diff options
author | Ian Romanick <ian.d.romanick@intel.com> | 2019-05-20 16:21:58 -0700 |
---|---|---|
committer | Ian Romanick <ian.d.romanick@intel.com> | 2019-05-31 08:47:04 -0700 |
commit | 9f2000db5bacb16f4355ed00a8d65b71c1f71611 (patch) | |
tree | cf3de76054dc64f1f0ab220ccda6eddebc10c3f1 | |
parent | 65df6122da9d171bc44f9ca969825478543b5384 (diff) |
nir: Rematerialize comparisons for nir_intrinsic_discard_if tooexperiment/rematerialize-compares
Shader-db results for this series rebased on top of MR !935:
Ice Lake
total instructions in shared programs: 17179795 -> 17179776 (<.01%)
instructions in affected programs: 10310 -> 10291 (-0.18%)
helped: 19
HURT: 0
helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
helped stats (rel) min: 0.08% max: 0.49% x̄: 0.28% x̃: 0.30%
95% mean confidence interval for instructions value: -1.00 -1.00
95% mean confidence interval for instructions %-change: -0.36% -0.21%
Instructions are helped.
total cycles in shared programs: 361083242 -> 361084861 (<.01%)
cycles in affected programs: 595261 -> 596880 (0.27%)
helped: 12
HURT: 7
helped stats (abs) min: 2 max: 211 x̄: 53.58 x̃: 6
helped stats (rel) min: 0.03% max: 0.89% x̄: 0.26% x̃: 0.06%
HURT stats (abs) min: 1 max: 1882 x̄: 323.14 x̃: 48
HURT stats (rel) min: 0.01% max: 11.34% x̄: 1.86% x̃: 0.06%
95% mean confidence interval for cycles value: -128.77 299.19
95% mean confidence interval for cycles %-change: -0.76% 1.80%
Inconclusive result (value mean confidence interval includes 0).
LOST: 0
GAINED: 1
-rw-r--r-- | src/compiler/nir/nir_opt_rematerialize_compares.c | 61 |
1 files changed, 40 insertions, 21 deletions
diff --git a/src/compiler/nir/nir_opt_rematerialize_compares.c b/src/compiler/nir/nir_opt_rematerialize_compares.c index 806dbd2f29a..16702fb23e6 100644 --- a/src/compiler/nir/nir_opt_rematerialize_compares.c +++ b/src/compiler/nir/nir_opt_rematerialize_compares.c @@ -58,25 +58,32 @@ is_two_src_comparison(const nir_alu_instr *instr) } static bool -all_uses_are_bcsel(const nir_alu_instr *instr) +all_uses_are_bcsel_or_discard(const nir_alu_instr *instr) { if (!instr->dest.dest.is_ssa) return false; nir_foreach_use(use, &instr->dest.dest.ssa) { - if (use->parent_instr->type != nir_instr_type_alu) - return false; - - nir_alu_instr *const alu = nir_instr_as_alu(use->parent_instr); - if (alu->op != nir_op_bcsel && - alu->op != nir_op_b32csel) - return false; - - /* Not only must the result be used by a bcsel, but it must be used as - * the first source (the condition). - */ - if (alu->src[0].src.ssa != &instr->dest.dest.ssa) + if (use->parent_instr->type == nir_instr_type_alu) { + nir_alu_instr *const alu = nir_instr_as_alu(use->parent_instr); + if (alu->op != nir_op_bcsel && + alu->op != nir_op_b32csel) + return false; + + /* Not only must the result be used by a bcsel, but it must be used as + * the first source (the condition). + */ + if (alu->src[0].src.ssa != &instr->dest.dest.ssa) + return false; + } else if (use->parent_instr->type == nir_instr_type_intrinsic) { + nir_intrinsic_instr *const intrin = + nir_instr_as_intrinsic(use->parent_instr); + + if (intrin->intrinsic != nir_intrinsic_discard_if) + return false; + } else { return false; + } } return true; @@ -96,7 +103,7 @@ nir_opt_rematerialize_compares_impl(nir_shader *shader, nir_function_impl *impl) if (!is_two_src_comparison(alu)) continue; - if (!all_uses_are_bcsel(alu)) + if (!all_uses_are_bcsel_or_discard(alu)) continue; /* At this point it is known that alu is a comparison instruction @@ -122,14 +129,26 @@ nir_opt_rematerialize_compares_impl(nir_shader *shader, nir_function_impl *impl) nir_instr_insert_before(use_instr, &clone->instr); - nir_alu_instr *const use_alu = nir_instr_as_alu(use_instr); - for (unsigned i = 0; i < nir_op_infos[use_alu->op].num_inputs; i++) { - if (use_alu->src[i].src.ssa == &alu->dest.dest.ssa) { - nir_instr_rewrite_src(&use_alu->instr, - &use_alu->src[i].src, - nir_src_for_ssa(&clone->dest.dest.ssa)); - progress = true; + if (use_instr->type == nir_instr_type_alu) { + nir_alu_instr *const use_alu = nir_instr_as_alu(use_instr); + for (unsigned i = 0; i < nir_op_infos[use_alu->op].num_inputs; i++) { + if (use_alu->src[i].src.ssa == &alu->dest.dest.ssa) { + nir_instr_rewrite_src(&use_alu->instr, + &use_alu->src[i].src, + nir_src_for_ssa(&clone->dest.dest.ssa)); + progress = true; + } } + } else { + assert(use_instr->type == nir_instr_type_intrinsic); + + nir_intrinsic_instr *const intrin = + nir_instr_as_intrinsic(use_instr); + + nir_instr_rewrite_src(&intrin->instr, + &intrin->src[0], + nir_src_for_ssa(&clone->dest.dest.ssa)); + progress = true; } } |