src/compiler/spirv/vtn_subgroup.c

   1 /*
   2  * Copyright © 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 "vtn_private.h"
  25
  26 static void
  27 vtn_build_subgroup_instr(struct vtn_builder *b,
  28                          nir_intrinsic_op nir_op,
  29                          struct vtn_ssa_value *dst,
  30                          struct vtn_ssa_value *src0,
  31                          nir_ssa_def *index,
  32                          unsigned const_idx0,
  33                          unsigned const_idx1)
  34 {
  35    /* Some of the subgroup operations take an index.  SPIR-V allows this to be
  36     * any integer type.  To make things simpler for drivers, we only support
  37     * 32-bit indices.
  38     */
  39    if (index && index->bit_size != 32)
  40       index = nir_u2u32(&b->nb, index);
  41
  42    vtn_assert(dst->type == src0->type);
  43    if (!glsl_type_is_vector_or_scalar(dst->type)) {
  44       for (unsigned i = 0; i < glsl_get_length(dst->type); i++) {
  45          vtn_build_subgroup_instr(b, nir_op, dst->elems[i],
  46                                   src0->elems[i], index,
  47                                   const_idx0, const_idx1);
  48       }
  49       return;
  50    }
  51
  52    nir_intrinsic_instr *intrin =
  53       nir_intrinsic_instr_create(b->nb.shader, nir_op);
  54    nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
  55                               dst->type, NULL);
  56    intrin->num_components = intrin->dest.ssa.num_components;
  57
  58    intrin->src[0] = nir_src_for_ssa(src0->def);
  59    if (index)
  60       intrin->src[1] = nir_src_for_ssa(index);
  61
  62    intrin->const_index[0] = const_idx0;
  63    intrin->const_index[1] = const_idx1;
  64
  65    nir_builder_instr_insert(&b->nb, &intrin->instr);
  66
  67    dst->def = &intrin->dest.ssa;
  68 }
  69
  70 void
  71 vtn_handle_subgroup(struct vtn_builder *b, SpvOp opcode,
  72                     const uint32_t *w, unsigned count)
  73 {
  74    struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
  75
  76    val->ssa = vtn_create_ssa_value(b, val->type->type);
  77
  78    switch (opcode) {
  79    case SpvOpGroupNonUniformElect: {
  80       vtn_fail_if(val->type->type != glsl_bool_type(),
  81                   "OpGroupNonUniformElect must return a Bool");
  82       nir_intrinsic_instr *elect =
  83          nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_elect);
  84       nir_ssa_dest_init_for_type(&elect->instr, &elect->dest,
  85                                  val->type->type, NULL);
  86       nir_builder_instr_insert(&b->nb, &elect->instr);
  87       val->ssa->def = &elect->dest.ssa;
  88       break;
  89    }
  90
  91    case SpvOpGroupNonUniformBallot: ++w;
  92    case SpvOpSubgroupBallotKHR: {
  93       vtn_fail_if(val->type->type != glsl_vector_type(GLSL_TYPE_UINT, 4),
  94                   "OpGroupNonUniformBallot must return a uvec4");
  95       nir_intrinsic_instr *ballot =
  96          nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_ballot);
  97       ballot->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
  98       nir_ssa_dest_init(&ballot->instr, &ballot->dest, 4, 32, NULL);
  99       ballot->num_components = 4;
 100       nir_builder_instr_insert(&b->nb, &ballot->instr);
 101       val->ssa->def = &ballot->dest.ssa;
 102       break;
 103    }
 104
 105    case SpvOpGroupNonUniformInverseBallot: {
 106       /* This one is just a BallotBitfieldExtract with subgroup invocation.
 107        * We could add a NIR intrinsic but it's easier to just lower it on the
 108        * spot.
