cycles %-change: -0.82% -0.41% Cycles are helped. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2018-02-07 01:27:53 +00:00
import itertools
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
# Convenience variables
a = 'a'
b = 'b'
c = 'c'
d = 'd'
nir/algebraic: Replace a pattern where iand with a Boolean is used as a bcsel All of the affected shaders are in Mad Max. I noticed this while looking at some other things. I tried a couple similar patterns, but the affect on cycles was general negative. It may be worth revisiting this later. v2: Rebase on 1-bit Boolean changes. All Gen7+ platforms had similar results. (Skylake shown) total instructions in shared programs: 15282073 -> 15282053 (<.01%) instructions in affected programs: 1192 -> 1172 (-1.68%) helped: 14 HURT: 0 helped stats (abs) min: 1 max: 2 x̄: 1.43 x̃: 1 helped stats (rel) min: 1.16% max: 2.17% x̄: 1.65% x̃: 1.39% 95% mean confidence interval for instructions value: -1.73 -1.13 95% mean confidence interval for instructions %-change: -1.91% -1.38% Instructions are helped. total cycles in shared programs: 372595954 -> 372594532 (<.01%) cycles in affected programs: 11477 -> 10055 (-12.39%) helped: 14 HURT: 0 helped stats (abs) min: 76 max: 122 x̄: 101.57 x̃: 104 helped stats (rel) min: 7.76% max: 15.62% x̄: 12.94% x̃: 14.78% 95% mean confidence interval for cycles value: -111.05 -92.09 95% mean confidence interval for cycles %-change: -14.90% -10.98% Cycles are helped. No changes on any Gen6 or earlier platforms. Reviewed-by: Matt Turner <mattst88@gmail.com>
2018-09-06 04:45:19 +01:00
e = 'e'
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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
nir/algebraic: Allow for flagging operations as being inexact Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:04:49 +00:00
# defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
# where each source is either an expression or a value. A value can be
nir/algebraic: Support specifying variable as constant or by type Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2015-01-29 00:42:20 +00:00
# either a numeric constant or a string representing a variable name.
#
nir/algebraic: Allow for flagging operations as being inexact Reviewed-by: Francisco Jerez <currojerez@riseup.net>
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.
#
nir/algebraic: support for power-of-two optimizations Some optimizations, like converting integer multiply/divide into left/ right shifts, have additional constraints on the search expression. Like requiring that a variable is a constant power of two. Support these cases by allowing a fxn name to be appended to the search var expression (ie. "a#32(is_power_of_two)"). Signed-off-by: Rob Clark <robclark@freedesktop.org> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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
nir/algebraic: Support specifying variable as constant or by type Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2015-01-29 00:42:20 +00:00
# the given variable will only match values from ALU instructions with the
nir/algebraic: support for power-of-two optimizations Some optimizations, like converting integer multiply/divide into left/ right shifts, have additional constraints on the search expression. Like requiring that a variable is a constant power of two. Support these cases by allowing a fxn name to be appended to the search var expression (ie. "a#32(is_power_of_two)"). Signed-off-by: Rob Clark <robclark@freedesktop.org> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2016-05-07 18:01:24 +01:00
# given output type, and (cond) specifies an additional condition function
# (see nir_search_helpers.h).
nir/algebraic: Support specifying variable as constant or by type Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
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
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
# opcodes.
nir/algebraic: Add a mechanism for specifying the bit size of a value Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
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
nir/algebraic: Don't put quotes around floating point literals The quotation marks around 1.0 cause it to be treated as a string instead of a floating point value. The generator then treats it as an arbitrary variable replacement, so any iand involving a ('ineg', ('b2i', a)) matches. v2: Remove misleading comment about sized literals (suggested by Timothy). Add assertion that the name of a varible is entierly alphabetic (suggested by Jason). Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> Tested-by: Timothy Arceri <tarceri@itsqueeze.com> [v1] Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com> [v1] Fixes: 6bcd2af0863 ("nir/algebraic: Add some optimizations for D3D-style Booleans") Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=109075
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.
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
optimizations = [
nir/algebraic: support for power-of-two optimizations Some optimizations, like converting integer multiply/divide into left/ right shifts, have additional constraints on the search expression. Like requiring that a variable is a constant power of two. Support these cases by allowing a fxn name to be appended to the search var expression (ie. "a#32(is_power_of_two)"). Signed-off-by: Rob Clark <robclark@freedesktop.org> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2016-05-07 18:01:24 +01:00
nir: add lower_bitshift option Add a "lower_bitshift" option, which disables optimizations introducing bitshifts and lowers ishl by constant to a multiply, so that we don't have to deal with bitshifts in int_to_float lowering. Signed-off-by: Jonathan Marek <jonathan@marek.ca> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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'),
nir/algebraic: Optimize low 32 bit extraction Optimize a situation where we only need lower 32 bits from 64 bit result. Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com> Suggested-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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)),
nir/glsl: Add another way of doing lower_imul64 for gen8+ On Gen 8 and 9, "mul" instruction supports 64 bit destination type. We can reduce our 64x64 int multiplication from 4 instructions to 3. Also instead of emitting two mul instructions, we can emit single mul instuction and extract low/high 32 bits from 64 bit result for [i/u]mulExtended v2: 1) Allow lower_mul_high64 to use new opcode (Jason Ekstrand) 2) Add lower_mul_2x32_64 flag (Matt Turner) 3) Remove associative property as bit size is different (Connor Abbott) v3: Fix indentation and variable naming convention (Jason Ekstrand) Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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'),
nir: Optimize integer division and modulus with 1 The previous power-of-two rules didn't catch idiv (because i965 doesn't set lower_idiv) and imod cases. The udiv and umod cases should have been caught, but I included them for orthogonality. This fixes silly code observed from compute shaders with local_size_[xy] = 1. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98299 Reviewed-by: Jordan Justen <jordan.l.justen@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2016-10-19 16:53:10 +01:00
(('udiv', a, 1), a),
(('idiv', a, 1), a),
(('umod', a, 1), 0),
(('imod', a, 1), 0),
nir: add lower_bitshift option Add a "lower_bitshift" option, which disables optimizations introducing bitshifts and lowers ishl by constant to a multiply, so that we don't have to deal with bitshifts in int_to_float lowering. Signed-off-by: Jonathan Marek <jonathan@marek.ca> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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'),
nir/algebraic: support for power-of-two optimizations Some optimizations, like converting integer multiply/divide into left/ right shifts, have additional constraints on the search expression. Like requiring that a variable is a constant power of two. Support these cases by allowing a fxn name to be appended to the search var expression (ie. "a#32(is_power_of_two)"). Signed-off-by: Rob Clark <robclark@freedesktop.org> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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))),
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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)),
nir/opt_algebraic: Drop bit-size suffixes from conversions Suffixes are dropped from a bunch of conversion opcodes when it makes sense to do so. Others are kept if we really do want the bit-size restriction. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2018-11-07 23:32:32 +00:00
(('fabs', ('u2f', a)), ('u2f', a)),
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
(('iabs', ('iabs', a)), ('iabs', a)),
(('iabs', ('ineg', a)), ('iabs', a)),
nir/algebraic: Optimize x2b(xneg(a)) -> a Shader-db results on Kaby Lake: total instructions in shared programs: 15072525 -> 15072525 (0.00%) instructions in affected programs: 0 -> 0 helped: 0 HURT: 0 This helps prevent regressions in later commits. Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-12-06 19:02:01 +00:00
(('f2b', ('fneg', a)), ('f2b', a)),
(('i2b', ('ineg', a)), ('i2b', a)),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~fadd', a, 0.0), a),
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
(('iadd', a, 0), a),
nir: Add opcodes for saturated vector math. This corresponds to instructions used on vc4 for its blending inside of shaders. I've seen these opcodes on other architectures before, but I think it's the first time these are needed in Mesa. v2: Rename to 'u' instead of 'i', since they're all 'u'norm (from review by jekstrand)
2015-08-19 06:38:34 +01:00
(('usadd_4x8', a, 0), a),
(('usadd_4x8', a, ~0), ~0),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
nir: Optimize (a*b)+(a*c) -> a*(b+c) Shader-db i965 instructions: total instructions in shared programs: 1715894 -> 1710802 (-0.30%) instructions in affected programs: 443080 -> 437988 (-1.15%) helped: 1502 HURT: 13 GAINED: 4 LOST: 4 Shader-db NIR instructions: total instructions in shared programs: 607710 -> 606187 (-0.25%) instructions in affected programs: 208285 -> 206762 (-0.73%) helped: 769 HURT: 8 GAINED: 0 LOST: 0 Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
2015-02-28 19:32:31 +00:00
(('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~fadd', ('fneg', a), a), 0.0),
nir: Optimize a + neg(a) Shader-db i965 instructions: total instructions in shared programs: 1711180 -> 1711159 (-0.00%) instructions in affected programs: 825 -> 804 (-2.55%) helped: 9 HURT: 0 GAINED: 3 LOST: 3 Shader-db NIR instructions: total instructions in shared programs: 606187 -> 606179 (-0.00%) instructions in affected programs: 298 -> 290 (-2.68%) helped: 4 HURT: 0 GAINED: 0 LOST: 0 Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
2015-02-28 19:32:32 +00:00
(('iadd', ('ineg', a), a), 0),
nir: Teach nir_opt_algebraic about adding and subtracting the same thing This optimizes a + b - b to just a. Modest shader-db results (BDW): total instructions in shared programs: 7842452 -> 7841862 (-0.01%) instructions in affected programs: 61938 -> 61348 (-0.95%) total loops in shared programs: 2131 -> 2131 (0.00%) helped: 263 HURT: 0 GAINED: 0 LOST: 0 but the optimization turns gl_VertexID - gl_BaseVertexARB into just a reference to SYSTEM_VALUE_VERTEX_ID_ZERO_BASE, which the i965 hardware supports natively. That means we can avoid using the internal vertex buffer for gl_BaseVertexARB in this case. Reviewed-by: Eduardo Lima Mitev <elima@igalia.com> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-12-15 01:44:23 +00:00
(('iadd', ('ineg', a), ('iadd', a, b)), b),
(('iadd', a, ('iadd', ('ineg', a), b)), b),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~fadd', ('fneg', a), ('fadd', a, b)), b),
(('~fadd', a, ('fadd', ('fneg', a), b)), b),
nir/algebraic: Recognize open-coded fsat with modifiers This change also enables a later change (nir/algebraic: Replace 1-fsat(a) with fsat(1-a)) to affect more shaders. Almost all of the affected shaders are in Bioshock Infinite, and all of those shaders all require GLSL 4.10. All Intel platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17228584 -> 17228376 (<.01%) instructions in affected programs: 31438 -> 31230 (-0.66%) helped: 105 HURT: 0 helped stats (abs) min: 1 max: 5 x̄: 1.98 x̃: 1 helped stats (rel) min: 0.08% max: 1.53% x̄: 0.73% x̃: 0.70% 95% mean confidence interval for instructions value: -2.20 -1.76 95% mean confidence interval for instructions %-change: -0.80% -0.67% Instructions are helped. total cycles in shared programs: 360936431 -> 360935690 (<.01%) cycles in affected programs: 420100 -> 419359 (-0.18%) helped: 71 HURT: 21 helped stats (abs) min: 1 max: 160 x̄: 19.28 x̃: 10 helped stats (rel) min: <.01% max: 9.78% x̄: 0.95% x̃: 0.48% HURT stats (abs) min: 1 max: 198 x̄: 29.90 x̃: 10 HURT stats (rel) min: 0.05% max: 8.36% x̄: 1.24% x̃: 0.90% 95% mean confidence interval for cycles value: -16.77 0.66 95% mean confidence interval for cycles %-change: -0.85% -0.06% Inconclusive result (value mean confidence interval includes 0). Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2019-03-05 20:46:11 +00:00
(('fadd', ('fsat', a), ('fsat', ('fneg', a))), ('fsat', ('fabs', a))),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~fmul', a, 0.0), 0.0),
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
(('imul', a, 0), 0),
nir: Add opcodes for saturated vector math. This corresponds to instructions used on vc4 for its blending inside of shaders. I've seen these opcodes on other architectures before, but I think it's the first time these are needed in Mesa. v2: Rename to 'u' instead of 'i', since they're all 'u'norm (from review by jekstrand)
2015-08-19 06:38:34 +01:00
(('umul_unorm_4x8', a, 0), 0),
(('umul_unorm_4x8', a, ~0), a),
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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)),
nir/algebraic: sign(x)*x*x is abs(x)*x shader-db results: All Gen7+ platforms had similar results. (Skylake shown) total instructions in shared programs: 15106023 -> 15105981 (<.01%) instructions in affected programs: 300 -> 258 (-14.00%) helped: 6 HURT: 0 helped stats (abs) min: 7 max: 7 x̄: 7.00 x̃: 7 helped stats (rel) min: 14.00% max: 14.00% x̄: 14.00% x̃: 14.00% 95% mean confidence interval for instructions value: -7.00 -7.00 95% mean confidence interval for instructions %-change: -14.00% -14.00% Instructions are helped. total cycles in shared programs: 566050327 -> 566050075 (<.01%) cycles in affected programs: 2826 -> 2574 (-8.92%) helped: 6 HURT: 0 helped stats (abs) min: 40 max: 44 x̄: 42.00 x̃: 42 helped stats (rel) min: 8.89% max: 8.94% x̄: 8.92% x̃: 8.92% 95% mean confidence interval for cycles value: -44.30 -39.70 95% mean confidence interval for cycles %-change: -8.95% -8.88% Cycles are helped. No changes on Gen6 or earlier. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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)),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~ffma', 0.0, a, b), b),
(('~ffma', a, b, 0.0), ('fmul', a, b)),
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
(('ffma', 1.0, a, b), ('fadd', a, b)),
nir/algebraic: Add missing ffma(-1, a, b) pattern All Gen7+ platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17229439 -> 17229377 (<.01%) instructions in affected programs: 9859 -> 9797 (-0.63%) helped: 41 HURT: 0 helped stats (abs) min: 1 max: 6 x̄: 1.51 x̃: 1 helped stats (rel) min: 0.08% max: 11.54% x̄: 1.65% x̃: 0.67% 95% mean confidence interval for instructions value: -1.88 -1.14 95% mean confidence interval for instructions %-change: -2.48% -0.81% Instructions are helped. total cycles in shared programs: 360944145 -> 360942989 (<.01%) cycles in affected programs: 178167 -> 177011 (-0.65%) helped: 36 HURT: 19 helped stats (abs) min: 1 max: 222 x̄: 38.03 x̃: 5 helped stats (rel) min: 0.01% max: 31.01% x̄: 4.01% x̃: 0.45% HURT stats (abs) min: 1 max: 34 x̄: 11.21 x̃: 6 HURT stats (rel) min: 0.03% max: 2.74% x̄: 0.72% x̃: 0.50% 95% mean confidence interval for cycles value: -36.01 -6.02 95% mean confidence interval for cycles %-change: -4.18% -0.57% Cycles are helped. Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-05-10 18:20:02 +01:00
(('ffma', -1.0, a, b), ('fadd', ('fneg', a), b)),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
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)),
nir/algebraic: Pull common addition out of flrp arguments v2: Augment the late optimization patterns with a couple pre-ffma pass patterns. All Gen7+ platforms had similar results. (Skylake shown) total instructions in shared programs: 15342982 -> 15342485 (<.01%) instructions in affected programs: 56304 -> 55807 (-0.88%) helped: 235 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 2.11 x̃: 1 helped stats (rel) min: 0.11% max: 8.82% x̄: 1.27% x̃: 0.74% 95% mean confidence interval for instructions value: -2.31 -1.92 95% mean confidence interval for instructions %-change: -1.46% -1.09% Instructions are helped. total cycles in shared programs: 355734740 -> 355734320 (<.01%) cycles in affected programs: 1028807 -> 1028387 (-0.04%) helped: 134 HURT: 104 helped stats (abs) min: 1 max: 212 x̄: 25.69 x̃: 8 helped stats (rel) min: <.01% max: 9.36% x̄: 1.33% x̃: 0.61% HURT stats (abs) min: 1 max: 203 x̄: 29.06 x̃: 8 HURT stats (rel) min: 0.02% max: 15.76% x̄: 1.76% x̃: 0.46% 95% mean confidence interval for cycles value: -8.51 4.98 95% mean confidence interval for cycles %-change: -0.35% 0.39% Inconclusive result (value mean confidence interval includes 0). Sandy Bridge total instructions in shared programs: 10886815 -> 10886390 (<.01%) instructions in affected programs: 36883 -> 36458 (-1.15%) helped: 147 HURT: 0 helped stats (abs) min: 1 max: 7 x̄: 2.89 x̃: 3 helped stats (rel) min: 0.35% max: 8.00% x̄: 1.60% x̃: 1.23% 95% mean confidence interval for instructions value: -3.12 -2.67 95% mean confidence interval for instructions %-change: -1.83% -1.38% Instructions are helped. total cycles in shared programs: 154188360 -> 154186902 (<.01%) cycles in affected programs: 388094 -> 386636 (-0.38%) helped: 90 HURT: 58 helped stats (abs) min: 1 max: 243 x̄: 36.80 x̃: 15 helped stats (rel) min: 0.04% max: 9.23% x̄: 1.26% x̃: 0.83% HURT stats (abs) min: 1 max: 684 x̄: 31.97 x̃: 10 HURT stats (rel) min: 0.03% max: 13.50% x̄: 1.15% x̃: 0.51% 95% mean confidence interval for cycles value: -22.62 2.92 95% mean confidence interval for cycles %-change: -0.68% 0.05% Inconclusive result (value mean confidence interval includes 0). Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 8221239 -> 8220357 (-0.01%) instructions in affected programs: 54560 -> 53678 (-1.62%) helped: 186 HURT: 0 helped stats (abs) min: 1 max: 14 x̄: 4.74 x̃: 3 helped stats (rel) min: 0.34% max: 10.77% x̄: 1.97% x̃: 1.17% 95% mean confidence interval for instructions value: -5.21 -4.28 95% mean confidence interval for instructions %-change: -2.23% -1.72% Instructions are helped. total cycles in shared programs: 188654442 -> 188650364 (<.01%) cycles in affected programs: 1454384 -> 1450306 (-0.28%) helped: 204 HURT: 0 helped stats (abs) min: 2 max: 84 x̄: 19.99 x̃: 18 helped stats (rel) min: 0.02% max: 4.69% x̄: 0.56% x̃: 0.22% 95% mean confidence interval for cycles value: -22.38 -17.60 95% mean confidence interval for cycles %-change: -0.67% -0.46% Cycles are helped. Reviewed-by: Matt Turner <mattst88@gmail.com>
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'),
nir/algebraic: Pull common multiplication out of flrp arguments All Intel platforms had similar results. (Skylake shown) total instructions in shared programs: 15342485 -> 15337495 (-0.03%) instructions in affected programs: 217456 -> 212466 (-2.29%) helped: 1539 HURT: 1 helped stats (abs) min: 1 max: 17 x̄: 3.24 x̃: 3 helped stats (rel) min: 0.22% max: 18.75% x̄: 3.10% x̃: 1.91% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.56% max: 0.56% x̄: 0.56% x̃: 0.56% 95% mean confidence interval for instructions value: -3.39 -3.09 95% mean confidence interval for instructions %-change: -3.24% -2.96% Instructions are helped. total cycles in shared programs: 355734320 -> 355728237 (<.01%) cycles in affected programs: 1851555 -> 1845472 (-0.33%) helped: 835 HURT: 575 helped stats (abs) min: 1 max: 658 x̄: 40.62 x̃: 14 helped stats (rel) min: <.01% max: 35.69% x̄: 3.78% x̃: 1.81% HURT stats (abs) min: 1 max: 322 x̄: 48.40 x̃: 14 HURT stats (rel) min: 0.04% max: 71.02% x̄: 8.06% x̃: 2.43% 95% mean confidence interval for cycles value: -8.50 -0.13 95% mean confidence interval for cycles %-change: 0.48% 1.62% Inconclusive result (value mean confidence interval and %-change mean confidence interval disagree). Reviewed-by: Matt Turner <mattst88@gmail.com>
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)),
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-12-06 03:36:10 +00:00
(('~flrp', a, b, ('b2f', 'c@1')), ('bcsel', c, b, a), 'options->lower_flrp32'),
nir: add imprecise flrp optimisation On BDW: total instructions in shared programs: 13061890 -> 13061877 (-0.00%) instructions in affected programs: 2441 -> 2428 (-0.53%) helped: 13 HURT: 0 total cycles in shared programs: 256612254 -> 256611784 (-0.00%) cycles in affected programs: 16418 -> 15948 (-2.86%) helped: 10 HURT: 2 V2: don't use ffma directly Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2017-01-03 10:10:33 +00:00
(('~flrp', a, 0.0, c), ('fadd', ('fmul', ('fneg', a), c), a)),
nir: add lower_ftrunc Port TGSI TRUNC lowering to nir Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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'),
nir: add lowering for ffloor Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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'),
nir: add lowering for ffract Signed-off-by: Rob Clark <robclark@freedesktop.org> Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2015-09-14 16:13:19 +01:00
(('ffract', a), ('fsub', a, ('ffloor', a)), 'options->lower_ffract'),
nir: add fceil lowering lowers ceil(x) as -floor(-x) Signed-off-by: Jonathan Marek <jonathan@marek.ca> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com> Signed-off-by: Rob Clark <robdclark@gmail.com>
2018-11-12 17:49:32 +00:00
(('fceil', a), ('fneg', ('ffloor', ('fneg', a))), 'options->lower_fceil'),
nir/algebraic: Recognize open-coded flrp(a, b, fsat(c)) All Gen6+ GPUs had similar results. (Skylake shown) total instructions in shared programs: 15336712 -> 15336622 (<.01%) instructions in affected programs: 3952 -> 3862 (-2.28%) helped: 24 HURT: 0 helped stats (abs) min: 3 max: 5 x̄: 3.75 x̃: 4 helped stats (rel) min: 1.75% max: 2.70% x̄: 2.34% x̃: 2.46% 95% mean confidence interval for instructions value: -4.06 -3.44 95% mean confidence interval for instructions %-change: -2.47% -2.22% Instructions are helped. total cycles in shared programs: 355722052 -> 355721235 (<.01%) cycles in affected programs: 27326 -> 26509 (-2.99%) helped: 20 HURT: 4 helped stats (abs) min: 1 max: 227 x̄: 44.75 x̃: 14 helped stats (rel) min: 0.12% max: 22.95% x̄: 3.83% x̃: 1.23% HURT stats (abs) min: 2 max: 64 x̄: 19.50 x̃: 6 HURT stats (rel) min: 0.21% max: 3.63% x̄: 1.24% x̃: 0.55% 95% mean confidence interval for cycles value: -61.61 -6.47 95% mean confidence interval for cycles %-change: -5.59% -0.39% Cycles are helped. No changes on Ice Lake, Iron Lake, or GM45. Reviewed-by: Matt Turner <mattst88@gmail.com>
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
nir/algebraic: Commute 1-fsat(a) to fsat(1-a) for all non-fmul instructions The goal is to avoid having an extra MOV instruction to perform the saturate. Doing the subtraction first allows the saturate to be applied to the ADD instruction making the MOV unnecessary. Values generated in different block and values from non-ALU instructions (e.g., texture instructions) almost always need the extra MOV. Multiply instructions are restricted because doing this rearrangement can interfere with the generation of flrp and ffma instructions. v2: Now that the final method has been selected, squash three commits into one. All Intel platforms has similar results. (Ice Lake shown) total instructions in shared programs: 17223214 -> 17219386 (-0.02%) instructions in affected programs: 1524376 -> 1520548 (-0.25%) helped: 2686 HURT: 26 helped stats (abs) min: 1 max: 32 x̄: 1.44 x̃: 1 helped stats (rel) min: 0.03% max: 16.67% x̄: 0.54% x̃: 0.37% HURT stats (abs) min: 1 max: 2 x̄: 1.69 x̃: 2 HURT stats (rel) min: 0.33% max: 1.67% x̄: 0.54% x̃: 0.35% 95% mean confidence interval for instructions value: -1.46 -1.36 95% mean confidence interval for instructions %-change: -0.56% -0.50% Instructions are helped. total cycles in shared programs: 360811571 -> 360791896 (<.01%) cycles in affected programs: 103650214 -> 103630539 (-0.02%) helped: 1557 HURT: 675 helped stats (abs) min: 1 max: 1773 x̄: 41.44 x̃: 16 helped stats (rel) min: <.01% max: 26.77% x̄: 1.37% x̃: 0.64% HURT stats (abs) min: 1 max: 1513 x̄: 66.44 x̃: 14 HURT stats (rel) min: <.01% max: 46.16% x̄: 2.00% x̃: 0.49% 95% mean confidence interval for cycles value: -14.82 -2.81 95% mean confidence interval for cycles %-change: -0.50% -0.20% Cycles are helped. LOST: 2 GAINED: 0 Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2018-03-28 06:57:07 +01:00
# 1-fsat(x) <=> fsat(1-x). See below.
nir/algebraic: Recognize open-coded flrp(a, b, fsat(c)) All Gen6+ GPUs had similar results. (Skylake shown) total instructions in shared programs: 15336712 -> 15336622 (<.01%) instructions in affected programs: 3952 -> 3862 (-2.28%) helped: 24 HURT: 0 helped stats (abs) min: 3 max: 5 x̄: 3.75 x̃: 4 helped stats (rel) min: 1.75% max: 2.70% x̄: 2.34% x̃: 2.46% 95% mean confidence interval for instructions value: -4.06 -3.44 95% mean confidence interval for instructions %-change: -2.47% -2.22% Instructions are helped. total cycles in shared programs: 355722052 -> 355721235 (<.01%) cycles in affected programs: 27326 -> 26509 (-2.99%) helped: 20 HURT: 4 helped stats (abs) min: 1 max: 227 x̄: 44.75 x̃: 14 helped stats (rel) min: 0.12% max: 22.95% x̄: 3.83% x̃: 1.23% HURT stats (abs) min: 2 max: 64 x̄: 19.50 x̃: 6 HURT stats (rel) min: 0.21% max: 3.63% x̄: 1.24% x̃: 0.55% 95% mean confidence interval for cycles value: -61.61 -6.47 95% mean confidence interval for cycles %-change: -5.59% -0.39% Cycles are helped. No changes on Ice Lake, Iron Lake, or GM45. Reviewed-by: Matt Turner <mattst88@gmail.com>
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'),
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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'),
nir: Add lrp lowering for doubles in opt_algebraic Some hardware (i965 on Broadwell generation, for example) does not support natively the execution of lrp instruction with double arguments. Add 'lower_flrp64' flag to lower this instruction in that case. v2: - Rename lower_flrp_double to lower_flrp64 (Jason) - Fix typo (Jason) - Adapt the code to define bit_size information in the opcodes. Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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'),
nir: Add a flag for lowering ffma. vc4 cse/algebraic-disabled stats: total uniforms in shared programs: 13966 -> 13791 (-1.25%) uniforms in affected programs: 435 -> 260 (-40.23%) total instructions in shared programs: 44732 -> 44356 (-0.84%) instructions in affected programs: 9599 -> 9223 (-3.92%) v2: Rebase to master (no TGSI->NIR present) Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
2015-01-30 21:49:48 +00:00
(('ffma', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma'),
nir/algebraic: Separate ffma lowering from fusing The i965 driver has its own pass for fusing mul+add combinations that's much smarter than what nir_opt_algebraic can do so we don't want to get the nir_opt_algebraic one just because we didn't set lower_ffma. Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
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
nir/algebraic: Replace a pattern where iand with a Boolean is used as a bcsel All of the affected shaders are in Mad Max. I noticed this while looking at some other things. I tried a couple similar patterns, but the affect on cycles was general negative. It may be worth revisiting this later. v2: Rebase on 1-bit Boolean changes. All Gen7+ platforms had similar results. (Skylake shown) total instructions in shared programs: 15282073 -> 15282053 (<.01%) instructions in affected programs: 1192 -> 1172 (-1.68%) helped: 14 HURT: 0 helped stats (abs) min: 1 max: 2 x̄: 1.43 x̃: 1 helped stats (rel) min: 1.16% max: 2.17% x̄: 1.65% x̃: 1.39% 95% mean confidence interval for instructions value: -1.73 -1.13 95% mean confidence interval for instructions %-change: -1.91% -1.38% Instructions are helped. total cycles in shared programs: 372595954 -> 372594532 (<.01%) cycles in affected programs: 11477 -> 10055 (-12.39%) helped: 14 HURT: 0 helped stats (abs) min: 76 max: 122 x̄: 101.57 x̃: 104 helped stats (rel) min: 7.76% max: 15.62% x̄: 12.94% x̃: 14.78% 95% mean confidence interval for cycles value: -111.05 -92.09 95% mean confidence interval for cycles %-change: -14.90% -10.98% Cycles are helped. No changes on any Gen6 or earlier platforms. Reviewed-by: Matt Turner <mattst88@gmail.com>
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))),
nir: Narrow some dot product operations On vector platforms, this helps elide some constant loads. v2: Reorder the transformations. No changes on Broadwell or Skylake. Haswell total instructions in shared programs: 13093793 -> 13060163 (-0.26%) instructions in affected programs: 1277532 -> 1243902 (-2.63%) helped: 13216 HURT: 95 helped stats (abs) min: 1 max: 18 x̄: 2.56 x̃: 2 helped stats (rel) min: 0.21% max: 20.00% x̄: 3.63% x̃: 2.78% HURT stats (abs) min: 1 max: 6 x̄: 1.77 x̃: 1 HURT stats (rel) min: 0.09% max: 5.56% x̄: 1.25% x̃: 1.19% 95% mean confidence interval for instructions value: -2.57 -2.49 95% mean confidence interval for instructions %-change: -3.65% -3.54% Instructions are helped. total cycles in shared programs: 409580819 -> 409268463 (-0.08%) cycles in affected programs: 71730652 -> 71418296 (-0.44%) helped: 9898 HURT: 2352 helped stats (abs) min: 2 max: 16014 x̄: 37.08 x̃: 16 helped stats (rel) min: <.01% max: 35.55% x̄: 6.26% x̃: 4.50% HURT stats (abs) min: 2 max: 276 x̄: 23.25 x̃: 6 HURT stats (rel) min: <.01% max: 40.00% x̄: 3.54% x̃: 1.97% 95% mean confidence interval for cycles value: -33.19 -17.80 95% mean confidence interval for cycles %-change: -4.50% -4.26% Cycles are helped. total fills in shared programs: 82059 -> 82052 (<.01%) fills in affected programs: 21 -> 14 (-33.33%) helped: 7 HURT: 0 Sandy Bridge and Ivy Bridge had similar results (Ivy Bridge shown) total instructions in shared programs: 11811851 -> 11780605 (-0.26%) instructions in affected programs: 1155007 -> 1123761 (-2.71%) helped: 12304 HURT: 95 helped stats (abs) min: 1 max: 18 x̄: 2.55 x̃: 2 helped stats (rel) min: 0.21% max: 20.00% x̄: 3.69% x̃: 2.86% HURT stats (abs) min: 1 max: 6 x̄: 1.77 x̃: 1 HURT stats (rel) min: 0.09% max: 5.56% x̄: 1.25% x̃: 1.19% 95% mean confidence interval for instructions value: -2.56 -2.48 95% mean confidence interval for instructions %-change: -3.71% -3.59% Instructions are helped. total cycles in shared programs: 257618409 -> 257316805 (-0.12%) cycles in affected programs: 71999580 -> 71697976 (-0.42%) helped: 9155 HURT: 2380 helped stats (abs) min: 2 max: 16014 x̄: 38.44 x̃: 16 helped stats (rel) min: <.01% max: 35.75% x̄: 6.39% x̃: 4.62% HURT stats (abs) min: 2 max: 290 x̄: 21.14 x̃: 4 HURT stats (rel) min: <.01% max: 41.55% x̄: 3.14% x̃: 1.33% 95% mean confidence interval for cycles value: -34.32 -17.97 95% mean confidence interval for cycles %-change: -4.55% -4.29% Cycles are helped. GM45 and Iron Lake had nearly identical results (Iron Lake shown) total instructions in shared programs: 7886750 -> 7879944 (-0.09%) instructions in affected programs: 373781 -> 366975 (-1.82%) helped: 3715 HURT: 47 helped stats (abs) min: 1 max: 8 x̄: 1.86 x̃: 1 helped stats (rel) min: 0.22% max: 16.67% x̄: 2.88% x̃: 2.06% HURT stats (abs) min: 1 max: 6 x̄: 2.55 x̃: 2 HURT stats (rel) min: 1.09% max: 5.00% x̄: 1.93% x̃: 2.35% 95% mean confidence interval for instructions value: -1.85 -1.77 95% mean confidence interval for instructions %-change: -2.91% -2.73% Instructions are helped. total cycles in shared programs: 178114636 -> 178095452 (-0.01%) cycles in affected programs: 7227666 -> 7208482 (-0.27%) helped: 3349 HURT: 301 helped stats (abs) min: 2 max: 90 x̄: 6.55 x̃: 4 helped stats (rel) min: <.01% max: 14.18% x̄: 0.95% x̃: 0.63% HURT stats (abs) min: 2 max: 42 x̄: 9.13 x̃: 10 HURT stats (rel) min: 0.01% max: 11.19% x̄: 1.22% x̃: 1.50% 95% mean confidence interval for cycles value: -5.52 -4.99 95% mean confidence interval for cycles %-change: -0.81% -0.73% Cycles are helped. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> [v1]
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)),
