src/intel/compiler/brw_nir_lower_cs_intrinsics.c

   1 /*
   2  * Copyright (c) 2016 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 "brw_nir.h"
  25 #include "compiler/nir/nir_builder.h"
  26
  27 struct lower_intrinsics_state {
  28    nir_shader *nir;
  29    unsigned dispatch_width;
  30    nir_function_impl *impl;
  31    bool progress;
  32    nir_builder builder;
  33    unsigned local_workgroup_size;
  34 };
  35
  36 static bool
  37 lower_cs_intrinsics_convert_block(struct lower_intrinsics_state *state,
  38                                   nir_block *block)
  39 {
  40    bool progress = false;
  41    nir_builder *b = &state->builder;
  42    nir_shader *nir = state->nir;
  43
  44    nir_foreach_instr_safe(instr, block) {
  45       if (instr->type != nir_instr_type_intrinsic)
  46          continue;
  47
  48       nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(instr);
  49
  50       b->cursor = nir_after_instr(&intrinsic->instr);
  51
  52       nir_ssa_def *sysval;
  53       switch (intrinsic->intrinsic) {
  54       case nir_intrinsic_load_local_invocation_index: {
  55          /* We construct the local invocation index from:
  56           *
  57           *    gl_LocalInvocationIndex =
  58           *       cs_thread_local_id + subgroup_invocation;
  59           */
  60          nir_ssa_def *subgroup_id;
  61          if (state->local_workgroup_size <= state->dispatch_width)
  62             subgroup_id = nir_imm_int(b, 0);
  63          else
  64             subgroup_id = nir_load_subgroup_id(b);
  65
  66          nir_ssa_def *thread_local_id =
  67             nir_imul(b, subgroup_id, nir_imm_int(b, state->dispatch_width));
  68          nir_ssa_def *channel = nir_load_subgroup_invocation(b);
  69          sysval = nir_iadd(b, channel, thread_local_id);
  70          break;
  71       }
  72
  73       case nir_intrinsic_load_local_invocation_id: {
  74          /* We lower gl_LocalInvocationID from gl_LocalInvocationIndex based
  75           * on this formula:
  76           *
  77           *    gl_LocalInvocationID.x =
  78           *       gl_LocalInvocationIndex % gl_WorkGroupSize.x;
  79           *    gl_LocalInvocationID.y =
  80           *       (gl_LocalInvocationIndex / gl_WorkGroupSize.x) %
  81           *       gl_WorkGroupSize.y;
  82           *    gl_LocalInvocationID.z =
  83           *       (gl_LocalInvocationIndex /
  84           *        (gl_WorkGroupSize.x * gl_WorkGroupSize.y)) %
  85           *       gl_WorkGroupSize.z;
  86           */
  87          unsigned *size = nir->info.cs.local_size;
  88
  89          nir_ssa_def *local_index = nir_load_local_invocation_index(b);
  90
  91          nir_const_value uvec3;
  92          memset(&uvec3, 0, sizeof(uvec3));
  93          uvec3.u32[0] = 1;
  94          uvec3.u32[1] = size[0];
  95          uvec3.u32[2] = size[0] * size[1];
  96          nir_ssa_def *div_val = nir_build_imm(b, 3, 32, uvec3);
  97          uvec3.u32[0] = size[0];
  98          uvec3.u32[1] = size[1];
  99          uvec3.u32[2] = size[2];
 100          nir_ssa_def *mod_val = nir_build_imm(b, 3, 32, uvec3);
 101
 102          sysval = nir_umod(b, nir_udiv(b, local_index, div_val), mod_val);
 103          break;
 104       }
 105
 106       case nir_intrinsic_load_subgroup_id:
 107          if (state->local_workgroup_size > 8)
 108             continue;
 109
 110          /* For small workgroup sizes, we know subgroup_id will be zero */
 111          sysval = nir_imm_int(b, 0);
 112          break;
 113
 114       case nir_intrinsic_load_num_subgroups: {
 115          unsigned local_workgroup_size =
 116             nir->info.cs.local_size[0] * nir->info.cs.local_size[1] *
 117             nir->info.cs.local_size[2];
 118          unsigned num_subgroups =
 119             DIV_ROUND_UP(local_workgroup_size, state->dispatch_width);
 120          sysval = nir_imm_int(b, num_subgroups);
 121          break;
 122       }
 123
 124       default:
 125          continue;
 126       }
 127
 128       nir_ssa_def_rewrite_uses(&intrinsic->dest.ssa, nir_src_for_ssa(sysval));
 129       nir_instr_remove(&intrinsic->instr);
 130
 131       state->progress = true;
 132    }
 133
 134    return progress;
 135 }
 136
 137 static void
 138 lower_cs_intrinsics_convert_impl(struct lower_intrinsics_state *state)
 139 {
 140    nir_builder_init(&state->builder, state->impl);
 141
 142    nir_foreach_block(block, state->impl) {
 143       lower_cs_intrinsics_convert_block(state, block);
 144    }
 145
 146    nir_metadata_preserve(state->impl,
 147                          nir_metadata_block_index | nir_metadata_dominance);
 148 }
 149
 150 bool
 151 brw_nir_lower_cs_intrinsics(nir_shader *nir,
 152                             unsigned dispatch_width)
 153 {
 154    assert(nir->info.stage == MESA_SHADER_COMPUTE);
 155
 156    bool progress = false;
 157    struct lower_intrinsics_state state;
 158    memset(&state, 0, sizeof(state));
 159    state.nir = nir;
 160    state.dispatch_width = dispatch_width;
 161    state.local_workgroup_size = nir->info.cs.local_size[0] *
 162                                 nir->info.cs.local_size[1] *
 163                                 nir->info.cs.local_size[2];
 164
 165    do {
 166       state.progress = false;
 167       nir_foreach_function(function, nir) {
 168          if (function->impl) {
 169             state.impl = function->impl;
 170             lower_cs_intrinsics_convert_impl(&state);
 171          }
 172       }
 173       progress |= state.progress;
 174    } while (state.progress);
 175
 176    return progress;
 177 }