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    nir_function_impl *impl;
  30    bool progress;
  31    nir_builder builder;
  32 };
  33
  34 static bool
  35 lower_cs_intrinsics_convert_block(struct lower_intrinsics_state *state,
  36                                   nir_block *block)
  37 {
  38    bool progress = false;
  39    nir_builder *b = &state->builder;
  40    nir_shader *nir = state->nir;
  41
  42    /* Reuse calculated values inside the block. */
  43    nir_ssa_def *local_index = NULL;
  44    nir_ssa_def *local_id = NULL;
  45
  46    nir_foreach_instr_safe(instr, block) {
  47       if (instr->type != nir_instr_type_intrinsic)
  48          continue;
  49
  50       nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(instr);
  51
  52       b->cursor = nir_after_instr(&intrinsic->instr);
  53
  54       nir_ssa_def *sysval;
  55       switch (intrinsic->intrinsic) {
  56       case nir_intrinsic_load_local_invocation_index:
  57       case nir_intrinsic_load_local_invocation_id: {
  58          /* First time we are using those, so let's calculate them. */
  59          if (!local_index) {
  60             assert(!local_id);
  61
  62             nir_ssa_def *subgroup_id = nir_load_subgroup_id(b);
  63
  64             nir_ssa_def *thread_local_id =
  65                nir_imul(b, subgroup_id, nir_load_simd_width_intel(b));
  66             nir_ssa_def *channel = nir_load_subgroup_invocation(b);
  67             nir_ssa_def *linear = nir_iadd(b, channel, thread_local_id);
  68
  69             nir_ssa_def *size_x;
  70             nir_ssa_def *size_y;
  71             if (state->nir->info.cs.local_size_variable) {
  72                nir_ssa_def *size_xyz = nir_load_local_group_size(b);
  73                size_x = nir_channel(b, size_xyz, 0);
  74                size_y = nir_channel(b, size_xyz, 1);
  75             } else {
  76                size_x = nir_imm_int(b, nir->info.cs.local_size[0]);
  77                size_y = nir_imm_int(b, nir->info.cs.local_size[1]);
  78             }
  79
  80             /* The local invocation index and ID must respect the following
  81              *
  82              *    gl_LocalInvocationID.x =
  83              *       gl_LocalInvocationIndex % gl_WorkGroupSize.x;
  84              *    gl_LocalInvocationID.y =
  85              *       (gl_LocalInvocationIndex / gl_WorkGroupSize.x) %
  86              *       gl_WorkGroupSize.y;
  87              *    gl_LocalInvocationID.z =
  88              *       (gl_LocalInvocationIndex /
  89              *        (gl_WorkGroupSize.x * gl_WorkGroupSize.y)) %
  90              *       gl_WorkGroupSize.z;
  91              *
  92              * However, the final % gl_WorkGroupSize.z does nothing unless we
  93              * accidentally end up with a gl_LocalInvocationIndex that is too
  94              * large so it can safely be omitted.
  95              */
  96
  97             if (state->nir->info.cs.derivative_group != DERIVATIVE_GROUP_QUADS) {
  98                /* If we are not grouping in quads, just set the local invocatio
  99                 * index linearly, and calculate local invocation ID from that.
 100                 */
 101                local_index = linear;
 102
 103                nir_ssa_def *id_x, *id_y, *id_z;
 104                id_x = nir_umod(b, local_index, size_x);
 105                id_y = nir_umod(b, nir_udiv(b, local_index, size_x), size_y);
 106                id_z = nir_udiv(b, local_index, nir_imul(b, size_x, size_y));
 107                local_id = nir_vec3(b, id_x, id_y, id_z);
 108             } else {
 109                /* For quads, first we figure out the 2x2 grid the invocation
 110                 * belongs to -- treating extra Z layers as just more rows.
 111                 * Then map that into local invocation ID (trivial) and local
 112                 * invocation index.  Skipping Z simplify index calculation.