nir/algebraic: Commute 1-fsat(a) to fsat(1-a) for all non-fmul instructions The goal is to avoid having an extra MOV instruction to perform the saturate. Doing the subtraction first allows the saturate to be applied to the ADD instruction making the MOV unnecessary. Values generated in different block and values from non-ALU instructions (e.g., texture instructions) almost always need the extra MOV. Multiply instructions are restricted because doing this rearrangement can interfere with the generation of flrp and ffma instructions. v2: Now that the final method has been selected, squash three commits into one. All Intel platforms has similar results. (Ice Lake shown) total instructions in shared programs: 17223214 -> 17219386 (-0.02%) instructions in affected programs: 1524376 -> 1520548 (-0.25%) helped: 2686 HURT: 26 helped stats (abs) min: 1 max: 32 x̄: 1.44 x̃: 1 helped stats (rel) min: 0.03% max: 16.67% x̄: 0.54% x̃: 0.37% HURT stats (abs) min: 1 max: 2 x̄: 1.69 x̃: 2 HURT stats (rel) min: 0.33% max: 1.67% x̄: 0.54% x̃: 0.35% 95% mean confidence interval for instructions value: -1.46 -1.36 95% mean confidence interval for instructions %-change: -0.56% -0.50% Instructions are helped. total cycles in shared programs: 360811571 -> 360791896 (<.01%) cycles in affected programs: 103650214 -> 103630539 (-0.02%) helped: 1557 HURT: 675 helped stats (abs) min: 1 max: 1773 x̄: 41.44 x̃: 16 helped stats (rel) min: <.01% max: 26.77% x̄: 1.37% x̃: 0.64% HURT stats (abs) min: 1 max: 1513 x̄: 66.44 x̃: 14 HURT stats (rel) min: <.01% max: 46.16% x̄: 2.00% x̃: 0.49% 95% mean confidence interval for cycles value: -14.82 -2.81 95% mean confidence interval for cycles %-change: -0.50% -0.20% Cycles are helped. LOST: 2 GAINED: 0 Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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))),
nir: Add algebraic optimizations for simplifying comparisons. The first batch removes bonus fnot/inot operations, possibly allowing other optimizations to better recognize patterns. The next batch replaces a fadd and constant 0.0 with an fneg - negation is usually free on GPUs, while addition is not. total NIR instructions in shared programs: 2020814 -> 2015593 (-0.26%) NIR instructions in affected programs: 411143 -> 405922 (-1.27%) helped: 2233 HURT: 214 A few shaders are hurt by a few instructions due to moving neg such that it has a constant operand, which is then folded, resulting in two distinct load_consts for x and -x. We can always clean that up later. total i965 instructions in shared programs: 6035392 -> 6025505 (-0.16%) i965 instructions in affected programs: 784980 -> 775093 (-1.26%) helped: 4508 HURT: 2 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-01-22 07:36:01 +00:00
# Comparison simplifications
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
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)),
nir: Add algebraic optimizations for simplifying comparisons. The first batch removes bonus fnot/inot operations, possibly allowing other optimizations to better recognize patterns. The next batch replaces a fadd and constant 0.0 with an fneg - negation is usually free on GPUs, while addition is not. total NIR instructions in shared programs: 2020814 -> 2015593 (-0.26%) NIR instructions in affected programs: 411143 -> 405922 (-1.27%) helped: 2233 HURT: 214 A few shaders are hurt by a few instructions due to moving neg such that it has a constant operand, which is then folded, resulting in two distinct load_consts for x and -x. We can always clean that up later. total i965 instructions in shared programs: 6035392 -> 6025505 (-0.16%) i965 instructions in affected programs: 784980 -> 775093 (-1.26%) helped: 4508 HURT: 2 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-01-22 07:36:01 +00:00
(('inot', ('ilt', a, b)), ('ige', a, b)),
nir: add unsigned comparison simplifications This avoids loop unrolling regressions in Wolfenstein II on DXVK with an upcoming optimisation series from Samuel. Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-05-30 11:32:24 +01:00
(('inot', ('ult', a, b)), ('uge', a, b)),
nir: Add algebraic optimizations for simplifying comparisons. The first batch removes bonus fnot/inot operations, possibly allowing other optimizations to better recognize patterns. The next batch replaces a fadd and constant 0.0 with an fneg - negation is usually free on GPUs, while addition is not. total NIR instructions in shared programs: 2020814 -> 2015593 (-0.26%) NIR instructions in affected programs: 411143 -> 405922 (-1.27%) helped: 2233 HURT: 214 A few shaders are hurt by a few instructions due to moving neg such that it has a constant operand, which is then folded, resulting in two distinct load_consts for x and -x. We can always clean that up later. total i965 instructions in shared programs: 6035392 -> 6025505 (-0.16%) i965 instructions in affected programs: 784980 -> 775093 (-1.26%) helped: 4508 HURT: 2 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-01-22 07:36:01 +00:00
(('inot', ('ige', a, b)), ('ilt', a, b)),
nir: add unsigned comparison simplifications This avoids loop unrolling regressions in Wolfenstein II on DXVK with an upcoming optimisation series from Samuel. Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-05-30 11:32:24 +01:00
(('inot', ('uge', a, b)), ('ult', a, b)),
nir: Allow feq/fne/ieq/ine to be optimized with inot. instructions in affected programs: 380 -> 376 (-1.05%) helped: 2 Reviewed-by: Jordan Justen <jordan.l.justen@intel.com> Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
2015-05-06 06:54:59 +01:00
(('inot', ('ieq', a, b)), ('ine', a, b)),
(('inot', ('ine', a, b)), ('ieq', a, b)),
nir: Simplify 0 >= b2f(a) This also prevented some regressions with other patches in my local tree. Broadwell / Skylake total instructions in shared programs: 8980835 -> 8980833 (-0.00%) instructions in affected programs: 45 -> 43 (-4.44%) helped: 1 HURT: 0 total cycles in shared programs: 70077904 -> 70077900 (-0.00%) cycles in affected programs: 122 -> 118 (-3.28%) helped: 1 HURT: 0 No changes on earlier platforms. v2: Modify the comments to look more like a proof. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-03-02 23:18:34 +00:00
nir/algebraic: Strip double negatives from comparison sources All Intel platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17224623 -> 17224337 (<.01%) instructions in affected programs: 32648 -> 32362 (-0.88%) helped: 148 HURT: 0 helped stats (abs) min: 1 max: 2 x̄: 1.93 x̃: 2 helped stats (rel) min: 0.16% max: 2.74% x̄: 1.07% x̃: 1.08% 95% mean confidence interval for instructions value: -1.97 -1.89 95% mean confidence interval for instructions %-change: -1.15% -1.00% Instructions are helped. total cycles in shared programs: 360828714 -> 360826090 (<.01%) cycles in affected programs: 347416 -> 344792 (-0.76%) helped: 148 HURT: 26 helped stats (abs) min: 1 max: 426 x̄: 26.33 x̃: 18 helped stats (rel) min: 0.03% max: 15.10% x̄: 1.78% x̃: 1.41% HURT stats (abs) min: 2 max: 337 x̄: 48.96 x̃: 6 HURT stats (rel) min: 0.04% max: 18.82% x̄: 2.15% x̃: 0.27% 95% mean confidence interval for cycles value: -23.78 -6.38 95% mean confidence interval for cycles %-change: -1.59% -0.79% Cycles are helped. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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)),
nir/algebraic: Eliminate useless fsat() on operand of comparison w/value in (0, 1) v2: Fix copy-and-paste bug in a cmp b vs b cmp a cases. All Gen7+ platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17224337 -> 17224269 (<.01%) instructions in affected programs: 13578 -> 13510 (-0.50%) helped: 68 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.31% max: 3.12% x̄: 0.84% x̃: 0.42% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -1.05% -0.63% Instructions are helped. total cycles in shared programs: 360826090 -> 360825137 (<.01%) cycles in affected programs: 94867 -> 93914 (-1.00%) helped: 58 HURT: 1 helped stats (abs) min: 2 max: 28 x̄: 17.74 x̃: 18 helped stats (rel) min: 0.08% max: 3.17% x̄: 1.39% x̃: 1.22% HURT stats (abs) min: 76 max: 76 x̄: 76.00 x̃: 76 HURT stats (rel) min: 2.86% max: 2.86% x̄: 2.86% x̃: 2.86% 95% mean confidence interval for cycles value: -19.53 -12.78 95% mean confidence interval for cycles %-change: -1.56% -1.08% Cycles are helped. No changes on any other Intel platform. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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)),
nir: Simplify 0 >= b2f(a) This also prevented some regressions with other patches in my local tree. Broadwell / Skylake total instructions in shared programs: 8980835 -> 8980833 (-0.00%) instructions in affected programs: 45 -> 43 (-4.44%) helped: 1 HURT: 0 total cycles in shared programs: 70077904 -> 70077900 (-0.00%) cycles in affected programs: 122 -> 118 (-3.28%) helped: 1 HURT: 0 No changes on earlier platforms. v2: Modify the comments to look more like a proof. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
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)
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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)),
nir: Transform expressions of b2f(a) and b2f(b) to a || b All Gen6+ platforms had pretty similar results. (Skylake shown) total instructions in shared programs: 14277184 -> 14276964 (<.01%) instructions in affected programs: 10082 -> 9862 (-2.18%) helped: 37 HURT: 1 helped stats (abs) min: 1 max: 30 x̄: 5.97 x̃: 4 helped stats (rel) min: 0.14% max: 16.00% x̄: 5.23% x̃: 2.04% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.70% max: 0.70% x̄: 0.70% x̃: 0.70% 95% mean confidence interval for instructions value: -7.87 -3.71 95% mean confidence interval for instructions %-change: -6.98% -3.16% Instructions are helped. total cycles in shared programs: 532577990 -> 532577062 (<.01%) cycles in affected programs: 170959 -> 170031 (-0.54%) helped: 33 HURT: 9 helped stats (abs) min: 2 max: 120 x̄: 30.91 x̃: 30 helped stats (rel) min: 0.02% max: 7.65% x̄: 2.66% x̃: 1.13% HURT stats (abs) min: 2 max: 24 x̄: 10.22 x̃: 8 HURT stats (rel) min: 0.09% max: 1.79% x̄: 0.61% x̃: 0.22% 95% mean confidence interval for cycles value: -31.23 -12.96 95% mean confidence interval for cycles %-change: -2.90% -1.02% Cycles are helped. Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 7781539 -> 7781301 (<.01%) instructions in affected programs: 10169 -> 9931 (-2.34%) helped: 32 HURT: 0 helped stats (abs) min: 2 max: 20 x̄: 7.44 x̃: 6 helped stats (rel) min: 0.47% max: 17.02% x̄: 4.03% x̃: 1.88% 95% mean confidence interval for instructions value: -9.53 -5.34 95% mean confidence interval for instructions %-change: -5.94% -2.12% Instructions are helped. total cycles in shared programs: 177878590 -> 177878932 (<.01%) cycles in affected programs: 78706 -> 79048 (0.43%) helped: 7 HURT: 21 helped stats (abs) min: 6 max: 34 x̄: 24.57 x̃: 28 helped stats (rel) min: 0.15% max: 8.33% x̄: 4.66% x̃: 6.37% HURT stats (abs) min: 2 max: 86 x̄: 24.48 x̃: 22 HURT stats (rel) min: 0.01% max: 4.28% x̄: 1.21% x̃: 0.70% 95% mean confidence interval for cycles value: 0.30 24.13 95% mean confidence interval for cycles %-change: -1.52% 1.01% Inconclusive result (%-change mean confidence interval includes 0). v2: s/fmin/fmax/. Noticed by Thomas Helland. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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)),
nir: Transform expressions of b2f(a) and b2f(b) to a || b All Gen6+ platforms had pretty similar results. (Skylake shown) total instructions in shared programs: 14277184 -> 14276964 (<.01%) instructions in affected programs: 10082 -> 9862 (-2.18%) helped: 37 HURT: 1 helped stats (abs) min: 1 max: 30 x̄: 5.97 x̃: 4 helped stats (rel) min: 0.14% max: 16.00% x̄: 5.23% x̃: 2.04% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.70% max: 0.70% x̄: 0.70% x̃: 0.70% 95% mean confidence interval for instructions value: -7.87 -3.71 95% mean confidence interval for instructions %-change: -6.98% -3.16% Instructions are helped. total cycles in shared programs: 532577990 -> 532577062 (<.01%) cycles in affected programs: 170959 -> 170031 (-0.54%) helped: 33 HURT: 9 helped stats (abs) min: 2 max: 120 x̄: 30.91 x̃: 30 helped stats (rel) min: 0.02% max: 7.65% x̄: 2.66% x̃: 1.13% HURT stats (abs) min: 2 max: 24 x̄: 10.22 x̃: 8 HURT stats (rel) min: 0.09% max: 1.79% x̄: 0.61% x̃: 0.22% 95% mean confidence interval for cycles value: -31.23 -12.96 95% mean confidence interval for cycles %-change: -2.90% -1.02% Cycles are helped. Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 7781539 -> 7781301 (<.01%) instructions in affected programs: 10169 -> 9931 (-2.34%) helped: 32 HURT: 0 helped stats (abs) min: 2 max: 20 x̄: 7.44 x̃: 6 helped stats (rel) min: 0.47% max: 17.02% x̄: 4.03% x̃: 1.88% 95% mean confidence interval for instructions value: -9.53 -5.34 95% mean confidence interval for instructions %-change: -5.94% -2.12% Instructions are helped. total cycles in shared programs: 177878590 -> 177878932 (<.01%) cycles in affected programs: 78706 -> 79048 (0.43%) helped: 7 HURT: 21 helped stats (abs) min: 6 max: 34 x̄: 24.57 x̃: 28 helped stats (rel) min: 0.15% max: 8.33% x̄: 4.66% x̃: 6.37% HURT stats (abs) min: 2 max: 86 x̄: 24.48 x̃: 22 HURT stats (rel) min: 0.01% max: 4.28% x̄: 1.21% x̃: 0.70% 95% mean confidence interval for cycles value: 0.30 24.13 95% mean confidence interval for cycles %-change: -1.52% 1.01% Inconclusive result (%-change mean confidence interval includes 0). v2: s/fmin/fmax/. Noticed by Thomas Helland. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2018-08-02 01:18:07 +01:00
nir: Transform expressions of b2f(a) and b2f(b) to !(a || b) All Gen6+ platforms had similar results. (Skylake shown) total instructions in shared programs: 14276961 -> 14276892 (<.01%) instructions in affected programs: 3215 -> 3146 (-2.15%) helped: 28 HURT: 0 helped stats (abs) min: 1 max: 6 x̄: 2.46 x̃: 2 helped stats (rel) min: 0.47% max: 9.52% x̄: 4.34% x̃: 1.92% 95% mean confidence interval for instructions value: -2.87 -2.06 95% mean confidence interval for instructions %-change: -5.73% -2.95% Instructions are helped. total cycles in shared programs: 532577068 -> 532578400 (<.01%) cycles in affected programs: 121864 -> 123196 (1.09%) helped: 35 HURT: 30 helped stats (abs) min: 2 max: 268 x̄: 42.34 x̃: 22 helped stats (rel) min: 0.12% max: 12.14% x̄: 3.22% x̃: 1.86% HURT stats (abs) min: 2 max: 246 x̄: 93.80 x̃: 36 HURT stats (rel) min: 0.09% max: 13.63% x̄: 4.47% x̃: 2.58% 95% mean confidence interval for cycles value: -5.02 46.01 95% mean confidence interval for cycles %-change: -0.99% 1.65% Inconclusive result (value mean confidence interval includes 0). Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 7781299 -> 7781342 (<.01%) instructions in affected programs: 22300 -> 22343 (0.19%) helped: 13 HURT: 40 helped stats (abs) min: 2 max: 3 x̄: 2.85 x̃: 3 helped stats (rel) min: 1.15% max: 7.69% x̄: 3.72% x̃: 3.33% HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 HURT stats (rel) min: 0.26% max: 1.30% x̄: 0.47% x̃: 0.43% 95% mean confidence interval for instructions value: 0.23 1.39 95% mean confidence interval for instructions %-change: -1.18% 0.07% Inconclusive result (%-change mean confidence interval includes 0). total cycles in shared programs: 177878928 -> 177879332 (<.01%) cycles in affected programs: 383298 -> 383702 (0.11%) helped: 7 HURT: 43 helped stats (abs) min: 2 max: 18 x̄: 10.00 x̃: 10 helped stats (rel) min: 0.17% max: 4.81% x̄: 2.62% x̃: 3.40% HURT stats (abs) min: 2 max: 38 x̄: 11.02 x̃: 12 HURT stats (rel) min: 0.08% max: 1.54% x̄: 0.25% x̃: 0.09% 95% mean confidence interval for cycles value: 5.21 10.95 95% mean confidence interval for cycles %-change: -0.51% 0.21% Inconclusive result (%-change mean confidence interval includes 0). v2: s/fmin/fmax/. Noticed by Thomas Helland. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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)
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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))),
nir: Transform expressions of b2f(a) and b2f(b) to !(a || b) All Gen6+ platforms had similar results. (Skylake shown) total instructions in shared programs: 14276961 -> 14276892 (<.01%) instructions in affected programs: 3215 -> 3146 (-2.15%) helped: 28 HURT: 0 helped stats (abs) min: 1 max: 6 x̄: 2.46 x̃: 2 helped stats (rel) min: 0.47% max: 9.52% x̄: 4.34% x̃: 1.92% 95% mean confidence interval for instructions value: -2.87 -2.06 95% mean confidence interval for instructions %-change: -5.73% -2.95% Instructions are helped. total cycles in shared programs: 532577068 -> 532578400 (<.01%) cycles in affected programs: 121864 -> 123196 (1.09%) helped: 35 HURT: 30 helped stats (abs) min: 2 max: 268 x̄: 42.34 x̃: 22 helped stats (rel) min: 0.12% max: 12.14% x̄: 3.22% x̃: 1.86% HURT stats (abs) min: 2 max: 246 x̄: 93.80 x̃: 36 HURT stats (rel) min: 0.09% max: 13.63% x̄: 4.47% x̃: 2.58% 95% mean confidence interval for cycles value: -5.02 46.01 95% mean confidence interval for cycles %-change: -0.99% 1.65% Inconclusive result (value mean confidence interval includes 0). Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 7781299 -> 7781342 (<.01%) instructions in affected programs: 22300 -> 22343 (0.19%) helped: 13 HURT: 40 helped stats (abs) min: 2 max: 3 x̄: 2.85 x̃: 3 helped stats (rel) min: 1.15% max: 7.69% x̄: 3.72% x̃: 3.33% HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 HURT stats (rel) min: 0.26% max: 1.30% x̄: 0.47% x̃: 0.43% 95% mean confidence interval for instructions value: 0.23 1.39 95% mean confidence interval for instructions %-change: -1.18% 0.07% Inconclusive result (%-change mean confidence interval includes 0). total cycles in shared programs: 177878928 -> 177879332 (<.01%) cycles in affected programs: 383298 -> 383702 (0.11%) helped: 7 HURT: 43 helped stats (abs) min: 2 max: 18 x̄: 10.00 x̃: 10 helped stats (rel) min: 0.17% max: 4.81% x̄: 2.62% x̃: 3.40% HURT stats (abs) min: 2 max: 38 x̄: 11.02 x̃: 12 HURT stats (rel) min: 0.08% max: 1.54% x̄: 0.25% x̃: 0.09% 95% mean confidence interval for cycles value: 5.21 10.95 95% mean confidence interval for cycles %-change: -0.51% 0.21% Inconclusive result (%-change mean confidence interval includes 0). v2: s/fmin/fmax/. Noticed by Thomas Helland. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2018-08-02 03:27:01 +01:00
nir/algebraic: Strength reduce some compares of x and -x Converting the x vs -x comparison to an x vs 0 comparison enable cmod propagation to help. The seems to be a win everywhere except Gen7. Skylake and Broadwell had similar results. (Broadwell shown) total instructions in shared programs: 15566733 -> 15566014 (<.01%) instructions in affected programs: 72617 -> 71898 (-0.99%) helped: 302 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 2.38 x̃: 2 helped stats (rel) min: 0.15% max: 7.69% x̄: 1.28% x̃: 0.98% 95% mean confidence interval for instructions value: -2.55 -2.21 95% mean confidence interval for instructions %-change: -1.40% -1.16% Instructions are helped. total cycles in shared programs: 413014786 -> 413015475 (<.01%) cycles in affected programs: 707594 -> 708283 (0.10%) helped: 227 HURT: 101 helped stats (abs) min: 1 max: 612 x̄: 36.07 x̃: 20 helped stats (rel) min: 0.04% max: 19.39% x̄: 2.25% x̃: 1.49% HURT stats (abs) min: 2 max: 334 x̄: 87.90 x̃: 45 HURT stats (rel) min: 0.07% max: 14.51% x̄: 4.54% x̃: 3.36% 95% mean confidence interval for cycles value: -8.12 12.32 95% mean confidence interval for cycles %-change: -0.67% 0.34% Inconclusive result (value mean confidence interval includes 0). Haswell and Ivy Bridge had similar results. (Haswell shown) total instructions in shared programs: 13828220 -> 13827881 (<.01%) instructions in affected programs: 60887 -> 60548 (-0.56%) helped: 253 HURT: 6 helped stats (abs) min: 1 max: 5 x̄: 1.36 x̃: 1 helped stats (rel) min: 0.16% max: 3.85% x̄: 0.81% x̃: 0.64% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.26% max: 0.89% x̄: 0.47% x̃: 0.27% 95% mean confidence interval for instructions value: -1.39 -1.23 95% mean confidence interval for instructions %-change: -0.85% -0.70% Instructions are helped. total cycles in shared programs: 386870095 -> 386894412 (<.01%) cycles in affected programs: 1537307 -> 1561624 (1.58%) helped: 127 HURT: 188 helped stats (abs) min: 1 max: 381 x̄: 17.89 x̃: 4 helped stats (rel) min: 0.02% max: 14.33% x̄: 1.00% x̃: 0.33% HURT stats (abs) min: 2 max: 5585 x̄: 141.43 x̃: 14 HURT stats (rel) min: 0.03% max: 11.50% x̄: 1.65% x̃: 1.06% 95% mean confidence interval for cycles value: 21.95 132.45 95% mean confidence interval for cycles %-change: 0.32% 0.85% Cycles are HURT. Sandy Bridge total instructions in shared programs: 10896339 -> 10896276 (<.01%) instructions in affected programs: 10757 -> 10694 (-0.59%) helped: 49 HURT: 0 helped stats (abs) min: 1 max: 2 x̄: 1.29 x̃: 1 helped stats (rel) min: 0.12% max: 1.85% x̄: 0.87% x̃: 0.89% 95% mean confidence interval for instructions value: -1.42 -1.15 95% mean confidence interval for instructions %-change: -1.03% -0.72% Instructions are helped. total cycles in shared programs: 155091003 -> 155090480 (<.01%) cycles in affected programs: 102761 -> 102238 (-0.51%) helped: 51 HURT: 0 helped stats (abs) min: 1 max: 36 x̄: 10.25 x̃: 4 helped stats (rel) min: 0.02% max: 2.57% x̄: 0.76% x̃: 0.36% 95% mean confidence interval for cycles value: -12.98 -7.53 95% mean confidence interval for cycles %-change: -0.97% -0.56% Cycles are helped. Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 8234667 -> 8234652 (<.01%) instructions in affected programs: 2063 -> 2048 (-0.73%) helped: 15 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.30% max: 1.56% x̄: 0.82% x̃: 0.81% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -0.97% -0.67% Instructions are helped. total cycles in shared programs: 188700906 -> 188700598 (<.01%) cycles in affected programs: 283480 -> 283172 (-0.11%) helped: 83 HURT: 3 helped stats (abs) min: 2 max: 8 x̄: 3.78 x̃: 4 helped stats (rel) min: 0.04% max: 0.55% x̄: 0.15% x̃: 0.12% HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 HURT stats (rel) min: 0.02% max: 0.04% x̄: 0.03% x̃: 0.04% 95% mean confidence interval for cycles value: -3.87 -3.29 95% mean confidence interval for cycles %-change: -0.16% -0.12% Cycles are helped. Reviewed-by: Matt Turner <mattst88@gmail.com>
2018-12-18 06:29:26 +00:00
(('flt', a, ('fneg', a)), ('flt', a, 0.0)),
(('fge', a, ('fneg', a)), ('fge', a, 0.0)),
nir: Transform expressions of b2f(a) and b2f(b) to a || b All Gen6+ platforms had pretty similar results. (Skylake shown) total instructions in shared programs: 14277184 -> 14276964 (<.01%) instructions in affected programs: 10082 -> 9862 (-2.18%) helped: 37 HURT: 1 helped stats (abs) min: 1 max: 30 x̄: 5.97 x̃: 4 helped stats (rel) min: 0.14% max: 16.00% x̄: 5.23% x̃: 2.04% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.70% max: 0.70% x̄: 0.70% x̃: 0.70% 95% mean confidence interval for instructions value: -7.87 -3.71 95% mean confidence interval for instructions %-change: -6.98% -3.16% Instructions are helped. total cycles in shared programs: 532577990 -> 532577062 (<.01%) cycles in affected programs: 170959 -> 170031 (-0.54%) helped: 33 HURT: 9 helped stats (abs) min: 2 max: 120 x̄: 30.91 x̃: 30 helped stats (rel) min: 0.02% max: 7.65% x̄: 2.66% x̃: 1.13% HURT stats (abs) min: 2 max: 24 x̄: 10.22 x̃: 8 HURT stats (rel) min: 0.09% max: 1.79% x̄: 0.61% x̃: 0.22% 95% mean confidence interval for cycles value: -31.23 -12.96 95% mean confidence interval for cycles %-change: -2.90% -1.02% Cycles are helped. Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 7781539 -> 7781301 (<.01%) instructions in affected programs: 10169 -> 9931 (-2.34%) helped: 32 HURT: 0 helped stats (abs) min: 2 max: 20 x̄: 7.44 x̃: 6 helped stats (rel) min: 0.47% max: 17.02% x̄: 4.03% x̃: 1.88% 95% mean confidence interval for instructions value: -9.53 -5.34 95% mean confidence interval for instructions %-change: -5.94% -2.12% Instructions are helped. total cycles in shared programs: 177878590 -> 177878932 (<.01%) cycles in affected programs: 78706 -> 79048 (0.43%) helped: 7 HURT: 21 helped stats (abs) min: 6 max: 34 x̄: 24.57 x̃: 28 helped stats (rel) min: 0.15% max: 8.33% x̄: 4.66% x̃: 6.37% HURT stats (abs) min: 2 max: 86 x̄: 24.48 x̃: 22 HURT stats (rel) min: 0.01% max: 4.28% x̄: 1.21% x̃: 0.70% 95% mean confidence interval for cycles value: 0.30 24.13 95% mean confidence interval for cycles %-change: -1.52% 1.01% Inconclusive result (%-change mean confidence interval includes 0). v2: s/fmin/fmax/. Noticed by Thomas Helland. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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.
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-12-06 03:36:10 +00:00
(('flt', ('fmin', c, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
nir: Transform expressions of b2f(a) and b2f(b) to a || b All Gen6+ platforms had pretty similar results. (Skylake shown) total instructions in shared programs: 14277184 -> 14276964 (<.01%) instructions in affected programs: 10082 -> 9862 (-2.18%) helped: 37 HURT: 1 helped stats (abs) min: 1 max: 30 x̄: 5.97 x̃: 4 helped stats (rel) min: 0.14% max: 16.00% x̄: 5.23% x̃: 2.04% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.70% max: 0.70% x̄: 0.70% x̃: 0.70% 95% mean confidence interval for instructions value: -7.87 -3.71 95% mean confidence interval for instructions %-change: -6.98% -3.16% Instructions are helped. total cycles in shared programs: 532577990 -> 532577062 (<.01%) cycles in affected programs: 170959 -> 170031 (-0.54%) helped: 33 HURT: 9 helped stats (abs) min: 2 max: 120 x̄: 30.91 x̃: 30 helped stats (rel) min: 0.02% max: 7.65% x̄: 2.66% x̃: 1.13% HURT stats (abs) min: 2 max: 24 x̄: 10.22 x̃: 8 HURT stats (rel) min: 0.09% max: 1.79% x̄: 0.61% x̃: 0.22% 95% mean confidence interval for cycles value: -31.23 -12.96 95% mean confidence interval for cycles %-change: -2.90% -1.02% Cycles are helped. Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 7781539 -> 7781301 (<.01%) instructions in affected programs: 10169 -> 9931 (-2.34%) helped: 32 HURT: 0 helped stats (abs) min: 2 max: 20 x̄: 7.44 x̃: 6 helped stats (rel) min: 0.47% max: 17.02% x̄: 4.03% x̃: 1.88% 95% mean confidence interval for instructions value: -9.53 -5.34 95% mean confidence interval for instructions %-change: -5.94% -2.12% Instructions are helped. total cycles in shared programs: 177878590 -> 177878932 (<.01%) cycles in affected programs: 78706 -> 79048 (0.43%) helped: 7 HURT: 21 helped stats (abs) min: 6 max: 34 x̄: 24.57 x̃: 28 helped stats (rel) min: 0.15% max: 8.33% x̄: 4.66% x̃: 6.37% HURT stats (abs) min: 2 max: 86 x̄: 24.48 x̃: 22 HURT stats (rel) min: 0.01% max: 4.28% x̄: 1.21% x̃: 0.70% 95% mean confidence interval for cycles value: 0.30 24.13 95% mean confidence interval for cycles %-change: -1.52% 1.01% Inconclusive result (%-change mean confidence interval includes 0). v2: s/fmin/fmax/. Noticed by Thomas Helland. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2018-08-02 01:18:07 +01:00
('ior', ('flt', c, 0.0), ('ior', a, b))),
nir: Simplify some comparisons like a+b < a All Gen7+ platforms had similar results. (Skylake shown) total instructions in shared programs: 14514555 -> 14514547 (<.01%) instructions in affected programs: 1972 -> 1964 (-0.41%) helped: 8 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.39% max: 0.42% x̄: 0.41% x̃: 0.41% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -0.41% -0.40% Instructions are helped. total cycles in shared programs: 533141444 -> 533136780 (<.01%) cycles in affected programs: 164728 -> 160064 (-2.83%) helped: 181 HURT: 3 helped stats (abs) min: 2 max: 94 x̄: 26.17 x̃: 30 helped stats (rel) min: 0.12% max: 5.33% x̄: 3.42% x̃: 3.80% HURT stats (abs) min: 4 max: 54 x̄: 24.00 x̃: 14 HURT stats (rel) min: 0.20% max: 2.39% x̄: 1.09% x̃: 0.68% 95% mean confidence interval for cycles value: -27.12 -23.58 95% mean confidence interval for cycles %-change: -3.54% -3.16% Cycles are helped. Sandy Bridge total instructions in shared programs: 10533667 -> 10533539 (<.01%) instructions in affected programs: 10148 -> 10020 (-1.26%) helped: 124 HURT: 0 helped stats (abs) min: 1 max: 2 x̄: 1.03 x̃: 1 helped stats (rel) min: 0.39% max: 4.35% x̄: 2.20% x̃: 2.04% 95% mean confidence interval for instructions value: -1.06 -1.00 95% mean confidence interval for instructions %-change: -2.46% -1.95% Instructions are helped. total cycles in shared programs: 146136887 -> 146132122 (<.01%) cycles in affected programs: 206382 -> 201617 (-2.31%) helped: 171 HURT: 0 helped stats (abs) min: 2 max: 40 x̄: 27.87 x̃: 30 helped stats (rel) min: 0.08% max: 5.73% x̄: 2.98% x̃: 2.67% 95% mean confidence interval for cycles value: -29.19 -26.54 95% mean confidence interval for cycles %-change: -3.20% -2.76% Cycles are helped. Iron Lake total instructions in shared programs: 7886515 -> 7886507 (<.01%) instructions in affected programs: 3016 -> 3008 (-0.27%) helped: 8 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.25% max: 0.28% x̄: 0.27% x̃: 0.27% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -0.27% -0.26% Instructions are helped. total cycles in shared programs: 178100396 -> 178100388 (<.01%) cycles in affected programs: 156128 -> 156120 (<.01%) helped: 4 HURT: 4 helped stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4 helped stats (rel) min: 0.02% max: 0.04% x̄: 0.03% x̃: 0.03% HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 HURT stats (rel) min: <.01% max: 0.01% x̄: <.01% x̃: <.01% 95% mean confidence interval for cycles value: -3.68 1.68 95% mean confidence interval for cycles %-change: -0.03% <.01% Inconclusive result (value mean confidence interval includes 0). GM45 total instructions in shared programs: 4857872 -> 4857868 (<.01%) instructions in affected programs: 1544 -> 1540 (-0.26%) helped: 4 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.25% max: 0.27% x̄: 0.26% x̃: 0.26% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -0.28% -0.24% Instructions are helped. total cycles in shared programs: 122167654 -> 122167662 (<.01%) cycles in affected programs: 96248 -> 96256 (<.01%) helped: 0 HURT: 4 HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 HURT stats (rel) min: <.01% max: 0.01% x̄: <.01% x̃: <.01% 95% mean confidence interval for cycles value: 2.00 2.00 95% mean confidence interval for cycles %-change: <.01% 0.02% Cycles are HURT. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
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)),
nir: Replace an odd comparison involving fmin of -b2f I noticed the fge version while looking at a shader for an unrelated reason. The feq version prevents a regression in a later change that performs strength reduction of some compares. Broadwell and Skylake had similar results. (Skylake shown) total instructions in shared programs: 14514808 -> 14514796 (<.01%) instructions in affected programs: 750 -> 738 (-1.60%) helped: 4 HURT: 0 helped stats (abs) min: 1 max: 5 x̄: 3.00 x̃: 3 helped stats (rel) min: 0.83% max: 1.96% x̄: 1.40% x̃: 1.40% 95% mean confidence interval for instructions value: -6.67 0.67 95% mean confidence interval for instructions %-change: -2.43% -0.36% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 533144939 -> 533144853 (<.01%) cycles in affected programs: 8911 -> 8825 (-0.97%) helped: 4 HURT: 0 helped stats (abs) min: 16 max: 32 x̄: 21.50 x̃: 19 helped stats (rel) min: 0.60% max: 1.89% x̄: 1.28% x̃: 1.31% 95% mean confidence interval for cycles value: -32.94 -10.06 95% mean confidence interval for cycles %-change: -2.30% -0.26% Cycles are helped. Haswell total instructions in shared programs: 13093785 -> 13093775 (<.01%) instructions in affected programs: 924 -> 914 (-1.08%) helped: 4 HURT: 2 helped stats (abs) min: 1 max: 5 x̄: 3.00 x̃: 3 helped stats (rel) min: 0.82% max: 1.95% x̄: 1.39% x̃: 1.39% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 1.19% max: 1.19% x̄: 1.19% x̃: 1.19% 95% mean confidence interval for instructions value: -4.53 1.20 95% mean confidence interval for instructions %-change: -2.02% 0.97% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 409580553 -> 409580118 (<.01%) cycles in affected programs: 10909 -> 10474 (-3.99%) helped: 5 HURT: 1 helped stats (abs) min: 6 max: 222 x̄: 89.60 x̃: 18 helped stats (rel) min: 0.16% max: 24.72% x̄: 9.54% x̃: 1.78% HURT stats (abs) min: 13 max: 13 x̄: 13.00 x̃: 13 HURT stats (rel) min: 0.39% max: 0.39% x̄: 0.39% x̃: 0.39% 95% mean confidence interval for cycles value: -180.68 35.68 95% mean confidence interval for cycles %-change: -19.55% 3.79% Inconclusive result (value mean confidence interval includes 0). Ivy Bridge total instructions in shared programs: 11811851 -> 11811840 (<.01%) instructions in affected programs: 1032 -> 1021 (-1.07%) helped: 5 HURT: 1 helped stats (abs) min: 1 max: 5 x̄: 2.40 x̃: 1 helped stats (rel) min: 0.63% max: 1.95% x̄: 1.13% x̃: 0.97% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 1.19% max: 1.19% x̄: 1.19% x̃: 1.19% 95% mean confidence interval for instructions value: -4.17 0.51 95% mean confidence interval for instructions %-change: -1.86% 0.36% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 257618403 -> 257618168 (<.01%) cycles in affected programs: 10784 -> 10549 (-2.18%) helped: 4 HURT: 2 helped stats (abs) min: 4 max: 220 x̄: 64.50 x̃: 17 helped stats (rel) min: 0.50% max: 24.34% x̄: 7.07% x̃: 1.72% HURT stats (abs) min: 9 max: 14 x̄: 11.50 x̃: 11 HURT stats (rel) min: 0.24% max: 0.42% x̄: 0.33% x̃: 0.33% 95% mean confidence interval for cycles value: -133.11 54.78 95% mean confidence interval for cycles %-change: -14.79% 5.59% Inconclusive result (value mean confidence interval includes 0). GM45, Iron Lake, and Sandy Bridge had similar results. (Sandy Bridge shown) total instructions in shared programs: 10533871 -> 10533859 (<.01%) instructions in affected programs: 865 -> 853 (-1.39%) helped: 4 HURT: 0 helped stats (abs) min: 1 max: 5 x̄: 3.00 x̃: 3 helped stats (rel) min: 0.63% max: 1.83% x̄: 1.22% x̃: 1.21% 95% mean confidence interval for instructions value: -6.67 0.67 95% mean confidence interval for instructions %-change: -2.16% -0.29% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 146139904 -> 146139852 (<.01%) cycles in affected programs: 15213 -> 15161 (-0.34%) helped: 4 HURT: 0 helped stats (abs) min: 3 max: 18 x̄: 13.00 x̃: 15 helped stats (rel) min: 0.15% max: 0.84% x̄: 0.39% x̃: 0.29% 95% mean confidence interval for cycles value: -23.79 -2.21 95% mean confidence interval for cycles %-change: -0.88% 0.09% Inconclusive result (%-change mean confidence interval includes 0). Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
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.
