From 62795475e8f45f92bb8f467d9e2318fdfdba6297 Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Mon, 15 Jul 2019 15:55:00 -0700 Subject: [PATCH] nir/algebraic: Distribute source modifiers into instructions MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit There are three main classes of cases that are helped by this change: 1. When the negation is applied to a value being type converted (e.g., float(-x)). This could possibly also be handled with more clever code generation. 2. When the negation is applied to a phi node source (e.g., x = -(...); at the end of a basic block). This was the original case that caught my attention while looking at shader-db dumps. 3. When the negation is applied to the source of an instruction that cannot have source modifiers. This includes texture instructions and math box instructions on pre-Gen7 platforms (see more details below). In many these cases the negation can be propagated into the instructions that generate the value (e.g., -(a*b) = (-a)*b). In addition to the operations implemtned in this patch, I also tried: - frcp - Helped 6 or fewer shaders on Gen7+, and hurt just as many on pre-Gen7. On Gen6 and earlier, frcp is a math box instruction, and math box instructions cannot have source modifiers. I suspect this is why so many more shaders are helped on Gen6 than on Gen5 or Gen7. Gen6 supports OpenGL 3.3, so a lot more shaders compile on it. A lot of these shaders may have things like cos(-x) or rcp(-x) that could result in an explicit negation instruction. - bcsel - Hurt a few shaders with none helped. bcsel operates on integer sources, so the fabs or fneg cannot be a source modifier in the bcsel itself. - Integer instructions - No changes on any Intel platform. Some notes about the shader-db results below. - On Tiger Lake, a single Deus Ex fragment shader is hurt for both spills and fills. - On Haswell, a different Deus Ex fragment shader is hurt for both spills and fills. - On GM45, the "LOST: 1" and "GAINED: 1" is a single Left4Dead 2 (very high graphics settings, lol) fragment shader that upgrades from SIMD8 to SIMD16. v2: Add support for fsign. Add some patterns that remove redundant negations and redundant absolute value rather than trying to push them down the tree. Tiger Lake total instructions in shared programs: 17611333 -> 17586465 (-0.14%) instructions in affected programs: 3033734 -> 3008866 (-0.82%) helped: 10310 HURT: 632 helped stats (abs) min: 1 max: 35 x̄: 2.61 x̃: 1 helped stats (rel) min: 0.04% max: 16.67% x̄: 1.43% x̃: 1.01% HURT stats (abs) min: 1 max: 47 x̄: 3.21 x̃: 2 HURT stats (rel) min: 0.04% max: 5.08% x̄: 0.88% x̃: 0.63% 95% mean confidence interval for instructions value: -2.33 -2.21 95% mean confidence interval for instructions %-change: -1.32% -1.27% Instructions are helped. total cycles in shared programs: 338365223 -> 338262252 (-0.03%) cycles in affected programs: 125291811 -> 125188840 (-0.08%) helped: 5224 HURT: 2031 helped stats (abs) min: 1 max: 5670 x̄: 46.73 x̃: 12 helped stats (rel) min: <.01% max: 34.78% x̄: 1.91% x̃: 0.97% HURT stats (abs) min: 1 max: 2882 x̄: 69.50 x̃: 14 HURT stats (rel) min: <.01% max: 44.93% x̄: 2.35% x̃: 0.74% 