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    /* Reuse calculated values inside the block. */
  45    nir_ssa_def *local_index = NULL;
  46    nir_ssa_def *local_id = NULL;
  47
  48    nir_foreach_instr_safe(instr, block) {
  49       if (instr->type != nir_instr_type_intrinsic)
  50          continue;
  51
  52       nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(instr);
  53
  54       b->cursor = nir_after_instr(&intrinsic->instr);
  55
  56       nir_ssa_def *sysval;
  57       switch (intrinsic->intrinsic) {
  58       case nir_intrinsic_barrier: {
  59          /* Our HW barrier instruction doesn't do a memory barrier for us but
  60           * the GLSL barrier() intrinsic does for shared memory.  Insert a
  61           * shared memory barrier before every barrier().
  62           */
  63          b->cursor = nir_before_instr(&intrinsic->instr);
  64
  65          nir_intrinsic_instr *shared_barrier =
  66             nir_intrinsic_instr_create(b->shader,
  67                                        nir_intrinsic_memory_barrier_shared);
  68          nir_builder_instr_insert(b, &shared_barrier->instr);
  69          continue;
  70       }
  71
  72       case nir_intrinsic_load_local_invocation_index:
  73       case nir_intrinsic_load_local_invocation_id: {
  74          /* First time we are using those, so let's calculate them. */
  75          if (!local_index) {
  76             assert(!local_id);
  77
  78             nir_ssa_def *subgroup_id;
  79             if (state->local_workgroup_size <= state->dispatch_width)
  80                subgroup_id = nir_imm_int(b, 0);
  81             else
  82                subgroup_id = nir_load_subgroup_id(b);
  83
  84             nir_ssa_def *thread_local_id =
  85                nir_imul_imm(b, subgroup_id, state->dispatch_width);
  86             nir_ssa_def *channel = nir_load_subgroup_invocation(b);
  87             nir_ssa_def *linear = nir_iadd(b, channel, thread_local_id);
  88
  89             nir_ssa_def *size_x = nir_imm_int(b, nir->info.cs.local_size[0]);
  90             nir_ssa_def *size_y = nir_imm_int(b, nir->info.cs.local_size[1]);
  91
  92             /* The local invocation index and ID must respect the following
  93              *
  94              *    gl_LocalInvocationID.x =
  95              *       gl_LocalInvocationIndex % gl_WorkGroupSize.x;
  96              *    gl_LocalInvocationID.y =
  97              *       (gl_LocalInvocationIndex / gl_WorkGroupSize.x) %
  98              *       gl_WorkGroupSize.y;
  99              *    gl_LocalInvocationID.z =
 100              *       (gl_LocalInvocationIndex /
 101              *        (gl_WorkGroupSize.x * gl_WorkGroupSize.y)) %
 102              *       gl_WorkGroupSize.z;
 103              *
 104              * However, the final % gl_WorkGroupSize.z does nothing unless we
 105              * accidentally end up with a gl_LocalInvocationIndex that is too
 106              * large so it can safely be omitted.
 107              */
 108
 109             if (state->nir->info.cs.derivative_group != DERIVATIVE_GROUP_QUADS) {
 110                /* If we are not grouping in quads, just set the local invocatio
 111                 * index linearly, and calculate local invocation ID from that.
 112                 */
 113                local_index = linear;
 114
 115                nir_ssa_def *id_x, *id_y, *id_z;
 116                id_x = nir_umod(b, local_index, size_x);
 117                id_y = nir_umod(b, nir_udiv(b, local_index, size_x), size_y);
 118                id_z = nir_udiv(b, local_index, nir_imul(b, size_x, size_y));
 119                local_id = nir_vec3(b, id_x, id_y, id_z);
 120             } else {
 121                /* For quads, first we figure out the 2x2 grid the invocation
 122                 * belongs to -- treating extra Z layers as just more rows.
 123                 * Then map that into local invocation ID (trivial) and local
 124                 * invocation index.  Skipping Z simplify index calculation.