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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))),
nir: Replace an odd comparison involving fmin of -b2f I noticed the fge version while looking at a shader for an unrelated reason. The feq version prevents a regression in a later change that performs strength reduction of some compares. Broadwell and Skylake had similar results. (Skylake shown) total instructions in shared programs: 14514808 -> 14514796 (<.01%) instructions in affected programs: 750 -> 738 (-1.60%) helped: 4 HURT: 0 helped stats (abs) min: 1 max: 5 x̄: 3.00 x̃: 3 helped stats (rel) min: 0.83% max: 1.96% x̄: 1.40% x̃: 1.40% 95% mean confidence interval for instructions value: -6.67 0.67 95% mean confidence interval for instructions %-change: -2.43% -0.36% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 533144939 -> 533144853 (<.01%) cycles in affected programs: 8911 -> 8825 (-0.97%) helped: 4 HURT: 0 helped stats (abs) min: 16 max: 32 x̄: 21.50 x̃: 19 helped stats (rel) min: 0.60% max: 1.89% x̄: 1.28% x̃: 1.31% 95% mean confidence interval for cycles value: -32.94 -10.06 95% mean confidence interval for cycles %-change: -2.30% -0.26% Cycles are helped. Haswell total instructions in shared programs: 13093785 -> 13093775 (<.01%) instructions in affected programs: 924 -> 914 (-1.08%) helped: 4 HURT: 2 helped stats (abs) min: 1 max: 5 x̄: 3.00 x̃: 3 helped stats (rel) min: 0.82% max: 1.95% x̄: 1.39% x̃: 1.39% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 1.19% max: 1.19% x̄: 1.19% x̃: 1.19% 95% mean confidence interval for instructions value: -4.53 1.20 95% mean confidence interval for instructions %-change: -2.02% 0.97% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 409580553 -> 409580118 (<.01%) cycles in affected programs: 10909 -> 10474 (-3.99%) helped: 5 HURT: 1 helped stats (abs) min: 6 max: 222 x̄: 89.60 x̃: 18 helped stats (rel) min: 0.16% max: 24.72% x̄: 9.54% x̃: 1.78% HURT stats (abs) min: 13 max: 13 x̄: 13.00 x̃: 13 HURT stats (rel) min: 0.39% max: 0.39% x̄: 0.39% x̃: 0.39% 95% mean confidence interval for cycles value: -180.68 35.68 95% mean confidence interval for cycles %-change: -19.55% 3.79% Inconclusive result (value mean confidence interval includes 0). Ivy Bridge total instructions in shared programs: 11811851 -> 11811840 (<.01%) instructions in affected programs: 1032 -> 1021 (-1.07%) helped: 5 HURT: 1 helped stats (abs) min: 1 max: 5 x̄: 2.40 x̃: 1 helped stats (rel) min: 0.63% max: 1.95% x̄: 1.13% x̃: 0.97% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 1.19% max: 1.19% x̄: 1.19% x̃: 1.19% 95% mean confidence interval for instructions value: -4.17 0.51 95% mean confidence interval for instructions %-change: -1.86% 0.36% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 257618403 -> 257618168 (<.01%) cycles in affected programs: 10784 -> 10549 (-2.18%) helped: 4 HURT: 2 helped stats (abs) min: 4 max: 220 x̄: 64.50 x̃: 17 helped stats (rel) min: 0.50% max: 24.34% x̄: 7.07% x̃: 1.72% HURT stats (abs) min: 9 max: 14 x̄: 11.50 x̃: 11 HURT stats (rel) min: 0.24% max: 0.42% x̄: 0.33% x̃: 0.33% 95% mean confidence interval for cycles value: -133.11 54.78 95% mean confidence interval for cycles %-change: -14.79% 5.59% Inconclusive result (value mean confidence interval includes 0). GM45, Iron Lake, and Sandy Bridge had similar results. (Sandy Bridge shown) total instructions in shared programs: 10533871 -> 10533859 (<.01%) instructions in affected programs: 865 -> 853 (-1.39%) helped: 4 HURT: 0 helped stats (abs) min: 1 max: 5 x̄: 3.00 x̃: 3 helped stats (rel) min: 0.63% max: 1.83% x̄: 1.22% x̃: 1.21% 95% mean confidence interval for instructions value: -6.67 0.67 95% mean confidence interval for instructions %-change: -2.16% -0.29% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 146139904 -> 146139852 (<.01%) cycles in affected programs: 15213 -> 15161 (-0.34%) helped: 4 HURT: 0 helped stats (abs) min: 3 max: 18 x̄: 13.00 x̃: 15 helped stats (rel) min: 0.15% max: 0.84% x̄: 0.39% x̃: 0.29% 95% mean confidence interval for cycles value: -23.79 -2.21 95% mean confidence interval for cycles %-change: -0.88% 0.09% Inconclusive result (%-change mean confidence interval includes 0). Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2018-01-31 19:11:02 +00:00
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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),
nir: Don't compare b2f or b2i with zero All of the shaders that had loops changed were in Tomb Raider. The one shader that lost SIMD16 is one of those. Skylake total instructions in shared programs: 14391653 -> 14390468 (<.01%) instructions in affected programs: 111891 -> 110706 (-1.06%) helped: 501 HURT: 0 helped stats (abs) min: 1 max: 155 x̄: 2.37 x̃: 1 helped stats (rel) min: 0.05% max: 21.54% x̄: 1.61% x̃: 1.01% 95% mean confidence interval for instructions value: -3.23 -1.50 95% mean confidence interval for instructions %-change: -1.77% -1.45% Instructions are helped. total cycles in shared programs: 532793024 -> 532776598 (<.01%) cycles in affected programs: 987682 -> 971256 (-1.66%) helped: 348 nnHURT: 41 helped stats (abs) min: 1 max: 3074 x̄: 54.91 x̃: 18 helped stats (rel) min: 0.05% max: 32.24% x̄: 3.36% x̃: 1.68% HURT stats (abs) min: 1 max: 422 x̄: 65.39 x̃: 24 HURT stats (rel) min: 0.09% max: 39.29% x̄: 9.50% x̃: 2.02% 95% mean confidence interval for cycles value: -64.08 -20.38 95% mean confidence interval for cycles %-change: -2.78% -1.23% Cycles are helped. total loops in shared programs: 4854 -> 4829 (-0.52%) loops in affected programs: 27 -> 2 (-92.59%) helped: 18 HURT: 0 LOST: 1 GAINED: 0 Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
2016-03-02 03:05:14 +00:00
nir/opt_algebraic: Drop bit-size suffixes from conversions Suffixes are dropped from a bunch of conversion opcodes when it makes sense to do so. Others are kept if we really do want the bit-size restriction. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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)),
nir: Don't compare i2f or u2i with zero Broadwell and Skylake had similar results. (Skylake shown) total instructions in shared programs: 14277620 -> 14277230 (<.01%) instructions in affected programs: 36905 -> 36515 (-1.06%) helped: 101 HURT: 6 helped stats (abs) min: 1 max: 6 x̄: 4.46 x̃: 6 helped stats (rel) min: 0.32% max: 7.69% x̄: 1.80% x̃: 1.51% HURT stats (abs) min: 1 max: 28 x̄: 10.00 x̃: 1 HURT stats (rel) min: 0.33% max: 1.74% x̄: 0.68% x̃: 0.47% 95% mean confidence interval for instructions value: -4.59 -2.70 95% mean confidence interval for instructions %-change: -1.90% -1.41% Instructions are helped. total cycles in shared programs: 532580716 -> 532577947 (<.01%) cycles in affected programs: 940575 -> 937806 (-0.29%) helped: 92 HURT: 12 helped stats (abs) min: 2 max: 158 x̄: 51.04 x̃: 62 helped stats (rel) min: 0.24% max: 3.99% x̄: 2.14% x̃: 2.41% HURT stats (abs) min: 10 max: 1112 x̄: 160.58 x̃: 63 HURT stats (rel) min: 0.06% max: 21.90% x̄: 4.22% x̃: 0.20% 95% mean confidence interval for cycles value: -50.66 -2.59 95% mean confidence interval for cycles %-change: -2.09% -0.73% Cycles are helped. total spills in shared programs: 8116 -> 8124 (0.10%) spills in affected programs: 200 -> 208 (4.00%) helped: 0 HURT: 2 total fills in shared programs: 11086 -> 11094 (0.07%) fills in affected programs: 436 -> 444 (1.83%) helped: 0 HURT: 2 Ivy Bridge and Haswell had similar results. (Haswell shown) total instructions in shared programs: 12979054 -> 12978067 (<.01%) instructions in affected programs: 33633 -> 32646 (-2.93%) helped: 120 HURT: 2 helped stats (abs) min: 1 max: 13 x̄: 8.53 x̃: 13 helped stats (rel) min: 0.30% max: 16.67% x̄: 4.55% x̃: 3.17% HURT stats (abs) min: 18 max: 18 x̄: 18.00 x̃: 18 HURT stats (rel) min: 1.15% max: 2.84% x̄: 2.00% x̃: 2.00% 95% mean confidence interval for instructions value: -9.19 -6.99 95% mean confidence interval for instructions %-change: -5.27% -3.62% Instructions are helped. total cycles in shared programs: 411212880 -> 411199636 (<.01%) cycles in affected programs: 696441 -> 683197 (-1.90%) helped: 107 HURT: 5 helped stats (abs) min: 2 max: 864 x̄: 124.90 x̃: 146 helped stats (rel) min: 0.03% max: 29.20% x̄: 8.58% x̃: 5.88% HURT stats (abs) min: 2 max: 50 x̄: 24.00 x̃: 22 HURT stats (rel) min: 0.01% max: 5.35% x̄: 1.29% x̃: 0.25% 95% mean confidence interval for cycles value: -136.96 -99.54 95% mean confidence interval for cycles %-change: -9.75% -6.53% Cycles are helped. total spills in shared programs: 78623 -> 78631 (0.01%) spills in affected programs: 66 -> 74 (12.12%) helped: 0 HURT: 2 total fills in shared programs: 80104 -> 80108 (<.01%) fills in affected programs: 133 -> 137 (3.01%) helped: 0 HURT: 2 No changes on Sandy Bridge, Iron Lake, or GM45. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2018-07-03 20:39:54 +01:00
nir: Simplify 0 < fabs(a) Sandy Bridge / Ivy Bridge / Haswell total instructions in shared programs: 8462180 -> 8462174 (-0.00%) instructions in affected programs: 564 -> 558 (-1.06%) helped: 6 HURT: 0 total cycles in shared programs: 117542462 -> 117542276 (-0.00%) cycles in affected programs: 9768 -> 9582 (-1.90%) helped: 12 HURT: 0 Broadwell / Skylake total instructions in shared programs: 8980833 -> 8980826 (-0.00%) instructions in affected programs: 626 -> 619 (-1.12%) helped: 7 HURT: 0 total cycles in shared programs: 70077900 -> 70077714 (-0.00%) cycles in affected programs: 9378 -> 9192 (-1.98%) helped: 12 HURT: 0 G45 and Ironlake showed no change. v2: Modify the comments to look more like a proof. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
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
nir: Mark the 0.0 < abs(a) transformation as imprecise Unlike the much older -abs(a) >= 0.0 transformation, this is not precise. The behavior changes if the source is NaN. No shader-db changes on any platform. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2018-08-01 18:13:55 +01:00
(('~flt', 0.0, ('fabs', a)), ('fne', a, 0.0)),
nir: Simplify 0 < fabs(a) Sandy Bridge / Ivy Bridge / Haswell total instructions in shared programs: 8462180 -> 8462174 (-0.00%) instructions in affected programs: 564 -> 558 (-1.06%) helped: 6 HURT: 0 total cycles in shared programs: 117542462 -> 117542276 (-0.00%) cycles in affected programs: 9768 -> 9582 (-1.90%) helped: 12 HURT: 0 Broadwell / Skylake total instructions in shared programs: 8980833 -> 8980826 (-0.00%) instructions in affected programs: 626 -> 619 (-1.12%) helped: 7 HURT: 0 total cycles in shared programs: 70077900 -> 70077714 (-0.00%) cycles in affected programs: 9378 -> 9192 (-1.98%) helped: 12 HURT: 0 G45 and Ironlake showed no change. v2: Modify the comments to look more like a proof. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-03-02 23:36:14 +00:00
nir: Transform -fabs(a) < 0 to a != 0 Unlike the much older -abs(a) >= 0.0 transformation, this is not precise. The behavior changes if a is NaN. All Gen platforms had pretty similar results. (Skylake shown) total instructions in shared programs: 14277216 -> 14277184 (<.01%) instructions in affected programs: 2300 -> 2268 (-1.39%) helped: 8 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 4.00 x̃: 3 helped stats (rel) min: 0.48% max: 15.15% x̄: 4.41% x̃: 1.01% 95% mean confidence interval for instructions value: -6.45 -1.55 95% mean confidence interval for instructions %-change: -9.96% 1.13% Inconclusive result (%-change mean confidence interval includes 0). total cycles in shared programs: 532577848 -> 532577990 (<.01%) cycles in affected programs: 17486 -> 17628 (0.81%) helped: 2 HURT: 5 helped stats (abs) min: 2 max: 6 x̄: 4.00 x̃: 4 helped stats (rel) min: 0.06% max: 1.81% x̄: 0.93% x̃: 0.93% HURT stats (abs) min: 6 max: 50 x̄: 30.00 x̃: 26 HURT stats (rel) min: 0.55% max: 2.17% x̄: 1.19% x̃: 1.02% 95% mean confidence interval for cycles value: -1.06 41.63 95% mean confidence interval for cycles %-change: -0.58% 1.74% Inconclusive result (value mean confidence interval includes 0). Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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)),
nir: Transform -fabs(a) >= 0 to a == 0 All Gen platforms had pretty similar results. (Skylake shown) total instructions in shared programs: 14276964 -> 14276961 (<.01%) instructions in affected programs: 411 -> 408 (-0.73%) helped: 3 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.47% max: 1.96% x̄: 1.04% x̃: 0.68% total cycles in shared programs: 532577062 -> 532577068 (<.01%) cycles in affected programs: 1093 -> 1099 (0.55%) helped: 1 HURT: 1 helped stats (abs) min: 16 max: 16 x̄: 16.00 x̃: 16 helped stats (rel) min: 7.77% max: 7.77% x̄: 7.77% x̃: 7.77% HURT stats (abs) min: 22 max: 22 x̄: 22.00 x̃: 22 HURT stats (rel) min: 2.48% max: 2.48% x̄: 2.48% x̃: 2.48% Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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)),
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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)))),
nir: Simplify min and max of b2f v2: Rebase on almost 2 years. Require that one of the arguments to fmin or fmax be used only once. This prevents some regressions. shader-db results: Skylake and Broadwell had similar results. Skylake shown. total instructions in shared programs: 14526021 -> 14525913 (<.01%) instructions in affected programs: 4613 -> 4505 (-2.34%) helped: 31 HURT: 0 helped stats (abs) min: 1 max: 4 x̄: 3.48 x̃: 4 helped stats (rel) min: 0.62% max: 6.67% x̄: 3.31% x̃: 2.42% total cycles in shared programs: 533118710 -> 533118403 (<.01%) cycles in affected programs: 34334 -> 34027 (-0.89%) helped: 24 HURT: 0 helped stats (abs) min: 4 max: 24 x̄: 12.79 x̃: 14 helped stats (rel) min: 0.25% max: 2.40% x̄: 1.08% x̃: 1.03% No changes on GM45, Iron Lake, Sandy Bridge, Ivy Bridge, or Haswell. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
2016-03-08 19:11:00 +00:00
nir: Replace fmin(b2f(a), b) with a bcsel All of the affected shaders are HDR mappers from Serious Sam 3. All Gen7+ platforms had similar results. (Skylake shown) total instructions in shared programs: 14516285 -> 14516273 (<.01%) instructions in affected programs: 348 -> 336 (-3.45%) helped: 12 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 2.08% max: 6.67% x̄: 4.31% x̃: 4.17% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -5.55% -3.06% Instructions are helped. total cycles in shared programs: 533163876 -> 533163808 (<.01%) cycles in affected programs: 1144 -> 1076 (-5.94%) helped: 4 HURT: 0 helped stats (abs) min: 16 max: 18 x̄: 17.00 x̃: 17 helped stats (rel) min: 5.80% max: 6.08% x̄: 5.94% x̃: 5.94% 95% mean confidence interval for cycles value: -18.84 -15.16 95% mean confidence interval for cycles %-change: -6.20% -5.68% Cycles are helped. Sandy Bridge total instructions in shared programs: 10533321 -> 10533309 (<.01%) instructions in affected programs: 372 -> 360 (-3.23%) helped: 12 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 2.00% max: 5.88% x̄: 3.91% x̃: 3.85% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -4.96% -2.86% Instructions are helped. total cycles in shared programs: 146136632 -> 146136428 (<.01%) cycles in affected programs: 11668 -> 11464 (-1.75%) helped: 12 HURT: 0 helped stats (abs) min: 16 max: 18 x̄: 17.00 x̃: 17 helped stats (rel) min: 0.99% max: 3.44% x̄: 2.20% x̃: 2.29% 95% mean confidence interval for cycles value: -17.66 -16.34 95% mean confidence interval for cycles %-change: -2.82% -1.58% Cycles are helped. Iron Lake total instructions in shared programs: 7886301 -> 7886277 (<.01%) instructions in affected programs: 576 -> 552 (-4.17%) helped: 12 HURT: 0 helped stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 helped stats (rel) min: 2.94% max: 6.06% x̄: 4.51% x̃: 4.65% 95% mean confidence interval for instructions value: -2.00 -2.00 95% mean confidence interval for instructions %-change: -5.30% -3.72% Instructions are helped. total cycles in shared programs: 178113176 -> 178113176 (0.00%) cycles in affected programs: 2116 -> 2116 (0.00%) helped: 2 HURT: 4 helped stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4 helped stats (rel) min: 1.14% max: 1.14% x̄: 1.14% x̃: 1.14% HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 HURT stats (rel) min: 0.50% max: 0.65% x̄: 0.58% x̃: 0.58% 95% mean confidence interval for cycles value: -3.25 3.25 95% mean confidence interval for cycles %-change: -0.93% 0.94% Inconclusive result (value mean confidence interval includes 0). GM45 total instructions in shared programs: 4857756 -> 4857744 (<.01%) instructions in affected programs: 294 -> 282 (-4.08%) helped: 6 HURT: 0 helped stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 helped stats (rel) min: 2.94% max: 5.71% x̄: 4.40% x̃: 4.55% 95% mean confidence interval for instructions value: -2.00 -2.00 95% mean confidence interval for instructions %-change: -5.71% -3.09% Instructions are helped. total cycles in shared programs: 122178730 -> 122178722 (<.01%) cycles in affected programs: 700 -> 692 (-1.14%) helped: 2 HURT: 0 Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
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.
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-12-06 03:36:10 +00:00
(('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a, ('fmin', b, 1.0), ('fmin', b, 0.0))),
nir: Replace fmin(b2f(a), b) with a bcsel All of the affected shaders are HDR mappers from Serious Sam 3. All Gen7+ platforms had similar results. (Skylake shown) total instructions in shared programs: 14516285 -> 14516273 (<.01%) instructions in affected programs: 348 -> 336 (-3.45%) helped: 12 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 2.08% max: 6.67% x̄: 4.31% x̃: 4.17% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -5.55% -3.06% Instructions are helped. total cycles in shared programs: 533163876 -> 533163808 (<.01%) cycles in affected programs: 1144 -> 1076 (-5.94%) helped: 4 HURT: 0 helped stats (abs) min: 16 max: 18 x̄: 17.00 x̃: 17 helped stats (rel) min: 5.80% max: 6.08% x̄: 5.94% x̃: 5.94% 95% mean confidence interval for cycles value: -18.84 -15.16 95% mean confidence interval for cycles %-change: -6.20% -5.68% Cycles are helped. Sandy Bridge total instructions in shared programs: 10533321 -> 10533309 (<.01%) instructions in affected programs: 372 -> 360 (-3.23%) helped: 12 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 2.00% max: 5.88% x̄: 3.91% x̃: 3.85% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -4.96% -2.86% Instructions are helped. total cycles in shared programs: 146136632 -> 146136428 (<.01%) cycles in affected programs: 11668 -> 11464 (-1.75%) helped: 12 HURT: 0 helped stats (abs) min: 16 max: 18 x̄: 17.00 x̃: 17 helped stats (rel) min: 0.99% max: 3.44% x̄: 2.20% x̃: 2.29% 95% mean confidence interval for cycles value: -17.66 -16.34 95% mean confidence interval for cycles %-change: -2.82% -1.58% Cycles are helped. Iron Lake total instructions in shared programs: 7886301 -> 7886277 (<.01%) instructions in affected programs: 576 -> 552 (-4.17%) helped: 12 HURT: 0 helped stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 helped stats (rel) min: 2.94% max: 6.06% x̄: 4.51% x̃: 4.65% 95% mean confidence interval for instructions value: -2.00 -2.00 95% mean confidence interval for instructions %-change: -5.30% -3.72% Instructions are helped. total cycles in shared programs: 178113176 -> 178113176 (0.00%) cycles in affected programs: 2116 -> 2116 (0.00%) helped: 2 HURT: 4 helped stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4 helped stats (rel) min: 1.14% max: 1.14% x̄: 1.14% x̃: 1.14% HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 HURT stats (rel) min: 0.50% max: 0.65% x̄: 0.58% x̃: 0.58% 95% mean confidence interval for cycles value: -3.25 3.25 95% mean confidence interval for cycles %-change: -0.93% 0.94% Inconclusive result (value mean confidence interval includes 0). GM45 total instructions in shared programs: 4857756 -> 4857744 (<.01%) instructions in affected programs: 294 -> 282 (-4.08%) helped: 6 HURT: 0 helped stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 helped stats (rel) min: 2.94% max: 5.71% x̄: 4.40% x̃: 4.55% 95% mean confidence interval for instructions value: -2.00 -2.00 95% mean confidence interval for instructions %-change: -5.71% -3.09% Instructions are helped. total cycles in shared programs: 122178730 -> 122178722 (<.01%) cycles in affected programs: 700 -> 692 (-1.14%) helped: 2 HURT: 0 Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2018-02-03 01:39:54 +00:00
nir: Don't condition 'a-b < 0' -> 'a < b' on is_not_used_by_conditional Now that i965 recognizes that a-b generates the same conditions as 'a < b', there is no reason to condition this transformation on 'is not used by conditional.' Since this was the only user of the is_not_used_by_conditional function, delete it. All Gen6+ platforms had similar results. (Skylake shown) total instructions in shared programs: 14400775 -> 14400595 (<.01%) instructions in affected programs: 36712 -> 36532 (-0.49%) helped: 182 HURT: 26 helped stats (abs) min: 1 max: 2 x̄: 1.13 x̃: 1 helped stats (rel) min: 0.15% max: 1.82% x̄: 0.70% x̃: 0.62% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.24% max: 1.02% x̄: 0.82% x̃: 0.90% 95% mean confidence interval for instructions value: -0.97 -0.76 95% mean confidence interval for instructions %-change: -0.59% -0.43% Instructions are helped. total cycles in shared programs: 532929592 -> 532926345 (<.01%) cycles in affected programs: 478660 -> 475413 (-0.68%) helped: 187 HURT: 22 helped stats (abs) min: 2 max: 200 x̄: 20.99 x̃: 18 helped stats (rel) min: 0.23% max: 24.10% x̄: 1.48% x̃: 1.03% HURT stats (abs) min: 1 max: 214 x̄: 30.86 x̃: 11 HURT stats (rel) min: 0.01% max: 23.06% x̄: 3.12% x̃: 0.86% 95% mean confidence interval for cycles value: -19.50 -11.57 95% mean confidence interval for cycles %-change: -1.42% -0.58% Cycles are helped. GM45 and Iron Lake had similar results. (Iron Lake shown) total cycles in shared programs: 177851578 -> 177851810 (<.01%) cycles in affected programs: 24408 -> 24640 (0.95%) helped: 2 HURT: 4 helped stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4 helped stats (rel) min: 0.42% max: 0.47% x̄: 0.44% x̃: 0.44% HURT stats (abs) min: 24 max: 108 x̄: 60.00 x̃: 54 HURT stats (rel) min: 0.52% max: 1.62% x̄: 1.04% x̃: 1.02% 95% mean confidence interval for cycles value: -7.75 85.08 95% mean confidence interval for cycles %-change: -0.39% 1.49% Inconclusive result (value mean confidence interval includes 0). Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2018-03-14 23:25:07 +00:00
(('flt', ('fadd(is_used_once)', a, ('fneg', b)), 0.0), ('flt', a, b)),
nir: add flt comparision simplification Didn't turn out as useful as I'd hoped, but it will help alot more on i965 by reducing regressions when we drop brw_do_channel_expressions() and brw_do_vector_splitting(). I'm not sure how much sense 'is_not_used_by_conditional' makes on platforms other than i965 but since this is a new opt it at least won't do any harm. shader-db BDW: total instructions in shared programs: 13029581 -> 13029415 (-0.00%) instructions in affected programs: 15268 -> 15102 (-1.09%) helped: 86 HURT: 0 total cycles in shared programs: 247038346 -> 247036198 (-0.00%) cycles in affected programs: 692634 -> 690486 (-0.31%) helped: 183 HURT: 27 Reviewed-by: Elie Tournier <elie.tournier@collabora.com> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2017-01-12 06:01:03 +00:00
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
(('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
nir: Mark bcsel-to-fmin (or fmax) transformations as inexact These transformations are inexact because section 4.7.1 (Range and Precision) says: Operations and built-in functions that operate on a NaN are not required to return a NaN as the result. The fmin or fmax might not return NaN in cases where the original expression would be required to return NaN. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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)),
nir: Recognize some more open-coded fmin / fmax This transformation is inexact because section 4.7.1 (Range and Precision) says: Operations and built-in functions that operate on a NaN are not required to return a NaN as the result. The fmin or fmax might not return NaN in cases where the original expression would be required to return NaN. v2: Reorder operands and mark as inexact. The latter suggested by Jason. shader-db results: Haswell, Broadwell, and Skylake had similar results. (Skylake shown) total instructions in shared programs: 14514817 -> 14514808 (<.01%) instructions in affected programs: 229 -> 220 (-3.93%) helped: 3 HURT: 0 helped stats (abs) min: 1 max: 4 x̄: 3.00 x̃: 4 helped stats (rel) min: 2.86% max: 4.12% x̄: 3.70% x̃: 4.12% total cycles in shared programs: 533145211 -> 533144939 (<.01%) cycles in affected programs: 37268 -> 36996 (-0.73%) helped: 8 HURT: 0 helped stats (abs) min: 2 max: 134 x̄: 34.00 x̃: 2 helped stats (rel) min: 0.02% max: 14.22% x̄: 3.53% x̃: 0.05% Sandy Bridge and Ivy Bridge had similar results. (Ivy Bridge shown) total cycles in shared programs: 257618409 -> 257618403 (<.01%) cycles in affected programs: 12582 -> 12576 (-0.05%) helped: 3 HURT: 0 helped stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 helped stats (rel) min: 0.05% max: 0.05% x̄: 0.05% x̃: 0.05% No changes on Iron Lake or GM45. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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)),
nir/algebraic: Replace i2b used by bcsel or if-statement with comparison All of the helped shaders are in Deus Ex. I looked at a couple shaders, and they have a pattern like: vec1 32 ssa_373 = i2b32 ssa_345.w vec1 32 ssa_374 = bcsel ssa_373, ssa_20, ssa_0 ... vec1 32 ssa_377 = ine ssa_345.w, ssa_0 if ssa_377 { ... vec1 32 ssa_416 = i2b32 ssa_385.w vec1 32 ssa_417 = bcsel ssa_416, ssa_386, ssa_374 ... } The massive help occurs because the i2b32 is removed, then other passes determine that ssa_374 must be ssa_20 inside the if-statement allowing the first bcsel to also be deleted. v2: Rebase on 1-bit Boolean changes. v3: Fix i2b32 vs ine problem in if-statement replacement. Noticed by Bas. Skylake total instructions in shared programs: 15241394 -> 15186287 (-0.36%) instructions in affected programs: 890583 -> 835476 (-6.19%) helped: 355 HURT: 0 helped stats (abs) min: 1 max: 497 x̄: 155.23 x̃: 149 helped stats (rel) min: 0.09% max: 16.49% x̄: 6.10% x̃: 6.59% 95% mean confidence interval for instructions value: -165.07 -145.39 95% mean confidence interval for instructions %-change: -6.42% -5.77% Instructions are helped. total cycles in shared programs: 373846583 -> 371023357 (-0.76%) cycles in affected programs: 118972102 -> 116148876 (-2.37%) helped: 343 HURT: 14 helped stats (abs) min: 45 max: 118284 x̄: 8332.32 x̃: 6089 helped stats (rel) min: 0.03% max: 38.19% x̄: 2.48% x̃: 1.77% HURT stats (abs) min: 120 max: 4126 x̄: 2482.79 x̃: 3019 HURT stats (rel) min: 0.16% max: 17.37% x̄: 2.13% x̃: 1.11% 95% mean confidence interval for cycles value: -8723.28 -7093.12 95% mean confidence interval for cycles %-change: -2.57% -2.02% Cycles are helped. total spills in shared programs: 32401 -> 23465 (-27.58%) spills in affected programs: 24457 -> 15521 (-36.54%) helped: 343 HURT: 0 total fills in shared programs: 37866 -> 31765 (-16.11%) fills in affected programs: 18889 -> 12788 (-32.30%) helped: 343 HURT: 0 Broadwell and Haswell had similar results. (Haswell shown) Haswell total instructions in shared programs: 13764783 -> 13750679 (-0.10%) instructions in affected programs: 1176256 -> 1162152 (-1.20%) helped: 334 HURT: 21 helped stats (abs) min: 1 max: 358 x̄: 42.59 x̃: 47 helped stats (rel) min: 0.09% max: 11.81% x̄: 1.30% x̃: 1.37% HURT stats (abs) min: 1 max: 61 x̄: 5.76 x̃: 1 HURT stats (rel) min: 0.03% max: 1.84% x̄: 0.17% x̃: 0.03% 95% mean confidence interval for instructions value: -43.99 -35.47 95% mean confidence interval for instructions %-change: -1.35% -1.08% Instructions are helped. total cycles in shared programs: 386511910 -> 385402528 (-0.29%) cycles in affected programs: 143831110 -> 142721728 (-0.77%) helped: 327 HURT: 39 helped stats (abs) min: 16 max: 25219 x̄: 3519.74 x̃: 3570 helped stats (rel) min: <.01% max: 10.26% x̄: 0.95% x̃: 0.96% HURT stats (abs) min: 16 max: 4881 x̄: 1065.95 x̃: 997 HURT stats (rel) min: <.01% max: 16.67% x̄: 0.70% x̃: 0.24% 95% mean confidence interval for cycles value: -3375.59 -2686.60 95% mean confidence interval for cycles %-change: -0.92% -0.64% Cycles are helped. total spills in shared programs: 100480 -> 97846 (-2.62%) spills in affected programs: 84702 -> 82068 (-3.11%) helped: 316 HURT: 21 total fills in shared programs: 96877 -> 94369 (-2.59%) fills in affected programs: 69167 -> 66659 (-3.63%) helped: 316 HURT: 9 No changes on Ivy Bridge or earlier platforms. Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-12-04 00:30:44 +00:00
(('bcsel', ('i2b', a), b, c), ('bcsel', ('ine', a, 0), b, c)),
nir: Rely on the fact that bcsel takes a well formed boolean. According to Connor, it's safe to assume that the first operand of bcsel, as well as the operand of b2f and b2i, must be well formed booleans. https://lists.freedesktop.org/archives/mesa-dev/2016-August/125658.html With the previous improvements to a@bool handling, this now has no change in shader-db instruction counts on Broadwell. Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-08-17 23:03:01 +01:00
(('bcsel', ('inot', a), b, c), ('bcsel', a, c, b)),
nir: Collapse repeated bcsels on the same argument. vc4 results: total instructions in shared programs: 39881 -> 39794 (-0.22%) instructions in affected programs: 6302 -> 6215 (-1.38%) Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-02-20 09:20:34 +00:00
(('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)),
nir: Collapse more repeated bcsels on the same argument All Gen platforms had pretty similar results. (Skylake shown) total instructions in shared programs: 14277230 -> 14277220 (<.01%) instructions in affected programs: 751 -> 741 (-1.33%) helped: 4 HURT: 0 helped stats (abs) min: 2 max: 3 x̄: 2.50 x̃: 2 helped stats (rel) min: 1.23% max: 1.40% x̄: 1.32% x̃: 1.32% 95% mean confidence interval for instructions value: -3.42 -1.58 95% mean confidence interval for instructions %-change: -1.47% -1.17% Instructions are helped. total cycles in shared programs: 532577947 -> 532577908 (<.01%) cycles in affected programs: 10641 -> 10602 (-0.37%) helped: 4 HURT: 3 helped stats (abs) min: 1 max: 40 x̄: 13.75 x̃: 7 helped stats (rel) min: 0.11% max: 3.08% x̄: 1.10% x̃: 0.60% HURT stats (abs) min: 2 max: 8 x̄: 5.33 x̃: 6 HURT stats (rel) min: 0.13% max: 0.55% x̄: 0.30% x̃: 0.23% 95% mean confidence interval for cycles value: -20.69 9.55 95% mean confidence interval for cycles %-change: -1.63% 0.63% Inconclusive result (value mean confidence interval includes 0). Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2018-08-01 17:58:19 +01:00
(('bcsel', a, b, ('bcsel', a, c, d)), ('bcsel', a, b, d)),
nir: Rearrange bcsel with two bcsel sources All Gen platforms had pretty similar results. (Skylake shown) total instructions in shared programs: 14277220 -> 14277216 (<.01%) instructions in affected programs: 422 -> 418 (-0.95%) helped: 2 HURT: 0 total cycles in shared programs: 532577908 -> 532577848 (<.01%) cycles in affected programs: 2800 -> 2740 (-2.14%) helped: 2 HURT: 0 Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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)),
nir/algebraic: Drop some @bool specifiers Now that we have one-bit booleans, we don't need to rely on looking at parent instructions in order to figure out if a value is a Boolean most of the time. We can drop these specifiers and now the optimizations will apply more generally. Shader-DB results on Kaby Lake: total instructions in shared programs: 15321168 -> 15321227 (<.01%) instructions in affected programs: 8836 -> 8895 (0.67%) helped: 1 HURT: 31 total cycles in shared programs: 357481781 -> 357481321 (<.01%) cycles in affected programs: 146524 -> 146064 (-0.31%) helped: 22 HURT: 10 total spills in shared programs: 23675 -> 23673 (<.01%) spills in affected programs: 11 -> 9 (-18.18%) helped: 1 HURT: 0 total fills in shared programs: 32040 -> 32036 (-0.01%) fills in affected programs: 27 -> 23 (-14.81%) helped: 1 HURT: 0 No change in VkPipeline-DB Looking at the instructions hurt, a bunch of them seem to be a case where doing exactly the right thing in NIR ends up doing the wrong-ish thing in the back-end because flags are dumb. In particular, there's a case where we have a MUL followed by a CMP followed by a SEL and when we turn that SEL into an OR, it uses the GRF result of the CMP rather than the flag result so the CMP can't be merged with the MUL. Those shaders appear to schedule better according to the cycle estimates so I guess it's a win? Also it helps spilling in one Car Chase compute shader. Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
2019-03-30 03:51:20 +00:00
(('bcsel', a, True, b), ('ior', a, b)),
nir/algebraic: Add some logical OR and AND patterns The new OR pattern has been seen in the wild and can end up being generated by GLSLang. Not sure about the other two new patterns but we may as well throw them in for completeness. While we're here, we can drop the '@bool' specifier from the one pattern because specifying True already implies 1-bit which basically implies boolean. Shader-db results on Kaby Lake: total instructions in shared programs: 15321227 -> 15321129 (<.01%) instructions in affected programs: 3594 -> 3496 (-2.73%) helped: 6 HURT: 0 total cycles in shared programs: 357481321 -> 357479725 (<.01%) cycles in affected programs: 44109 -> 42513 (-3.62%) helped: 6 HURT: 0 VkPipeline-DB results on Kaby Lake: total instructions in shared programs: 3770504 -> 3769734 (-0.02%) instructions in affected programs: 19058 -> 18288 (-4.04%) helped: 163 HURT: 0 total cycles in shared programs: 1417583701 -> 1417569727 (<.01%) cycles in affected programs: 750958 -> 736984 (-1.86%) helped: 158 HURT: 1 Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
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)),
nir: Recognize trivial min/max. No changes, but does prevent some regressions in the next commit. Reviewed-by: Jordan Justen <jordan.l.justen@intel.com> Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
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),
nir/algebraic: Add some max/min optimizations Found by inspection. This doesn't help much now but we'll see this pattern with images if you load UNORM and then store UNORM. Shader-db results on Kaby Lake: total instructions in shared programs: 15166916 -> 15166910 (<.01%) instructions in affected programs: 761 -> 755 (-0.79%) helped: 6 HURT: 0 Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
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)),
nir: Add a couple trivial abs optimizations Spotted in a shader in Batman: Arkham City. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
2018-07-23 07:57:07 +01:00
(('fmax', a, ('fneg', a)), ('fabs', a)),
(('imax', a, ('ineg', a)), ('iabs', a)),
nir: add min/max optimisation Add the following optimisations: min(x, -x) = -abs(x) min(x, -abs(x)) = -abs(x) min(x, abs(x)) = x max(x, -abs(x)) = x max(x, abs(x)) = abs(x) max(x, -x) = abs(x) shader-db: total instructions in shared programs: 13067779 -> 13067775 (-0.00%) instructions in affected programs: 249 -> 245 (-1.61%) helped: 4 HURT: 0 total cycles in shared programs: 252054838 -> 252054806 (-0.00%) cycles in affected programs: 504 -> 472 (-6.35%) helped: 2 HURT: 0 Signed-off-by: Elie Tournier <tournier.elie@gmail.com> Reviewed-by: Plamena Manolova <plamena.manolova@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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)),
nir/algebraic: Simplify max(abs(a), 0.0) -> abs(a) This pattern was noticed in glmark's jellyfish scene. v2: Add inexact qualifier due to NaN behaviour. Minimal shader-db changes (slightly helped). Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com> Reviewed-by: Elie Tournier <tournier.elie@gmail.com>
2019-06-04 20:54:12 +01:00
(('~fmax', ('fabs', a), 0.0), ('fabs', a)),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
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'),
nir/algebraic: Recognize open-coded fsat with modifiers This change also enables a later change (nir/algebraic: Replace 1-fsat(a) with fsat(1-a)) to affect more shaders. Almost all of the affected shaders are in Bioshock Infinite, and all of those shaders all require GLSL 4.10. All Intel platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17228584 -> 17228376 (<.01%) instructions in affected programs: 31438 -> 31230 (-0.66%) helped: 105 HURT: 0 helped stats (abs) min: 1 max: 5 x̄: 1.98 x̃: 1 helped stats (rel) min: 0.08% max: 1.53% x̄: 0.73% x̃: 0.70% 95% mean confidence interval for instructions value: -2.20 -1.76 95% mean confidence interval for instructions %-change: -0.80% -0.67% Instructions are helped. total cycles in shared programs: 360936431 -> 360935690 (<.01%) cycles in affected programs: 420100 -> 419359 (-0.18%) helped: 71 HURT: 21 helped stats (abs) min: 1 max: 160 x̄: 19.28 x̃: 10 helped stats (rel) min: <.01% max: 9.78% x̄: 0.95% x̃: 0.48% HURT stats (abs) min: 1 max: 198 x̄: 29.90 x̃: 10 HURT stats (rel) min: 0.05% max: 8.36% x̄: 1.24% x̃: 0.90% 95% mean confidence interval for cycles value: -16.77 0.66 95% mean confidence interval for cycles %-change: -0.85% -0.06% Inconclusive result (value mean confidence interval includes 0). Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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'),
nir/algebraic: Simplify fsat of fsign These allows us to not support fsign.sat in the Intel compiler backend, and that will simplify some later changes. No shader-db changes on any Intel platform. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2018-09-09 21:40:02 +01:00
(('fsat', ('fsign', a)), ('b2f', ('flt', 0.0, a))),
nir/algebraic: Optimize away an fsat of a b2f The b2f can only produce 0.0 or 1.0, so the fsat does nothing. Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2019-03-01 22:41:59 +00:00
(('fsat', ('b2f', a)), ('b2f', a)),
nir: Add a flag for lowering fsat. vc4 cse/algebraic-disabled stats: total instructions in shared programs: 44356 -> 44354 (-0.00%) instructions in affected programs: 55 -> 53 (-3.64%) v2: Rebase to master (no TGSI->NIR present) Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
2015-01-30 21:53:39 +00:00
(('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'),
nir: Recognize and reduce duplicated fsats. No effect on vc4 shader-db. v2: Rebase to master (no TGSI->NIR present) Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
2015-01-30 22:10:58 +00:00
(('fsat', ('fsat', a)), ('fsat', a)),
nir/algebraic: Push unary operations into source operands of fsat source Pushing a unary operation, like fneg, into the operation that generates its operand allows the fsat to be applied to the inner instruction instead of on a separate instruction that performs the unary operation. This changes fmul ssa_100, ssa_99, ssa_98 fmov.sat ssa_101, -ssa_100 into fmul.sat ssa_100, -ssa_99, ssa_98 Ice Lake, Skylake, and Broadwell had similar results. (Ice Lake shown) total instructions in shared programs: 17228658 -> 17228584 (<.01%) instructions in affected programs: 3163 -> 3089 (-2.34%) helped: 49 HURT: 0 helped stats (abs) min: 1 max: 2 x̄: 1.51 x̃: 2 helped stats (rel) min: 0.58% max: 9.09% x̄: 3.69% x̃: 3.51% 95% mean confidence interval for instructions value: -1.66 -1.37 95% mean confidence interval for instructions %-change: -4.37% -3.00% Instructions are helped. total cycles in shared programs: 360937144 -> 360936431 (<.01%) cycles in affected programs: 24029 -> 23316 (-2.97%) helped: 47 HURT: 2 helped stats (abs) min: 4 max: 18 x̄: 15.34 x̃: 16 helped stats (rel) min: 0.69% max: 6.18% x̄: 3.78% x̃: 4.27% HURT stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4 HURT stats (rel) min: 0.34% max: 0.67% x̄: 0.50% x̃: 0.50% 95% mean confidence interval for cycles value: -16.05 -13.05 95% mean confidence interval for cycles %-change: -4.07% -3.15% Cycles are helped. All Gen7 and earlier platforms had similar results. (Haswell shown) total instructions in shared programs: 13536059 -> 13535884 (<.01%) instructions in affected programs: 8797 -> 8622 (-1.99%) helped: 150 HURT: 0 helped stats (abs) min: 1 max: 2 x̄: 1.17 x̃: 1 helped stats (rel) min: 0.40% max: 11.11% x̄: 3.51% x̃: 1.96% 95% mean confidence interval for instructions value: -1.23 -1.11 95% mean confidence interval for instructions %-change: -3.97% -3.05% Instructions are helped. total cycles in shared programs: 357696119 -> 357694193 (<.01%) cycles in affected programs: 50216 -> 48290 (-3.84%) helped: 109 HURT: 14 helped stats (abs) min: 2 max: 92 x̄: 18.97 x̃: 16 helped stats (rel) min: 0.26% max: 19.09% x̄: 7.37% x̃: 5.37% HURT stats (abs) min: 2 max: 26 x̄: 10.14 x̃: 5 HURT stats (rel) min: 0.18% max: 4.73% x̄: 1.84% x̃: 0.92% 95% mean confidence interval for cycles value: -19.27 -12.05 95% mean confidence interval for cycles %-change: -7.34% -5.31% Cycles are helped. Reviewed-by: Matt Turner <mattst88@gmail.com>
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'),