95% mean confidence interval for cycles value: -18.71 -9.68 95% mean confidence interval for cycles %-change: -0.80% -0.63% Cycles are helped. total spills in shared programs: 8942 -> 8946 (0.04%) spills in affected programs: 8 -> 12 (50.00%) helped: 0 HURT: 1 total fills in shared programs: 9399 -> 9401 (0.02%) fills in affected programs: 21 -> 23 (9.52%) helped: 0 HURT: 1 Ice Lake total instructions in shared programs: 16124348 -> 16102258 (-0.14%) instructions in affected programs: 2830928 -> 2808838 (-0.78%) helped: 11294 HURT: 2 helped stats (abs) min: 1 max: 12 x̄: 1.96 x̃: 1 helped stats (rel) min: 0.07% max: 17.65% x̄: 1.32% x̃: 0.93% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 3.45% max: 4.00% x̄: 3.72% x̃: 3.72% 95% mean confidence interval for instructions value: -1.99 -1.93 95% mean confidence interval for instructions %-change: -1.34% -1.29% Instructions are helped. total cycles in shared programs: 335393932 -> 335325794 (-0.02%) cycles in affected programs: 123834609 -> 123766471 (-0.06%) helped: 5034 HURT: 2128 helped stats (abs) min: 1 max: 3256 x̄: 43.39 x̃: 11 helped stats (rel) min: <.01% max: 35.79% x̄: 1.98% x̃: 1.00% HURT stats (abs) min: 1 max: 2634 x̄: 70.63 x̃: 16 HURT stats (rel) min: <.01% max: 49.49% x̄: 2.73% x̃: 0.62% 95% mean confidence interval for cycles value: -13.66 -5.37 95% mean confidence interval for cycles %-change: -0.69% -0.48% Cycles are helped. LOST: 0 GAINED: 2 Skylake total instructions in shared programs: 14949240 -> 14927930 (-0.14%) instructions in affected programs: 2594756 -> 2573446 (-0.82%) helped: 11000 HURT: 2 helped stats (abs) min: 1 max: 12 x̄: 1.94 x̃: 1 helped stats (rel) min: 0.07% max: 18.75% x̄: 1.39% x̃: 0.94% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 4.76% max: 4.76% x̄: 4.76% x̃: 4.76% 95% mean confidence interval for instructions value: -1.97 -1.91 95% mean confidence interval for instructions %-change: -1.42% -1.37% Instructions are helped. total cycles in shared programs: 324829346 -> 324821596 (<.01%) cycles in affected programs: 121566087 -> 121558337 (<.01%) helped: 4611 HURT: 2147 helped stats (abs) min: 1 max: 3715 x̄: 33.29 x̃: 10 helped stats (rel) min: <.01% max: 36.08% x̄: 1.94% x̃: 1.00% HURT stats (abs) min: 1 max: 2551 x̄: 67.88 x̃: 16 HURT stats (rel) min: <.01% max: 53.79% x̄: 3.69% x̃: 0.89% 95% mean confidence interval for cycles value: -4.25 1.96 95% mean confidence interval for cycles %-change: -0.28% -0.02% Inconclusive result (value mean confidence interval includes 0). Broadwell total instructions in shared programs: 14971203 -> 14949957 (-0.14%) instructions in affected programs: 2635699 -> 2614453 (-0.81%) helped: 10982 HURT: 2 helped stats (abs) min: 1 max: 12 x̄: 1.93 x̃: 1 helped stats (rel) min: 0.07% max: 18.75% x̄: 1.39% x̃: 0.94% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 4.76% max: 4.76% x̄: 4.76% x̃: 4.76% 95% mean confidence interval for instructions value: -1.97 -1.90 95% mean confidence interval for instructions %-change: -1.42% -1.37% Instructions are helped. total cycles in shared programs: 336215033 -> 336086458 (-0.04%) cycles in affected programs: 127383198 -> 127254623 (-0.10%) helped: 4884 