 113                 */
 114
 115                nir_ssa_def *one = nir_imm_int(b, 1);
 116                nir_ssa_def *double_size_x = nir_ishl(b, size_x, one);
 117
 118                /* ID within a pair of rows, where each group of 4 is 2x2 quad. */
 119                nir_ssa_def *row_pair_id = nir_umod(b, linear, double_size_x);
 120                nir_ssa_def *y_row_pairs = nir_udiv(b, linear, double_size_x);
 121
 122                nir_ssa_def *x =
 123                   nir_ior(b,
 124                           nir_iand(b, row_pair_id, one),
 125                           nir_iand(b, nir_ishr(b, row_pair_id, one),
 126                                    nir_imm_int(b, 0xfffffffe)));
 127                nir_ssa_def *y =
 128                   nir_ior(b,
 129                           nir_ishl(b, y_row_pairs, one),
 130                           nir_iand(b, nir_ishr(b, row_pair_id, one), one));
 131
 132                local_id = nir_vec3(b, x,
 133                                    nir_umod(b, y, size_y),
 134                                    nir_udiv(b, y, size_y));
 135                local_index = nir_iadd(b, x, nir_imul(b, y, size_x));
 136             }
 137          }
 138
 139          assert(local_id);
 140          assert(local_index);
 141          if (intrinsic->intrinsic == nir_intrinsic_load_local_invocation_id)
 142             sysval = local_id;
 143          else
 144             sysval = local_index;
 145          break;
 146       }
 147
 148       case nir_intrinsic_load_num_subgroups: {
 149          nir_ssa_def *size;
 150          if (state->nir->info.cs.local_size_variable) {
 151             nir_ssa_def *size_xyz = nir_load_local_group_size(b);
 152             nir_ssa_def *size_x = nir_channel(b, size_xyz, 0);
 153             nir_ssa_def *size_y = nir_channel(b, size_xyz, 1);
 154             nir_ssa_def *size_z = nir_channel(b, size_xyz, 2);
 155             size = nir_imul(b, nir_imul(b, size_x, size_y), size_z);
 156          } else {
 157             size = nir_imm_int(b, nir->info.cs.local_size[0] *
 158                                   nir->info.cs.local_size[1] *
 159                                   nir->info.cs.local_size[2]);
 160          }
 161
 162          /* Calculate the equivalent of DIV_ROUND_UP. */
 163          nir_ssa_def *simd_width = nir_load_simd_width_intel(b);
 164          sysval =
 165             nir_udiv(b, nir_iadd_imm(b, nir_iadd(b, size, simd_width), -1),
 166                         simd_width);
 167          break;
 168       }
 169
 170       default:
 171          continue;
 172       }
 173
 174       nir_ssa_def_rewrite_uses(&intrinsic->dest.ssa, nir_src_for_ssa(sysval));
 175       nir_instr_remove(&intrinsic->instr);
 176
 177       state->progress = true;
 178    }
 179
 180    return progress;
 181 }
 182
 183 static void
 184 lower_cs_intrinsics_convert_impl(struct lower_intrinsics_state *state)
 185 {
 186    nir_builder_init(&state->builder, state->impl);
 187
 188    nir_foreach_block(block, state->impl) {
 189       lower_cs_intrinsics_convert_block(state, block);
 190    }
 191
 192    nir_metadata_preserve(state->impl,
 193                          nir_metadata_block_index | nir_metadata_dominance);
 194 }
 195
 196 bool
 197 brw_nir_lower_cs_intrinsics(nir_shader *nir)
 198 {
 199    assert(nir->info.stage == MESA_SHADER_COMPUTE);
 200
 201    struct lower_intrinsics_state state = {
 202       .nir = nir,
 203    };
 204
 205    /* Constraints from NV_compute_shader_derivatives. */
 206    if (!nir->info.cs.local_size_variable) {
 207       if (nir->info.cs.derivative_group == DERIVATIVE_GROUP_QUADS) {
 208          assert(nir->info.cs.local_size[0] % 2 == 0);
 209          assert(nir->info.cs.local_size[1] % 2 == 0);
 210       } else if (nir->info.cs.derivative_group == DERIVATIVE_GROUP_LINEAR) {
 211          ASSERTED unsigned local_workgroup_size =
 212             nir->info.cs.local_size[0] *
 213             nir->info.cs.local_size[1] *
 214             nir->info.cs.local_size[2];
 215          assert(local_workgroup_size % 4 == 0);
 216       }
 217    }
 218
 219    nir_foreach_function(function, nir) {
 220       if (function->impl) {
 221          state.impl = function->impl;
 222          lower_cs_intrinsics_convert_impl(&state);
 223       }
 224    }
 225
 226    return state.progress;
 227 }