nir: Merge redudant integer clamping. Dolphin uses them a lot. Range tracking would be better in the long term, but this two lines works fine for now. Signed-off-by: Markus Wick <markus@selfnet.de> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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)),
nir: Move fsat outside of fmin/fmax if second arg is 0 to 1. instructions in affected programs: 550 -> 544 (-1.09%) helped: 6 cycles in affected programs: 6952 -> 6850 (-1.47%) helped: 6 Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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))),
nir: Merge redudant integer clamping. Dolphin uses them a lot. Range tracking would be better in the long term, but this two lines works fine for now. Signed-off-by: Markus Wick <markus@selfnet.de> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2016-04-09 02:31:16 +01:00
(('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)),
nir: Be more conservative about rearranging or-compounded compares If both comparisons are used as sources for instructions other than the ior, this transformation is detrimental. If the non-identical value in both compares is constant, the fmin or fmax will be constant-folded away, so the transformation is always a win. shader-db results: Skylake total instructions in shared programs: 14526147 -> 14525898 (<.01%) instructions in affected programs: 70239 -> 69990 (-0.35%) helped: 102 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 2.44 x̃: 1 helped stats (rel) min: 0.07% max: 2.30% x̄: 0.38% x̃: 0.20% 95% mean confidence interval for instructions value: -2.86 -2.02 95% mean confidence interval for instructions %-change: -0.46% -0.31% Instructions are helped. total cycles in shared programs: 533120531 -> 533119892 (<.01%) cycles in affected programs: 994875 -> 994236 (-0.06%) helped: 76 HURT: 26 helped stats (abs) min: 1 max: 324 x̄: 27.09 x̃: 13 helped stats (rel) min: <.01% max: 4.21% x̄: 0.45% x̃: 0.18% HURT stats (abs) min: 1 max: 167 x̄: 54.62 x̃: 26 HURT stats (rel) min: <.01% max: 4.36% x̄: 1.01% x̃: 0.39% 95% mean confidence interval for cycles value: -19.44 6.91 95% mean confidence interval for cycles %-change: -0.30% 0.15% Inconclusive result (value mean confidence interval includes 0). Broadwell total instructions in shared programs: 14816005 -> 14815787 (<.01%) instructions in affected programs: 64658 -> 64440 (-0.34%) helped: 97 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 2.25 x̃: 1 helped stats (rel) min: 0.07% max: 2.30% x̄: 0.38% x̃: 0.20% 95% mean confidence interval for instructions value: -2.62 -1.87 95% mean confidence interval for instructions %-change: -0.45% -0.30% Instructions are helped. total cycles in shared programs: 559340386 -> 559339907 (<.01%) cycles in affected programs: 1090491 -> 1090012 (-0.04%) helped: 66 HURT: 28 helped stats (abs) min: 2 max: 198 x̄: 23.83 x̃: 16 helped stats (rel) min: 0.01% max: 4.21% x̄: 0.47% x̃: 0.27% HURT stats (abs) min: 2 max: 226 x̄: 39.07 x̃: 11 HURT stats (rel) min: <.01% max: 4.61% x̄: 0.64% x̃: 0.20% 95% mean confidence interval for cycles value: -15.94 5.75 95% mean confidence interval for cycles %-change: -0.35% 0.07% Inconclusive result (value mean confidence interval includes 0). LOST: 0 GAINED: 1 Haswell total instructions in shared programs: 9034106 -> 9033948 (<.01%) instructions in affected programs: 24096 -> 23938 (-0.66%) helped: 38 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 4.16 x̃: 4 helped stats (rel) min: 0.42% max: 2.29% x̄: 0.71% x̃: 0.64% 95% mean confidence interval for instructions value: -4.71 -3.60 95% mean confidence interval for instructions %-change: -0.84% -0.58% Instructions are helped. total cycles in shared programs: 84631628 -> 84631402 (<.01%) cycles in affected programs: 148674 -> 148448 (-0.15%) helped: 14 HURT: 14 helped stats (abs) min: 1 max: 114 x̄: 22.14 x̃: 12 helped stats (rel) min: 0.02% max: 2.98% x̄: 0.66% x̃: 0.21% HURT stats (abs) min: 1 max: 10 x̄: 6.00 x̃: 5 HURT stats (rel) min: 0.01% max: 0.20% x̄: 0.12% x̃: 0.11% 95% mean confidence interval for cycles value: -19.42 3.28 95% mean confidence interval for cycles %-change: -0.59% 0.05% Inconclusive result (value mean confidence interval includes 0). Ivy Bridge total instructions in shared programs: 10015456 -> 10015293 (<.01%) instructions in affected programs: 27701 -> 27538 (-0.59%) helped: 38 HURT: 0 helped stats (abs) min: 1 max: 9 x̄: 4.29 x̃: 4 helped stats (rel) min: 0.33% max: 2.79% x̄: 0.66% x̃: 0.52% 95% mean confidence interval for instructions value: -4.87 -3.71 95% mean confidence interval for instructions %-change: -0.82% -0.51% Instructions are helped. total cycles in shared programs: 87524771 -> 87524569 (<.01%) cycles in affected programs: 112324 -> 112122 (-0.18%) helped: 6 HURT: 12 helped stats (abs) min: 2 max: 111 x̄: 44.67 x̃: 20 helped stats (rel) min: 0.02% max: 2.94% x̄: 1.45% x̃: 1.26% HURT stats (abs) min: 1 max: 16 x̄: 5.50 x̃: 5 HURT stats (rel) min: <.01% max: 0.16% x̄: 0.08% x̃: 0.08% 95% mean confidence interval for cycles value: -29.14 6.69 95% mean confidence interval for cycles %-change: -0.93% 0.08% Inconclusive result (value mean confidence interval includes 0). LOST: 0 GAINED: 2 Sandy Bridge total instructions in shared programs: 10545655 -> 10545465 (<.01%) instructions in affected programs: 37198 -> 37008 (-0.51%) helped: 42 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 4.52 x̃: 4 helped stats (rel) min: 0.31% max: 2.15% x̄: 0.58% x̃: 0.49% 95% mean confidence interval for instructions value: -5.14 -3.91 95% mean confidence interval for instructions %-change: -0.68% -0.47% Instructions are helped. total cycles in shared programs: 146113059 -> 146112427 (<.01%) cycles in affected programs: 423514 -> 422882 (-0.15%) helped: 32 HURT: 10 helped stats (abs) min: 4 max: 162 x̄: 24.34 x̃: 12 helped stats (rel) min: 0.06% max: 2.74% x̄: 0.37% x̃: 0.11% HURT stats (abs) min: 12 max: 19 x̄: 14.70 x̃: 14 HURT stats (rel) min: 0.10% max: 0.18% x̄: 0.16% x̃: 0.14% 95% mean confidence interval for cycles value: -26.03 -4.07 95% mean confidence interval for cycles %-change: -0.43% -0.05% Cycles are helped. Iron Lake total instructions in shared programs: 7886959 -> 7886925 (<.01%) instructions in affected programs: 1340 -> 1306 (-2.54%) helped: 4 HURT: 0 helped stats (abs) min: 2 max: 15 x̄: 8.50 x̃: 8 helped stats (rel) min: 0.63% max: 4.30% x̄: 2.45% x̃: 2.43% 95% mean confidence interval for instructions value: -20.44 3.44 95% mean confidence interval for instructions %-change: -5.78% 0.89% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 178116996 -> 178116888 (<.01%) cycles in affected programs: 6262 -> 6154 (-1.72%) helped: 2 HURT: 2 helped stats (abs) min: 44 max: 78 x̄: 61.00 x̃: 61 helped stats (rel) min: 3.31% max: 3.94% x̄: 3.62% x̃: 3.62% HURT stats (abs) min: 6 max: 8 x̄: 7.00 x̃: 7 HURT stats (rel) min: 0.34% max: 0.68% x̄: 0.51% x̃: 0.51% 95% mean confidence interval for cycles value: -93.27 39.27 95% mean confidence interval for cycles %-change: -5.38% 2.27% Inconclusive result (value mean confidence interval includes 0). GM45 total instructions in shared programs: 4857887 -> 4857870 (<.01%) instructions in affected programs: 674 -> 657 (-2.52%) helped: 2 HURT: 0 total cycles in shared programs: 122180816 -> 122180744 (<.01%) cycles in affected programs: 3764 -> 3692 (-1.91%) helped: 1 HURT: 1 helped stats (abs) min: 78 max: 78 x̄: 78.00 x̃: 78 helped stats (rel) min: 3.94% max: 3.94% x̄: 3.94% x̃: 3.94% HURT stats (abs) min: 6 max: 6 x̄: 6.00 x̃: 6 HURT stats (rel) min: 0.34% max: 0.34% x̄: 0.34% x̃: 0.34% Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
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)),
nir: Rearrange and-compounded float compares If both comparisons are used as sources for instructions other than the iand, this transformation is detrimental. If the non-identical value in both compares is constant, the fmin or fmax will be constant-folded away, so the transformation is always a win. It is interesting to me that on Iron Lake only 81 shaders have instruction counts changed, but 726 shaders have cycle counts changed. shader-db results: Skylake total instructions in shared programs: 14525728 -> 14521017 (-0.03%) instructions in affected programs: 1164726 -> 1160015 (-0.40%) helped: 1692 HURT: 5 helped stats (abs) min: 1 max: 637 x̄: 2.79 x̃: 2 helped stats (rel) min: 0.07% max: 16.36% x̄: 0.81% x̃: 0.33% HURT stats (abs) min: 1 max: 12 x̄: 3.20 x̃: 1 HURT stats (rel) min: 0.38% max: 2.86% x̄: 2.36% x̃: 2.86% 95% mean confidence interval for instructions value: -3.52 -2.03 95% mean confidence interval for instructions %-change: -0.86% -0.74% Instructions are helped. total cycles in shared programs: 533115449 -> 532991404 (-0.02%) cycles in affected programs: 119401803 -> 119277758 (-0.10%) helped: 1145 HURT: 467 helped stats (abs) min: 1 max: 34644 x̄: 145.92 x̃: 18 helped stats (rel) min: <.01% max: 45.33% x̄: 1.58% x̃: 0.42% HURT stats (abs) min: 1 max: 1590 x̄: 92.15 x̃: 15 HURT stats (rel) min: <.01% max: 13.48% x̄: 1.26% x̃: 0.39% 95% mean confidence interval for cycles value: -122.16 -31.74 95% mean confidence interval for cycles %-change: -0.94% -0.57% Cycles are helped. total spills in shared programs: 9597 -> 9534 (-0.66%) spills in affected programs: 403 -> 340 (-15.63%) helped: 1 HURT: 1 total fills in shared programs: 13904 -> 13790 (-0.82%) fills in affected programs: 1627 -> 1513 (-7.01%) helped: 2 HURT: 1 LOST: 0 GAINED: 2 Broadwell total instructions in shared programs: 14816966 -> 14812590 (-0.03%) instructions in affected programs: 1499885 -> 1495509 (-0.29%) helped: 1672 HURT: 15 helped stats (abs) min: 1 max: 455 x̄: 2.70 x̃: 2 helped stats (rel) min: 0.05% max: 16.36% x̄: 0.81% x̃: 0.33% HURT stats (abs) min: 1 max: 21 x̄: 9.20 x̃: 8 HURT stats (rel) min: 0.08% max: 2.86% x̄: 1.06% x̃: 0.53% 95% mean confidence interval for instructions value: -3.14 -2.05 95% mean confidence interval for instructions %-change: -0.85% -0.73% Instructions are helped. total cycles in shared programs: 559353622 -> 559345595 (<.01%) cycles in affected programs: 139893703 -> 139885676 (<.01%) helped: 921 HURT: 697 helped stats (abs) min: 1 max: 42424 x̄: 143.45 x̃: 18 helped stats (rel) min: <.01% max: 36.23% x̄: 2.02% x̃: 0.87% HURT stats (abs) min: 1 max: 2370 x̄: 178.03 x̃: 38 HURT stats (rel) min: <.01% max: 17.35% x̄: 0.71% x̃: 0.14% 95% mean confidence interval for cycles value: -59.64 49.72 95% mean confidence interval for cycles %-change: -1.02% -0.66% Inconclusive result (value mean confidence interval includes 0). total spills in shared programs: 78902 -> 78861 (-0.05%) spills in affected programs: 2418 -> 2377 (-1.70%) helped: 1 HURT: 11 total fills in shared programs: 83782 -> 83678 (-0.12%) fills in affected programs: 3515 -> 3411 (-2.96%) helped: 2 HURT: 11 LOST: 0 GAINED: 5 Haswell and Ivy Bridge had similar results. Haswell shown. total instructions in shared programs: 9033898 -> 9032010 (-0.02%) instructions in affected programs: 308064 -> 306176 (-0.61%) helped: 921 HURT: 4 helped stats (abs) min: 1 max: 20 x̄: 2.05 x̃: 1 helped stats (rel) min: 0.17% max: 17.54% x̄: 0.80% x̃: 0.35% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 3.23% max: 3.23% x̄: 3.23% x̃: 3.23% 95% mean confidence interval for instructions value: -2.21 -1.87 95% mean confidence interval for instructions %-change: -0.88% -0.68% Instructions are helped. total cycles in shared programs: 84628949 -> 84620520 (<.01%) cycles in affected programs: 2164913 -> 2156484 (-0.39%) helped: 518 HURT: 359 helped stats (abs) min: 1 max: 440 x̄: 41.52 x̃: 20 helped stats (rel) min: <.01% max: 17.17% x̄: 1.95% x̃: 1.01% HURT stats (abs) min: 1 max: 586 x̄: 36.43 x̃: 8 HURT stats (rel) min: 0.04% max: 18.65% x̄: 1.47% x̃: 0.40% 95% mean confidence interval for cycles value: -15.17 -4.05 95% mean confidence interval for cycles %-change: -0.77% -0.32% Cycles are helped. LOST: 0 GAINED: 4 Sandy Bridge total instructions in shared programs: 10544860 -> 10542933 (-0.02%) instructions in affected programs: 360019 -> 358092 (-0.54%) helped: 931 HURT: 4 helped stats (abs) min: 1 max: 20 x̄: 2.07 x̃: 1 helped stats (rel) min: 0.11% max: 15.52% x̄: 0.68% x̃: 0.30% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 3.33% max: 3.33% x̄: 3.33% x̃: 3.33% 95% mean confidence interval for instructions value: -2.23 -1.89 95% mean confidence interval for instructions %-change: -0.76% -0.58% Instructions are helped. total cycles in shared programs: 146106820 -> 146098397 (<.01%) cycles in affected programs: 3435047 -> 3426624 (-0.25%) helped: 572 HURT: 329 helped stats (abs) min: 1 max: 1289 x̄: 32.52 x̃: 15 helped stats (rel) min: <.01% max: 26.29% x̄: 0.97% x̃: 0.33% HURT stats (abs) min: 1 max: 1714 x̄: 30.93 x̃: 6 HURT stats (rel) min: 0.02% max: 41.31% x̄: 1.13% x̃: 0.19% 95% mean confidence interval for cycles value: -16.85 -1.85 95% mean confidence interval for cycles %-change: -0.39% -0.01% Cycles are helped. LOST: 1 GAINED: 0 Iron Lake total instructions in shared programs: 7886925 -> 7886711 (<.01%) instructions in affected programs: 25763 -> 25549 (-0.83%) helped: 75 HURT: 6 helped stats (abs) min: 1 max: 13 x̄: 3.33 x̃: 1 helped stats (rel) min: 0.35% max: 17.57% x̄: 1.96% x̃: 0.53% HURT stats (abs) min: 1 max: 16 x̄: 6.00 x̃: 1 HURT stats (rel) min: 2.86% max: 4.79% x̄: 3.49% x̃: 2.86% 95% mean confidence interval for instructions value: -3.69 -1.60 95% mean confidence interval for instructions %-change: -2.54% -0.57% Instructions are helped. total cycles in shared programs: 178116888 -> 178115324 (<.01%) cycles in affected programs: 5858790 -> 5857226 (-0.03%) helped: 484 HURT: 242 helped stats (abs) min: 2 max: 76 x̄: 5.27 x̃: 6 helped stats (rel) min: 0.01% max: 10.70% x̄: 0.18% x̃: 0.06% HURT stats (abs) min: 2 max: 76 x̄: 4.07 x̃: 2 HURT stats (rel) min: 0.01% max: 3.99% x̄: 0.19% x̃: 0.03% 95% mean confidence interval for cycles value: -2.76 -1.55 95% mean confidence interval for cycles %-change: -0.12% 0.01% Inconclusive result (%-change mean confidence interval includes 0). GM45 total instructions in shared programs: 4857870 -> 4857762 (<.01%) instructions in affected programs: 13994 -> 13886 (-0.77%) helped: 39 HURT: 5 helped stats (abs) min: 1 max: 13 x̄: 3.28 x̃: 2 helped stats (rel) min: 0.33% max: 17.11% x̄: 1.86% x̃: 0.48% HURT stats (abs) min: 1 max: 16 x̄: 4.00 x̃: 1 HURT stats (rel) min: 2.86% max: 4.71% x̄: 3.23% x̃: 2.86% 95% mean confidence interval for instructions value: -3.86 -1.05 95% mean confidence interval for instructions %-change: -2.61% 0.04% Inconclusive result (%-change mean confidence interval includes 0). total cycles in shared programs: 122180744 -> 122179674 (<.01%) cycles in affected programs: 3686646 -> 3685576 (-0.03%) helped: 273 HURT: 141 helped stats (abs) min: 2 max: 76 x̄: 5.81 x̃: 6 helped stats (rel) min: 0.01% max: 10.70% x̄: 0.18% x̃: 0.06% HURT stats (abs) min: 2 max: 76 x̄: 3.66 x̃: 2 HURT stats (rel) min: 0.01% max: 3.99% x̄: 0.16% x̃: 0.02% 95% mean confidence interval for cycles value: -3.42 -1.75 95% mean confidence interval for cycles %-change: -0.15% 0.03% Inconclusive result (%-change mean confidence interval includes 0). Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
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)),
nir: Be more conservative about rearranging or-compounded compares If both comparisons are used as sources for instructions other than the ior, this transformation is detrimental. If the non-identical value in both compares is constant, the fmin or fmax will be constant-folded away, so the transformation is always a win. shader-db results: Skylake total instructions in shared programs: 14526147 -> 14525898 (<.01%) instructions in affected programs: 70239 -> 69990 (-0.35%) helped: 102 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 2.44 x̃: 1 helped stats (rel) min: 0.07% max: 2.30% x̄: 0.38% x̃: 0.20% 95% mean confidence interval for instructions value: -2.86 -2.02 95% mean confidence interval for instructions %-change: -0.46% -0.31% Instructions are helped. total cycles in shared programs: 533120531 -> 533119892 (<.01%) cycles in affected programs: 994875 -> 994236 (-0.06%) helped: 76 HURT: 26 helped stats (abs) min: 1 max: 324 x̄: 27.09 x̃: 13 helped stats (rel) min: <.01% max: 4.21% x̄: 0.45% x̃: 0.18% HURT stats (abs) min: 1 max: 167 x̄: 54.62 x̃: 26 HURT stats (rel) min: <.01% max: 4.36% x̄: 1.01% x̃: 0.39% 95% mean confidence interval for cycles value: -19.44 6.91 95% mean confidence interval for cycles %-change: -0.30% 0.15% Inconclusive result (value mean confidence interval includes 0). Broadwell total instructions in shared programs: 14816005 -> 14815787 (<.01%) instructions in affected programs: 64658 -> 64440 (-0.34%) helped: 97 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 2.25 x̃: 1 helped stats (rel) min: 0.07% max: 2.30% x̄: 0.38% x̃: 0.20% 95% mean confidence interval for instructions value: -2.62 -1.87 95% mean confidence interval for instructions %-change: -0.45% -0.30% Instructions are helped. total cycles in shared programs: 559340386 -> 559339907 (<.01%) cycles in affected programs: 1090491 -> 1090012 (-0.04%) helped: 66 HURT: 28 helped stats (abs) min: 2 max: 198 x̄: 23.83 x̃: 16 helped stats (rel) min: 0.01% max: 4.21% x̄: 0.47% x̃: 0.27% HURT stats (abs) min: 2 max: 226 x̄: 39.07 x̃: 11 HURT stats (rel) min: <.01% max: 4.61% x̄: 0.64% x̃: 0.20% 95% mean confidence interval for cycles value: -15.94 5.75 95% mean confidence interval for cycles %-change: -0.35% 0.07% Inconclusive result (value mean confidence interval includes 0). LOST: 0 GAINED: 1 Haswell total instructions in shared programs: 9034106 -> 9033948 (<.01%) instructions in affected programs: 24096 -> 23938 (-0.66%) helped: 38 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 4.16 x̃: 4 helped stats (rel) min: 0.42% max: 2.29% x̄: 0.71% x̃: 0.64% 95% mean confidence interval for instructions value: -4.71 -3.60 95% mean confidence interval for instructions %-change: -0.84% -0.58% Instructions are helped. total cycles in shared programs: 84631628 -> 84631402 (<.01%) cycles in affected programs: 148674 -> 148448 (-0.15%) helped: 14 HURT: 14 helped stats (abs) min: 1 max: 114 x̄: 22.14 x̃: 12 helped stats (rel) min: 0.02% max: 2.98% x̄: 0.66% x̃: 0.21% HURT stats (abs) min: 1 max: 10 x̄: 6.00 x̃: 5 HURT stats (rel) min: 0.01% max: 0.20% x̄: 0.12% x̃: 0.11% 95% mean confidence interval for cycles value: -19.42 3.28 95% mean confidence interval for cycles %-change: -0.59% 0.05% Inconclusive result (value mean confidence interval includes 0). Ivy Bridge total instructions in shared programs: 10015456 -> 10015293 (<.01%) instructions in affected programs: 27701 -> 27538 (-0.59%) helped: 38 HURT: 0 helped stats (abs) min: 1 max: 9 x̄: 4.29 x̃: 4 helped stats (rel) min: 0.33% max: 2.79% x̄: 0.66% x̃: 0.52% 95% mean confidence interval for instructions value: -4.87 -3.71 95% mean confidence interval for instructions %-change: -0.82% -0.51% Instructions are helped. total cycles in shared programs: 87524771 -> 87524569 (<.01%) cycles in affected programs: 112324 -> 112122 (-0.18%) helped: 6 HURT: 12 helped stats (abs) min: 2 max: 111 x̄: 44.67 x̃: 20 helped stats (rel) min: 0.02% max: 2.94% x̄: 1.45% x̃: 1.26% HURT stats (abs) min: 1 max: 16 x̄: 5.50 x̃: 5 HURT stats (rel) min: <.01% max: 0.16% x̄: 0.08% x̃: 0.08% 95% mean confidence interval for cycles value: -29.14 6.69 95% mean confidence interval for cycles %-change: -0.93% 0.08% Inconclusive result (value mean confidence interval includes 0). LOST: 0 GAINED: 2 Sandy Bridge total instructions in shared programs: 10545655 -> 10545465 (<.01%) instructions in affected programs: 37198 -> 37008 (-0.51%) helped: 42 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 4.52 x̃: 4 helped stats (rel) min: 0.31% max: 2.15% x̄: 0.58% x̃: 0.49% 95% mean confidence interval for instructions value: -5.14 -3.91 95% mean confidence interval for instructions %-change: -0.68% -0.47% Instructions are helped. total cycles in shared programs: 146113059 -> 146112427 (<.01%) cycles in affected programs: 423514 -> 422882 (-0.15%) helped: 32 HURT: 10 helped stats (abs) min: 4 max: 162 x̄: 24.34 x̃: 12 helped stats (rel) min: 0.06% max: 2.74% x̄: 0.37% x̃: 0.11% HURT stats (abs) min: 12 max: 19 x̄: 14.70 x̃: 14 HURT stats (rel) min: 0.10% max: 0.18% x̄: 0.16% x̃: 0.14% 95% mean confidence interval for cycles value: -26.03 -4.07 95% mean confidence interval for cycles %-change: -0.43% -0.05% Cycles are helped. Iron Lake total instructions in shared programs: 7886959 -> 7886925 (<.01%) instructions in affected programs: 1340 -> 1306 (-2.54%) helped: 4 HURT: 0 helped stats (abs) min: 2 max: 15 x̄: 8.50 x̃: 8 helped stats (rel) min: 0.63% max: 4.30% x̄: 2.45% x̃: 2.43% 95% mean confidence interval for instructions value: -20.44 3.44 95% mean confidence interval for instructions %-change: -5.78% 0.89% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 178116996 -> 178116888 (<.01%) cycles in affected programs: 6262 -> 6154 (-1.72%) helped: 2 HURT: 2 helped stats (abs) min: 44 max: 78 x̄: 61.00 x̃: 61 helped stats (rel) min: 3.31% max: 3.94% x̄: 3.62% x̃: 3.62% HURT stats (abs) min: 6 max: 8 x̄: 7.00 x̃: 7 HURT stats (rel) min: 0.34% max: 0.68% x̄: 0.51% x̃: 0.51% 95% mean confidence interval for cycles value: -93.27 39.27 95% mean confidence interval for cycles %-change: -5.38% 2.27% Inconclusive result (value mean confidence interval includes 0). GM45 total instructions in shared programs: 4857887 -> 4857870 (<.01%) instructions in affected programs: 674 -> 657 (-2.52%) helped: 2 HURT: 0 total cycles in shared programs: 122180816 -> 122180744 (<.01%) cycles in affected programs: 3764 -> 3692 (-1.91%) helped: 1 HURT: 1 helped stats (abs) min: 78 max: 78 x̄: 78.00 x̃: 78 helped stats (rel) min: 3.94% max: 3.94% x̄: 3.94% x̃: 3.94% HURT stats (abs) min: 6 max: 6 x̄: 6.00 x̃: 6 HURT stats (rel) min: 0.34% max: 0.34% x̄: 0.34% x̃: 0.34% Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
2018-01-04 21:30:49 +00:00
nir: Rearrange logic op-compounded integer compares Skylake and Broadwell had similar results. Skylake shown. total instructions in shared programs: 14521769 -> 14521753 (<.01%) instructions in affected programs: 8782 -> 8766 (-0.18%) helped: 16 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.12% max: 0.40% x̄: 0.20% x̃: 0.18% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -0.23% -0.16% Instructions are helped. total cycles in shared programs: 533000376 -> 533000205 (<.01%) cycles in affected programs: 447035 -> 446864 (-0.04%) helped: 9 HURT: 9 helped stats (abs) min: 2 max: 40 x̄: 35.78 x̃: 40 helped stats (rel) min: 0.02% max: 0.18% x̄: 0.10% x̃: 0.09% HURT stats (abs) min: 1 max: 52 x̄: 16.78 x̃: 10 HURT stats (rel) min: <.01% max: 1.11% x̄: 0.29% x̃: 0.12% 95% mean confidence interval for cycles value: -25.07 6.07 95% mean confidence interval for cycles %-change: -0.08% 0.27% Inconclusive result (value mean confidence interval includes 0). No changes on GM45, Iron Lake, Sandy Bridge, Ivy Bridge, or Haswell. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
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)),
nir/algebraic: Simplify comparison with sequential integers starting with 0 All of the affected shaders are Unreal4 demos. All Gen6+ platforms had similar results. (Skylake shown) total instructions in shared programs: 15437170 -> 15437001 (<.01%) instructions in affected programs: 21536 -> 21367 (-0.78%) helped: 43 HURT: 0 helped stats (abs) min: 1 max: 4 x̄: 3.93 x̃: 4 helped stats (rel) min: 0.68% max: 1.01% x̄: 0.80% x̃: 0.80% 95% mean confidence interval for instructions value: -4.07 -3.79 95% mean confidence interval for instructions %-change: -0.83% -0.77% Instructions are helped. total cycles in shared programs: 383007896 -> 383007378 (<.01%) cycles in affected programs: 158640 -> 158122 (-0.33%) helped: 38 HURT: 4 helped stats (abs) min: 1 max: 48 x̄: 13.89 x̃: 6 helped stats (rel) min: 0.03% max: 1.01% x̄: 0.33% x̃: 0.19% HURT stats (abs) min: 2 max: 3 x̄: 2.50 x̃: 2 HURT stats (rel) min: 0.06% max: 0.09% x̄: 0.08% x̃: 0.08% 95% mean confidence interval for cycles value: -16.90 -7.77 95% mean confidence interval for cycles %-change: -0.39% -0.19% Cycles are helped. Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 8213746 -> 8213745 (<.01%) instructions in affected programs: 127 -> 126 (-0.79%) helped: 1 HURT: 0 total cycles in shared programs: 187734146 -> 187734144 (<.01%) cycles in affected programs: 2132 -> 2130 (-0.09%) helped: 1 HURT: 0 Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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)),
nir/algebraic: Recognize open-coded copysign(1.0, a) All of the affected shaders are in Mad Max. The inner part of the pattern is itself an open-coded sign(a). I tried using that as a pattern, but the results were not good. A bunch of shaders were helped for instructions, but overall cycles, spill, and fills were hurt. v2: Rebase on 1-bit Boolean changes. v3: Fix order of copysign() parameters in comments and commit message. Noticed by Matt. All Gen7+ platforms had similar results. (Skylake shown) total instructions in shared programs: 15282141 -> 15282073 (<.01%) instructions in affected programs: 6106 -> 6038 (-1.11%) helped: 17 HURT: 0 helped stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4 helped stats (rel) min: 1.02% max: 2.20% x̄: 1.15% x̃: 1.06% 95% mean confidence interval for instructions value: -4.00 -4.00 95% mean confidence interval for instructions %-change: -1.30% -1.00% Instructions are helped. total cycles in shared programs: 372597886 -> 372595954 (<.01%) cycles in affected programs: 32701 -> 30769 (-5.91%) helped: 17 HURT: 0 helped stats (abs) min: 6 max: 216 x̄: 113.65 x̃: 118 helped stats (rel) min: 0.40% max: 21.86% x̄: 6.20% x̃: 5.83% 95% mean confidence interval for cycles value: -152.84 -74.45 95% mean confidence interval for cycles %-change: -8.89% -3.51% Cycles are helped. No changes on any Gen6 or earlier platforms. Reviewed-by: Matt Turner <mattst88@gmail.com>
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))),
nir/algebraic: Drop some @bool specifiers Now that we have one-bit booleans, we don't need to rely on looking at parent instructions in order to figure out if a value is a Boolean most of the time. We can drop these specifiers and now the optimizations will apply more generally. Shader-DB results on Kaby Lake: total instructions in shared programs: 15321168 -> 15321227 (<.01%) instructions in affected programs: 8836 -> 8895 (0.67%) helped: 1 HURT: 31 total cycles in shared programs: 357481781 -> 357481321 (<.01%) cycles in affected programs: 146524 -> 146064 (-0.31%) helped: 22 HURT: 10 total spills in shared programs: 23675 -> 23673 (<.01%) spills in affected programs: 11 -> 9 (-18.18%) helped: 1 HURT: 0 total fills in shared programs: 32040 -> 32036 (-0.01%) fills in affected programs: 27 -> 23 (-14.81%) helped: 1 HURT: 0 No change in VkPipeline-DB Looking at the instructions hurt, a bunch of them seem to be a case where doing exactly the right thing in NIR ends up doing the wrong-ish thing in the back-end because flags are dumb. In particular, there's a case where we have a MUL followed by a CMP followed by a SEL and when we turn that SEL into an OR, it uses the GRF result of the CMP rather than the flag result so the CMP can't be merged with the MUL. Those shaders appear to schedule better according to the cycle estimates so I guess it's a win? Also it helps spilling in one Car Chase compute shader. Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
2019-03-30 03:51:20 +00:00
(('ior', a, ('ieq', a, False)), True),
nir/algebraic: Generalize an optimization There's nothing boolean about (a | ~a) ~> -1 Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2018-10-19 18:43:43 +01:00
(('ior', a, ('inot', a)), -1),
nir: add a couple of ior opts to nir_opt_algebraic One of these was seen in a Deus Ex shader. Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2018-07-16 06:19:29 +01:00
nir/algebraic: Fix some 1-bit Boolean weirdness Skylake, Broadwell, and Haswell had similar results. (Skylake shown) total cycles in shared programs: 372594532 -> 372594460 (<.01%) cycles in affected programs: 46854 -> 46782 (-0.15%) helped: 9 HURT: 0 helped stats (abs) min: 2 max: 22 x̄: 8.00 x̃: 2 helped stats (rel) min: 0.02% max: 0.41% x̄: 0.16% x̃: 0.09% 95% mean confidence interval for cycles value: -14.34 -1.66 95% mean confidence interval for cycles %-change: -0.28% -0.04% Cycles are helped. Ivy Bridge total instructions in shared programs: 12038379 -> 12038373 (<.01%) instructions in affected programs: 1278 -> 1272 (-0.47%) helped: 3 HURT: 0 helped stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 helped stats (rel) min: 0.31% max: 0.77% x̄: 0.54% x̃: 0.55% total cycles in shared programs: 180889027 -> 180888997 (<.01%) cycles in affected programs: 29979 -> 29949 (-0.10%) helped: 5 HURT: 0 helped stats (abs) min: 1 max: 16 x̄: 6.00 x̃: 5 helped stats (rel) min: 0.02% max: 0.34% x̄: 0.11% x̃: 0.07% 95% mean confidence interval for cycles value: -13.40 1.40 95% mean confidence interval for cycles %-change: -0.27% 0.05% Inconclusive result (value mean confidence interval includes 0). Sandy Bridge total cycles in shared programs: 155091021 -> 155091003 (<.01%) cycles in affected programs: 8842 -> 8824 (-0.20%) helped: 2 HURT: 0 No changes on Iron Lake or GM45. Reviewed-by: Matt Turner <mattst88@gmail.com>
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))),
nir: optimize iand(ieq(a, 0), ieq(b, 0)) to ieq(ior(a, b), 0) Totals from affected shaders: SGPRS: 80 -> 80 (0.00 %) VGPRS: 48 -> 48 (0.00 %) Code Size: 2120 -> 2096 (-1.13 %) bytes Max Waves: 16 -> 16 (0.00 %) Only two Rise of Tomb Raider shaders are affected on my side. Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
2018-05-30 09:48:31 +01:00
(('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
nir: Undo possible damage caused by rearranging or-compounded float compares shader-db results: Skylake and Broadwell had similar results (Skylake shown) total instructions in shared programs: 14525898 -> 14525836 (<.01%) instructions in affected programs: 1964 -> 1902 (-3.16%) helped: 14 HURT: 0 helped stats (abs) min: 1 max: 25 x̄: 4.43 x̃: 1 helped stats (rel) min: 0.68% max: 9.77% x̄: 2.10% x̃: 0.86% 95% mean confidence interval for instructions value: -9.46 0.60 95% mean confidence interval for instructions %-change: -3.97% -0.24% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 533119892 -> 533115756 (<.01%) cycles in affected programs: 96061 -> 91925 (-4.31%) helped: 13 HURT: 1 helped stats (abs) min: 60 max: 596 x̄: 318.77 x̃: 300 helped stats (rel) min: 1.15% max: 5.49% x̄: 4.27% x̃: 4.42% HURT stats (abs) min: 8 max: 8 x̄: 8.00 x̃: 8 HURT stats (rel) min: 0.46% max: 0.46% x̄: 0.46% x̃: 0.46% 95% mean confidence interval for cycles value: -379.43 -211.43 95% mean confidence interval for cycles %-change: -4.84% -3.01% Cycles are helped. Haswell, Ivy Bridge and Sandy Bridge had similar results (Haswell shown). total instructions in shared programs: 9033948 -> 9033898 (<.01%) instructions in affected programs: 535 -> 485 (-9.35%) helped: 2 HURT: 0 total cycles in shared programs: 84631402 -> 84628949 (<.01%) cycles in affected programs: 63197 -> 60744 (-3.88%) helped: 13 HURT: 2 helped stats (abs) min: 1 max: 594 x̄: 189.62 x̃: 140 helped stats (rel) min: 0.07% max: 5.04% x̄: 3.79% x̃: 4.01% HURT stats (abs) min: 4 max: 8 x̄: 6.00 x̃: 6 HURT stats (rel) min: 0.17% max: 0.45% x̄: 0.31% x̃: 0.31% 95% mean confidence interval for cycles value: -253.40 -73.67 95% mean confidence interval for cycles %-change: -4.24% -2.25% Cycles are helped. No changes on GM45 or Iron Lake. v2: Add a couple more tautological compares. Suggested by Elie. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
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)),
nir: Undo possible damage caused by rearranging or-compounded float compares shader-db results: Skylake and Broadwell had similar results (Skylake shown) total instructions in shared programs: 14525898 -> 14525836 (<.01%) instructions in affected programs: 1964 -> 1902 (-3.16%) helped: 14 HURT: 0 helped stats (abs) min: 1 max: 25 x̄: 4.43 x̃: 1 helped stats (rel) min: 0.68% max: 9.77% x̄: 2.10% x̃: 0.86% 95% mean confidence interval for instructions value: -9.46 0.60 95% mean confidence interval for instructions %-change: -3.97% -0.24% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 533119892 -> 533115756 (<.01%) cycles in affected programs: 96061 -> 91925 (-4.31%) helped: 13 HURT: 1 helped stats (abs) min: 60 max: 596 x̄: 318.77 x̃: 300 helped stats (rel) min: 1.15% max: 5.49% x̄: 4.27% x̃: 4.42% HURT stats (abs) min: 8 max: 8 x̄: 8.00 x̃: 8 HURT stats (rel) min: 0.46% max: 0.46% x̄: 0.46% x̃: 0.46% 95% mean confidence interval for cycles value: -379.43 -211.43 95% mean confidence interval for cycles %-change: -4.84% -3.01% Cycles are helped. Haswell, Ivy Bridge and Sandy Bridge had similar results (Haswell shown). total instructions in shared programs: 9033948 -> 9033898 (<.01%) instructions in affected programs: 535 -> 485 (-9.35%) helped: 2 HURT: 0 total cycles in shared programs: 84631402 -> 84628949 (<.01%) cycles in affected programs: 63197 -> 60744 (-3.88%) helped: 13 HURT: 2 helped stats (abs) min: 1 max: 594 x̄: 189.62 x̃: 140 helped stats (rel) min: 0.07% max: 5.04% x̄: 3.79% x̃: 4.01% HURT stats (abs) min: 4 max: 8 x̄: 6.00 x̃: 6 HURT stats (rel) min: 0.17% max: 0.45% x̄: 0.31% x̃: 0.31% 95% mean confidence interval for cycles value: -253.40 -73.67 95% mean confidence interval for cycles %-change: -4.24% -2.25% Cycles are helped. No changes on GM45 or Iron Lake. v2: Add a couple more tautological compares. Suggested by Elie. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
2018-01-05 21:29:26 +00:00
nir: Rearrange logic op-compounded integer compares Skylake and Broadwell had similar results. Skylake shown. total instructions in shared programs: 14521769 -> 14521753 (<.01%) instructions in affected programs: 8782 -> 8766 (-0.18%) helped: 16 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.12% max: 0.40% x̄: 0.20% x̃: 0.18% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -0.23% -0.16% Instructions are helped. total cycles in shared programs: 533000376 -> 533000205 (<.01%) cycles in affected programs: 447035 -> 446864 (-0.04%) helped: 9 HURT: 9 helped stats (abs) min: 2 max: 40 x̄: 35.78 x̃: 40 helped stats (rel) min: 0.02% max: 0.18% x̄: 0.10% x̃: 0.09% HURT stats (abs) min: 1 max: 52 x̄: 16.78 x̃: 10 HURT stats (rel) min: <.01% max: 1.11% x̄: 0.29% x̃: 0.12% 95% mean confidence interval for cycles value: -25.07 6.07 95% mean confidence interval for cycles %-change: -0.08% 0.27% Inconclusive result (value mean confidence interval includes 0). No changes on GM45, Iron Lake, Sandy Bridge, Ivy Bridge, or Haswell. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
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)),
nir/algebraic: add optimization pattern for ('ult', a, ('and', b, a)) and friends. These optimizations are based on the fact that 'and(a,b) <= umin(a,b)'. For AMD, this series moves the optimization from LLVM to NIR, so currently no vkpipeline-db changes here. Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
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),
nir: Rearrange logic op-compounded integer compares Skylake and Broadwell had similar results. Skylake shown. total instructions in shared programs: 14521769 -> 14521753 (<.01%) instructions in affected programs: 8782 -> 8766 (-0.18%) helped: 16 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.12% max: 0.40% x̄: 0.20% x̃: 0.18% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -0.23% -0.16% Instructions are helped. total cycles in shared programs: 533000376 -> 533000205 (<.01%) cycles in affected programs: 447035 -> 446864 (-0.04%) helped: 9 HURT: 9 helped stats (abs) min: 2 max: 40 x̄: 35.78 x̃: 40 helped stats (rel) min: 0.02% max: 0.18% x̄: 0.10% x̃: 0.09% HURT stats (abs) min: 1 max: 52 x̄: 16.78 x̃: 10 HURT stats (rel) min: <.01% max: 1.11% x̄: 0.29% x̃: 0.12% 95% mean confidence interval for cycles value: -25.07 6.07 95% mean confidence interval for cycles %-change: -0.08% 0.27% Inconclusive result (value mean confidence interval includes 0). No changes on GM45, Iron Lake, Sandy Bridge, Ivy Bridge, or Haswell. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
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))),
nir: Rearrange logic op-compounded integer compares Skylake and Broadwell had similar results. Skylake shown. total instructions in shared programs: 14521769 -> 14521753 (<.01%) instructions in affected programs: 8782 -> 8766 (-0.18%) helped: 16 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.12% max: 0.40% x̄: 0.20% x̃: 0.18% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -0.23% -0.16% Instructions are helped. total cycles in shared programs: 533000376 -> 533000205 (<.01%) cycles in affected programs: 447035 -> 446864 (-0.04%) helped: 9 HURT: 9 helped stats (abs) min: 2 max: 40 x̄: 35.78 x̃: 40 helped stats (rel) min: 0.02% max: 0.18% x̄: 0.10% x̃: 0.09% HURT stats (abs) min: 1 max: 52 x̄: 16.78 x̃: 10 HURT stats (rel) min: <.01% max: 1.11% x̄: 0.29% x̃: 0.12% 95% mean confidence interval for cycles value: -25.07 6.07 95% mean confidence interval for cycles %-change: -0.08% 0.27% Inconclusive result (value mean confidence interval includes 0). No changes on GM45, Iron Lake, Sandy Bridge, Ivy Bridge, or Haswell. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
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))),
nir/algebraic: Add a couple optimizations for iabs and ishr Shader-db results on Kaby Lake: total instructions in shared programs: 15225213 -> 15222365 (-0.02%) instructions in affected programs: 43524 -> 40676 (-6.54%) helped: 203 HURT: 0 Lots of shaders in Shadow Warrior had this pattern along with Deus Ex, Civ, Shadow of Mordor, and several others. Reviewed-by: Kristian H. Kristensen <hoegsberg@chromium.org>
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)),
nir: Don't abs slt and friends No shader-db changes, but this is symmetric with the previous commit. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
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)),
nir: add option to lower slt/sge/seq/sne In freedreno these get implemented as the matching f* instruction plus a u2f to convert the result to float 1.0/0.0. But less lines of code to just let nir_opt_algebraic handle this for us, plus opens up some small window for other opt passes to improve (ie. if some shader ended up with both a flt and slt with same src args, for example). v2: use b2f rather than u2f Signed-off-by: Rob Clark <robclark@freedesktop.org> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Matt Turner <mattst88@gmail.com>
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'),
nir: Simplify fne(fneg(a), a) -> fne(a, 0.0) -NaN != NaN, and -Inf != Inf, so this should be safe. Found while working on my VRP pass. Shader-db results on my IVB: total instructions in shared programs: 1698267 -> 1698067 (-0.01%) instructions in affected programs: 15785 -> 15585 (-1.27%) helped: 36 HURT: 0 GAINED: 0 LOST: 0 Some shaders was found to have the following pattern in NIR: vec1 ssa_26 = fneg ssa_21 vec1 ssa_27 = fne ssa_21, ssa_26 Make that: vec1 ssa_27 = fne ssa_21, 0.0f This is found in Dota2 and Brutal Legend. One shader is cut by 8%, from 323 -> 296 instructons in SIMD8 Signed-off-by: Thomas Helland <thomashelland90@gmail.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-08-06 12:36:04 +01:00
(('fne', ('fneg', a), a), ('fne', a, 0.0)),
nir: Simplify feq(fneg(a), a)) -> feq(a, 0.0) The positive and negative value of a float can only be equal to each other if it is -0.0f and 0.0f. This is safe for Nan and Inf, as -Nan != Nan, and -Inf != Inf This gives no changes in my shader-db Signed-off-by: Thomas Helland <thomashelland90@gmail.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-08-06 12:36:05 +01:00
(('feq', ('fneg', a), a), ('feq', a, 0.0)),
nir: Recognize mul(b2f(a), b2f(b)) as a logical AND. Transform this into b2f(and(a, b)). total instructions in shared programs: 6205448 -> 6204391 (-0.02%) instructions in affected programs: 284030 -> 282973 (-0.37%) helped: 903 HURT: 6 Acked-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-03-18 06:30:32 +00:00
# Emulating booleans
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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))),
nir/algebraic: Make an optimization more specific Later in this series, bool is not going to imply 32-bit. Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-12-06 03:32:11 +00:00
(('iand', 'a@bool32', 1.0), ('b2f', a)),
nir: Convert ineg(b2i(a)) to a if it's a boolean. On BDW: total instructions in shared programs: 13071119 -> 13070371 (-0.01%) instructions in affected programs: 83424 -> 82676 (-0.90%) helped: 505 HURT: 45 (all TCS, all hurt by a single instruction) total cycles in shared programs: 256601322 -> 256588932 (-0.00%) cycles in affected programs: 819410 -> 807020 (-1.51%) helped: 450 HURT: 57 total loops in shared programs: 2950 -> 2942 (-0.27%) loops in affected programs: 8 -> 0 helped: 7 HURT: 0 v2: Drop unnecessary 'a@bool' annotation (Connor, Eric). Add a comment explaining the rule (Ian). Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> [v1] Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Matt Turner <mattst88@gmail.com>
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).