 109        */
 110       nir_intrinsic_instr *intrin =
 111          nir_intrinsic_instr_create(b->nb.shader,
 112                                     nir_intrinsic_ballot_bitfield_extract);
 113
 114       intrin->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
 115       intrin->src[1] = nir_src_for_ssa(nir_load_subgroup_invocation(&b->nb));
 116
 117       nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
 118                                  val->type->type, NULL);
 119       nir_builder_instr_insert(&b->nb, &intrin->instr);
 120
 121       val->ssa->def = &intrin->dest.ssa;
 122       break;
 123    }
 124
 125    case SpvOpGroupNonUniformBallotBitExtract:
 126    case SpvOpGroupNonUniformBallotBitCount:
 127    case SpvOpGroupNonUniformBallotFindLSB:
 128    case SpvOpGroupNonUniformBallotFindMSB: {
 129       nir_ssa_def *src0, *src1 = NULL;
 130       nir_intrinsic_op op;
 131       switch (opcode) {
 132       case SpvOpGroupNonUniformBallotBitExtract:
 133          op = nir_intrinsic_ballot_bitfield_extract;
 134          src0 = vtn_ssa_value(b, w[4])->def;
 135          src1 = vtn_ssa_value(b, w[5])->def;
 136          break;
 137       case SpvOpGroupNonUniformBallotBitCount:
 138          switch ((SpvGroupOperation)w[4]) {
 139          case SpvGroupOperationReduce:
 140             op = nir_intrinsic_ballot_bit_count_reduce;
 141             break;
 142          case SpvGroupOperationInclusiveScan:
 143             op = nir_intrinsic_ballot_bit_count_inclusive;
 144             break;
 145          case SpvGroupOperationExclusiveScan:
 146             op = nir_intrinsic_ballot_bit_count_exclusive;
 147             break;
 148          default:
 149             unreachable("Invalid group operation");
 150          }
 151          src0 = vtn_ssa_value(b, w[5])->def;
 152          break;
 153       case SpvOpGroupNonUniformBallotFindLSB:
 154          op = nir_intrinsic_ballot_find_lsb;
 155          src0 = vtn_ssa_value(b, w[4])->def;
 156          break;
 157       case SpvOpGroupNonUniformBallotFindMSB:
 158          op = nir_intrinsic_ballot_find_msb;
 159          src0 = vtn_ssa_value(b, w[4])->def;
 160          break;
 161       default:
 162          unreachable("Unhandled opcode");
 163       }
 164
 165       nir_intrinsic_instr *intrin =
 166          nir_intrinsic_instr_create(b->nb.shader, op);
 167
 168       intrin->src[0] = nir_src_for_ssa(src0);
 169       if (src1)
 170          intrin->src[1] = nir_src_for_ssa(src1);
 171
 172       nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
 173                                  val->type->type, NULL);
 174       nir_builder_instr_insert(&b->nb, &intrin->instr);
 175
 176       val->ssa->def = &intrin->dest.ssa;
 177       break;
 178    }
 179
 180    case SpvOpGroupNonUniformBroadcastFirst: ++w;
 181    case SpvOpSubgroupFirstInvocationKHR:
 182       vtn_build_subgroup_instr(b, nir_intrinsic_read_first_invocation,
 183                                val->ssa, vtn_ssa_value(b, w[3]), NULL, 0, 0);
 184       break;
 185
 186    case SpvOpGroupNonUniformBroadcast: ++w;
 187    case SpvOpSubgroupReadInvocationKHR:
 188       vtn_build_subgroup_instr(b, nir_intrinsic_read_invocation,
 189                                val->ssa, vtn_ssa_value(b, w[3]),
 190                                vtn_ssa_value(b, w[4])->def, 0, 0);
 191       break;
 192
 193    case SpvOpGroupNonUniformAll:
 194    case SpvOpGroupNonUniformAny:
 195    case SpvOpGroupNonUniformAllEqual:
 196    case SpvOpSubgroupAllKHR:
 197    case SpvOpSubgroupAnyKHR:
 198    case SpvOpSubgroupAllEqualKHR: {
 199       vtn_fail_if(val->type->type != glsl_bool_type(),