HURT: 1963 helped stats (abs) min: 1 max: 25696 x̄: 51.78 x̃: 12 helped stats (rel) min: <.01% max: 58.28% x̄: 2.00% x̃: 1.05% HURT stats (abs) min: 1 max: 3401 x̄: 63.33 x̃: 16 HURT stats (rel) min: <.01% max: 39.95% x̄: 2.20% x̃: 0.70% 95% mean confidence interval for cycles value: -29.99 -7.57 95% mean confidence interval for cycles %-change: -0.89% -0.71% Cycles are helped. total fills in shared programs: 24905 -> 24901 (-0.02%) fills in affected programs: 117 -> 113 (-3.42%) helped: 4 HURT: 0 LOST: 0 GAINED: 16 Haswell total instructions in shared programs: 13148927 -> 13131528 (-0.13%) instructions in affected programs: 2220941 -> 2203542 (-0.78%) helped: 8017 HURT: 4 helped stats (abs) min: 1 max: 12 x̄: 2.17 x̃: 1 helped stats (rel) min: 0.07% max: 15.25% x̄: 1.40% x̃: 0.93% HURT stats (abs) min: 1 max: 7 x̄: 2.50 x̃: 1 HURT stats (rel) min: 0.33% max: 4.76% x̄: 2.73% x̃: 2.91% 95% mean confidence interval for instructions value: -2.21 -2.13 95% mean confidence interval for instructions %-change: -1.43% -1.37% Instructions are helped. total cycles in shared programs: 321221791 -> 321079870 (-0.04%) cycles in affected programs: 126886055 -> 126744134 (-0.11%) helped: 4674 HURT: 1729 helped stats (abs) min: 1 max: 23654 x̄: 56.47 x̃: 16 helped stats (rel) min: <.01% max: 53.22% x̄: 2.13% x̃: 1.05% HURT stats (abs) min: 1 max: 3694 x̄: 70.58 x̃: 18 HURT stats (rel) min: <.01% max: 63.06% x̄: 2.48% x̃: 0.90% 95% mean confidence interval for cycles value: -33.31 -11.02 95% mean confidence interval for cycles %-change: -0.99% -0.78% Cycles are helped. total spills in shared programs: 19872 -> 19874 (0.01%) spills in affected programs: 21 -> 23 (9.52%) helped: 0 HURT: 1 total fills in shared programs: 20941 -> 20941 (0.00%) fills in affected programs: 62 -> 62 (0.00%) helped: 1 HURT: 1 LOST: 0 GAINED: 8 Ivy Bridge total instructions in shared programs: 11875553 -> 11853839 (-0.18%) instructions in affected programs: 1553112 -> 1531398 (-1.40%) helped: 7304 HURT: 3 helped stats (abs) min: 1 max: 16 x̄: 2.97 x̃: 2 helped stats (rel) min: 0.07% max: 15.25% x̄: 1.62% x̃: 1.15% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 1.05% max: 3.33% x̄: 2.44% x̃: 2.94% 95% mean confidence interval for instructions value: -3.04 -2.90 95% mean confidence interval for instructions %-change: -1.65% -1.59% Instructions are helped. total cycles in shared programs: 178246425 -> 178184484 (-0.03%) cycles in affected programs: 13702146 -> 13640205 (-0.45%) helped: 4409 HURT: 1566 helped stats (abs) min: 1 max: 531 x̄: 24.52 x̃: 13 helped stats (rel) min: <.01% max: 38.67% x̄: 2.14% x̃: 1.02% HURT stats (abs) min: 1 max: 356 x̄: 29.48 x̃: 10 HURT stats (rel) min: <.01% max: 64.73% x̄: 1.87% x̃: 0.70% 95% mean confidence interval for cycles value: -11.60 -9.14 95% mean confidence interval for cycles %-change: -1.19% -0.99% Cycles are helped. LOST: 0 GAINED: 10 Sandy Bridge total instructions in shared programs: 10695740 -> 10667483 (-0.26%) instructions in affected programs: 2337607 -> 2309350 (-1.21%) helped: 10720 HURT: 1 helped stats (abs) min: 1 max: 49 x̄: 2.64 x̃: 2 helped stats (rel) min: 0.07% max: 20.00% x̄: 1.54% x̃: 