nir: Make boolean conversions sized just like the others Instead of a single i2b and b2i, we now have i2b32 and b2iN where N is one if 8, 16, 32, or 64. This leads to having a few more opcodes but now everything is consistent and booleans aren't a weird special case anymore. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2018-11-07 19:43:40 +00:00
(('ineg', ('b2i32', 'a@32')), a),
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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.
nir: Add algebraic opt for int comparisons with identical operands. No change on shader-db on i965. v2: Reword the comment due to feedback from Erik Faye-Lund Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v1) Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> (v1)
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),
nir: Add a bunch of algebraic optimizations on logic/bit operations. Matt and I noticed a bunch of "val <- ior a a" operations in a shader, so we decided to add an algebraic optimization for that. While there, I decided to add a bunch more of them. v2: Delete bogus fand/for optimizations (caught by Jason). total NIR instructions in shared programs: 2023511 -> 2020814 (-0.13%) NIR instructions in affected programs: 149634 -> 146937 (-1.80%) helped: 1032 total i965 instructions in shared programs: 6035392 -> 6035397 (0.00%) i965 instructions in affected programs: 537 -> 542 (0.93%) HURT: 2 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-01-19 22:57:38 +00:00
# Logical and bit operations
(('fand', a, 0.0), 0.0),
(('iand', a, a), a),
nir: Add algebraic opt for no-op iand. I lazily generated some of these in VC4 NIR lowering. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2015-07-31 17:12:48 +01:00
(('iand', a, ~0), a),
nir: Add a bunch of algebraic optimizations on logic/bit operations. Matt and I noticed a bunch of "val <- ior a a" operations in a shader, so we decided to add an algebraic optimization for that. While there, I decided to add a bunch more of them. v2: Delete bogus fand/for optimizations (caught by Jason). total NIR instructions in shared programs: 2023511 -> 2020814 (-0.13%) NIR instructions in affected programs: 149634 -> 146937 (-1.80%) helped: 1032 total i965 instructions in shared programs: 6035392 -> 6035397 (0.00%) i965 instructions in affected programs: 537 -> 542 (0.93%) HURT: 2 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-01-19 22:57:38 +00:00
(('iand', a, 0), 0),
(('ior', a, a), a),
(('ior', a, 0), a),
nir: Add optimization for (a || True == True) This was appearing in vc4 VS/CS in mupen64, due to vertex attrib lowering producing some constants that were getting compared. total instructions in shared programs: 112276 -> 112198 (-0.07%) instructions in affected programs: 2239 -> 2161 (-3.48%) total estimated cycles in shared programs: 283102 -> 283038 (-0.02%) estimated cycles in affected programs: 2365 -> 2301 (-2.71%) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2016-07-06 23:00:45 +01:00
(('ior', a, True), True),
nir: Add a bunch of algebraic optimizations on logic/bit operations. Matt and I noticed a bunch of "val <- ior a a" operations in a shader, so we decided to add an algebraic optimization for that. While there, I decided to add a bunch more of them. v2: Delete bogus fand/for optimizations (caught by Jason). total NIR instructions in shared programs: 2023511 -> 2020814 (-0.13%) NIR instructions in affected programs: 149634 -> 146937 (-1.80%) helped: 1032 total i965 instructions in shared programs: 6035392 -> 6035397 (0.00%) i965 instructions in affected programs: 537 -> 542 (0.93%) HURT: 2 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-01-19 22:57:38 +00:00
(('fxor', a, a), 0.0),
(('ixor', a, a), 0),
nir: Add opt_algebraic rules for xor with zero. instructions in affected programs: 668 -> 664 (-0.60%) helped: 4 Reviewed-by: Eduardo Lima Mitev <elima@igalia.com> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
2016-02-02 00:34:46 +00:00
(('ixor', a, 0), a),
nir: Add a bunch of algebraic optimizations on logic/bit operations. Matt and I noticed a bunch of "val <- ior a a" operations in a shader, so we decided to add an algebraic optimization for that. While there, I decided to add a bunch more of them. v2: Delete bogus fand/for optimizations (caught by Jason). total NIR instructions in shared programs: 2023511 -> 2020814 (-0.13%) NIR instructions in affected programs: 149634 -> 146937 (-1.80%) helped: 1032 total i965 instructions in shared programs: 6035392 -> 6035397 (0.00%) i965 instructions in affected programs: 537 -> 542 (0.93%) HURT: 2 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-01-19 22:57:38 +00:00
(('inot', ('inot', a)), a),
nir: Add a couple of iand/ior optimizations Spotted in a shader in Batman: Arkham City. Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
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)),
nir: Add a bunch of algebraic optimizations on logic/bit operations. Matt and I noticed a bunch of "val <- ior a a" operations in a shader, so we decided to add an algebraic optimization for that. While there, I decided to add a bunch more of them. v2: Delete bogus fand/for optimizations (caught by Jason). total NIR instructions in shared programs: 2023511 -> 2020814 (-0.13%) NIR instructions in affected programs: 149634 -> 146937 (-1.80%) helped: 1032 total i965 instructions in shared programs: 6035392 -> 6035397 (0.00%) i965 instructions in affected programs: 537 -> 542 (0.93%) HURT: 2 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
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))),
nir: Add algebraic optimizations for pointless shifts. The GLSL IR optimization pass contained these; we may as well include them too. v2: Fix a >> 0 and a << 0 optimizations (caught by Matt). No change in the number of NIR instructions on a shader-db run. total i965 instructions in shared programs: 6035397 -> 6035392 (-0.00%) i965 instructions in affected programs: 542 -> 537 (-0.92%) helped: 2 (in glamor) Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
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),
nir: Fix typo in "ushr by 0" algebraic replacement Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Jordan Justen <jordan.l.justen@intel.com> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com> Cc: "10.5" <mesa-stable@lists.freedestkop.org>
2015-04-14 16:40:22 +01:00
(('ushr', a, 0), a),
nir: fix algebraic optimizations The optimizations are only valid for 32-bit integers. They were mistakenly firing for 64-bit integers as well. Cc: mesa-stable@lists.freedesktop.org Reviewed-by: Matt Turner <mattst88@gmail.com>
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)),
nir: Add algebraic optimizations for exponential/logarithmic functions. Most of these exist in the GLSL IR algebraic pass already. However, SSA allows us to find more instances of the patterns. total NIR instructions in shared programs: 2015593 -> 2011430 (-0.21%) NIR instructions in affected programs: 124189 -> 120026 (-3.35%) helped: 604 total i965 instructions in shared programs: 6025505 -> 6018717 (-0.11%) i965 instructions in affected programs: 261295 -> 254507 (-2.60%) helped: 1295 HURT: 3 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-01-22 07:47:06 +00:00
# Exponential/logarithmic identities
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a
(('~flog2', ('fexp2', a)), a), # lg2(2^a) = a
nir: Add lowering of POW instructions if the lower flag is set. This could be done in a separate pass like we do in GLSL IR, but it seems to me like having the definitions of the transformations in the two directions next to each other makes a lot of sense. v2: Reorder the comment about the transformation. Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
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)
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
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
nir: lower fexp2(fmul(flog2(a), 2)) to fmul(a, a) Similar for the 4 case. Suggested by Bas. Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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))),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
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))),
nir: add is_used_once for fmul(fexp2(a), fexp2(b)) to fexp2(fadd(a, b)) Otherwise the code size increases because the original fexp2() instructions can't be deleted. Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-02-05 15:07:45 +00:00
(('~fmul', ('fexp2(is_used_once)', a), ('fexp2(is_used_once)', b)), ('fexp2', ('fadd', a, b))),
nir/algebraic: Simplify fsqrt domain guard All Gen7+ platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17228376 -> 17225365 (-0.02%) instructions in affected programs: 280732 -> 277721 (-1.07%) helped: 1072 HURT: 0 helped stats (abs) min: 1 max: 12 x̄: 2.81 x̃: 2 helped stats (rel) min: 0.16% max: 5.10% x̄: 1.43% x̃: 1.07% 95% mean confidence interval for instructions value: -2.92 -2.70 95% mean confidence interval for instructions %-change: -1.48% -1.37% Instructions are helped. total cycles in shared programs: 360935690 -> 360842788 (-0.03%) cycles in affected programs: 7838017 -> 7745115 (-1.19%) helped: 1569 HURT: 69 helped stats (abs) min: 1 max: 1198 x̄: 63.53 x̃: 20 helped stats (rel) min: 0.06% max: 26.17% x̄: 3.44% x̃: 2.12% HURT stats (abs) min: 1 max: 2820 x̄: 98.22 x̃: 47 HURT stats (rel) min: 0.05% max: 16.67% x̄: 3.50% x̃: 2.31% 95% mean confidence interval for cycles value: -63.55 -49.89 95% mean confidence interval for cycles %-change: -3.33% -2.96% Cycles are helped. No changes on any other platform. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2018-10-25 03:48:49 +01:00
(('bcsel', ('flt', a, 0.0), 0.0, ('fsqrt', a)), ('fsqrt', ('fmax', a, 0.0))),
nir: Add algebraic optimizations for division and reciprocal. These also exist in opt_algebraic.cpp. total NIR instructions in shared programs: 2011430 -> 2011211 (-0.01%) NIR instructions in affected programs: 42221 -> 42002 (-0.52%) helped: 198 total i965 instructions in shared programs: 6020553 -> 6020116 (-0.01%) i965 instructions in affected programs: 84322 -> 83885 (-0.52%) helped: 394 HURT: 1 (by 1 instruction) Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-01-22 07:52:17 +00:00
# Division and reciprocal
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
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'),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~frcp', ('frcp', a)), a),
(('~frcp', ('fsqrt', a)), ('frsq', a)),
nir: Add a flag for lowering fsqrt(x) to frcp(frsqrt(x)). vc4 cse/algebraic-disabled stats: total uniforms in shared programs: 13972 -> 13966 (-0.04%) uniforms in affected programs: 408 -> 402 (-1.47%) total instructions in shared programs: 44973 -> 44911 (-0.14%) instructions in affected programs: 1551 -> 1489 (-4.00%) v2: Rebase to master (no TGSI->NIR present) Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
2015-01-28 18:39:29 +00:00
(('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
nir/opt_algebraic: Add some boolean simplifications total instructions in shared programs: 5998321 -> 5998287 (-0.00%) instructions in affected programs: 4520 -> 4486 (-0.75%) helped: 8 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2015-01-29 00:55:03 +00:00
# Boolean simplifications
nir/algebraic: Replace i2b used by bcsel or if-statement with comparison All of the helped shaders are in Deus Ex. I looked at a couple shaders, and they have a pattern like: vec1 32 ssa_373 = i2b32 ssa_345.w vec1 32 ssa_374 = bcsel ssa_373, ssa_20, ssa_0 ... vec1 32 ssa_377 = ine ssa_345.w, ssa_0 if ssa_377 { ... vec1 32 ssa_416 = i2b32 ssa_385.w vec1 32 ssa_417 = bcsel ssa_416, ssa_386, ssa_374 ... } The massive help occurs because the i2b32 is removed, then other passes determine that ssa_374 must be ssa_20 inside the if-statement allowing the first bcsel to also be deleted. v2: Rebase on 1-bit Boolean changes. v3: Fix i2b32 vs ine problem in if-statement replacement. Noticed by Bas. Skylake total instructions in shared programs: 15241394 -> 15186287 (-0.36%) instructions in affected programs: 890583 -> 835476 (-6.19%) helped: 355 HURT: 0 helped stats (abs) min: 1 max: 497 x̄: 155.23 x̃: 149 helped stats (rel) min: 0.09% max: 16.49% x̄: 6.10% x̃: 6.59% 95% mean confidence interval for instructions value: -165.07 -145.39 95% mean confidence interval for instructions %-change: -6.42% -5.77% Instructions are helped. total cycles in shared programs: 373846583 -> 371023357 (-0.76%) cycles in affected programs: 118972102 -> 116148876 (-2.37%) helped: 343 HURT: 14 helped stats (abs) min: 45 max: 118284 x̄: 8332.32 x̃: 6089 helped stats (rel) min: 0.03% max: 38.19% x̄: 2.48% x̃: 1.77% HURT stats (abs) min: 120 max: 4126 x̄: 2482.79 x̃: 3019 HURT stats (rel) min: 0.16% max: 17.37% x̄: 2.13% x̃: 1.11% 95% mean confidence interval for cycles value: -8723.28 -7093.12 95% mean confidence interval for cycles %-change: -2.57% -2.02% Cycles are helped. total spills in shared programs: 32401 -> 23465 (-27.58%) spills in affected programs: 24457 -> 15521 (-36.54%) helped: 343 HURT: 0 total fills in shared programs: 37866 -> 31765 (-16.11%) fills in affected programs: 18889 -> 12788 (-32.30%) helped: 343 HURT: 0 Broadwell and Haswell had similar results. (Haswell shown) Haswell total instructions in shared programs: 13764783 -> 13750679 (-0.10%) instructions in affected programs: 1176256 -> 1162152 (-1.20%) helped: 334 HURT: 21 helped stats (abs) min: 1 max: 358 x̄: 42.59 x̃: 47 helped stats (rel) min: 0.09% max: 11.81% x̄: 1.30% x̃: 1.37% HURT stats (abs) min: 1 max: 61 x̄: 5.76 x̃: 1 HURT stats (rel) min: 0.03% max: 1.84% x̄: 0.17% x̃: 0.03% 95% mean confidence interval for instructions value: -43.99 -35.47 95% mean confidence interval for instructions %-change: -1.35% -1.08% Instructions are helped. total cycles in shared programs: 386511910 -> 385402528 (-0.29%) cycles in affected programs: 143831110 -> 142721728 (-0.77%) helped: 327 HURT: 39 helped stats (abs) min: 16 max: 25219 x̄: 3519.74 x̃: 3570 helped stats (rel) min: <.01% max: 10.26% x̄: 0.95% x̃: 0.96% HURT stats (abs) min: 16 max: 4881 x̄: 1065.95 x̃: 997 HURT stats (rel) min: <.01% max: 16.67% x̄: 0.70% x̃: 0.24% 95% mean confidence interval for cycles value: -3375.59 -2686.60 95% mean confidence interval for cycles %-change: -0.92% -0.64% Cycles are helped. total spills in shared programs: 100480 -> 97846 (-2.62%) spills in affected programs: 84702 -> 82068 (-3.11%) helped: 316 HURT: 21 total fills in shared programs: 96877 -> 94369 (-2.59%) fills in affected programs: 69167 -> 66659 (-3.63%) helped: 316 HURT: 9 No changes on Ivy Bridge or earlier platforms. Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
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)),
nir/algebraic: Drop some @bool specifiers Now that we have one-bit booleans, we don't need to rely on looking at parent instructions in order to figure out if a value is a Boolean most of the time. We can drop these specifiers and now the optimizations will apply more generally. Shader-DB results on Kaby Lake: total instructions in shared programs: 15321168 -> 15321227 (<.01%) instructions in affected programs: 8836 -> 8895 (0.67%) helped: 1 HURT: 31 total cycles in shared programs: 357481781 -> 357481321 (<.01%) cycles in affected programs: 146524 -> 146064 (-0.31%) helped: 22 HURT: 10 total spills in shared programs: 23675 -> 23673 (<.01%) spills in affected programs: 11 -> 9 (-18.18%) helped: 1 HURT: 0 total fills in shared programs: 32040 -> 32036 (-0.01%) fills in affected programs: 27 -> 23 (-14.81%) helped: 1 HURT: 0 No change in VkPipeline-DB Looking at the instructions hurt, a bunch of them seem to be a case where doing exactly the right thing in NIR ends up doing the wrong-ish thing in the back-end because flags are dumb. In particular, there's a case where we have a MUL followed by a CMP followed by a SEL and when we turn that SEL into an OR, it uses the GRF result of the CMP rather than the flag result so the CMP can't be merged with the MUL. Those shaders appear to schedule better according to the cycle estimates so I guess it's a win? Also it helps spilling in one Car Chase compute shader. Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
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')),
nir: Rely on the fact that bcsel takes a well formed boolean. According to Connor, it's safe to assume that the first operand of bcsel, as well as the operand of b2f and b2i, must be well formed booleans. https://lists.freedesktop.org/archives/mesa-dev/2016-August/125658.html With the previous improvements to a@bool handling, this now has no change in shader-db instruction counts on Broadwell. Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-08-17 23:03:01 +01:00
(('bcsel', a, True, False), a),
(('bcsel', a, False, True), ('inot', a)),
nir: Turn bcsel of +/- 1.0 and 0.0 into b2f sequences. On BDW: total instructions in shared programs: 13074882 -> 13068703 (-0.05%) instructions in affected programs: 1823116 -> 1816937 (-0.34%) helped: 4187 HURT: 537 total cycles in shared programs: 256622718 -> 256425382 (-0.08%) cycles in affected programs: 123790120 -> 123592784 (-0.16%) helped: 3823 HURT: 2037 total spills in shared programs: 15276 -> 14929 (-2.27%) spills in affected programs: 9446 -> 9099 (-3.67%) helped: 352 HURT: 1 total fills in shared programs: 20496 -> 20144 (-1.72%) fills in affected programs: 13040 -> 12688 (-2.70%) helped: 352 HURT: 1 LOST: 2 GAINED: 21 v2: Rely on 'a' being a well-formed boolean (Connor, Eric). Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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)))),
nir/opt_algebraic: Add some constant bcsel reductions total instructions in shared programs: 5998190 -> 5997603 (-0.01%) instructions in affected programs: 54276 -> 53689 (-1.08%) helped: 293 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2015-01-29 00:53:51 +00:00
(('bcsel', True, b, c), b),
(('bcsel', False, b, c), c),
nir/opt_algebraic: Add 32-bit specifiers to a bunch of booleans Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2018-11-09 00:08:37 +00:00
(('bcsel', a, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a, b, c))),
nir: Add a bunch of algebraic optimizations on logic/bit operations. Matt and I noticed a bunch of "val <- ior a a" operations in a shader, so we decided to add an algebraic optimization for that. While there, I decided to add a bunch more of them. v2: Delete bogus fand/for optimizations (caught by Jason). total NIR instructions in shared programs: 2023511 -> 2020814 (-0.13%) NIR instructions in affected programs: 149634 -> 146937 (-1.80%) helped: 1032 total i965 instructions in shared programs: 6035392 -> 6035397 (0.00%) i965 instructions in affected programs: 537 -> 542 (0.93%) HURT: 2 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-01-19 22:57:38 +00:00
nir: Add a couple of simplifications of csel operations. vc4 was already cleaning these up, but it does shave 4 NIR instructions in shader-db. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-01-29 23:50:18 +00:00
(('bcsel', a, b, b), b),
(('fcsel', a, b, b), b),
nir/algebraic: Add some optimizations for D3D-style Booleans D3D Booleans use a 32-bit 0/-1 representation. Because this previously matched NIR exactly, we didn't have to really optimize for it. Now that we have 1-bit Booleans, we need some specific optimizations to chew through the D3D12-style Booleans. Shader-db results on Kaby Lake: total instructions in shared programs: 15136811 -> 14967944 (-1.12%) instructions in affected programs: 2457021 -> 2288154 (-6.87%) helped: 8318 HURT: 10 total cycles in shared programs: 373544524 -> 359701825 (-3.71%) cycles in affected programs: 151029683 -> 137186984 (-9.17%) helped: 7749 HURT: 682 total loops in shared programs: 4431 -> 4399 (-0.72%) loops in affected programs: 32 -> 0 helped: 21 HURT: 0 total spills in shared programs: 10290 -> 10051 (-2.32%) spills in affected programs: 2532 -> 2293 (-9.44%) helped: 18 HURT: 18 total fills in shared programs: 22203 -> 21732 (-2.12%) fills in affected programs: 3319 -> 2848 (-14.19%) helped: 18 HURT: 18 Note that a large chunk of the improvement fixing regressions caused by switching to 1-bit Booleans. Previously, our ability to optimize D3D booleans was improved by using the D3D representation directly in NIR. Now that NIR does 1-bit bools, we need a few more optimizations. Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Reviewed-by: Eric Anholt <eric@anholt.net> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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)),
nir/algebraic: Don't put quotes around floating point literals The quotation marks around 1.0 cause it to be treated as a string instead of a floating point value. The generator then treats it as an arbitrary variable replacement, so any iand involving a ('ineg', ('b2i', a)) matches. v2: Remove misleading comment about sized literals (suggested by Timothy). Add assertion that the name of a varible is entierly alphabetic (suggested by Jason). Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> Tested-by: Timothy Arceri <tarceri@itsqueeze.com> [v1] Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com> [v1] Fixes: 6bcd2af0863 ("nir/algebraic: Add some optimizations for D3D-style Booleans") Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=109075
2018-12-18 21:28:22 +00:00
(('iand', ('ineg', ('b2i', a)), 1.0), ('b2f', a)),
nir/algebraic: Add some optimizations for D3D-style Booleans D3D Booleans use a 32-bit 0/-1 representation. Because this previously matched NIR exactly, we didn't have to really optimize for it. Now that we have 1-bit Booleans, we need some specific optimizations to chew through the D3D12-style Booleans. Shader-db results on Kaby Lake: total instructions in shared programs: 15136811 -> 14967944 (-1.12%) instructions in affected programs: 2457021 -> 2288154 (-6.87%) helped: 8318 HURT: 10 total cycles in shared programs: 373544524 -> 359701825 (-3.71%) cycles in affected programs: 151029683 -> 137186984 (-9.17%) helped: 7749 HURT: 682 total loops in shared programs: 4431 -> 4399 (-0.72%) loops in affected programs: 32 -> 0 helped: 21 HURT: 0 total spills in shared programs: 10290 -> 10051 (-2.32%) spills in affected programs: 2532 -> 2293 (-9.44%) helped: 18 HURT: 18 total fills in shared programs: 22203 -> 21732 (-2.12%) fills in affected programs: 3319 -> 2848 (-14.19%) helped: 18 HURT: 18 Note that a large chunk of the improvement fixing regressions caused by switching to 1-bit Booleans. Previously, our ability to optimize D3D booleans was improved by using the D3D representation directly in NIR. Now that NIR does 1-bit bools, we need a few more optimizations. Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Reviewed-by: Eric Anholt <eric@anholt.net> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-10-19 22:58:36 +01:00
nir: Use SM5 properties to optimize shift(a@32, iand(31, b)) This is a common pattern from HLSL->SPIRV translation and supported in HW by all current NIR backends. vkpipeline-db results anv (SKL): total instructions in shared programs: 6403130 -> 6402380 (-0.01%) instructions in affected programs: 204084 -> 203334 (-0.37%) helped: 208 HURT: 0 total cycles in shared programs: 1915629582 -> 1918198408 (0.13%) cycles in affected programs: 1158892682 -> 1161461508 (0.22%) helped: 107 HURT: 86 shader-db results on i965 (KBL): total instructions in shared programs: 15284592 -> 15284568 (<.01%) instructions in affected programs: 81683 -> 81659 (-0.03%) helped: 24 HURT: 0 total cycles in shared programs: 375013622 -> 375013932 (<.01%) cycles in affected programs: 40169618 -> 40169928 (<.01%) helped: 13 HURT: 9 Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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)),
nir: Remove useless ftrunc inside f2i/f2u. No shader-db changes, probably because they're all removed by the GLSL compiler optimization added in commit 69ad5fd4. Reviewed-by: Eric Anholt <eric@anholt.net>
2015-03-26 17:09:42 +00:00
# Conversions
nir: Make boolean conversions sized just like the others Instead of a single i2b and b2i, we now have i2b32 and b2iN where N is one if 8, 16, 32, or 64. This leads to having a few more opcodes but now everything is consistent and booleans aren't a weird special case anymore. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2018-11-07 19:43:40 +00:00
(('i2b32', ('b2i', 'a@32')), a),
nir/opt_algebraic: Drop bit-size suffixes from conversions Suffixes are dropped from a bunch of conversion opcodes when it makes sense to do so. Others are kept if we really do want the bit-size restriction. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2018-11-07 23:32:32 +00:00
(('f2i', ('ftrunc', a)), ('f2i', a)),
(('f2u', ('ftrunc', a)), ('f2u', a)),
nir: Simplify i2b with negated or abs operand This enables removing ssa_201 and ssa_202 in sequences like: vec1 ssa_200 = flt ssa_199, ssa_194 vec1 ssa_201 = b2i ssa_200 vec1 ssa_202 = i2b -ssa_201 shader-db results: Sandy Bridge total instructions in shared programs: 8462257 -> 8462180 (-0.00%) instructions in affected programs: 3846 -> 3769 (-2.00%) helped: 35 HURT: 0 total cycles in shared programs: 117542934 -> 117542462 (-0.00%) cycles in affected programs: 20072 -> 19600 (-2.35%) helped: 20 HURT: 1 Ivy Bridge total instructions in shared programs: 7775252 -> 7775137 (-0.00%) instructions in affected programs: 3645 -> 3530 (-3.16%) helped: 35 HURT: 0 total cycles in shared programs: 65760522 -> 65760068 (-0.00%) cycles in affected programs: 21082 -> 20628 (-2.15%) helped: 25 HURT: 2 Haswell total instructions in shared programs: 7108666 -> 7108589 (-0.00%) instructions in affected programs: 3253 -> 3176 (-2.37%) helped: 35 HURT: 0 total cycles in shared programs: 64675726 -> 64675272 (-0.00%) cycles in affected programs: 21034 -> 20580 (-2.16%) helped: 26 HURT: 1 Broadwell / Skylake total instructions in shared programs: 8980912 -> 8980835 (-0.00%) instructions in affected programs: 3223 -> 3146 (-2.39%) helped: 35 HURT: 0 total cycles in shared programs: 70077926 -> 70077904 (-0.00%) cycles in affected programs: 21886 -> 21864 (-0.10%) helped: 21 HURT: 6 G45 and Ironlake showed no change. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Suggested-by: Jason Ekstrand <jason.ekstrand@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-03-02 02:59:57 +00:00
(('i2b', ('ineg', a)), ('i2b', a)),
(('i2b', ('iabs', a)), ('i2b', a)),
nir: Don't abs the result of b2f or b2i In the results below, 2 SIMD16 shaders in Trine are lost. G4X total instructions in shared programs: 4012279 -> 4011108 (-0.03%) instructions in affected programs: 116776 -> 115605 (-1.00%) helped: 339 HURT: 0 total cycles in shared programs: 84315862 -> 84313584 (-0.00%) cycles in affected programs: 1767232 -> 1764954 (-0.13%) helped: 274 HURT: 81 Ironlake total instructions in shared programs: 6399073 -> 6396998 (-0.03%) instructions in affected programs: 218050 -> 215975 (-0.95%) helped: 600 HURT: 0 total cycles in shared programs: 128892088 -> 128888810 (-0.00%) cycles in affected programs: 2867452 -> 2864174 (-0.11%) helped: 422 HURT: 137 Sandy Bridge total instructions in shared programs: 8462174 -> 8460759 (-0.02%) instructions in affected programs: 178529 -> 177114 (-0.79%) helped: 596 HURT: 0 total cycles in shared programs: 117542276 -> 117534098 (-0.01%) cycles in affected programs: 1239166 -> 1230988 (-0.66%) helped: 369 HURT: 150 Ivy Bridge total instructions in shared programs: 7775131 -> 7773410 (-0.02%) instructions in affected programs: 162903 -> 161182 (-1.06%) helped: 590 HURT: 0 total cycles in shared programs: 65759882 -> 65747268 (-0.02%) cycles in affected programs: 1004354 -> 991740 (-1.26%) helped: 467 HURT: 141 Haswell total instructions in shared programs: 7107786 -> 7106327 (-0.02%) instructions in affected programs: 140954 -> 139495 (-1.04%) helped: 590 HURT: 0 total cycles in shared programs: 64668028 -> 64655322 (-0.02%) cycles in affected programs: 967080 -> 954374 (-1.31%) helped: 452 HURT: 149 LOST: 2 GAINED: 0 Broadwell total instructions in shared programs: 8980029 -> 8978287 (-0.02%) instructions in affected programs: 197232 -> 195490 (-0.88%) helped: 715 HURT: 0 total cycles in shared programs: 70070448 -> 70055970 (-0.02%) cycles in affected programs: 975724 -> 961246 (-1.48%) helped: 471 HURT: 111 LOST: 2 GAINED: 0 Skylake total instructions in shared programs: 9115178 -> 9113436 (-0.02%) instructions in affected programs: 203012 -> 201270 (-0.86%) helped: 715 HURT: 0 total cycles in shared programs: 68848660 -> 68834004 (-0.02%) cycles in affected programs: 993888 -> 979232 (-1.47%) helped: 473 HURT: 116 LOST: 2 GAINED: 0 Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-03-02 21:46:50 +00:00
(('fabs', ('b2f', a)), ('b2f', a)),
(('iabs', ('b2i', a)), ('b2i', a)),
nir/algebraic: Optimize 1-bit Booleans Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
2018-12-06 03:36:10 +00:00
(('inot', ('f2b1', a)), ('feq', a, 0.0)),
nir: Remove useless ftrunc inside f2i/f2u. No shader-db changes, probably because they're all removed by the GLSL compiler optimization added in commit 69ad5fd4. Reviewed-by: Eric Anholt <eric@anholt.net>
2015-03-26 17:09:42 +00:00
nir: Remove f2i(i2f(x)) conversions Broadwell and Skylake had similar results. (Skylake shown) total instructions in shared programs: 14277978 -> 14277620 (<.01%) instructions in affected programs: 36957 -> 36599 (-0.97%) helped: 76 HURT: 1 helped stats (abs) min: 2 max: 90 x̄: 4.89 x̃: 4 helped stats (rel) min: 0.44% max: 5.88% x̄: 1.04% x̃: 0.87% HURT stats (abs) min: 14 max: 14 x̄: 14.00 x̃: 14 HURT stats (rel) min: 0.36% max: 0.36% x̄: 0.36% x̃: 0.36% 95% mean confidence interval for instructions value: -7.06 -2.24 95% mean confidence interval for instructions %-change: -1.28% -0.77% Instructions are helped. total cycles in shared programs: 532584581 -> 532580716 (<.01%) cycles in affected programs: 973591 -> 969726 (-0.40%) helped: 76 HURT: 1 helped stats (abs) min: 2 max: 9940 x̄: 159.80 x̃: 32 helped stats (rel) min: <.01% max: 8.70% x̄: 1.15% x̃: 1.19% HURT stats (abs) min: 8280 max: 8280 x̄: 8280.00 x̃: 8280 HURT stats (rel) min: 2.10% max: 2.10% x̄: 2.10% x̃: 2.10% 95% mean confidence interval for cycles value: -386.98 286.59 95% mean confidence interval for cycles %-change: -1.41% -0.81% Inconclusive result (value mean confidence interval includes 0). total spills in shared programs: 8127 -> 8116 (-0.14%) spills in affected programs: 108 -> 97 (-10.19%) helped: 1 HURT: 0 total fills in shared programs: 11090 -> 11086 (-0.04%) fills in affected programs: 440 -> 436 (-0.91%) helped: 1 HURT: 1 Haswell total instructions in shared programs: 12979174 -> 12979054 (<.01%) instructions in affected programs: 9040 -> 8920 (-1.33%) helped: 14 HURT: 1 helped stats (abs) min: 2 max: 34 x̄: 8.79 x̃: 6 helped stats (rel) min: 0.41% max: 7.04% x̄: 2.66% x̃: 1.14% HURT stats (abs) min: 3 max: 3 x̄: 3.00 x̃: 3 HURT stats (rel) min: 0.19% max: 0.19% x̄: 0.19% x̃: 0.19% 95% mean confidence interval for instructions value: -13.58 -2.42 95% mean confidence interval for instructions %-change: -3.94% -1.01% Instructions are helped. total cycles in shared programs: 411227148 -> 411212880 (<.01%) cycles in affected programs: 630506 -> 616238 (-2.26%) helped: 15 HURT: 0 helped stats (abs) min: 2 max: 11192 x̄: 951.20 x̃: 38 helped stats (rel) min: <.01% max: 16.01% x̄: 3.92% x̃: 0.17% 95% mean confidence interval for cycles value: -2544.28 641.88 95% mean confidence interval for cycles %-change: -6.89% -0.94% Inconclusive result (value mean confidence interval includes 0). total spills in shared programs: 78626 -> 78623 (<.01%) spills in affected programs: 42 -> 39 (-7.14%) helped: 1 HURT: 0 total fills in shared programs: 80111 -> 80104 (<.01%) fills in affected programs: 140 -> 133 (-5.00%) helped: 1 HURT: 1 Ivy Bridge total instructions in shared programs: 11684101 -> 11684030 (<.01%) instructions in affected programs: 3080 -> 3009 (-2.31%) helped: 4 HURT: 1 helped stats (abs) min: 5 max: 59 x̄: 18.50 x̃: 5 helped stats (rel) min: 6.47% max: 7.04% x̄: 6.87% x̃: 6.99% HURT stats (abs) min: 3 max: 3 x̄: 3.00 x̃: 3 HURT stats (rel) min: 0.15% max: 0.15% x̄: 0.15% x̃: 0.15% 95% mean confidence interval for instructions value: -45.59 17.19 95% mean confidence interval for instructions %-change: -9.38% -1.56% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 258407697 -> 258389653 (<.01%) cycles in affected programs: 328323 -> 310279 (-5.50%) helped: 5 HURT: 0 helped stats (abs) min: 32 max: 14908 x̄: 3608.80 x̃: 32 helped stats (rel) min: 1.26% max: 17.22% x̄: 9.30% x̃: 10.60% 95% mean confidence interval for cycles value: -11616.71 4399.11 95% mean confidence interval for cycles %-change: -16.56% -2.03% Inconclusive result (value mean confidence interval includes 0). total spills in shared programs: 4537 -> 4528 (-0.20%) spills in affected programs: 64 -> 55 (-14.06%) helped: 1 HURT: 0 total fills in shared programs: 4823 -> 4815 (-0.17%) fills in affected programs: 189 -> 181 (-4.23%) helped: 1 HURT: 1 Sandy Bridge total instructions in shared programs: 10488464 -> 10488449 (<.01%) instructions in affected programs: 272 -> 257 (-5.51%) helped: 3 HURT: 0 helped stats (abs) min: 5 max: 5 x̄: 5.00 x̃: 5 helped stats (rel) min: 5.49% max: 5.56% x̄: 5.51% x̃: 5.49% total cycles in shared programs: 150263359 -> 150263263 (<.01%) cycles in affected programs: 7978 -> 7882 (-1.20%) helped: 3 HURT: 0 helped stats (abs) min: 32 max: 32 x̄: 32.00 x̃: 32 helped stats (rel) min: 1.15% max: 1.23% x̄: 1.20% x̃: 1.23% No changes on Iron Lake or GM45. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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).