 200                   "OpGroupNonUniform(All|Any|AllEqual) must return a bool");
 201       nir_intrinsic_op op;
 202       switch (opcode) {
 203       case SpvOpGroupNonUniformAll:
 204       case SpvOpSubgroupAllKHR:
 205          op = nir_intrinsic_vote_all;
 206          break;
 207       case SpvOpGroupNonUniformAny:
 208       case SpvOpSubgroupAnyKHR:
 209          op = nir_intrinsic_vote_any;
 210          break;
 211       case SpvOpGroupNonUniformAllEqual:
 212       case SpvOpSubgroupAllEqualKHR: {
 213          switch (glsl_get_base_type(val->type->type)) {
 214          case GLSL_TYPE_FLOAT:
 215          case GLSL_TYPE_DOUBLE:
 216             op = nir_intrinsic_vote_feq;
 217             break;
 218          case GLSL_TYPE_UINT:
 219          case GLSL_TYPE_INT:
 220          case GLSL_TYPE_UINT64:
 221          case GLSL_TYPE_INT64:
 222          case GLSL_TYPE_BOOL:
 223             op = nir_intrinsic_vote_ieq;
 224             break;
 225          default:
 226             unreachable("Unhandled type");
 227          }
 228          break;
 229       }
 230       default:
 231          unreachable("Unhandled opcode");
 232       }
 233
 234       nir_ssa_def *src0;
 235       if (opcode == SpvOpGroupNonUniformAll ||
 236           opcode == SpvOpGroupNonUniformAny ||
 237           opcode == SpvOpGroupNonUniformAllEqual) {
 238          src0 = vtn_ssa_value(b, w[4])->def;
 239       } else {
 240          src0 = vtn_ssa_value(b, w[3])->def;
 241       }
 242       nir_intrinsic_instr *intrin =
 243          nir_intrinsic_instr_create(b->nb.shader, op);
 244       intrin->num_components = src0->num_components;
 245       intrin->src[0] = nir_src_for_ssa(src0);
 246       nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
 247                                  val->type->type, NULL);
 248       nir_builder_instr_insert(&b->nb, &intrin->instr);
 249
 250       val->ssa->def = &intrin->dest.ssa;
 251       break;
 252    }
 253
 254    case SpvOpGroupNonUniformShuffle:
 255    case SpvOpGroupNonUniformShuffleXor:
 256    case SpvOpGroupNonUniformShuffleUp:
 257    case SpvOpGroupNonUniformShuffleDown: {
 258       nir_intrinsic_op op;
 259       switch (opcode) {
 260       case SpvOpGroupNonUniformShuffle:
 261          op = nir_intrinsic_shuffle;
 262          break;
 263       case SpvOpGroupNonUniformShuffleXor:
 264          op = nir_intrinsic_shuffle_xor;
 265          break;
 266       case SpvOpGroupNonUniformShuffleUp:
 267          op = nir_intrinsic_shuffle_up;
 268          break;
 269       case SpvOpGroupNonUniformShuffleDown:
 270          op = nir_intrinsic_shuffle_down;
 271          break;
 272       default:
 273          unreachable("Invalid opcode");
 274       }
 275       vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[4]),
 276                                vtn_ssa_value(b, w[5])->def, 0, 0);
 277       break;
 278    }
 279
 280    case SpvOpGroupNonUniformQuadBroadcast:
 281       vtn_build_subgroup_instr(b, nir_intrinsic_quad_broadcast,
 282                                val->ssa, vtn_ssa_value(b, w[4]),
 283                                vtn_ssa_value(b, w[5])->def, 0, 0);
 284       break;
 285
 286    case SpvOpGroupNonUniformQuadSwap: {
 287       unsigned direction = vtn_constant_uint(b, w[5]);
 288       nir_intrinsic_op op;
 289       switch (direction) {
 290       case 0:
 291          op = nir_intrinsic_quad_swap_horizontal;
 292          break;
 293       case 1:
 294          op = nir_intrinsic_quad_swap_vertical;
 295          break;
 296       case 2:
 297          op = nir_intrinsic_quad_swap_diagonal;