 125                 */
 126
 127                nir_ssa_def *one = nir_imm_int(b, 1);
 128                nir_ssa_def *double_size_x = nir_ishl(b, size_x, one);
 129
 130                /* ID within a pair of rows, where each group of 4 is 2x2 quad. */
 131                nir_ssa_def *row_pair_id = nir_umod(b, linear, double_size_x);
 132                nir_ssa_def *y_row_pairs = nir_udiv(b, linear, double_size_x);
 133
 134                nir_ssa_def *x =
 135                   nir_ior(b,
 136                           nir_iand(b, row_pair_id, one),
 137                           nir_iand(b, nir_ishr(b, row_pair_id, one),
 138                                    nir_imm_int(b, 0xfffffffe)));
 139                nir_ssa_def *y =
 140                   nir_ior(b,
 141                           nir_ishl(b, y_row_pairs, one),
 142                           nir_iand(b, nir_ishr(b, row_pair_id, one), one));
 143
 144                local_id = nir_vec3(b, x,
 145                                    nir_umod(b, y, size_y),
 146                                    nir_udiv(b, y, size_y));
 147                local_index = nir_iadd(b, x, nir_imul(b, y, size_x));
 148             }
 149          }
 150
 151          assert(local_id);
 152          assert(local_index);
 153          if (intrinsic->intrinsic == nir_intrinsic_load_local_invocation_id)
 154             sysval = local_id;
 155          else
 156             sysval = local_index;
 157          break;
 158       }
 159
 160       case nir_intrinsic_load_subgroup_id:
 161          if (state->local_workgroup_size > 8)
 162             continue;
 163
 164          /* For small workgroup sizes, we know subgroup_id will be zero */
 165          sysval = nir_imm_int(b, 0);
 166          break;
 167
 168       case nir_intrinsic_load_num_subgroups: {
 169          unsigned local_workgroup_size =
 170             nir->info.cs.local_size[0] * nir->info.cs.local_size[1] *
 171             nir->info.cs.local_size[2];
 172          unsigned num_subgroups =
 173             DIV_ROUND_UP(local_workgroup_size, state->dispatch_width);
 174          sysval = nir_imm_int(b, num_subgroups);
 175          break;
 176       }
 177
 178       default:
 179          continue;
 180       }
 181
 182       nir_ssa_def_rewrite_uses(&intrinsic->dest.ssa, nir_src_for_ssa(sysval));
 183       nir_instr_remove(&intrinsic->instr);
 184
 185       state->progress = true;
 186    }
 187
 188    return progress;
 189 }
 190
 191 static void
 192 lower_cs_intrinsics_convert_impl(struct lower_intrinsics_state *state)
 193 {
 194    nir_builder_init(&state->builder, state->impl);
 195
 196    nir_foreach_block(block, state->impl) {
 197       lower_cs_intrinsics_convert_block(state, block);
 198    }
 199
 200    nir_metadata_preserve(state->impl,
 201                          nir_metadata_block_index | nir_metadata_dominance);
 202 }
 203
 204 bool
 205 brw_nir_lower_cs_intrinsics(nir_shader *nir,
 206                             unsigned dispatch_width)
 207 {
 208    assert(nir->info.stage == MESA_SHADER_COMPUTE);
 209
 210    struct lower_intrinsics_state state = {
 211       .nir = nir,
 212       .dispatch_width = dispatch_width,
 213    };
 214
 215    assert(!nir->info.cs.local_size_variable);
 216    state.local_workgroup_size = nir->info.cs.local_size[0] *
 217                                 nir->info.cs.local_size[1] *
 218                                 nir->info.cs.local_size[2];
 219
 220    /* Constraints from NV_compute_shader_derivatives. */
 221    if (nir->info.cs.derivative_group == DERIVATIVE_GROUP_QUADS) {
 222       assert(nir->info.cs.local_size[0] % 2 == 0);
 223       assert(nir->info.cs.local_size[1] % 2 == 0);
 224    } else if (nir->info.cs.derivative_group == DERIVATIVE_GROUP_LINEAR) {
 225       assert(state.local_workgroup_size % 4 == 0);
 226    }
 227
 228    nir_foreach_function(function, nir) {
 229       if (function->impl) {
 230          state.impl = function->impl;
 231          lower_cs_intrinsics_convert_impl(&state);
 232       }
 233    }
 234
 235    return state.progress;
 236 }