nir/opt_algebraic: Drop bit-size suffixes from conversions Suffixes are dropped from a bunch of conversion opcodes when it makes sense to do so. Others are kept if we really do want the bit-size restriction. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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),
nir: Remove f2i(i2f(x)) conversions Broadwell and Skylake had similar results. (Skylake shown) total instructions in shared programs: 14277978 -> 14277620 (<.01%) instructions in affected programs: 36957 -> 36599 (-0.97%) helped: 76 HURT: 1 helped stats (abs) min: 2 max: 90 x̄: 4.89 x̃: 4 helped stats (rel) min: 0.44% max: 5.88% x̄: 1.04% x̃: 0.87% HURT stats (abs) min: 14 max: 14 x̄: 14.00 x̃: 14 HURT stats (rel) min: 0.36% max: 0.36% x̄: 0.36% x̃: 0.36% 95% mean confidence interval for instructions value: -7.06 -2.24 95% mean confidence interval for instructions %-change: -1.28% -0.77% Instructions are helped. total cycles in shared programs: 532584581 -> 532580716 (<.01%) cycles in affected programs: 973591 -> 969726 (-0.40%) helped: 76 HURT: 1 helped stats (abs) min: 2 max: 9940 x̄: 159.80 x̃: 32 helped stats (rel) min: <.01% max: 8.70% x̄: 1.15% x̃: 1.19% HURT stats (abs) min: 8280 max: 8280 x̄: 8280.00 x̃: 8280 HURT stats (rel) min: 2.10% max: 2.10% x̄: 2.10% x̃: 2.10% 95% mean confidence interval for cycles value: -386.98 286.59 95% mean confidence interval for cycles %-change: -1.41% -0.81% Inconclusive result (value mean confidence interval includes 0). total spills in shared programs: 8127 -> 8116 (-0.14%) spills in affected programs: 108 -> 97 (-10.19%) helped: 1 HURT: 0 total fills in shared programs: 11090 -> 11086 (-0.04%) fills in affected programs: 440 -> 436 (-0.91%) helped: 1 HURT: 1 Haswell total instructions in shared programs: 12979174 -> 12979054 (<.01%) instructions in affected programs: 9040 -> 8920 (-1.33%) helped: 14 HURT: 1 helped stats (abs) min: 2 max: 34 x̄: 8.79 x̃: 6 helped stats (rel) min: 0.41% max: 7.04% x̄: 2.66% x̃: 1.14% HURT stats (abs) min: 3 max: 3 x̄: 3.00 x̃: 3 HURT stats (rel) min: 0.19% max: 0.19% x̄: 0.19% x̃: 0.19% 95% mean confidence interval for instructions value: -13.58 -2.42 95% mean confidence interval for instructions %-change: -3.94% -1.01% Instructions are helped. total cycles in shared programs: 411227148 -> 411212880 (<.01%) cycles in affected programs: 630506 -> 616238 (-2.26%) helped: 15 HURT: 0 helped stats (abs) min: 2 max: 11192 x̄: 951.20 x̃: 38 helped stats (rel) min: <.01% max: 16.01% x̄: 3.92% x̃: 0.17% 95% mean confidence interval for cycles value: -2544.28 641.88 95% mean confidence interval for cycles %-change: -6.89% -0.94% Inconclusive result (value mean confidence interval includes 0). total spills in shared programs: 78626 -> 78623 (<.01%) spills in affected programs: 42 -> 39 (-7.14%) helped: 1 HURT: 0 total fills in shared programs: 80111 -> 80104 (<.01%) fills in affected programs: 140 -> 133 (-5.00%) helped: 1 HURT: 1 Ivy Bridge total instructions in shared programs: 11684101 -> 11684030 (<.01%) instructions in affected programs: 3080 -> 3009 (-2.31%) helped: 4 HURT: 1 helped stats (abs) min: 5 max: 59 x̄: 18.50 x̃: 5 helped stats (rel) min: 6.47% max: 7.04% x̄: 6.87% x̃: 6.99% HURT stats (abs) min: 3 max: 3 x̄: 3.00 x̃: 3 HURT stats (rel) min: 0.15% max: 0.15% x̄: 0.15% x̃: 0.15% 95% mean confidence interval for instructions value: -45.59 17.19 95% mean confidence interval for instructions %-change: -9.38% -1.56% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 258407697 -> 258389653 (<.01%) cycles in affected programs: 328323 -> 310279 (-5.50%) helped: 5 HURT: 0 helped stats (abs) min: 32 max: 14908 x̄: 3608.80 x̃: 32 helped stats (rel) min: 1.26% max: 17.22% x̄: 9.30% x̃: 10.60% 95% mean confidence interval for cycles value: -11616.71 4399.11 95% mean confidence interval for cycles %-change: -16.56% -2.03% Inconclusive result (value mean confidence interval includes 0). total spills in shared programs: 4537 -> 4528 (-0.20%) spills in affected programs: 64 -> 55 (-14.06%) helped: 1 HURT: 0 total fills in shared programs: 4823 -> 4815 (-0.17%) fills in affected programs: 189 -> 181 (-4.23%) helped: 1 HURT: 1 Sandy Bridge total instructions in shared programs: 10488464 -> 10488449 (<.01%) instructions in affected programs: 272 -> 257 (-5.51%) helped: 3 HURT: 0 helped stats (abs) min: 5 max: 5 x̄: 5.00 x̃: 5 helped stats (rel) min: 5.49% max: 5.56% x̄: 5.51% x̃: 5.49% total cycles in shared programs: 150263359 -> 150263263 (<.01%) cycles in affected programs: 7978 -> 7882 (-1.20%) helped: 3 HURT: 0 helped stats (abs) min: 32 max: 32 x̄: 32.00 x̃: 32 helped stats (rel) min: 1.15% max: 1.23% x̄: 1.20% x̃: 1.23% No changes on Iron Lake or GM45. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2018-07-03 21:57:06 +01:00
nir/algebraic: Convert some f2u to f2i 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. This basically undoes the small amount of damage done by "intel/compiler: Avoid propagating inequality cmods if types are different". v2: Replicate part of the commit message as a comment in the code. Suggested by Jason. shader-db results compairing *before* "intel/compiler: Avoid propagating inequality cmods if types are different" and after this commit: Skylake total cycles in shared programs: 383007996 -> 383007896 (<.01%) cycles in affected programs: 85208 -> 85108 (-0.12%) helped: 13 HURT: 8 helped stats (abs) min: 2 max: 26 x̄: 10.77 x̃: 6 helped stats (rel) min: 0.09% max: 0.65% x̄: 0.28% x̃: 0.14% HURT stats (abs) min: 2 max: 12 x̄: 5.00 x̃: 3 HURT stats (rel) min: 0.04% max: 0.32% x̄: 0.12% x̃: 0.07% 95% mean confidence interval for cycles value: -9.31 -0.21 95% mean confidence interval for cycles %-change: -0.24% <.01% Cycles are helped. Broadwell total cycles in shared programs: 415251194 -> 415251370 (<.01%) cycles in affected programs: 83750 -> 83926 (0.21%) helped: 7 HURT: 13 helped stats (abs) min: 10 max: 12 x̄: 11.43 x̃: 12 helped stats (rel) min: 0.30% max: 0.30% x̄: 0.30% x̃: 0.30% HURT stats (abs) min: 2 max: 36 x̄: 19.69 x̃: 22 HURT stats (rel) min: 0.05% max: 0.89% x̄: 0.44% x̃: 0.47% 95% mean confidence interval for cycles value: 0.76 16.84 95% mean confidence interval for cycles %-change: <.01% 0.37% Inconclusive result (%-change mean confidence interval includes 0). Haswell total instructions in shared programs: 13823885 -> 13823886 (<.01%) instructions in affected programs: 2249 -> 2250 (0.04%) helped: 0 HURT: 1 total cycles in shared programs: 390094243 -> 390094001 (<.01%) cycles in affected programs: 85640 -> 85398 (-0.28%) helped: 15 HURT: 6 helped stats (abs) min: 4 max: 26 x̄: 18.53 x̃: 18 helped stats (rel) min: 0.09% max: 0.66% x̄: 0.47% x̃: 0.42% HURT stats (abs) min: 2 max: 14 x̄: 6.00 x̃: 2 HURT stats (rel) min: 0.04% max: 0.37% x̄: 0.15% x̃: 0.04% 95% mean confidence interval for cycles value: -17.36 -5.69 95% mean confidence interval for cycles %-change: -0.44% -0.14% Cycles are helped. Ivy Bridge total cycles in shared programs: 180986448 -> 180986552 (<.01%) cycles in affected programs: 34835 -> 34939 (0.30%) helped: 0 HURT: 10 HURT stats (abs) min: 2 max: 18 x̄: 10.40 x̃: 10 HURT stats (rel) min: 0.06% max: 0.36% x̄: 0.28% x̃: 0.30% 95% mean confidence interval for cycles value: 4.67 16.13 95% mean confidence interval for cycles %-change: 0.20% 0.35% Cycles are HURT. Sandy Bridge total cycles in shared programs: 154603969 -> 154603970 (<.01%) cycles in affected programs: 171514 -> 171515 (<.01%) helped: 25 HURT: 14 helped stats (abs) min: 1 max: 4 x̄: 1.80 x̃: 1 helped stats (rel) min: 0.02% max: 0.10% x̄: 0.04% x̃: 0.04% HURT stats (abs) min: 1 max: 8 x̄: 3.29 x̃: 3 HURT stats (rel) min: 0.03% max: 0.28% x̄: 0.10% x̃: 0.11% 95% mean confidence interval for cycles value: -0.91 0.96 95% mean confidence interval for cycles %-change: -0.02% 0.04% Inconclusive result (value mean confidence interval includes 0). No changes on Iron Lake or GM45. Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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))),
nir/algebraic: Recognize open-coded fsat with modifiers This change also enables a later change (nir/algebraic: Replace 1-fsat(a) with fsat(1-a)) to affect more shaders. Almost all of the affected shaders are in Bioshock Infinite, and all of those shaders all require GLSL 4.10. All Intel platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17228584 -> 17228376 (<.01%) instructions in affected programs: 31438 -> 31230 (-0.66%) helped: 105 HURT: 0 helped stats (abs) min: 1 max: 5 x̄: 1.98 x̃: 1 helped stats (rel) min: 0.08% max: 1.53% x̄: 0.73% x̃: 0.70% 95% mean confidence interval for instructions value: -2.20 -1.76 95% mean confidence interval for instructions %-change: -0.80% -0.67% Instructions are helped. total cycles in shared programs: 360936431 -> 360935690 (<.01%) cycles in affected programs: 420100 -> 419359 (-0.18%) helped: 71 HURT: 21 helped stats (abs) min: 1 max: 160 x̄: 19.28 x̃: 10 helped stats (rel) min: <.01% max: 9.78% x̄: 0.95% x̃: 0.48% HURT stats (abs) min: 1 max: 198 x̄: 29.90 x̃: 10 HURT stats (rel) min: 0.05% max: 8.36% x̄: 1.24% x̃: 0.90% 95% mean confidence interval for cycles value: -16.77 0.66 95% mean confidence interval for cycles %-change: -0.85% -0.06% Inconclusive result (value mean confidence interval includes 0). Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
2019-03-05 20:46:11 +00:00
(('~fmin', ('fabs', a), 1.0), ('fsat', ('fabs', a)), '!options->lower_fsat'),
nir/algebraic: Eliminate a tautological compare The value-range tracking pass that is coming is not clever enough to know that the result of the ffma must be non-negative. Making it that smart will require quite a bit of work. It might be possible to add a special case that detects that a whole tree of fadd(fmul(fsat(a), fneg(fsat(a))), 1.0) cannot be negative. For cases when the comparison is used in the domain guard for a square-root (see nir/algebraic: Simplify fsqrt domain guard), the compare may be converted to a fmax. This patch also handles that case. All of the affected cases are in DiRT: Showdown. All Gen7+ platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17225365 -> 17225303 (<.01%) instructions in affected programs: 40051 -> 39989 (-0.15%) helped: 62 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.07% max: 0.66% x̄: 0.27% x̃: 0.26% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -0.31% -0.22% Instructions are helped. total cycles in shared programs: 360842788 -> 360842595 (<.01%) cycles in affected programs: 1818081 -> 1817888 (-0.01%) helped: 29 HURT: 22 helped stats (abs) min: 1 max: 206 x̄: 20.66 x̃: 14 helped stats (rel) min: <.01% max: 9.55% x̄: 0.87% x̃: 0.42% HURT stats (abs) min: 1 max: 108 x̄: 18.45 x̃: 7 HURT stats (rel) min: <.01% max: 4.48% x̄: 0.56% x̃: 0.19% 95% mean confidence interval for cycles value: -14.48 6.91 95% mean confidence interval for cycles %-change: -0.71% 0.21% Inconclusive result (value mean confidence interval includes 0). No changes on any other Intel platform. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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)),
nir/algebraic: Optimize 64bit pack/unpack This reduces the instruction count in some fp64 and int64 piglit tests Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
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
nir/algebraic: Add missing 16-bit extract_[iu]8 patterns No shader-db changes on any Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
2019-02-28 04:15:32 +00:00
(('ushr', 'a@16', 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
nir/algebraic: Be more careful converting ushr to extract_u8/16 If it's not the right bit-size, it may not actually be the correct extraction. For now, we'll only worry about 32-bit versions. Fixes: 905ff8619824 "nir: Recognize open-coded extract_u16" Fixes: 76289fbfa84a "nir: Recognize open-coded extract_u8" Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-14 21:10:22 +01:00
(('ushr', 'a@32', 24), ('extract_u8', a, 3), '!options->lower_extract_byte'),
nir/algebraic: Add missing 64-bit extract_[iu]8 patterns No shader-db changes on any Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
2019-02-28 04:12:46 +00:00
(('ushr', 'a@64', 56), ('extract_u8', a, 7), '!options->lower_extract_byte'),
nir/algebraic: Add missing 16-bit extract_[iu]8 patterns No shader-db changes on any Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
2019-02-28 04:15:32 +00:00
(('ishr', 'a@16', 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
nir/algebraic: Add more extract_[iu](8|16) optimizations This adds the "(a << N) >> M" family of mask or sign-extensions. Not a huge win right now but this pattern will soon be generated by NIR format lowering code. Shader-db results on Kaby Lake: total instructions in shared programs: 15166918 -> 15166916 (<.01%) instructions in affected programs: 36 -> 34 (-5.56%) helped: 2 HURT: 0 Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-14 20:37:39 +01:00
(('ishr', 'a@32', 24), ('extract_i8', a, 3), '!options->lower_extract_byte'),
nir/algebraic: Add missing 64-bit extract_[iu]8 patterns No shader-db changes on any Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
2019-02-28 04:12:46 +00:00
(('ishr', 'a@64', 56), ('extract_i8', a, 7), '!options->lower_extract_byte'),
nir/algebraic: Fix up extract_[iu]8 after loop unrolling Skylake, Broadwell, and Haswell had similar results. (Skylake shown) total instructions in shared programs: 15256840 -> 15256837 (<.01%) instructions in affected programs: 4713 -> 4710 (-0.06%) helped: 3 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.06% max: 0.08% x̄: 0.06% x̃: 0.06% total cycles in shared programs: 372286583 -> 372286583 (0.00%) cycles in affected programs: 198516 -> 198516 (0.00%) helped: 1 HURT: 1 helped stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 helped stats (rel) min: <.01% max: <.01% x̄: <.01% x̃: <.01% HURT stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 HURT stats (rel) min: 0.01% max: 0.01% x̄: 0.01% x̃: 0.01% No changes on any other Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
2019-02-28 03:52:12 +00:00
(('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte')
]
nir/algebraic: Remove redundant extract_[iu]8 patterns No shader-db changes on any Intel platform. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Dylan Baker <dylan@pnwbakers.com> Acked-by: Jason Ekstrand <jason@jlekstrand.net>
2019-02-28 04:08:38 +00:00
# After the ('extract_u8', a, 0) pattern, above, triggers, there will be
# patterns like those below.
nir/algebraic: Fix up extract_[iu]8 after loop unrolling Skylake, Broadwell, and Haswell had similar results. (Skylake shown) total instructions in shared programs: 15256840 -> 15256837 (<.01%) instructions in affected programs: 4713 -> 4710 (-0.06%) helped: 3 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.06% max: 0.08% x̄: 0.06% x̃: 0.06% total cycles in shared programs: 372286583 -> 372286583 (0.00%) cycles in affected programs: 198516 -> 198516 (0.00%) helped: 1 HURT: 1 helped stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 helped stats (rel) min: <.01% max: <.01% x̄: <.01% x̃: <.01% HURT stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 HURT stats (rel) min: 0.01% max: 0.01% x̄: 0.01% x̃: 0.01% No changes on any other Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
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
nir/algebraic: Remove redundant extract_[iu]8 patterns No shader-db changes on any Intel platform. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Dylan Baker <dylan@pnwbakers.com> Acked-by: Jason Ekstrand <jason@jlekstrand.net>
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.
nir/algebraic: Fix up extract_[iu]8 after loop unrolling Skylake, Broadwell, and Haswell had similar results. (Skylake shown) total instructions in shared programs: 15256840 -> 15256837 (<.01%) instructions in affected programs: 4713 -> 4710 (-0.06%) helped: 3 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.06% max: 0.08% x̄: 0.06% x̃: 0.06% total cycles in shared programs: 372286583 -> 372286583 (0.00%) cycles in affected programs: 198516 -> 198516 (0.00%) helped: 1 HURT: 1 helped stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 helped stats (rel) min: <.01% max: <.01% x̄: <.01% x̃: <.01% HURT stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 HURT stats (rel) min: 0.01% max: 0.01% x̄: 0.01% x̃: 0.01% No changes on any other Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
2019-02-28 03:52:12 +00:00
for op in ('extract_u8', 'extract_i8'):
nir/algebraic: Add missing 16-bit extract_[iu]8 patterns No shader-db changes on any Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
2019-02-28 04:15:32 +00:00
optimizations.extend([((op, ('ishl', 'a@16', 8), 1), (op, a, 0))])
nir/algebraic: Fix up extract_[iu]8 after loop unrolling Skylake, Broadwell, and Haswell had similar results. (Skylake shown) total instructions in shared programs: 15256840 -> 15256837 (<.01%) instructions in affected programs: 4713 -> 4710 (-0.06%) helped: 3 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.06% max: 0.08% x̄: 0.06% x̃: 0.06% total cycles in shared programs: 372286583 -> 372286583 (0.00%) cycles in affected programs: 198516 -> 198516 (0.00%) helped: 1 HURT: 1 helped stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 helped stats (rel) min: <.01% max: <.01% x̄: <.01% x̃: <.01% HURT stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 HURT stats (rel) min: 0.01% max: 0.01% x̄: 0.01% x̃: 0.01% No changes on any other Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
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)])
nir/algebraic: Add missing 64-bit extract_[iu]8 patterns No shader-db changes on any Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
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)])
nir/algebraic: Fix up extract_[iu]8 after loop unrolling Skylake, Broadwell, and Haswell had similar results. (Skylake shown) total instructions in shared programs: 15256840 -> 15256837 (<.01%) instructions in affected programs: 4713 -> 4710 (-0.06%) helped: 3 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.06% max: 0.08% x̄: 0.06% x̃: 0.06% total cycles in shared programs: 372286583 -> 372286583 (0.00%) cycles in affected programs: 198516 -> 198516 (0.00%) helped: 1 HURT: 1 helped stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 helped stats (rel) min: <.01% max: <.01% x̄: <.01% x̃: <.01% HURT stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 HURT stats (rel) min: 0.01% max: 0.01% x̄: 0.01% x̃: 0.01% No changes on any other Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
2019-02-28 03:52:12 +00:00
optimizations.extend([
nir: Recognize open-coded extract_u16. No shader-db changes, but does recognize some extract_u16 which enables the next patch to optimize some code. Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-02-03 22:28:31 +00:00
# Word extraction
nir/algebraic: Add more extract_[iu](8|16) optimizations This adds the "(a << N) >> M" family of mask or sign-extensions. Not a huge win right now but this pattern will soon be generated by NIR format lowering code. Shader-db results on Kaby Lake: total instructions in shared programs: 15166918 -> 15166916 (<.01%) instructions in affected programs: 36 -> 34 (-5.56%) helped: 2 HURT: 0 Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-14 20:37:39 +01:00
(('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
nir/algebraic: Be more careful converting ushr to extract_u8/16 If it's not the right bit-size, it may not actually be the correct extraction. For now, we'll only worry about 32-bit versions. Fixes: 905ff8619824 "nir: Recognize open-coded extract_u16" Fixes: 76289fbfa84a "nir: Recognize open-coded extract_u8" Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-14 21:10:22 +01:00
(('ushr', 'a@32', 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
nir/algebraic: Add more extract_[iu](8|16) optimizations This adds the "(a << N) >> M" family of mask or sign-extensions. Not a huge win right now but this pattern will soon be generated by NIR format lowering code. Shader-db results on Kaby Lake: total instructions in shared programs: 15166918 -> 15166916 (<.01%) instructions in affected programs: 36 -> 34 (-5.56%) helped: 2 HURT: 0 Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
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'),
nir: Recognize open-coded extract_u16. No shader-db changes, but does recognize some extract_u16 which enables the next patch to optimize some code. Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-02-03 22:28:31 +00:00
(('iand', 0xffff, a), ('extract_u16', a, 0), '!options->lower_extract_word'),
nir: Add a flag for lowering fneg/ineg. vc4 cse/algebraic-disabled stats: total instructions in shared programs: 44911 -> 44732 (-0.40%) instructions in affected programs: 11371 -> 11192 (-1.57%) v2: Fix broken iabs(isub(0, a)) transformation. v3: Rebase to master (no TGSI->NIR present) Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
2015-01-28 18:52:53 +00:00
# Subtracts
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~fsub', a, ('fsub', 0.0, b)), ('fadd', a, b)),
nir: Generalize the optimization of subs of subs from 0. I initially wrote this based on the "(('fneg', ('fneg', a)), a)" above, but we can generalize it and make it more potentially useful. In the specific original case of a 0 for our new 'a' argument, it'll get further algebraic optimization once the 0 is an argument to the new add. No shader-db effects. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-02-20 08:00:27 +00:00
(('isub', a, ('isub', 0, b)), ('iadd', a, b)),
nir: Add opcodes for saturated vector math. This corresponds to instructions used on vc4 for its blending inside of shaders. I've seen these opcodes on other architectures before, but I think it's the first time these are needed in Mesa. v2: Rename to 'u' instead of 'i', since they're all 'u'norm (from review by jekstrand)
2015-08-19 06:38:34 +01:00
(('ussub_4x8', a, 0), a),
(('ussub_4x8', a, ~0), 0),
nir: split out lower_sub from lower_negate Originally you had to have one or the other. But actually I don't want either. (Or rather I want whatever is the minimum # of instructions.) TODO: not sure where the best place to insert a check that driver hasn't set *both* lower_negate and lower_sub? Signed-off-by: Rob Clark <robclark@freedesktop.org>
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'),
nir: Add a flag for lowering fneg/ineg. vc4 cse/algebraic-disabled stats: total instructions in shared programs: 44911 -> 44732 (-0.40%) instructions in affected programs: 11371 -> 11192 (-1.57%) v2: Fix broken iabs(isub(0, a)) transformation. v3: Rebase to master (no TGSI->NIR present) Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
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'),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
2016-03-17 18:31:48 +00:00
(('~fadd', a, ('fsub', 0.0, b)), ('fsub', a, b)),
nir: Add a flag for lowering fneg/ineg. vc4 cse/algebraic-disabled stats: total instructions in shared programs: 44911 -> 44732 (-0.40%) instructions in affected programs: 11371 -> 11192 (-1.57%) v2: Fix broken iabs(isub(0, a)) transformation. v3: Rebase to master (no TGSI->NIR present) Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
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)),
nir/algebraic: Add more lowering This commit adds lowering options for the following opcodes: - nir_op_fmod - nir_op_bitfield_insert - nir_op_uadd_carry - nir_op_usub_borrow Reviewed-by: Iago Toral Quiroga <itoral@igalia.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-01-06 23:30:38 +00:00
nir: Propagate negates up multiplication chains. total instructions in shared programs: 7112159 -> 7088092 (-0.34%) instructions in affected programs: 1374915 -> 1350848 (-1.75%) helped: 7392 HURT: 621 GAINED: 2 LOST: 2
2016-02-22 23:16:49 +00:00
# Propagate negation up multiplication chains
nir/algebraic: Push unary operations into source operands of fsat source Pushing a unary operation, like fneg, into the operation that generates its operand allows the fsat to be applied to the inner instruction instead of on a separate instruction that performs the unary operation. This changes fmul ssa_100, ssa_99, ssa_98 fmov.sat ssa_101, -ssa_100 into fmul.sat ssa_100, -ssa_99, ssa_98 Ice Lake, Skylake, and Broadwell had similar results. (Ice Lake shown) total instructions in shared programs: 17228658 -> 17228584 (<.01%) instructions in affected programs: 3163 -> 3089 (-2.34%) helped: 49 HURT: 0 helped stats (abs) min: 1 max: 2 x̄: 1.51 x̃: 2 helped stats (rel) min: 0.58% max: 9.09% x̄: 3.69% x̃: 3.51% 95% mean confidence interval for instructions value: -1.66 -1.37 95% mean confidence interval for instructions %-change: -4.37% -3.00% Instructions are helped. total cycles in shared programs: 360937144 -> 360936431 (<.01%) cycles in affected programs: 24029 -> 23316 (-2.97%) helped: 47 HURT: 2 helped stats (abs) min: 4 max: 18 x̄: 15.34 x̃: 16 helped stats (rel) min: 0.69% max: 6.18% x̄: 3.78% x̃: 4.27% HURT stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4 HURT stats (rel) min: 0.34% max: 0.67% x̄: 0.50% x̃: 0.50% 95% mean confidence interval for cycles value: -16.05 -13.05 95% mean confidence interval for cycles %-change: -4.07% -3.15% Cycles are helped. All Gen7 and earlier platforms had similar results. (Haswell shown) total instructions in shared programs: 13536059 -> 13535884 (<.01%) instructions in affected programs: 8797 -> 8622 (-1.99%) helped: 150 HURT: 0 helped stats (abs) min: 1 max: 2 x̄: 1.17 x̃: 1 helped stats (rel) min: 0.40% max: 11.11% x̄: 3.51% x̃: 1.96% 95% mean confidence interval for instructions value: -1.23 -1.11 95% mean confidence interval for instructions %-change: -3.97% -3.05% Instructions are helped. total cycles in shared programs: 357696119 -> 357694193 (<.01%) cycles in affected programs: 50216 -> 48290 (-3.84%) helped: 109 HURT: 14 helped stats (abs) min: 2 max: 92 x̄: 18.97 x̃: 16 helped stats (rel) min: 0.26% max: 19.09% x̄: 7.37% x̃: 5.37% HURT stats (abs) min: 2 max: 26 x̄: 10.14 x̃: 5 HURT stats (rel) min: 0.18% max: 4.73% x̄: 1.84% x̃: 0.92% 95% mean confidence interval for cycles value: -19.27 -12.05 95% mean confidence interval for cycles %-change: -7.34% -5.31% Cycles are helped. Reviewed-by: Matt Turner <mattst88@gmail.com>
2018-10-23 22:30:41 +01:00
(('fmul(is_used_by_non_fsat)', ('fneg', a), b), ('fneg', ('fmul', a, b))),
nir: Fix typo from commit 6702f1acde9.
2016-03-31 03:18:16 +01:00
(('imul', ('ineg', a), b), ('ineg', ('imul', a, b))),
nir: Propagate negates up multiplication chains. total instructions in shared programs: 7112159 -> 7088092 (-0.34%) instructions in affected programs: 1374915 -> 1350848 (-1.75%) helped: 7392 HURT: 621 GAINED: 2 LOST: 2
2016-02-22 23:16:49 +00:00
nir: shuffle constants to the top V2: mark float opts as inexact If one of the inputs to an mul/add is the result of another mul/add there is a chance that we can reuse the result of that mul/add in other calls if we do the multiplication in the right order. Also by attempting to move all constants to the top we increase the chance of constant folding. For example it is a fairly common pattern for shaders to do something similar to this: const float a = 0.5; in vec4 b; in float c; ... b.x = b.x * c; b.y = b.y * c; ... b.x = b.x * a + a; b.y = b.y * a + a; So by simply detecting that constant a is part of the multiplication in ffma and switching it with previous fmul that updates b we end up with: ... c = a * c; ... b.x = b.x * c + a; b.y = b.y * c + a; Shader-db results BDW: total instructions in shared programs: 13011050 -> 12967888 (-0.33%) instructions in affected programs: 4118366 -> 4075204 (-1.05%) helped: 17739 HURT: 1343 total cycles in shared programs: 246717952 -> 246410716 (-0.12%) cycles in affected programs: 166870802 -> 166563566 (-0.18%) helped: 18493 HURT: 7965 total spills in shared programs: 14937 -> 14560 (-2.52%) spills in affected programs: 9331 -> 8954 (-4.04%) helped: 284 HURT: 33 total fills in shared programs: 20211 -> 19671 (-2.67%) fills in affected programs: 12586 -> 12046 (-4.29%) helped: 286 HURT: 33 LOST: 39 GAINED: 33 Some of the hurt will go away when we shuffle things back down to the bottom in the following patch. It's also noteworthy that almost all of the spill changes are in Deus Ex both hurt and helped. Reviewed-by: Elie Tournier <elie.tournier@collabora.com> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
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)),
nir: Do basic constant reassociation. Many shaders contain expression trees of the form: const_1 * (value * const_2) Reorganizing these to (const_1 * const_2) * value will allow constant folding to combine the constants. Sometimes, these constants are 2 and 0.5, so we can remove a multiply altogether. Other times, it can create more immediate constants, which can actually hurt. Finding a good balance here is tricky. While much more could be done, this simple patch seems to have a lot of positive benefit while having a low downside. shader-db results on Broadwell: total instructions in shared programs: 8963768 -> 8961369 (-0.03%) instructions in affected programs: 438318 -> 435919 (-0.55%) helped: 1502 HURT: 245 total cycles in shared programs: 71527354 -> 71421516 (-0.15%) cycles in affected programs: 11541788 -> 11435950 (-0.92%) helped: 3445 HURT: 1224 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Eduardo Lima Mitev <elima@igalia.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-04-07 06:34:39 +01:00
# Reassociate constants in add/mul chains so they can be folded together.
nir: shuffle constants to the top V2: mark float opts as inexact If one of the inputs to an mul/add is the result of another mul/add there is a chance that we can reuse the result of that mul/add in other calls if we do the multiplication in the right order. Also by attempting to move all constants to the top we increase the chance of constant folding. For example it is a fairly common pattern for shaders to do something similar to this: const float a = 0.5; in vec4 b; in float c; ... b.x = b.x * c; b.y = b.y * c; ... b.x = b.x * a + a; b.y = b.y * a + a; So by simply detecting that constant a is part of the multiplication in ffma and switching it with previous fmul that updates b we end up with: ... c = a * c; ... b.x = b.x * c + a; b.y = b.y * c + a; Shader-db results BDW: total instructions in shared programs: 13011050 -> 12967888 (-0.33%) instructions in affected programs: 4118366 -> 4075204 (-1.05%) helped: 17739 HURT: 1343 total cycles in shared programs: 246717952 -> 246410716 (-0.12%) cycles in affected programs: 166870802 -> 166563566 (-0.18%) helped: 18493 HURT: 7965 total spills in shared programs: 14937 -> 14560 (-2.52%) spills in affected programs: 9331 -> 8954 (-4.04%) helped: 284 HURT: 33 total fills in shared programs: 20211 -> 19671 (-2.67%) fills in affected programs: 12586 -> 12046 (-4.29%) helped: 286 HURT: 33 LOST: 39 GAINED: 33 Some of the hurt will go away when we shuffle things back down to the bottom in the following patch. It's also noteworthy that almost all of the spill changes are in Deus Ex both hurt and helped. Reviewed-by: Elie Tournier <elie.tournier@collabora.com> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2017-01-12 02:10:55 +00:00
# For now, we mostly only handle cases where the constants are separated by
nir: Do basic constant reassociation. Many shaders contain expression trees of the form: const_1 * (value * const_2) Reorganizing these to (const_1 * const_2) * value will allow constant folding to combine the constants. Sometimes, these constants are 2 and 0.5, so we can remove a multiply altogether. Other times, it can create more immediate constants, which can actually hurt. Finding a good balance here is tricky. While much more could be done, this simple patch seems to have a lot of positive benefit while having a low downside. shader-db results on Broadwell: total instructions in shared programs: 8963768 -> 8961369 (-0.03%) instructions in affected programs: 438318 -> 435919 (-0.55%) helped: 1502 HURT: 245 total cycles in shared programs: 71527354 -> 71421516 (-0.15%) cycles in affected programs: 11541788 -> 11435950 (-0.92%) helped: 3445 HURT: 1224 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Eduardo Lima Mitev <elima@igalia.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-04-07 06:34:39 +01:00
# a single non-constant. We could do better eventually.
nir: Don't reassociate add/mul chains containing only constants The idea here is to reassociate a * (b * c) into (a * c) * b, when b is a non-constant value, but a and c are constants, allowing them to be combined. But nothing was enforcing that 'b' must be non-constant, which meant that running opt_algebraic in a loop would never terminate if the IR contained non-folded constant expressions like 256 * 0.5 * 2. Normally, we call constant folding in such a loop too, but IMO it's better for nir_opt_algebraic to be robust and not rely on that. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=109581 Fixes: 32e266a9a58 i965: Compile fp64 funcs only if we do not have 64-bit hardware support Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
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)),
nir: Do basic constant reassociation. Many shaders contain expression trees of the form: const_1 * (value * const_2) Reorganizing these to (const_1 * const_2) * value will allow constant folding to combine the constants. Sometimes, these constants are 2 and 0.5, so we can remove a multiply altogether. Other times, it can create more immediate constants, which can actually hurt. Finding a good balance here is tricky. While much more could be done, this simple patch seems to have a lot of positive benefit while having a low downside. shader-db results on Broadwell: total instructions in shared programs: 8963768 -> 8961369 (-0.03%) instructions in affected programs: 438318 -> 435919 (-0.55%) helped: 1502 HURT: 245 total cycles in shared programs: 71527354 -> 71421516 (-0.15%) cycles in affected programs: 11541788 -> 11435950 (-0.92%) helped: 3445 HURT: 1224 Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Eduardo Lima Mitev <elima@igalia.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-04-07 06:34:39 +01:00
nir: Optimize find_lsb/imsb/umsb error checks Two of the ARB_shader_ballot piglit tests hit the find_lsb case, removing some of the noise allowed me to better debug the test when it was failing. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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)),
nir/algebraic: Add more lowering This commit adds lowering options for the following opcodes: - nir_op_fmod - nir_op_bitfield_insert - nir_op_uadd_carry - nir_op_usub_borrow Reviewed-by: Iago Toral Quiroga <itoral@igalia.com> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-01-06 23:30:38 +00:00
# Misc. lowering
nir: Combine lower_fmod16/32 back into a single lower_fmod. We originally had a single lower_fmod option. In commit 2ab2d2e5, Sam split 32 and 64-bit lowering into separate flags, with the rationale that some drivers might want different options there. This left 16-bit unhandled, so Iago added a lower_fmod16 option in commit ca31df6f. Now that lower_fmod64 is gone (in favor of nir_lower_doubles and nir_lower_dmod), we re-combine lower_fmod16 and lower_fmod32 into a single lower_fmod flag again. I'm not aware of any hardware which need lowering for one bitsize and not the other. Reviewed-by: Marek Olšák <marek.olsak@amd.com>
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'),
nir/opt_algebraic: Fix some expressions with ambiguous bit sizes Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
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'),
nir: Handle <bits>=32 case in bitfield_insert lowering. The OpenGL specifications for bitfieldInsert() says: The result will be undefined if <offset> or <bits> is negative, or if the sum of <offset> and <bits> is greater than the number of bits used to store the operand. Therefore passing bits=32, offset=0 is legal and defined in GLSL. But the earlier SM5 bfi opcode is specified to accept a bitfield width ranging from 0-31. As such, Intel and AMD instructions read only the low 5 bits of the width operand, making them not able to implement the GLSL-specified behavior directly. This commit fixes the lowering of bitfield_insert to handle the trivial case of <bits> = 32 as bitfieldInsert: bits > 31 ? insert : bfi(bfm(bits, offset), insert, base) Fixes: ES31-CTS.shader_bitfield_operation.bitfieldInsert.uint_2 ES31-CTS.shader_bitfield_operation.bitfieldInsert.uvec4_3 Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92595 Reviewed-by: Connor Abbott <cwabbott0@gmail.com> Tested-by: Marta Lofstedt <marta.lofstedt@intel.com>
2016-01-13 19:08:37 +00:00
(('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
nir/algebraic: Use unsigned comparison when lowering bitfield insert/extract This lets us use the optimization pattern (('ult', 31, ('iand', b, 31)), False) to remove the bcsel instruction for code originating in D3D shaders. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2019-05-02 14:28:59 +01:00
('bcsel', ('ult', 31, 'bits'), 'insert',
nir: Handle <bits>=32 case in bitfield_insert lowering. The OpenGL specifications for bitfieldInsert() says: The result will be undefined if <offset> or <bits> is negative, or if the sum of <offset> and <bits> is greater than the number of bits used to store the operand. Therefore passing bits=32, offset=0 is legal and defined in GLSL. But the earlier SM5 bfi opcode is specified to accept a bitfield width ranging from 0-31. As such, Intel and AMD instructions read only the low 5 bits of the width operand, making them not able to implement the GLSL-specified behavior directly. This commit fixes the lowering of bitfield_insert to handle the trivial case of <bits> = 32 as bitfieldInsert: bits > 31 ? insert : bfi(bfm(bits, offset), insert, base) Fixes: ES31-CTS.shader_bitfield_operation.bitfieldInsert.uint_2 ES31-CTS.shader_bitfield_operation.bitfieldInsert.uvec4_3 Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92595 Reviewed-by: Connor Abbott <cwabbott0@gmail.com> Tested-by: Marta Lofstedt <marta.lofstedt@intel.com>
2016-01-13 19:08:37 +00:00
('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
'options->lower_bitfield_insert'),
nir/spirv: initial handling of OpenCL.std extension opcodes Not complete, mostly just adding things as I encounter them in CTS. But not getting far enough yet to hit most of the OpenCL.std instructions. Anyway, this is better than nothing and covers the most common builtins. v2: add hadd proof from Jason move some of the lowering into opt_algebraic and create new nir opcodes simplify nextafter lowering fix normalize lowering for inf rework upsample to use nir_pack_bits add missing files to build systems v3: split lines of iadd/sub_sat expressions Signed-off-by: Karol Herbst <kherbst@redhat.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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'),
nir: Lower bitfield_extract. The OpenGL specifications for bitfieldExtract() says: The result will be undefined if <offset> or <bits> is negative, or if the sum of <offset> and <bits> is greater than the number of bits used to store the operand. Therefore passing bits=32, offset=0 is legal and defined in GLSL. But the earlier SM5 ubfe/ibfe opcodes are specified to accept a bitfield width ranging from 0-31. As such, Intel and AMD instructions read only the low 5 bits of the width operand, making them not able to implement the GLSL-specified behavior directly. This commit adds ubfe/ibfe operations from SM5 and a lowering pass for bitfield_extract to to handle the trivial case of <bits> = 32 as bitfieldExtract: bits > 31 ? value : bfe(value, offset, bits) Fixes: ES31-CTS.shader_bitfield_operation.bitfieldExtract.uvec3_0 Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92595 Reviewed-by: Connor Abbott <cwabbott0@gmail.com> Tested-by: Marta Lofstedt <marta.lofstedt@intel.com>
2016-01-13 19:09:11 +00:00
nir: Add lowering for bitfieldInsert without using bfi. If you don't have HW to do bfi, then lowering bitfieldInsert to bfi makes things harder than keeping the "bits" argument around. This still uses bfm, but I've added the obvious lowering of bfm if you need it. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