 298          break;
 299       default:
 300          vtn_fail("Invalid constant value in OpGroupNonUniformQuadSwap");
 301       }
 302       vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[4]),
 303                                NULL, 0, 0);
 304       break;
 305    }
 306
 307    case SpvOpGroupNonUniformIAdd:
 308    case SpvOpGroupNonUniformFAdd:
 309    case SpvOpGroupNonUniformIMul:
 310    case SpvOpGroupNonUniformFMul:
 311    case SpvOpGroupNonUniformSMin:
 312    case SpvOpGroupNonUniformUMin:
 313    case SpvOpGroupNonUniformFMin:
 314    case SpvOpGroupNonUniformSMax:
 315    case SpvOpGroupNonUniformUMax:
 316    case SpvOpGroupNonUniformFMax:
 317    case SpvOpGroupNonUniformBitwiseAnd:
 318    case SpvOpGroupNonUniformBitwiseOr:
 319    case SpvOpGroupNonUniformBitwiseXor:
 320    case SpvOpGroupNonUniformLogicalAnd:
 321    case SpvOpGroupNonUniformLogicalOr:
 322    case SpvOpGroupNonUniformLogicalXor: {
 323       nir_op reduction_op;
 324       switch (opcode) {
 325       case SpvOpGroupNonUniformIAdd:
 326          reduction_op = nir_op_iadd;
 327          break;
 328       case SpvOpGroupNonUniformFAdd:
 329          reduction_op = nir_op_fadd;
 330          break;
 331       case SpvOpGroupNonUniformIMul:
 332          reduction_op = nir_op_imul;
 333          break;
 334       case SpvOpGroupNonUniformFMul:
 335          reduction_op = nir_op_fmul;
 336          break;
 337       case SpvOpGroupNonUniformSMin:
 338          reduction_op = nir_op_imin;
 339          break;
 340       case SpvOpGroupNonUniformUMin:
 341          reduction_op = nir_op_umin;
 342          break;
 343       case SpvOpGroupNonUniformFMin:
 344          reduction_op = nir_op_fmin;
 345          break;
 346       case SpvOpGroupNonUniformSMax:
 347          reduction_op = nir_op_imax;
 348          break;
 349       case SpvOpGroupNonUniformUMax:
 350          reduction_op = nir_op_umax;
 351          break;
 352       case SpvOpGroupNonUniformFMax:
 353          reduction_op = nir_op_fmax;
 354          break;
 355       case SpvOpGroupNonUniformBitwiseAnd:
 356       case SpvOpGroupNonUniformLogicalAnd:
 357          reduction_op = nir_op_iand;
 358          break;
 359       case SpvOpGroupNonUniformBitwiseOr:
 360       case SpvOpGroupNonUniformLogicalOr:
 361          reduction_op = nir_op_ior;
 362          break;
 363       case SpvOpGroupNonUniformBitwiseXor:
 364       case SpvOpGroupNonUniformLogicalXor:
 365          reduction_op = nir_op_ixor;
 366          break;
 367       default:
 368          unreachable("Invalid reduction operation");
 369       }
 370
 371       nir_intrinsic_op op;
 372       unsigned cluster_size = 0;
 373       switch ((SpvGroupOperation)w[4]) {
 374       case SpvGroupOperationReduce:
 375          op = nir_intrinsic_reduce;
 376          break;
 377       case SpvGroupOperationInclusiveScan:
 378          op = nir_intrinsic_inclusive_scan;
 379          break;
 380       case SpvGroupOperationExclusiveScan:
 381          op = nir_intrinsic_exclusive_scan;
 382          break;
 383       case SpvGroupOperationClusteredReduce:
 384          op = nir_intrinsic_reduce;
 385          assert(count == 7);
 386          cluster_size = vtn_constant_uint(b, w[6]);
 387          break;
 388       default:
 389          unreachable("Invalid group operation");
 390       }
 391
 392       vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[5]),
 393                                NULL, reduction_op, cluster_size);
 394       break;
 395    }
 396
 397    default:
 398       unreachable("Invalid SPIR-V opcode");
 399    }
 400 }