src/compiler/nir/nir_lower_subgroups.c

   1 /*
   2  * Copyright © 2017 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "nir.h"
  25 #include "nir_builder.h"
  26
  27 /**
  28  * \file nir_opt_intrinsics.c
  29  */
  30
  31 static nir_ssa_def *
  32 lower_read_invocation_to_scalar(nir_builder *b, nir_intrinsic_instr *intrin)
  33 {
  34    /* This is safe to call on scalar things but it would be silly */
  35    assert(intrin->dest.ssa.num_components > 1);
  36
  37    nir_ssa_def *value = nir_ssa_for_src(b, intrin->src[0],
  38                                            intrin->num_components);
  39    nir_ssa_def *reads[4];
  40
  41    for (unsigned i = 0; i < intrin->num_components; i++) {
  42       nir_intrinsic_instr *chan_intrin =
  43          nir_intrinsic_instr_create(b->shader, intrin->intrinsic);
  44       nir_ssa_dest_init(&chan_intrin->instr, &chan_intrin->dest,
  45                         1, intrin->dest.ssa.bit_size, NULL);
  46       chan_intrin->num_components = 1;
  47
  48       /* value */
  49       chan_intrin->src[0] = nir_src_for_ssa(nir_channel(b, value, i));
  50       /* invocation */
  51       if (intrin->intrinsic == nir_intrinsic_read_invocation)
  52          nir_src_copy(&chan_intrin->src[1], &intrin->src[1], chan_intrin);
  53
  54       nir_builder_instr_insert(b, &chan_intrin->instr);
  55
  56       reads[i] = &chan_intrin->dest.ssa;
  57    }
  58
  59    return nir_vec(b, reads, intrin->num_components);
  60 }
  61
  62 static nir_ssa_def *
  63 high_subgroup_mask(nir_builder *b,
  64                    nir_ssa_def *count,
  65                    uint64_t base_mask)
  66 {
  67    /* group_mask could probably be calculated more efficiently but we want to
  68     * be sure not to shift by 64 if the subgroup size is 64 because the GLSL
  69     * shift operator is undefined in that case. In any case if we were worried
  70     * about efficency this should probably be done further down because the
  71     * subgroup size is likely to be known at compile time.
  72     */
  73    nir_ssa_def *subgroup_size = nir_load_subgroup_size(b);
  74    nir_ssa_def *all_bits = nir_imm_int64(b, ~0ull);
  75    nir_ssa_def *shift = nir_isub(b, nir_imm_int(b, 64), subgroup_size);
  76    nir_ssa_def *group_mask = nir_ushr(b, all_bits, shift);
  77    nir_ssa_def *higher_bits = nir_ishl(b, nir_imm_int64(b, base_mask), count);
  78
  79    return nir_iand(b, higher_bits, group_mask);
  80 }
  81
  82 static nir_ssa_def *
  83 lower_subgroups_intrin(nir_builder *b, nir_intrinsic_instr *intrin,
  84                        const nir_lower_subgroups_options *options)
  85 {
  86    switch (intrin->intrinsic) {
  87    case nir_intrinsic_vote_any:
  88    case nir_intrinsic_vote_all:
  89       if (options->lower_vote_trivial)
  90          return nir_ssa_for_src(b, intrin->src[0], 1);
  91       break;
  92
  93    case nir_intrinsic_vote_eq:
  94       if (options->lower_vote_trivial)
  95          return nir_imm_int(b, NIR_TRUE);
  96       break;
  97
  98    case nir_intrinsic_read_invocation:
  99    case nir_intrinsic_read_first_invocation:
 100       if (options->lower_to_scalar && intrin->num_components > 1)
 101          return lower_read_invocation_to_scalar(b, intrin);
 102       break;
 103
 104    case nir_intrinsic_load_subgroup_eq_mask:
 105    case nir_intrinsic_load_subgroup_ge_mask:
 106    case nir_intrinsic_load_subgroup_gt_mask:
 107    case nir_intrinsic_load_subgroup_le_mask:
 108    case nir_intrinsic_load_subgroup_lt_mask: {
 109       if (!options->lower_subgroup_masks)
 110          return NULL;
 111
 112       nir_ssa_def *count = nir_load_subgroup_invocation(b);
 113
 114       switch (intrin->intrinsic) {
 115       case nir_intrinsic_load_subgroup_eq_mask:
 116          return nir_ishl(b, nir_imm_int64(b, 1ull), count);
 117       case nir_intrinsic_load_subgroup_ge_mask:
 118          return high_subgroup_mask(b, count, ~0ull);
 119       case nir_intrinsic_load_subgroup_gt_mask:
 120          return high_subgroup_mask(b, count, ~1ull);
 121       case nir_intrinsic_load_subgroup_le_mask:
 122          return nir_inot(b, nir_ishl(b, nir_imm_int64(b, ~1ull), count));
 123       case nir_intrinsic_load_subgroup_lt_mask:
 124          return nir_inot(b, nir_ishl(b, nir_imm_int64(b, ~0ull), count));
 125       default:
 126          unreachable("you seriously can't tell this is unreachable?");
 127       }
 128       break;
 129    }
 130    default:
 131       break;
 132    }
 133
 134    return NULL;
 135 }
 136
 137 static bool
 138 lower_subgroups_impl(nir_function_impl *impl,
 139                      const nir_lower_subgroups_options *options)
 140 {
 141    nir_builder b;
 142    nir_builder_init(&b, impl);
 143    bool progress = false;
 144
 145    nir_foreach_block(block, impl) {
 146       nir_foreach_instr_safe(instr, block) {
 147          if (instr->type != nir_instr_type_intrinsic)
 148             continue;
 149
 150          nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 151          b.cursor = nir_before_instr(instr);
 152
 153          nir_ssa_def *lower = lower_subgroups_intrin(&b, intrin, options);
 154          if (!lower)
 155             continue;
 156
 157          nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(lower));
 158          nir_instr_remove(instr);
 159          progress = true;
 160       }
 161    }
 162
 163    return progress;
 164 }
 165
 166 bool
 167 nir_lower_subgroups(nir_shader *shader,
 168                     const nir_lower_subgroups_options *options)
 169 {
 170    bool progress = false;
 171
 172    nir_foreach_function(function, shader) {
 173       if (!function->impl)
 174          continue;
 175
 176       if (lower_subgroups_impl(function->impl, options)) {
 177          progress = true;
 178          nir_metadata_preserve(function->impl, nir_metadata_block_index |
 179                                                nir_metadata_dominance);
 180       }
 181    }
 182
 183    return progress;
 184 }