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