2018-05-02 22:13:23 +01:00
# Alternative lowering that doesn't rely on bfi.
(('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
nir: define behavior of nir_op_bfm and nir_op_u/ibfe according to SM5 spec. That is: the five least significant bits provide the values of 'bits' and 'offset' which is the case for all hardware currently supported by NIR and using the bfm/bfe instructions. This patch also changes the lowering of bitfield_insert/extract using shifts to not use bfm and removes the flag 'lower_bfm'. Tested-by: Eric Anholt <eric@anholt.net> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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'))),
nir: Add lowering for bitfieldInsert without using bfi. If you don't have HW to do bfi, then lowering bitfieldInsert to bfi makes things harder than keeping the "bits" argument around. This still uses bfm, but I've added the obvious lowering of bfm if you need it. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
2018-05-02 22:13:23 +01:00
'options->lower_bitfield_insert_to_shifts'),
nir: Lower bitfield_extract. The OpenGL specifications for bitfieldExtract() says: The result will be undefined if <offset> or <bits> is negative, or if the sum of <offset> and <bits> is greater than the number of bits used to store the operand. Therefore passing bits=32, offset=0 is legal and defined in GLSL. But the earlier SM5 ubfe/ibfe opcodes are specified to accept a bitfield width ranging from 0-31. As such, Intel and AMD instructions read only the low 5 bits of the width operand, making them not able to implement the GLSL-specified behavior directly. This commit adds ubfe/ibfe operations from SM5 and a lowering pass for bitfield_extract to to handle the trivial case of <bits> = 32 as bitfieldExtract: bits > 31 ? value : bfe(value, offset, bits) Fixes: ES31-CTS.shader_bitfield_operation.bitfieldExtract.uvec3_0 Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92595 Reviewed-by: Connor Abbott <cwabbott0@gmail.com> Tested-by: Marta Lofstedt <marta.lofstedt@intel.com>
2016-01-13 19:09:11 +00:00
(('ibitfield_extract', 'value', 'offset', 'bits'),
nir/algebraic: Use unsigned comparison when lowering bitfield insert/extract This lets us use the optimization pattern (('ult', 31, ('iand', b, 31)), False) to remove the bcsel instruction for code originating in D3D shaders. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2019-05-02 14:28:59 +01:00
('bcsel', ('ult', 31, 'bits'), 'value',
nir: Lower bitfield_extract. The OpenGL specifications for bitfieldExtract() says: The result will be undefined if <offset> or <bits> is negative, or if the sum of <offset> and <bits> is greater than the number of bits used to store the operand. Therefore passing bits=32, offset=0 is legal and defined in GLSL. But the earlier SM5 ubfe/ibfe opcodes are specified to accept a bitfield width ranging from 0-31. As such, Intel and AMD instructions read only the low 5 bits of the width operand, making them not able to implement the GLSL-specified behavior directly. This commit adds ubfe/ibfe operations from SM5 and a lowering pass for bitfield_extract to to handle the trivial case of <bits> = 32 as bitfieldExtract: bits > 31 ? value : bfe(value, offset, bits) Fixes: ES31-CTS.shader_bitfield_operation.bitfieldExtract.uvec3_0 Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92595 Reviewed-by: Connor Abbott <cwabbott0@gmail.com> Tested-by: Marta Lofstedt <marta.lofstedt@intel.com>
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'),
nir: Add opcodes to extract bytes or words. The uint versions zero extend while the int versions sign extend. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-21 17:09:29 +00:00
nir/algebraic: Remove unnecessary iand of [iu]bfe and bfm sources The [iu]bfe and bfm instructions are defined to only use the five least significant bits. This optimizes a common pattern from D3D -> SPIR-V translation. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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')),
nir: Add lowering from ibitfield_extract/ubitfield_extract to shifts. V3D doesn't have opcodes for ibfe/ubfe, so we need to lower similarly to glsl/lower_instructions.cpp. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
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,
nir: define behavior of nir_op_bfm and nir_op_u/ibfe according to SM5 spec. That is: the five least significant bits provide the values of 'bits' and 'offset' which is the case for all hardware currently supported by NIR and using the bfm/bfe instructions. This patch also changes the lowering of bitfield_insert/extract using shifts to not use bfm and removes the flag 'lower_bfm'. Tested-by: Eric Anholt <eric@anholt.net> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2019-01-26 08:12:46 +00:00
('isub', ('ishl', 1, 'bits'), 1))),
nir: Add lowering from ibitfield_extract/ubitfield_extract to shifts. V3D doesn't have opcodes for ibfe/ubfe, so we need to lower similarly to glsl/lower_instructions.cpp. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
2018-05-02 21:02:21 +01:00
'options->lower_bitfield_extract_to_shifts'),
nir: Add lowering for ifind_msb to ufind_msb. ufind_msb is easily expressed in terms of clz, and we can reduce ifind_msb to that. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
2018-05-04 21:33:47 +01:00
(('ifind_msb', 'value'),
('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
'options->lower_ifind_msb'),
nir: Add lowering for find_lsb. There is a fairly simple relation to turn this into ufind_msb. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
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'),
nir: Add opcodes to extract bytes or words. The uint versions zero extend while the int versions sign extend. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
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'),
nir: Add opcodes to extract bytes or words. The uint versions zero extend while the int versions sign extend. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
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'),
nir: Add opcodes to extract bytes or words. The uint versions zero extend while the int versions sign extend. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
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'),
nir: Add opcodes to extract bytes or words. The uint versions zero extend while the int versions sign extend. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-21 17:09:29 +00:00
('iand', ('ushr', a, ('imul', b, 16)), 0xffff),
'options->lower_extract_word'),
nir: Add lowering support for packing opcodes. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-25 19:05:52 +00:00
(('pack_unorm_2x16', 'v'),
('pack_uvec2_to_uint',
nir: Rework conversion opcodes The NIR story on conversion opcodes is a mess. We've had way too many of them, naming is inconsistent, and which ones have explicit sizes was sort-of random. This commit re-organizes things and makes them all consistent: - All non-bool conversion opcodes now have the explicit size in the destination and are named <src_type>2<dst_type><size>. - Integer <-> integer conversion opcodes now only come in i2i and u2u forms (i2u and u2i have been removed) since the only difference between the different integer conversions is whether or not they sign-extend when up-converting. - Boolean conversion opcodes all have the explicit size on the bool and are named <src_type>2<dst_type>. Making things consistent also allows nir_type_conversion_op to be moved to nir_opcodes.c and auto-generated using mako. This will make adding int8, int16, and float16 versions much easier when the time comes. Reviewed-by: Eric Anholt <eric@anholt.net>
2017-03-08 03:54:37 +00:00
('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
nir: Add lowering support for packing opcodes. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-25 19:05:52 +00:00
'options->lower_pack_unorm_2x16'),
(('pack_unorm_4x8', 'v'),
('pack_uvec4_to_uint',
nir: Rework conversion opcodes The NIR story on conversion opcodes is a mess. We've had way too many of them, naming is inconsistent, and which ones have explicit sizes was sort-of random. This commit re-organizes things and makes them all consistent: - All non-bool conversion opcodes now have the explicit size in the destination and are named <src_type>2<dst_type><size>. - Integer <-> integer conversion opcodes now only come in i2i and u2u forms (i2u and u2i have been removed) since the only difference between the different integer conversions is whether or not they sign-extend when up-converting. - Boolean conversion opcodes all have the explicit size on the bool and are named <src_type>2<dst_type>. Making things consistent also allows nir_type_conversion_op to be moved to nir_opcodes.c and auto-generated using mako. This will make adding int8, int16, and float16 versions much easier when the time comes. Reviewed-by: Eric Anholt <eric@anholt.net>
2017-03-08 03:54:37 +00:00
('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
nir: Add lowering support for packing opcodes. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-25 19:05:52 +00:00
'options->lower_pack_unorm_4x8'),
(('pack_snorm_2x16', 'v'),
('pack_uvec2_to_uint',
nir: Rework conversion opcodes The NIR story on conversion opcodes is a mess. We've had way too many of them, naming is inconsistent, and which ones have explicit sizes was sort-of random. This commit re-organizes things and makes them all consistent: - All non-bool conversion opcodes now have the explicit size in the destination and are named <src_type>2<dst_type><size>. - Integer <-> integer conversion opcodes now only come in i2i and u2u forms (i2u and u2i have been removed) since the only difference between the different integer conversions is whether or not they sign-extend when up-converting. - Boolean conversion opcodes all have the explicit size on the bool and are named <src_type>2<dst_type>. Making things consistent also allows nir_type_conversion_op to be moved to nir_opcodes.c and auto-generated using mako. This will make adding int8, int16, and float16 versions much easier when the time comes. Reviewed-by: Eric Anholt <eric@anholt.net>
2017-03-08 03:54:37 +00:00
('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
nir: Add lowering support for packing opcodes. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-25 19:05:52 +00:00
'options->lower_pack_snorm_2x16'),
(('pack_snorm_4x8', 'v'),
('pack_uvec4_to_uint',
nir: Rework conversion opcodes The NIR story on conversion opcodes is a mess. We've had way too many of them, naming is inconsistent, and which ones have explicit sizes was sort-of random. This commit re-organizes things and makes them all consistent: - All non-bool conversion opcodes now have the explicit size in the destination and are named <src_type>2<dst_type><size>. - Integer <-> integer conversion opcodes now only come in i2i and u2u forms (i2u and u2i have been removed) since the only difference between the different integer conversions is whether or not they sign-extend when up-converting. - Boolean conversion opcodes all have the explicit size on the bool and are named <src_type>2<dst_type>. Making things consistent also allows nir_type_conversion_op to be moved to nir_opcodes.c and auto-generated using mako. This will make adding int8, int16, and float16 versions much easier when the time comes. Reviewed-by: Eric Anholt <eric@anholt.net>
2017-03-08 03:54:37 +00:00
('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
nir: Add lowering support for packing opcodes. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-25 19:05:52 +00:00
'options->lower_pack_snorm_4x8'),
nir: Add lowering support for unpacking opcodes. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-25 19:07:02 +00:00
(('unpack_unorm_2x16', 'v'),
nir: Rework conversion opcodes The NIR story on conversion opcodes is a mess. We've had way too many of them, naming is inconsistent, and which ones have explicit sizes was sort-of random. This commit re-organizes things and makes them all consistent: - All non-bool conversion opcodes now have the explicit size in the destination and are named <src_type>2<dst_type><size>. - Integer <-> integer conversion opcodes now only come in i2i and u2u forms (i2u and u2i have been removed) since the only difference between the different integer conversions is whether or not they sign-extend when up-converting. - Boolean conversion opcodes all have the explicit size on the bool and are named <src_type>2<dst_type>. Making things consistent also allows nir_type_conversion_op to be moved to nir_opcodes.c and auto-generated using mako. This will make adding int8, int16, and float16 versions much easier when the time comes. Reviewed-by: Eric Anholt <eric@anholt.net>
2017-03-08 03:54:37 +00:00
('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
nir/opt_algebraic: Drop bit-size suffixes from conversions Suffixes are dropped from a bunch of conversion opcodes when it makes sense to do so. Others are kept if we really do want the bit-size restriction. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2018-11-07 23:32:32 +00:00
('extract_u16', 'v', 1))),
nir: Add lowering support for unpacking opcodes. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-25 19:07:02 +00:00
65535.0),
'options->lower_unpack_unorm_2x16'),
(('unpack_unorm_4x8', 'v'),
nir: Rework conversion opcodes The NIR story on conversion opcodes is a mess. We've had way too many of them, naming is inconsistent, and which ones have explicit sizes was sort-of random. This commit re-organizes things and makes them all consistent: - All non-bool conversion opcodes now have the explicit size in the destination and are named <src_type>2<dst_type><size>. - Integer <-> integer conversion opcodes now only come in i2i and u2u forms (i2u and u2i have been removed) since the only difference between the different integer conversions is whether or not they sign-extend when up-converting. - Boolean conversion opcodes all have the explicit size on the bool and are named <src_type>2<dst_type>. Making things consistent also allows nir_type_conversion_op to be moved to nir_opcodes.c and auto-generated using mako. This will make adding int8, int16, and float16 versions much easier when the time comes. Reviewed-by: Eric Anholt <eric@anholt.net>
2017-03-08 03:54:37 +00:00
('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
nir/opt_algebraic: Drop bit-size suffixes from conversions Suffixes are dropped from a bunch of conversion opcodes when it makes sense to do so. Others are kept if we really do want the bit-size restriction. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2018-11-07 23:32:32 +00:00
('extract_u8', 'v', 1),
('extract_u8', 'v', 2),
('extract_u8', 'v', 3))),
nir: Add lowering support for unpacking opcodes. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-25 19:07:02 +00:00
255.0),
'options->lower_unpack_unorm_4x8'),
(('unpack_snorm_2x16', 'v'),
nir/opt_algebraic: Drop bit-size suffixes from conversions Suffixes are dropped from a bunch of conversion opcodes when it makes sense to do so. Others are kept if we really do want the bit-size restriction. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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))),
nir: Add lowering support for unpacking opcodes. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-25 19:07:02 +00:00
32767.0))),
'options->lower_unpack_snorm_2x16'),
(('unpack_snorm_4x8', 'v'),
nir/opt_algebraic: Drop bit-size suffixes from conversions Suffixes are dropped from a bunch of conversion opcodes when it makes sense to do so. Others are kept if we really do want the bit-size restriction. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
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))),
nir: Add lowering support for unpacking opcodes. Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-01-25 19:07:02 +00:00
127.0))),
'options->lower_unpack_snorm_4x8'),
nir: Move panfrost's isign lowering to nir_opt_algebraic. I wanted to reuse this from v3d. Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io> Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
2019-02-06 21:12:25 +00:00
(('isign', a), ('imin', ('imax', a, -1), 1), 'options->lower_isign'),
nir/algebraic: add lowering for fsign The mali utgard pp doesn't support a sign instruction. In the ARM offline shader compiler, the sign function is implemented using sub(gt(0.0, a), lt(0.0, a)). This is a generic optimization, so implement it in the nir level when lower_fsign is set, alongside the lowering for isign. Signed-off-by: Erico Nunes <nunes.erico@gmail.com> Reviewed-by: Eric Anholt <eric@anholt.net>
2019-04-16 21:49:41 +01:00
(('fsign', a), ('fsub', ('b2f', ('flt', 0.0, a)), ('b2f', ('flt', a, 0.0))), 'options->lower_fsign'),
nir/algebraic: Fix up extract_[iu]8 after loop unrolling Skylake, Broadwell, and Haswell had similar results. (Skylake shown) total instructions in shared programs: 15256840 -> 15256837 (<.01%) instructions in affected programs: 4713 -> 4710 (-0.06%) helped: 3 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.06% max: 0.08% x̄: 0.06% x̃: 0.06% total cycles in shared programs: 372286583 -> 372286583 (0.00%) cycles in affected programs: 198516 -> 198516 (0.00%) helped: 1 HURT: 1 helped stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 helped stats (rel) min: <.01% max: <.01% x̄: <.01% x̃: <.01% HURT stats (abs) min: 10 max: 10 x̄: 10.00 x̃: 10 HURT stats (rel) min: 0.01% max: 0.01% x̄: 0.01% x̃: 0.01% No changes on any other Intel platform. v2: Use a loop to generate patterns. Suggested by Jason. v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would replace an extract_i8 with and extract_u8. This broke ~180 tests. This bug was introduced in v2. Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2] Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
2019-02-28 03:52:12 +00:00
])
nir: Add an algebraic optimization pass This pass uses the previously built algebraic transformations framework and should act as an example for anyone else wanting to make an algebraic transformation pass for NIR. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-12-12 19:13:10 +00:00
nir/spirv: initial handling of OpenCL.std extension opcodes Not complete, mostly just adding things as I encounter them in CTS. But not getting far enough yet to hit most of the OpenCL.std instructions. Anyway, this is better than nothing and covers the most common builtins. v2: add hadd proof from Jason move some of the lowering into opt_algebraic and create new nir opcodes simplify nextafter lowering fix normalize lowering for inf rework upsample to use nir_pack_bits add missing files to build systems v3: split lines of iadd/sub_sat expressions Signed-off-by: Karol Herbst <kherbst@redhat.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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'),
]
python: Stabilize some script outputs In Python, dictionaries and sets are unordered, and as a result their is no guarantee that running this script twice will produce the same output. Using ordered dicts and explicitly sorting items makes the build more reproducible, and will make it possible to verify that we're not breaking anything when we move the build scripts to Python 3. Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
2018-06-27 11:37:38 +01:00
invert = OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
nir: Use De Morgan's Law on logic compounded comparisons The replacement of the comparison operators must happen during this step. If it does not, the next pass of nir_opt_algebraic will reapply De Morgan's Law in the "opposite direction" before performing dead code elimination. The resulting infinite loop will eventually get OOM killed. Haswell, Broadwell, and Skylake had similar results. (Broadwell shown) total instructions in shared programs: 14808185 -> 14808036 (<.01%) instructions in affected programs: 13758 -> 13609 (-1.08%) helped: 39 HURT: 0 helped stats (abs) min: 1 max: 10 x̄: 3.82 x̃: 3 helped stats (rel) min: 0.44% max: 1.55% x̄: 0.98% x̃: 1.01% 95% mean confidence interval for instructions value: -4.67 -2.97 95% mean confidence interval for instructions %-change: -1.09% -0.88% Instructions are helped. total cycles in shared programs: 559438333 -> 559435832 (<.01%) cycles in affected programs: 199160 -> 196659 (-1.26%) helped: 42 HURT: 3 helped stats (abs) min: 2 max: 184 x̄: 61.50 x̃: 51 helped stats (rel) min: 0.02% max: 6.94% x̄: 1.41% x̃: 1.40% HURT stats (abs) min: 2 max: 40 x̄: 27.33 x̃: 40 HURT stats (rel) min: 0.05% max: 0.74% x̄: 0.51% x̃: 0.74% 95% mean confidence interval for cycles value: -71.47 -39.69 95% mean confidence interval for cycles %-change: -1.64% -0.93% Cycles are helped. Sandy Bridge and Ivy Bridge had similar results. (Ivy Bridge shown) total instructions in shared programs: 11811776 -> 11811553 (<.01%) instructions in affected programs: 15201 -> 14978 (-1.47%) helped: 39 HURT: 0 helped stats (abs) min: 1 max: 20 x̄: 5.72 x̃: 6 helped stats (rel) min: 0.44% max: 2.53% x̄: 1.30% x̃: 1.26% 95% mean confidence interval for instructions value: -7.21 -4.23 95% mean confidence interval for instructions %-change: -1.48% -1.12% Instructions are helped. total cycles in shared programs: 257617270 -> 257614589 (<.01%) cycles in affected programs: 212107 -> 209426 (-1.26%) helped: 45 HURT: 0 helped stats (abs) min: 2 max: 180 x̄: 59.58 x̃: 54 helped stats (rel) min: 0.02% max: 6.02% x̄: 1.30% x̃: 1.32% 95% mean confidence interval for cycles value: -74.02 -45.14 95% mean confidence interval for cycles %-change: -1.59% -1.01% Cycles are helped. Iron Lake total instructions in shared programs: 7886648 -> 7886515 (<.01%) instructions in affected programs: 14106 -> 13973 (-0.94%) helped: 29 HURT: 0 helped stats (abs) min: 1 max: 10 x̄: 4.59 x̃: 4 helped stats (rel) min: 0.35% max: 1.83% x̄: 0.90% x̃: 0.81% 95% mean confidence interval for instructions value: -5.65 -3.52 95% mean confidence interval for instructions %-change: -1.03% -0.76% Instructions are helped. total cycles in shared programs: 178100812 -> 178100396 (<.01%) cycles in affected programs: 67970 -> 67554 (-0.61%) helped: 29 HURT: 0 helped stats (abs) min: 2 max: 40 x̄: 14.34 x̃: 12 helped stats (rel) min: 0.15% max: 1.69% x̄: 0.58% x̃: 0.54% 95% mean confidence interval for cycles value: -18.30 -10.39 95% mean confidence interval for cycles %-change: -0.71% -0.45% Cycles are helped. GM45 total instructions in shared programs: 4857939 -> 4857872 (<.01%) instructions in affected programs: 7426 -> 7359 (-0.90%) helped: 15 HURT: 0 helped stats (abs) min: 1 max: 10 x̄: 4.47 x̃: 4 helped stats (rel) min: 0.33% max: 1.80% x̄: 0.87% x̃: 0.77% 95% mean confidence interval for instructions value: -6.06 -2.87 95% mean confidence interval for instructions %-change: -1.06% -0.67% Instructions are helped. total cycles in shared programs: 122167930 -> 122167654 (<.01%) cycles in affected programs: 43118 -> 42842 (-0.64%) helped: 15 HURT: 0 helped stats (abs) min: 4 max: 40 x̄: 18.40 x̃: 16 helped stats (rel) min: 0.15% max: 1.69% x̄: 0.62% x̃: 0.54% 95% mean confidence interval for cycles value: -25.03 -11.77 95% mean confidence interval for cycles %-change: -0.82% -0.41% Cycles are helped. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2018-02-07 01:27:53 +00:00
python: Use the right function for the job The code was just reimplementing itertools.combinations_with_replacement in a less efficient way. This does change the order of the results slightly, but it should be ok. Signed-off-by: Mathieu Bridon <bochecha@daitauha.fr> Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
2018-08-09 09:27:19 +01:00
for left, right in itertools.combinations_with_replacement(invert.keys(), 2):
nir: Use De Morgan's Law on logic compounded comparisons The replacement of the comparison operators must happen during this step. If it does not, the next pass of nir_opt_algebraic will reapply De Morgan's Law in the "opposite direction" before performing dead code elimination. The resulting infinite loop will eventually get OOM killed. Haswell, Broadwell, and Skylake had similar results. (Broadwell shown) total instructions in shared programs: 14808185 -> 14808036 (<.01%) instructions in affected programs: 13758 -> 13609 (-1.08%) helped: 39 HURT: 0 helped stats (abs) min: 1 max: 10 x̄: 3.82 x̃: 3 helped stats (rel) min: 0.44% max: 1.55% x̄: 0.98% x̃: 1.01% 95% mean confidence interval for instructions value: -4.67 -2.97 95% mean confidence interval for instructions %-change: -1.09% -0.88% Instructions are helped. total cycles in shared programs: 559438333 -> 559435832 (<.01%) cycles in affected programs: 199160 -> 196659 (-1.26%) helped: 42 HURT: 3 helped stats (abs) min: 2 max: 184 x̄: 61.50 x̃: 51 helped stats (rel) min: 0.02% max: 6.94% x̄: 1.41% x̃: 1.40% HURT stats (abs) min: 2 max: 40 x̄: 27.33 x̃: 40 HURT stats (rel) min: 0.05% max: 0.74% x̄: 0.51% x̃: 0.74% 95% mean confidence interval for cycles value: -71.47 -39.69 95% mean confidence interval for cycles %-change: -1.64% -0.93% Cycles are helped. Sandy Bridge and Ivy Bridge had similar results. (Ivy Bridge shown) total instructions in shared programs: 11811776 -> 11811553 (<.01%) instructions in affected programs: 15201 -> 14978 (-1.47%) helped: 39 HURT: 0 helped stats (abs) min: 1 max: 20 x̄: 5.72 x̃: 6 helped stats (rel) min: 0.44% max: 2.53% x̄: 1.30% x̃: 1.26% 95% mean confidence interval for instructions value: -7.21 -4.23 95% mean confidence interval for instructions %-change: -1.48% -1.12% Instructions are helped. total cycles in shared programs: 257617270 -> 257614589 (<.01%) cycles in affected programs: 212107 -> 209426 (-1.26%) helped: 45 HURT: 0 helped stats (abs) min: 2 max: 180 x̄: 59.58 x̃: 54 helped stats (rel) min: 0.02% max: 6.02% x̄: 1.30% x̃: 1.32% 95% mean confidence interval for cycles value: -74.02 -45.14 95% mean confidence interval for cycles %-change: -1.59% -1.01% Cycles are helped. Iron Lake total instructions in shared programs: 7886648 -> 7886515 (<.01%) instructions in affected programs: 14106 -> 13973 (-0.94%) helped: 29 HURT: 0 helped stats (abs) min: 1 max: 10 x̄: 4.59 x̃: 4 helped stats (rel) min: 0.35% max: 1.83% x̄: 0.90% x̃: 0.81% 95% mean confidence interval for instructions value: -5.65 -3.52 95% mean confidence interval for instructions %-change: -1.03% -0.76% Instructions are helped. total cycles in shared programs: 178100812 -> 178100396 (<.01%) cycles in affected programs: 67970 -> 67554 (-0.61%) helped: 29 HURT: 0 helped stats (abs) min: 2 max: 40 x̄: 14.34 x̃: 12 helped stats (rel) min: 0.15% max: 1.69% x̄: 0.58% x̃: 0.54% 95% mean confidence interval for cycles value: -18.30 -10.39 95% mean confidence interval for cycles %-change: -0.71% -0.45% Cycles are helped. GM45 total instructions in shared programs: 4857939 -> 4857872 (<.01%) instructions in affected programs: 7426 -> 7359 (-0.90%) helped: 15 HURT: 0 helped stats (abs) min: 1 max: 10 x̄: 4.47 x̃: 4 helped stats (rel) min: 0.33% max: 1.80% x̄: 0.87% x̃: 0.77% 95% mean confidence interval for instructions value: -6.06 -2.87 95% mean confidence interval for instructions %-change: -1.06% -0.67% Instructions are helped. total cycles in shared programs: 122167930 -> 122167654 (<.01%) cycles in affected programs: 43118 -> 42842 (-0.64%) helped: 15 HURT: 0 helped stats (abs) min: 4 max: 40 x̄: 18.40 x̃: 16 helped stats (rel) min: 0.15% max: 1.69% x̄: 0.62% x̃: 0.54% 95% mean confidence interval for cycles value: -25.03 -11.77 95% mean confidence interval for cycles %-change: -0.82% -0.41% Cycles are helped. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
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))))
nir/algebraic: Generalize an optimization This just makes it nicely scale across bit sizes. Reviewed-by: Eric Anholt <eric@anholt.net> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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))
nir: Add a bunch of b2[if] optimizations The b2f and b2i conversions always produce zero or one which are both representable in every type and size. Since b2i and b2f support all bit sizes, we can just get rid of the conversion opcode. total instructions in shared programs: 15089335 -> 15084368 (-0.03%) instructions in affected programs: 212564 -> 207597 (-2.34%) helped: 896 HURT: 0 total cycles in shared programs: 369831123 -> 369826267 (<.01%) cycles in affected programs: 2008647 -> 2003791 (-0.24%) helped: 693 HURT: 216 Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
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'
nir/opt_algebraic: Simplify an optimization using the new search ops Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2018-11-07 22:25:31 +00:00
x2yN = '{}2{}'.format(x, y)
optimizations.append(((x2yN, (b2x, a)), (b2y, a)))
nir: Add a bunch of b2[if] optimizations The b2f and b2i conversions always produce zero or one which are both representable in every type and size. Since b2i and b2f support all bit sizes, we can just get rid of the conversion opcode. total instructions in shared programs: 15089335 -> 15084368 (-0.03%) instructions in affected programs: 212564 -> 207597 (-2.34%) helped: 896 HURT: 0 total cycles in shared programs: 369831123 -> 369826267 (<.01%) cycles in affected programs: 2008647 -> 2003791 (-0.24%) helped: 693 HURT: 216 Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
2018-10-11 04:36:52 +01:00
nir/algebraic: Optimize integer cast-of-cast These have been popping up more and more with the OpenCL work and other bits causing extra conversions to/from 64-bit. Reviewed-by: Karol Herbst <kherbst@redhat.com>
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
nir/algebraic: Support lowering for both 64 and 32-bit ldexp Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2016-04-27 19:18:04 +01:00
def fexp2i(exp, bits):
# We assume that exp is already in the right range.
compiler/nir: add lowering for 16-bit ldexp v2 (Topi): - Make bit-size handling order be 16-bit, 32-bit, 64-bit - Clamp lower exponent range at -28 instead of -30. Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2018-05-24 08:54:45 +01:00
if bits == 16:
return ('i2i16', ('ishl', ('iadd', exp, 15), 10))
elif bits == 32:
nir/algebraic: Support lowering for both 64 and 32-bit ldexp Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2016-04-27 19:18:04 +01:00
return ('ishl', ('iadd', exp, 127), 23)
elif bits == 64:
nir: Combine the int and double [un]pack opcodes NIR is a typeless IR and the two opcodes, when considered bitwise, do exactly the same thing. There's no reason to have two versions. Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2017-02-15 06:15:16 +00:00
return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20))
nir/algebraic: Support lowering for both 64 and 32-bit ldexp Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2016-04-27 19:18:04 +01:00
else:
assert False
def ldexp(f, exp, bits):
nir/algebraic: Add lowering for ldexp The algorithm used is different from both the naive suggestion from the GLSL spec and the one used in GLSL IR today. Unfortunately, the GLSL IR implementation that we have today doesn't handle denormals (for those that care) or the case where the float source is +-inf. Reviewed-by: Matt Turner <mattst88@gmail.com>
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.
compiler/nir: add lowering for 16-bit ldexp v2 (Topi): - Make bit-size handling order be 16-bit, 32-bit, 64-bit - Clamp lower exponent range at -28 instead of -30. Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2018-05-24 08:54:45 +01:00
if bits == 16:
exp = ('imin', ('imax', exp, -28), 30)
elif bits == 32:
nir/algebraic: Support lowering for both 64 and 32-bit ldexp Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
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
nir/algebraic: Add lowering for ldexp The algorithm used is different from both the naive suggestion from the GLSL spec and the one used in GLSL IR today. Unfortunately, the GLSL IR implementation that we have today doesn't handle denormals (for those that care) or the case where the float source is +-inf. Reviewed-by: Matt Turner <mattst88@gmail.com>
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.