1.13% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 1.04% max: 1.04% x̄: 1.04% x̃: 1.04% 95% mean confidence interval for instructions value: -2.69 -2.58 95% mean confidence interval for instructions %-change: -1.57% -1.51% Instructions are helped. total cycles in shared programs: 153478839 -> 153416223 (-0.04%) cycles in affected programs: 22050900 -> 21988284 (-0.28%) helped: 5342 HURT: 2200 helped stats (abs) min: 1 max: 1020 x̄: 20.34 x̃: 16 helped stats (rel) min: <.01% max: 24.05% x̄: 1.51% x̃: 0.86% HURT stats (abs) min: 1 max: 335 x̄: 20.93 x̃: 6 HURT stats (rel) min: <.01% max: 20.18% x̄: 1.03% x̃: 0.30% 95% mean confidence interval for cycles value: -9.18 -7.42 95% mean confidence interval for cycles %-change: -0.82% -0.71% Cycles are helped. Iron Lake total instructions in shared programs: 8114882 -> 8105574 (-0.11%) instructions in affected programs: 1232504 -> 1223196 (-0.76%) helped: 4109 HURT: 2 helped stats (abs) min: 1 max: 6 x̄: 2.27 x̃: 1 helped stats (rel) min: 0.05% max: 8.33% x̄: 0.99% x̃: 0.66% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.94% max: 4.35% x̄: 2.65% x̃: 2.65% 95% mean confidence interval for instructions value: -2.31 -2.21 95% mean confidence interval for instructions %-change: -1.01% -0.96% Instructions are helped. total cycles in shared programs: 188504036 -> 188466296 (-0.02%) cycles in affected programs: 31203798 -> 31166058 (-0.12%) helped: 3447 HURT: 36 helped stats (abs) min: 2 max: 92 x̄: 11.03 x̃: 8 helped stats (rel) min: <.01% max: 5.41% x̄: 0.21% x̃: 0.13% HURT stats (abs) min: 2 max: 30 x̄: 7.33 x̃: 6 HURT stats (rel) min: 0.01% max: 1.65% x̄: 0.18% x̃: 0.10% 95% mean confidence interval for cycles value: -11.16 -10.51 95% mean confidence interval for cycles %-change: -0.22% -0.20% Cycles are helped. LOST: 0 GAINED: 1 GM45 total instructions in shared programs: 4989697 -> 4984531 (-0.10%) instructions in affected programs: 703952 -> 698786 (-0.73%) helped: 2493 HURT: 2 helped stats (abs) min: 1 max: 6 x̄: 2.07 x̃: 1 helped stats (rel) min: 0.05% max: 8.33% x̄: 1.03% x̃: 0.66% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 0.95% max: 4.35% x̄: 2.65% x̃: 2.65% 95% mean confidence interval for instructions value: -2.13 -2.01 95% mean confidence interval for instructions %-change: -1.07% -0.99% Instructions are helped. total cycles in shared programs: 128929136 -> 128903886 (-0.02%) cycles in affected programs: 21583096 -> 21557846 (-0.12%) helped: 2214 HURT: 17 helped stats (abs) min: 2 max: 92 x̄: 11.44 x̃: 8 helped stats (rel) min: <.01% max: 5.41% x̄: 0.24% x̃: 0.13% HURT stats (abs) min: 2 max: 8 x̄: 4.24 x̃: 4 HURT stats (rel) min: 0.01% max: 1.65% x̄: 0.20% x̃: 0.09% 95% mean confidence interval for cycles value: -11.75 -10.88 95% mean confidence interval for cycles %-change: -0.25% -0.22% Cycles are helped. LOST: 1 GAINED: 1 Reviewed-by: Matt Turner Part-of: --- src/compiler/nir/nir.h | 1 + src/compiler/nir/nir_opt_algebraic.py | 35 +++++++++++++++++++++++++++ src/intel/compiler/brw_nir.c | 7 ++++++ 3 files changed, 43 insertions(+) diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h index 5d29b2f64cc..c44513e5ac6 100644 --- a/src/compiler/nir/nir.h +++ b/src/compiler/nir/nir.h @@ -4255,6 +4255,7 @@ bool nir_opt_access(nir_shader *shader); bool nir_opt_algebraic(nir_shader *shader); bool nir_opt_algebraic_before_ffma(nir_shader *shader); bool nir_opt_algebraic_late(nir_shader *shader); +bool nir_opt_algebraic_distribute_src_mods(nir_shader *shader); bool nir_opt_constant_folding(nir_shader *shader); /* Try to combine a and b into a. Return true if combination was possible, diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py index 472db765026..76ec1acc056 100644 --- a/src/compiler/nir/nir_opt_algebraic.py +++ b/src/compiler/nir/nir_opt_algebraic.py @@ -1853,8 +1853,43 @@ for op in ['ffma']: (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, e, d)), (op, b, ('bcsel', a, c, e), d)), ] +distribute_src_mods = [ + # Try to remove some spurious negations rather than pushing them down. + (('fmul', ('fneg', a), ('fneg', b)), ('fmul', a, b)), + (('ffma', ('fneg', a), ('fneg', b), c), ('ffma', a, b, c)), + (('fdot_replicated2', ('fneg', a), ('fneg', b)), ('fdot_replicated2', a, b)), + (('fdot_replicated3', ('fneg', a), ('fneg', b)), ('fdot_replicated3', a, b)), + (('fdot_replicated4', ('fneg', a), ('fneg', b)), ('fdot_replicated4', a, b)), + (('fneg', ('fneg', a)), a), + + (('fneg', ('ffma(is_used_once)', a, b, c)), ('ffma', ('fneg', a), b, ('fneg', c))), + (('fneg', ('flrp(is_used_once)', a, b, c)), ('flrp', ('fneg', a), ('fneg', b), c)), + (('fneg', ('fadd(is_used_once)', a, b)), ('fadd', ('fneg', a), ('fneg', b))), + + # Note that fmin <-> fmax. I don't think there is a way to distribute + # fabs() into fmin or fmax. + (('fneg', ('fmin(is_used_once)', a, b)), ('fmax', ('fneg', a), ('fneg', b))), + (('fneg', ('fmax(is_used_once)', a, b)), ('fmin', ('fneg', a), ('fneg', b))), + + # fdph works mostly like fdot, but to get the correct result, the negation + # must be applied to the second source. + (('fneg', ('fdph_replicated(is_used_once)', a, b)), ('fdph_replicated', a, ('fneg', b))), + (('fabs', ('fdph_replicated(is_used_once)', a, b)), ('fdph_replicated', ('fabs', a), ('fabs', b))), + + (('fneg', ('fsign(is_used_once)', a)), ('fsign', ('fneg', a))), + (('fabs', ('fsign(is_used_once)', a)), ('fsign', ('fabs', a))), +] + +for op in ['fmul', 'fdot_replicated2', 'fdot_replicated3', 'fdot_replicated4']: + distribute_src_mods.extend([ + (('fneg', (op + '(is_used_once)', a, b)), (op, ('fneg', a), b)), + (('fabs', (op + '(is_used_once)', a, b)), (op, ('fabs', a), ('fabs', b))), + ]) + 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()) +print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_distribute_src_mods", + distribute_src_mods).render()) diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c index 8c0be0d8a26..c7e9c0e1129 100644 --- a/src/intel/compiler/brw_nir.c +++ b/src/intel/compiler/brw_nir.c @@ -921,6 +921,13 @@ brw_postprocess_nir(nir_shader *nir, const struct brw_compiler *compiler, if (is_scalar) OPT(nir_lower_alu_to_scalar, NULL, NULL); + + while (OPT(nir_opt_algebraic_distribute_src_mods)) { + OPT(nir_copy_prop); + OPT(nir_opt_dce); + OPT(nir_opt_cse); + } + OPT(nir_lower_to_source_mods, nir_lower_all_source_mods); OPT(nir_copy_prop); OPT(nir_opt_dce); -- 2.30.2