nir/algebraic: Support lowering for both 64 and 32-bit ldexp Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2016-04-27 19:18:04 +01:00
pow2_1 = fexp2i(('ishr', exp, 1), bits)
pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits)
nir/algebraic: Add lowering for ldexp The algorithm used is different from both the naive suggestion from the GLSL spec and the one used in GLSL IR today. Unfortunately, the GLSL IR implementation that we have today doesn't handle denormals (for those that care) or the case where the float source is +-inf. Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-03-25 19:12:38 +00:00
return ('fmul', ('fmul', f, pow2_1), pow2_2)
nir/algebraic: Support lowering for both 64 and 32-bit ldexp Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2016-04-27 19:18:04 +01:00
optimizations += [
compiler/nir: add lowering for 16-bit ldexp v2 (Topi): - Make bit-size handling order be 16-bit, 32-bit, 64-bit - Clamp lower exponent range at -28 instead of -30. Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2018-05-24 08:54:45 +01:00
(('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
nir: add lower_ldexp to nir compiler options Reviewed-by: Marek Olšák <marek.olsak@amd.com>
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'),
nir/algebraic: Support lowering for both 64 and 32-bit ldexp Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
2016-04-27 19:18:04 +01:00
]
nir/algebraic: Add lowering for ldexp The algorithm used is different from both the naive suggestion from the GLSL spec and the one used in GLSL IR today. Unfortunately, the GLSL IR implementation that we have today doesn't handle denormals (for those that care) or the case where the float source is +-inf. Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-03-25 19:12:38 +00:00
nir: Recognize open-coded bitfield_reverse. Helps 11 shaders in UnrealEngine4 demos. I seriously hope they would have given us bitfieldReverse() if we exposed GL 4.0 (but we do expose ARB_gpu_shader5, so why not use that anyway?). instructions in affected programs: 4875 -> 4633 (-4.96%) cycles in affected programs: 270516 -> 244516 (-9.61%) I suspect there's a *lot* of room to improve nir_search/opt_algebraic's handling of this. We'd actually like to match, e.g., step2 by matching step1 once and then doing a pointer comparison for the second instance of step1, but unfortunately we generate an enormous tuple for instead. The .text size increases by 6.5% and the .data by 17.5%. text data bss dec hex filename 22957 45224 0 68181 10a55 nir_libnir_la-nir_opt_algebraic.o 24461 53160 0 77621 12f35 nir_libnir_la-nir_opt_algebraic.o I'd be happy to remove this if Unreal4 uses bitfieldReverse() if it is in a GL 4.0 context once we expose GL 4.0. Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
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
nir/opt_algebraic: Fix some expressions with ambiguous bit sizes Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
2016-04-26 04:59:06 +01:00
optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
nir: Recognize open-coded bitfield_reverse. Helps 11 shaders in UnrealEngine4 demos. I seriously hope they would have given us bitfieldReverse() if we exposed GL 4.0 (but we do expose ARB_gpu_shader5, so why not use that anyway?). instructions in affected programs: 4875 -> 4633 (-4.96%) cycles in affected programs: 270516 -> 244516 (-9.61%) I suspect there's a *lot* of room to improve nir_search/opt_algebraic's handling of this. We'd actually like to match, e.g., step2 by matching step1 once and then doing a pointer comparison for the second instance of step1, but unfortunately we generate an enormous tuple for instead. The .text size increases by 6.5% and the .data by 17.5%. text data bss dec hex filename 22957 45224 0 68181 10a55 nir_libnir_la-nir_opt_algebraic.o 24461 53160 0 77621 12f35 nir_libnir_la-nir_opt_algebraic.o I'd be happy to remove this if Unreal4 uses bitfieldReverse() if it is in a GL 4.0 context once we expose GL 4.0. Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
2016-02-02 00:44:18 +00:00
nir/algebraic: Add optimizations for "a == a && a CMP b" This sequence shows up The Talos Principal, at least under Vulkan, and prevents loop analysis from properly computing trip counts in a few loops. Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
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)),
]
nir: Recognize open-coded bitfield_reverse. Helps 11 shaders in UnrealEngine4 demos. I seriously hope they would have given us bitfieldReverse() if we exposed GL 4.0 (but we do expose ARB_gpu_shader5, so why not use that anyway?). instructions in affected programs: 4875 -> 4633 (-4.96%) cycles in affected programs: 270516 -> 244516 (-9.61%) I suspect there's a *lot* of room to improve nir_search/opt_algebraic's handling of this. We'd actually like to match, e.g., step2 by matching step1 once and then doing a pointer comparison for the second instance of step1, but unfortunately we generate an enormous tuple for instead. The .text size increases by 6.5% and the .data by 17.5%. text data bss dec hex filename 22957 45224 0 68181 10a55 nir_libnir_la-nir_opt_algebraic.o 24461 53160 0 77621 12f35 nir_libnir_la-nir_opt_algebraic.o I'd be happy to remove this if Unreal4 uses bitfieldReverse() if it is in a GL 4.0 context once we expose GL 4.0. Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
2016-02-02 00:44:18 +00:00
nir/opt_algebraic: Add some constant bcsel reductions total instructions in shared programs: 5998190 -> 5997603 (-0.01%) instructions in affected programs: 54276 -> 53689 (-1.08%) helped: 293 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
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'))),
]
nir: Distribute binary operations with constants into bcsel This was specifically designed to simplify 1+mix(0, a-1, condition) to mix(1, a, condition) by pushing the 1+ inside. Skylake, Broadwell, and Haswell had similar results. Skylake shown. total instructions in shared programs: 14521753 -> 14521716 (<.01%) instructions in affected programs: 10619 -> 10582 (-0.35%) helped: 51 HURT: 14 helped stats (abs) min: 1 max: 12 x̄: 1.43 x̃: 1 helped stats (rel) min: 0.20% max: 3.58% x̄: 1.01% x̃: 0.95% HURT stats (abs) min: 1 max: 11 x̄: 2.57 x̃: 1 HURT stats (rel) min: 0.22% max: 1.75% x̄: 1.20% x̃: 1.32% 95% mean confidence interval for instructions value: -1.31 0.17 95% mean confidence interval for instructions %-change: -0.80% -0.27% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 533000205 -> 533003533 (<.01%) cycles in affected programs: 110610 -> 113938 (3.01%) helped: 43 HURT: 28 helped stats (abs) min: 6 max: 440 x̄: 27.12 x̃: 16 helped stats (rel) min: 0.39% max: 4.84% x̄: 1.60% x̃: 1.67% HURT stats (abs) min: 2 max: 3066 x̄: 160.50 x̃: 14 HURT stats (rel) min: 0.08% max: 77.78% x̄: 5.16% x̃: 0.62% 95% mean confidence interval for cycles value: -43.81 137.56 95% mean confidence interval for cycles %-change: -1.47% 3.60% Inconclusive result (value mean confidence interval includes 0). Ivy Bridge total instructions in shared programs: 10018840 -> 10018713 (<.01%) instructions in affected programs: 9431 -> 9304 (-1.35%) helped: 51 HURT: 3 helped stats (abs) min: 1 max: 80 x̄: 2.76 x̃: 1 helped stats (rel) min: 0.20% max: 16.43% x̄: 1.16% x̃: 0.81% HURT stats (abs) min: 1 max: 12 x̄: 4.67 x̃: 1 HURT stats (rel) min: 0.22% max: 1.33% x̄: 0.59% x̃: 0.22% 95% mean confidence interval for instructions value: -5.36 0.66 95% mean confidence interval for instructions %-change: -1.66% -0.46% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 87571944 -> 87572785 (<.01%) cycles in affected programs: 117234 -> 118075 (0.72%) helped: 42 HURT: 23 helped stats (abs) min: 2 max: 114 x̄: 51.90 x̃: 30 helped stats (rel) min: 0.11% max: 11.01% x̄: 4.45% x̃: 2.74% HURT stats (abs) min: 1 max: 2341 x̄: 131.35 x̃: 10 HURT stats (rel) min: 0.06% max: 37.11% x̄: 2.75% x̃: 0.61% 95% mean confidence interval for cycles value: -61.05 86.93 95% mean confidence interval for cycles %-change: -3.47% -0.33% Inconclusive result (value mean confidence interval includes 0). Sandy Bridge total instructions in shared programs: 10542933 -> 10542844 (<.01%) instructions in affected programs: 11487 -> 11398 (-0.77%) helped: 52 HURT: 3 helped stats (abs) min: 1 max: 40 x̄: 1.96 x̃: 1 helped stats (rel) min: 0.08% max: 8.16% x̄: 0.90% x̃: 0.72% HURT stats (abs) min: 1 max: 11 x̄: 4.33 x̃: 1 HURT stats (rel) min: 0.22% max: 1.22% x̄: 0.55% x̃: 0.22% 95% mean confidence interval for instructions value: -3.17 -0.07 95% mean confidence interval for instructions %-change: -1.13% -0.52% Instructions are helped. total cycles in shared programs: 146098397 -> 146097094 (<.01%) cycles in affected programs: 128140 -> 126837 (-1.02%) helped: 47 HURT: 8 helped stats (abs) min: 2 max: 333 x̄: 29.21 x̃: 18 helped stats (rel) min: 0.13% max: 5.04% x̄: 1.18% x̃: 0.95% HURT stats (abs) min: 1 max: 16 x̄: 8.75 x̃: 9 HURT stats (rel) min: 0.08% max: 0.43% x̄: 0.30% x̃: 0.34% 95% mean confidence interval for cycles value: -37.49 -9.90 95% mean confidence interval for cycles %-change: -1.22% -0.71% Cycles are helped. Iron Lake total instructions in shared programs: 7886711 -> 7886509 (<.01%) instructions in affected programs: 10425 -> 10223 (-1.94%) helped: 50 HURT: 2 helped stats (abs) min: 1 max: 78 x̄: 4.08 x̃: 1 helped stats (rel) min: 0.34% max: 15.38% x̄: 1.12% x̃: 0.54% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.86% max: 0.91% x̄: 0.89% x̃: 0.89% 95% mean confidence interval for instructions value: -8.05 0.28 95% mean confidence interval for instructions %-change: -1.83% -0.26% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 178115324 -> 178114612 (<.01%) cycles in affected programs: 765726 -> 765014 (-0.09%) helped: 39 HURT: 1 helped stats (abs) min: 2 max: 276 x̄: 18.31 x̃: 8 helped stats (rel) min: <.01% max: 8.47% x̄: 0.39% x̃: 0.04% HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 HURT stats (rel) min: 0.03% max: 0.03% x̄: 0.03% x̃: 0.03% 95% mean confidence interval for cycles value: -32.07 -3.53 95% mean confidence interval for cycles %-change: -0.86% 0.10% Inconclusive result (%-change mean confidence interval includes 0). GM45 total instructions in shared programs: 4857762 -> 4857661 (<.01%) instructions in affected programs: 5523 -> 5422 (-1.83%) helped: 25 HURT: 1 helped stats (abs) min: 1 max: 78 x̄: 4.08 x̃: 1 helped stats (rel) min: 0.34% max: 13.61% x̄: 1.04% x̃: 0.52% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.86% max: 0.86% x̄: 0.86% x̃: 0.86% 95% mean confidence interval for instructions value: -9.99 2.22 95% mean confidence interval for instructions %-change: -2.01% 0.08% Inconclusive result (value mean confidence interval includes 0). total cycles in shared programs: 122179674 -> 122179194 (<.01%) cycles in affected programs: 530162 -> 529682 (-0.09%) helped: 22 HURT: 1 helped stats (abs) min: 2 max: 292 x̄: 21.91 x̃: 7 helped stats (rel) min: <.01% max: 8.65% x̄: 0.44% x̃: 0.04% HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 HURT stats (rel) min: 0.03% max: 0.03% x̄: 0.03% x̃: 0.03% 95% mean confidence interval for cycles value: -46.56 4.82 95% mean confidence interval for cycles %-change: -1.20% 0.36% Inconclusive result (value mean confidence interval includes 0). Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
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)))
]
nir/algebraic: Lower CS derivatives to zero when no group defined In compute shaders if no derivative group is defined, the derivatives will always be zero. Specified in NV_compute_shader_derivatives. To make the check more convenient, add a "info" local variable to the generated code so we can refer to it in the Python rules. (Jason) Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
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')
]
nir_algebraic: Add basic optimizations for umul_low and imadsh_mix16 For umul_low (al * bl), zero is returned if the low 16-bits word of either source is zero. for imadsh_mix16 (ah * bl << 16 + c), c is returned if either 'ah' or 'bl' is zero. A couple of nir_search_helpers are added: is_upper_half_zero() returns true if the highest word of all components of an integer NIR alu src are zero. is_lower_half_zero() returns true if the lowest word of all components of an integer nir alu src are zero. Reviewed-by: Eric Anholt <eric@anholt.net>
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'.
nir/opt_algebraic: Fix rules for imadsh_mix16 The rules added in patch 3addd7c are inverted: It should be: (al * bh) << 16 + c instead of: (ah * bl) << 16 + c Fixes a number of regressions under dEQP-GLES31.functional.draw_indirect.compute_interop.large.* on Freedreno. Reviewed-by: Rob Clark <robdclark@gmail.com>
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')),
nir_algebraic: Add basic optimizations for umul_low and imadsh_mix16 For umul_low (al * bl), zero is returned if the low 16-bits word of either source is zero. for imadsh_mix16 (ah * bl << 16 + c), c is returned if either 'ah' or 'bl' is zero. A couple of nir_search_helpers are added: is_upper_half_zero() returns true if the highest word of all components of an integer NIR alu src are zero. is_lower_half_zero() returns true if the lowest word of all components of an integer nir alu src are zero. Reviewed-by: Eric Anholt <eric@anholt.net>
2019-05-12 23:09:38 +01:00
]
nir/i965: add before ffma algebraic opts This shuffles constants down in the reverse of what the previous patch does and applies some simpilifications that may be made possible from doing so. Shader-db results BDW: total instructions in shared programs: 12980814 -> 12977822 (-0.02%) instructions in affected programs: 281889 -> 278897 (-1.06%) helped: 1231 HURT: 128 total cycles in shared programs: 246562852 -> 246567288 (0.00%) cycles in affected programs: 11271524 -> 11275960 (0.04%) helped: 1630 HURT: 1378 V2: mark float opts as inexact Reviewed-by: Elie Tournier <elie.tournier@collabora.com> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
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),
nir/algebraic: Pull common addition out of flrp arguments v2: Augment the late optimization patterns with a couple pre-ffma pass patterns. All Gen7+ platforms had similar results. (Skylake shown) total instructions in shared programs: 15342982 -> 15342485 (<.01%) instructions in affected programs: 56304 -> 55807 (-0.88%) helped: 235 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 2.11 x̃: 1 helped stats (rel) min: 0.11% max: 8.82% x̄: 1.27% x̃: 0.74% 95% mean confidence interval for instructions value: -2.31 -1.92 95% mean confidence interval for instructions %-change: -1.46% -1.09% Instructions are helped. total cycles in shared programs: 355734740 -> 355734320 (<.01%) cycles in affected programs: 1028807 -> 1028387 (-0.04%) helped: 134 HURT: 104 helped stats (abs) min: 1 max: 212 x̄: 25.69 x̃: 8 helped stats (rel) min: <.01% max: 9.36% x̄: 1.33% x̃: 0.61% HURT stats (abs) min: 1 max: 203 x̄: 29.06 x̃: 8 HURT stats (rel) min: 0.02% max: 15.76% x̄: 1.76% x̃: 0.46% 95% mean confidence interval for cycles value: -8.51 4.98 95% mean confidence interval for cycles %-change: -0.35% 0.39% Inconclusive result (value mean confidence interval includes 0). Sandy Bridge total instructions in shared programs: 10886815 -> 10886390 (<.01%) instructions in affected programs: 36883 -> 36458 (-1.15%) helped: 147 HURT: 0 helped stats (abs) min: 1 max: 7 x̄: 2.89 x̃: 3 helped stats (rel) min: 0.35% max: 8.00% x̄: 1.60% x̃: 1.23% 95% mean confidence interval for instructions value: -3.12 -2.67 95% mean confidence interval for instructions %-change: -1.83% -1.38% Instructions are helped. total cycles in shared programs: 154188360 -> 154186902 (<.01%) cycles in affected programs: 388094 -> 386636 (-0.38%) helped: 90 HURT: 58 helped stats (abs) min: 1 max: 243 x̄: 36.80 x̃: 15 helped stats (rel) min: 0.04% max: 9.23% x̄: 1.26% x̃: 0.83% HURT stats (abs) min: 1 max: 684 x̄: 31.97 x̃: 10 HURT stats (rel) min: 0.03% max: 13.50% x̄: 1.15% x̃: 0.51% 95% mean confidence interval for cycles value: -22.62 2.92 95% mean confidence interval for cycles %-change: -0.68% 0.05% Inconclusive result (value mean confidence interval includes 0). Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 8221239 -> 8220357 (-0.01%) instructions in affected programs: 54560 -> 53678 (-1.62%) helped: 186 HURT: 0 helped stats (abs) min: 1 max: 14 x̄: 4.74 x̃: 3 helped stats (rel) min: 0.34% max: 10.77% x̄: 1.97% x̃: 1.17% 95% mean confidence interval for instructions value: -5.21 -4.28 95% mean confidence interval for instructions %-change: -2.23% -1.72% Instructions are helped. total cycles in shared programs: 188654442 -> 188650364 (<.01%) cycles in affected programs: 1454384 -> 1450306 (-0.28%) helped: 204 HURT: 0 helped stats (abs) min: 2 max: 84 x̄: 19.99 x̃: 18 helped stats (rel) min: 0.02% max: 4.69% x̄: 0.56% x̃: 0.22% 95% mean confidence interval for cycles value: -22.38 -17.60 95% mean confidence interval for cycles %-change: -0.67% -0.46% Cycles are helped. Reviewed-by: Matt Turner <mattst88@gmail.com>
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))),
nir/i965: add before ffma algebraic opts This shuffles constants down in the reverse of what the previous patch does and applies some simpilifications that may be made possible from doing so. Shader-db results BDW: total instructions in shared programs: 12980814 -> 12977822 (-0.02%) instructions in affected programs: 281889 -> 278897 (-1.06%) helped: 1231 HURT: 128 total cycles in shared programs: 246562852 -> 246567288 (0.00%) cycles in affected programs: 11271524 -> 11275960 (0.04%) helped: 1630 HURT: 1378 V2: mark float opts as inexact Reviewed-by: Elie Tournier <elie.tournier@collabora.com> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2017-01-13 06:25:11 +00:00
]
nir/algebraic: Add a seperate section for "late" optimizations i965/nir: Use the late optimizations Reviewed-by: Matt Turner <mattst88@gmail.com>
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 = [
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
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.
nir: See through an fneg to apply existing optimizations Doing the same for the existing feq and fne transformations didn't help anything in shader-db. shader-db results: Broadwell and Skylake (Skylake shown) total instructions in shared programs: 14529463 -> 14526147 (-0.02%) instructions in affected programs: 402420 -> 399104 (-0.82%) helped: 2136 HURT: 131 helped stats (abs) min: 1 max: 10 x̄: 1.61 x̃: 1 helped stats (rel) min: 0.03% max: 16.22% x̄: 3.14% x̃: 1.12% HURT stats (abs) min: 1 max: 2 x̄: 1.01 x̃: 1 HURT stats (rel) min: 0.13% max: 7.69% x̄: 0.75% x̃: 0.57% 95% mean confidence interval for instructions value: -1.51 -1.41 95% mean confidence interval for instructions %-change: -3.06% -2.78% Instructions are helped. total cycles in shared programs: 533146915 -> 533120531 (<.01%) cycles in affected programs: 10356261 -> 10329877 (-0.25%) helped: 1933 HURT: 844 helped stats (abs) min: 1 max: 490 x̄: 29.44 x̃: 16 helped stats (rel) min: <.01% max: 28.57% x̄: 3.43% x̃: 1.88% HURT stats (abs) min: 1 max: 423 x̄: 36.17 x̃: 12 HURT stats (rel) min: <.01% max: 23.75% x̄: 1.90% x̃: 0.59% 95% mean confidence interval for cycles value: -11.78 -7.22 95% mean confidence interval for cycles %-change: -1.98% -1.65% Cycles are helped. Haswell total instructions in shared programs: 9037416 -> 9034106 (-0.04%) instructions in affected programs: 389831 -> 386521 (-0.85%) helped: 2184 HURT: 120 helped stats (abs) min: 1 max: 11 x̄: 1.57 x̃: 1 helped stats (rel) min: 0.03% max: 25.00% x̄: 2.73% x̃: 1.02% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.19% max: 7.69% x̄: 0.81% x̃: 0.57% 95% mean confidence interval for instructions value: -1.49 -1.39 95% mean confidence interval for instructions %-change: -2.68% -2.41% Instructions are helped. total cycles in shared programs: 84636243 -> 84631628 (<.01%) cycles in affected programs: 4745058 -> 4740443 (-0.10%) helped: 1904 HURT: 960 helped stats (abs) min: 1 max: 466 x̄: 30.21 x̃: 18 helped stats (rel) min: 0.02% max: 36.36% x̄: 3.57% x̃: 2.38% HURT stats (abs) min: 1 max: 1080 x̄: 55.11 x̃: 14 HURT stats (rel) min: 0.02% max: 51.33% x̄: 2.77% x̃: 0.81% 95% mean confidence interval for cycles value: -4.51 1.29 95% mean confidence interval for cycles %-change: -1.64% -1.25% Inconclusive result (value mean confidence interval includes 0). LOST: 1 GAINED: 0 Sandy Bridge and Ivy Bridge (Ivy Bridge shown) total instructions in shared programs: 10018873 -> 10015456 (-0.03%) instructions in affected programs: 512820 -> 509403 (-0.67%) helped: 2268 HURT: 162 helped stats (abs) min: 1 max: 11 x̄: 1.62 x̃: 1 helped stats (rel) min: 0.03% max: 25.00% x̄: 2.47% x̃: 0.88% HURT stats (abs) min: 1 max: 4 x̄: 1.59 x̃: 1 HURT stats (rel) min: 0.09% max: 7.69% x̄: 0.86% x̃: 0.50% 95% mean confidence interval for instructions value: -1.46 -1.35 95% mean confidence interval for instructions %-change: -2.38% -2.12% Instructions are helped. total cycles in shared programs: 87538223 -> 87524771 (-0.02%) cycles in affected programs: 5435520 -> 5422068 (-0.25%) helped: 1916 HURT: 946 helped stats (abs) min: 1 max: 1392 x̄: 29.44 x̃: 18 helped stats (rel) min: <.01% max: 34.51% x̄: 3.34% x̃: 1.97% HURT stats (abs) min: 1 max: 633 x̄: 45.41 x̃: 11 HURT stats (rel) min: 0.02% max: 25.95% x̄: 2.41% x̃: 0.62% 95% mean confidence interval for cycles value: -7.34 -2.06 95% mean confidence interval for cycles %-change: -1.62% -1.26% Cycles are helped. LOST: 1 GAINED: 0 Iron Lake total instructions in shared programs: 7888446 -> 7886959 (-0.02%) instructions in affected programs: 331581 -> 330094 (-0.45%) helped: 1160 HURT: 97 helped stats (abs) min: 1 max: 10 x̄: 1.37 x̃: 1 helped stats (rel) min: 0.02% max: 9.68% x̄: 0.93% x̃: 0.43% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.17% max: 4.17% x̄: 0.37% x̃: 0.25% 95% mean confidence interval for instructions value: -1.25 -1.12 95% mean confidence interval for instructions %-change: -0.91% -0.75% Instructions are helped. total cycles in shared programs: 178130766 -> 178116996 (<.01%) cycles in affected programs: 12534564 -> 12520794 (-0.11%) helped: 1856 HURT: 187 helped stats (abs) min: 2 max: 202 x̄: 7.78 x̃: 4 helped stats (rel) min: <.01% max: 6.47% x̄: 0.28% x̃: 0.11% HURT stats (abs) min: 2 max: 26 x̄: 3.55 x̃: 2 HURT stats (rel) min: 0.01% max: 2.14% x̄: 0.08% x̃: 0.02% 95% mean confidence interval for cycles value: -7.41 -6.07 95% mean confidence interval for cycles %-change: -0.28% -0.22% Cycles are helped. GM45 total instructions in shared programs: 4858912 -> 4857887 (-0.02%) instructions in affected programs: 237565 -> 236540 (-0.43%) helped: 867 HURT: 57 helped stats (abs) min: 1 max: 10 x̄: 1.25 x̃: 1 helped stats (rel) min: 0.02% max: 9.38% x̄: 0.87% x̃: 0.43% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.16% max: 3.85% x̄: 0.34% x̃: 0.22% 95% mean confidence interval for instructions value: -1.18 -1.04 95% mean confidence interval for instructions %-change: -0.88% -0.71% Instructions are helped. total cycles in shared programs: 122189118 -> 122180816 (<.01%) cycles in affected programs: 8776418 -> 8768116 (-0.09%) helped: 1213 HURT: 166 helped stats (abs) min: 2 max: 202 x̄: 7.30 x̃: 4 helped stats (rel) min: <.01% max: 6.43% x̄: 0.25% x̃: 0.11% HURT stats (abs) min: 2 max: 26 x̄: 3.35 x̃: 2 HURT stats (rel) min: 0.01% max: 2.14% x̄: 0.06% x̃: 0.02% 95% mean confidence interval for cycles value: -6.78 -5.26 95% mean confidence interval for cycles %-change: -0.24% -0.18% Cycles are helped. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
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)),
nir/algebraic: Flag inexact optimizations Many of our optimizations, while great for cutting shaders down to size, aren't really precision-safe. This commit tries to flag all of the inexact floating-point optimizations so they don't get run on values that are flagged "exact". It's a bit conservative and maybe flags some safe optimizations as unsafe but that's better than missing one. Reviewed-by: Francisco Jerez <currojerez@riseup.net>
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))),
intel/compiler: Repeat nir_opt_algebraic_late A tiny bit of help seems to come from nir_copy_prop. Future patches will benefit from this change. Doing more copy propagation on the vec4 backend led to a disaster in hurt cycles. v2: Fix typo in comment. Noticed by Matt. All Gen8+ platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17224634 -> 17224623 (<.01%) instructions in affected programs: 4586 -> 4575 (-0.24%) helped: 11 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.19% max: 0.53% x̄: 0.27% x̃: 0.23% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -0.36% -0.19% Instructions are helped. total cycles in shared programs: 360828542 -> 360828714 (<.01%) cycles in affected programs: 151159 -> 151331 (0.11%) helped: 49 HURT: 28 helped stats (abs) min: 1 max: 254 x̄: 26.41 x̃: 6 helped stats (rel) min: 0.06% max: 12.02% x̄: 1.34% x̃: 0.42% HURT stats (abs) min: 1 max: 196 x̄: 52.36 x̃: 15 HURT stats (rel) min: 0.05% max: 10.74% x̄: 2.55% x̃: 0.88% 95% mean confidence interval for cycles value: -13.48 17.95 95% mean confidence interval for cycles %-change: -0.69% 0.84% Inconclusive result (value mean confidence interval includes 0). Haswell, Ivy Bridge, and Sandy Bridge had similar results. (Haswell shown) total instructions in shared programs: 13529544 -> 13529542 (<.01%) instructions in affected programs: 358 -> 356 (-0.56%) helped: 2 HURT: 0 total cycles in shared programs: 357290311 -> 357289678 (<.01%) cycles in affected programs: 178324 -> 177691 (-0.35%) helped: 48 HURT: 40 helped stats (abs) min: 1 max: 201 x̄: 31.52 x̃: 13 helped stats (rel) min: 0.06% max: 10.92% x̄: 1.71% x̃: 0.66% HURT stats (abs) min: 1 max: 224 x̄: 22.00 x̃: 6 HURT stats (rel) min: 0.05% max: 15.84% x̄: 1.29% x̃: 0.31% 95% mean confidence interval for cycles value: -18.28 3.89 95% mean confidence interval for cycles %-change: -1.01% 0.32% Inconclusive result (value mean confidence interval includes 0). Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 8159110 -> 8158980 (<.01%) instructions in affected programs: 22719 -> 22589 (-0.57%) helped: 65 HURT: 0 helped stats (abs) min: 1 max: 3 x̄: 2.00 x̃: 2 helped stats (rel) min: 0.07% max: 1.05% x̄: 0.73% x̃: 0.74% 95% mean confidence interval for instructions value: -2.06 -1.94 95% mean confidence interval for instructions %-change: -0.78% -0.68% Instructions are helped. total cycles in shared programs: 188609448 -> 188609214 (<.01%) cycles in affected programs: 1875852 -> 1875618 (-0.01%) helped: 109 HURT: 104 helped stats (abs) min: 2 max: 46 x̄: 5.30 x̃: 4 helped stats (rel) min: 0.02% max: 0.90% x̄: 0.09% x̃: 0.07% HURT stats (abs) min: 2 max: 20 x̄: 3.31 x̃: 2 HURT stats (rel) min: 0.01% max: 0.26% x̄: 0.04% x̃: 0.02% 95% mean confidence interval for cycles value: -1.95 -0.25 95% mean confidence interval for cycles %-change: -0.04% -0.01% Cycles are helped. Reviewed-by: Matt Turner <mattst88@gmail.com>
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),
nir/algebraic: Eliminate useless fsat() on operand of comparison w/value in (0, 1) v2: Fix copy-and-paste bug in a cmp b vs b cmp a cases. All Gen7+ platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17224337 -> 17224269 (<.01%) instructions in affected programs: 13578 -> 13510 (-0.50%) helped: 68 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 0.31% max: 3.12% x̄: 0.84% x̃: 0.42% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -1.05% -0.63% Instructions are helped. total cycles in shared programs: 360826090 -> 360825137 (<.01%) cycles in affected programs: 94867 -> 93914 (-1.00%) helped: 58 HURT: 1 helped stats (abs) min: 2 max: 28 x̄: 17.74 x̃: 18 helped stats (rel) min: 0.08% max: 3.17% x̄: 1.39% x̃: 1.22% HURT stats (abs) min: 76 max: 76 x̄: 76.00 x̃: 76 HURT stats (rel) min: 2.86% max: 2.86% x̄: 2.86% x̃: 2.86% 95% mean confidence interval for cycles value: -19.53 -12.78 95% mean confidence interval for cycles %-change: -1.56% -1.08% Cycles are helped. No changes on any other Intel platform. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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)),
nir: Separate a weird compare with zero to two compares with zero min(a+b, c+d) >= 0 becomes (a+b >= 0 && c+d >= 0). No shader-db changes, but it does prevent 6 to 12 instruction regressions in the next patch on all measured Intel platforms. Signed-off-by: Ian Romanick <ian.d.romanick@intel.com> Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com> Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
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)))),
nir/algebraic: Strip double negatives from comparison sources All Intel platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17224623 -> 17224337 (<.01%) instructions in affected programs: 32648 -> 32362 (-0.88%) helped: 148 HURT: 0 helped stats (abs) min: 1 max: 2 x̄: 1.93 x̃: 2 helped stats (rel) min: 0.16% max: 2.74% x̄: 1.07% x̃: 1.08% 95% mean confidence interval for instructions value: -1.97 -1.89 95% mean confidence interval for instructions %-change: -1.15% -1.00% Instructions are helped. total cycles in shared programs: 360828714 -> 360826090 (<.01%) cycles in affected programs: 347416 -> 344792 (-0.76%) helped: 148 HURT: 26 helped stats (abs) min: 1 max: 426 x̄: 26.33 x̃: 18 helped stats (rel) min: 0.03% max: 15.10% x̄: 1.78% x̃: 1.41% HURT stats (abs) min: 2 max: 337 x̄: 48.96 x̃: 6 HURT stats (rel) min: 0.04% max: 18.82% x̄: 2.15% x̃: 0.27% 95% mean confidence interval for cycles value: -23.78 -6.38 95% mean confidence interval for cycles %-change: -1.59% -0.79% Cycles are helped. Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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),
nir/algebraic: Commute 1-fsat(a) to fsat(1-a) for all non-fmul instructions The goal is to avoid having an extra MOV instruction to perform the saturate. Doing the subtraction first allows the saturate to be applied to the ADD instruction making the MOV unnecessary. Values generated in different block and values from non-ALU instructions (e.g., texture instructions) almost always need the extra MOV. Multiply instructions are restricted because doing this rearrangement can interfere with the generation of flrp and ffma instructions. v2: Now that the final method has been selected, squash three commits into one. All Intel platforms has similar results. (Ice Lake shown) total instructions in shared programs: 17223214 -> 17219386 (-0.02%) instructions in affected programs: 1524376 -> 1520548 (-0.25%) helped: 2686 HURT: 26 helped stats (abs) min: 1 max: 32 x̄: 1.44 x̃: 1 helped stats (rel) min: 0.03% max: 16.67% x̄: 0.54% x̃: 0.37% HURT stats (abs) min: 1 max: 2 x̄: 1.69 x̃: 2 HURT stats (rel) min: 0.33% max: 1.67% x̄: 0.54% x̃: 0.35% 95% mean confidence interval for instructions value: -1.46 -1.36 95% mean confidence interval for instructions %-change: -0.56% -0.50% Instructions are helped. total cycles in shared programs: 360811571 -> 360791896 (<.01%) cycles in affected programs: 103650214 -> 103630539 (-0.02%) helped: 1557 HURT: 675 helped stats (abs) min: 1 max: 1773 x̄: 41.44 x̃: 16 helped stats (rel) min: <.01% max: 26.77% x̄: 1.37% x̃: 0.64% HURT stats (abs) min: 1 max: 1513 x̄: 66.44 x̃: 14 HURT stats (rel) min: <.01% max: 46.16% x̄: 2.00% x̃: 0.49% 95% mean confidence interval for cycles value: -14.82 -2.81 95% mean confidence interval for cycles %-change: -0.50% -0.20% Cycles are helped. LOST: 2 GAINED: 0 Reviewed-by: Matt Turner <mattst88@gmail.com> [v1] Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
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)))),
nir: Add a fdot instruction that replicates the result to a vec4 Fortunately, nir_constant_expr already auto-splats if "dst" never shows up in the constant expression field so we don't need to do anything there. Reviewed-by: Connor Abbott <cwabbott0@gmail.com> Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
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'),
nir: Add fdph and fdph_replicated opcodes Reviewed-by: Matt Turner <mattst88@gmail.com> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2015-09-23 00:54:27 +01:00
(('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'),
nir: add late opt to turn inot/b2f combos back to bcsel We turn these from bcsel into inot/b2f combos in order for other optimisation passes to get further. Once we have finished turn the ones that remain and are used in more than a single expression back into a bcsel. On BDW: total instructions in shared programs: 13060965 -> 13060297 (-0.01%) instructions in affected programs: 835701 -> 835033 (-0.08%) helped: 670 HURT: 2 total cycles in shared programs: 256599536 -> 256598006 (-0.00%) cycles in affected programs: 114655488 -> 114653958 (-0.00%) helped: 419 HURT: 240 LOST: 0 GAINED: 1 The 2 HURT is because inserting bcsel creates the only use of const 1.0 in two shaders from tri-of-friendship-and-madness. Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2017-01-03 11:54:48 +00:00
nir/algebraic: Pull common addition out of flrp arguments v2: Augment the late optimization patterns with a couple pre-ffma pass patterns. All Gen7+ platforms had similar results. (Skylake shown) total instructions in shared programs: 15342982 -> 15342485 (<.01%) instructions in affected programs: 56304 -> 55807 (-0.88%) helped: 235 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 2.11 x̃: 1 helped stats (rel) min: 0.11% max: 8.82% x̄: 1.27% x̃: 0.74% 95% mean confidence interval for instructions value: -2.31 -1.92 95% mean confidence interval for instructions %-change: -1.46% -1.09% Instructions are helped. total cycles in shared programs: 355734740 -> 355734320 (<.01%) cycles in affected programs: 1028807 -> 1028387 (-0.04%) helped: 134 HURT: 104 helped stats (abs) min: 1 max: 212 x̄: 25.69 x̃: 8 helped stats (rel) min: <.01% max: 9.36% x̄: 1.33% x̃: 0.61% HURT stats (abs) min: 1 max: 203 x̄: 29.06 x̃: 8 HURT stats (rel) min: 0.02% max: 15.76% x̄: 1.76% x̃: 0.46% 95% mean confidence interval for cycles value: -8.51 4.98 95% mean confidence interval for cycles %-change: -0.35% 0.39% Inconclusive result (value mean confidence interval includes 0). Sandy Bridge total instructions in shared programs: 10886815 -> 10886390 (<.01%) instructions in affected programs: 36883 -> 36458 (-1.15%) helped: 147 HURT: 0 helped stats (abs) min: 1 max: 7 x̄: 2.89 x̃: 3 helped stats (rel) min: 0.35% max: 8.00% x̄: 1.60% x̃: 1.23% 95% mean confidence interval for instructions value: -3.12 -2.67 95% mean confidence interval for instructions %-change: -1.83% -1.38% Instructions are helped. total cycles in shared programs: 154188360 -> 154186902 (<.01%) cycles in affected programs: 388094 -> 386636 (-0.38%) helped: 90 HURT: 58 helped stats (abs) min: 1 max: 243 x̄: 36.80 x̃: 15 helped stats (rel) min: 0.04% max: 9.23% x̄: 1.26% x̃: 0.83% HURT stats (abs) min: 1 max: 684 x̄: 31.97 x̃: 10 HURT stats (rel) min: 0.03% max: 13.50% x̄: 1.15% x̃: 0.51% 95% mean confidence interval for cycles value: -22.62 2.92 95% mean confidence interval for cycles %-change: -0.68% 0.05% Inconclusive result (value mean confidence interval includes 0). Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 8221239 -> 8220357 (-0.01%) instructions in affected programs: 54560 -> 53678 (-1.62%) helped: 186 HURT: 0 helped stats (abs) min: 1 max: 14 x̄: 4.74 x̃: 3 helped stats (rel) min: 0.34% max: 10.77% x̄: 1.97% x̃: 1.17% 95% mean confidence interval for instructions value: -5.21 -4.28 95% mean confidence interval for instructions %-change: -2.23% -1.72% Instructions are helped. total cycles in shared programs: 188654442 -> 188650364 (<.01%) cycles in affected programs: 1454384 -> 1450306 (-0.28%) helped: 204 HURT: 0 helped stats (abs) min: 2 max: 84 x̄: 19.99 x̃: 18 helped stats (rel) min: 0.02% max: 4.69% x̄: 0.56% x̃: 0.22% 95% mean confidence interval for cycles value: -22.38 -17.60 95% mean confidence interval for cycles %-change: -0.67% -0.46% Cycles are helped. Reviewed-by: Matt Turner <mattst88@gmail.com>
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'),
nir: optimise min/max fadd combos shader-db results BDW: total instructions in shared programs: 13060410 -> 13060313 (-0.00%) instructions in affected programs: 24533 -> 24436 (-0.40%) helped: 88 HURT: 0 total cycles in shared programs: 256585692 -> 256586698 (0.00%) cycles in affected programs: 647290 -> 648296 (0.16%) helped: 35 HURT: 30 Reviewed-by: Matt Turner <mattst88@gmail.com>
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))),
nir/algebraic: Replace a bcsel of a b2f sources with a b2f(!(a || b)) I have not investigated the result of doing this during code generation. That should be possible, but it would be a bit more effort. All Gen6+ platforms had nearly identical results. (Skylake shown) total cycles in shared programs: 370961508 -> 370961367 (<.01%) cycles in affected programs: 5174 -> 5033 (-2.73%) helped: 2 HURT: 0 Iron Lake and GM45 had similar results. (Iron Lake shown) total instructions in shared programs: 8206587 -> 8206589 (<.01%) instructions in affected programs: 1325 -> 1327 (0.15%) helped: 0 HURT: 2 total cycles in shared programs: 187657422 -> 187657428 (<.01%) cycles in affected programs: 11566 -> 11572 (0.05%) helped: 0 HURT: 2 This change has almost no effect right now. However, removing this patch (but leaving the patch "intel/fs: Generate if instructions with inverted conditions") after adding a patch that removes !(a < b) -> (a >= b) optimizations (like https://patchwork.freedesktop.org/patch/264787/) has the following results on Skylake: Skylake total instructions in shared programs: 15071804 -> 15071806 (<.01%) instructions in affected programs: 640 -> 642 (0.31%) helped: 0 HURT: 2 total cycles in shared programs: 369914348 -> 369916569 (<.01%) cycles in affected programs: 27900 -> 30121 (7.96%) helped: 4 HURT: 15 helped stats (abs) min: 2 max: 112 x̄: 30.00 x̃: 3 helped stats (rel) min: 0.28% max: 12.28% x̄: 3.34% x̃: 0.40% HURT stats (abs) min: 2 max: 758 x̄: 156.07 x̃: 81 HURT stats (rel) min: 0.20% max: 74.30% x̄: 16.29% x̃: 16.91% 95% mean confidence interval for cycles value: 12.68 221.11 95% mean confidence interval for cycles %-change: 3.09% 21.23% Cycles are HURT. Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-12-03 22:41:07 +00:00
nir/algebraic: Drop some @bool specifiers Now that we have one-bit booleans, we don't need to rely on looking at parent instructions in order to figure out if a value is a Boolean most of the time. We can drop these specifiers and now the optimizations will apply more generally. Shader-DB results on Kaby Lake: total instructions in shared programs: 15321168 -> 15321227 (<.01%) instructions in affected programs: 8836 -> 8895 (0.67%) helped: 1 HURT: 31 total cycles in shared programs: 357481781 -> 357481321 (<.01%) cycles in affected programs: 146524 -> 146064 (-0.31%) helped: 22 HURT: 10 total spills in shared programs: 23675 -> 23673 (<.01%) spills in affected programs: 11 -> 9 (-18.18%) helped: 1 HURT: 0 total fills in shared programs: 32040 -> 32036 (-0.01%) fills in affected programs: 27 -> 23 (-14.81%) helped: 1 HURT: 0 No change in VkPipeline-DB Looking at the instructions hurt, a bunch of them seem to be a case where doing exactly the right thing in NIR ends up doing the wrong-ish thing in the back-end because flags are dumb. In particular, there's a case where we have a MUL followed by a CMP followed by a SEL and when we turn that SEL into an OR, it uses the GRF result of the CMP rather than the flag result so the CMP can't be merged with the MUL. Those shaders appear to schedule better according to the cycle estimates so I guess it's a win? Also it helps spilling in one Car Chase compute shader. Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
2019-03-30 03:51:20 +00:00
(('bcsel', a, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a, b)))),
nir/algebraic: Add a seperate section for "late" optimizations i965/nir: Use the late optimizations Reviewed-by: Matt Turner <mattst88@gmail.com>
2015-03-24 00:11:49 +00:00
]
python: Use the print function In Python 2, `print` was a statement, but it became a function in Python 3. Using print functions everywhere makes the script compatible with Python versions >= 2.6, including Python 3. Signed-off-by: Mathieu Bridon <bochecha@daitauha.fr> Acked-by: Eric Engestrom <eric.engestrom@intel.com> Acked-by: Dylan Baker <dylan@pnwbakers.com>
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())