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:
 187    case SpvOpGroupBroadcast: ++w;
 188    case SpvOpSubgroupReadInvocationKHR:
 189       vtn_build_subgroup_instr(b, nir_intrinsic_read_invocation,
 190                                val->ssa, vtn_ssa_value(b, w[3]),
 191                                vtn_ssa_value(b, w[4])->def, 0, 0);
 192       break;
 193
 194    case SpvOpGroupNonUniformAll:
 195    case SpvOpGroupNonUniformAny:
 196    case SpvOpGroupNonUniformAllEqual:
 197    case SpvOpGroupAll:
 198    case SpvOpGroupAny:
 199    case SpvOpSubgroupAllKHR:
 200    case SpvOpSubgroupAnyKHR:
 201    case SpvOpSubgroupAllEqualKHR: {
 202       vtn_fail_if(val->type->type != glsl_bool_type(),
 203                   "OpGroupNonUniform(All|Any|AllEqual) must return a bool");
 204       nir_intrinsic_op op;
 205       switch (opcode) {
 206       case SpvOpGroupNonUniformAll:
 207       case SpvOpGroupAll:
 208       case SpvOpSubgroupAllKHR:
 209          op = nir_intrinsic_vote_all;
 210          break;
 211       case SpvOpGroupNonUniformAny:
 212       case SpvOpGroupAny:
 213       case SpvOpSubgroupAnyKHR:
 214          op = nir_intrinsic_vote_any;
 215          break;
 216       case SpvOpGroupNonUniformAllEqual:
 217       case SpvOpSubgroupAllEqualKHR: {
 218          switch (glsl_get_base_type(val->type->type)) {
 219          case GLSL_TYPE_FLOAT:
 220          case GLSL_TYPE_DOUBLE:
 221             op = nir_intrinsic_vote_feq;
 222             break;
 223          case GLSL_TYPE_UINT:
 224          case GLSL_TYPE_INT:
 225          case GLSL_TYPE_UINT64:
 226          case GLSL_TYPE_INT64:
 227          case GLSL_TYPE_BOOL:
 228             op = nir_intrinsic_vote_ieq;
 229             break;
 230          default:
 231             unreachable("Unhandled type");
 232          }
 233          break;
 234       }
 235       default:
 236          unreachable("Unhandled opcode");
 237       }
 238
 239       nir_ssa_def *src0;
 240       if (opcode == SpvOpGroupNonUniformAll || opcode == SpvOpGroupAll ||
 241           opcode == SpvOpGroupNonUniformAny || opcode == SpvOpGroupAny ||
 242           opcode == SpvOpGroupNonUniformAllEqual) {
 243          src0 = vtn_ssa_value(b, w[4])->def;
 244       } else {
 245          src0 = vtn_ssa_value(b, w[3])->def;
 246       }
 247       nir_intrinsic_instr *intrin =
 248          nir_intrinsic_instr_create(b->nb.shader, op);
 249       intrin->num_components = src0->num_components;
 250       intrin->src[0] = nir_src_for_ssa(src0);
 251       nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
 252                                  val->type->type, NULL);
 253       nir_builder_instr_insert(&b->nb, &intrin->instr);
 254
 255       val->ssa->def = &intrin->dest.ssa;
 256       break;
 257    }
 258
 259    case SpvOpGroupNonUniformShuffle:
 260    case SpvOpGroupNonUniformShuffleXor:
 261    case SpvOpGroupNonUniformShuffleUp:
 262    case SpvOpGroupNonUniformShuffleDown: {
 263       nir_intrinsic_op op;
 264       switch (opcode) {
 265       case SpvOpGroupNonUniformShuffle:
 266          op = nir_intrinsic_shuffle;
 267          break;
 268       case SpvOpGroupNonUniformShuffleXor:
 269          op = nir_intrinsic_shuffle_xor;
 270          break;
 271       case SpvOpGroupNonUniformShuffleUp:
 272          op = nir_intrinsic_shuffle_up;
 273          break;
 274       case SpvOpGroupNonUniformShuffleDown:
 275          op = nir_intrinsic_shuffle_down;
 276          break;
 277       default:
 278          unreachable("Invalid opcode");
 279       }
 280       vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[4]),
 281                                vtn_ssa_value(b, w[5])->def, 0, 0);
 282       break;
 283    }
 284
 285    case SpvOpGroupNonUniformQuadBroadcast:
 286       vtn_build_subgroup_instr(b, nir_intrinsic_quad_broadcast,
 287                                val->ssa, vtn_ssa_value(b, w[4]),
 288                                vtn_ssa_value(b, w[5])->def, 0, 0);
 289       break;
 290
 291    case SpvOpGroupNonUniformQuadSwap: {
 292       unsigned direction = vtn_constant_uint(b, w[5]);
 293       nir_intrinsic_op op;
 294       switch (direction) {
 295       case 0:
 296          op = nir_intrinsic_quad_swap_horizontal;
 297          break;
 298       case 1:
 299          op = nir_intrinsic_quad_swap_vertical;
 300          break;
 301       case 2:
 302          op = nir_intrinsic_quad_swap_diagonal;
 303          break;
 304       default:
 305          vtn_fail("Invalid constant value in OpGroupNonUniformQuadSwap");
 306       }
 307       vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[4]),
 308                                NULL, 0, 0);
 309       break;
 310    }
 311
 312    case SpvOpGroupNonUniformIAdd:
 313    case SpvOpGroupNonUniformFAdd:
 314    case SpvOpGroupNonUniformIMul:
 315    case SpvOpGroupNonUniformFMul:
 316    case SpvOpGroupNonUniformSMin:
 317    case SpvOpGroupNonUniformUMin:
 318    case SpvOpGroupNonUniformFMin:
 319    case SpvOpGroupNonUniformSMax:
 320    case SpvOpGroupNonUniformUMax:
 321    case SpvOpGroupNonUniformFMax:
 322    case SpvOpGroupNonUniformBitwiseAnd:
 323    case SpvOpGroupNonUniformBitwiseOr:
 324    case SpvOpGroupNonUniformBitwiseXor:
 325    case SpvOpGroupNonUniformLogicalAnd:
 326    case SpvOpGroupNonUniformLogicalOr:
 327    case SpvOpGroupNonUniformLogicalXor:
 328    case SpvOpGroupIAdd:
 329    case SpvOpGroupFAdd:
 330    case SpvOpGroupFMin:
 331    case SpvOpGroupUMin:
 332    case SpvOpGroupSMin:
 333    case SpvOpGroupFMax:
 334    case SpvOpGroupUMax:
 335    case SpvOpGroupSMax:
 336    case SpvOpGroupIAddNonUniformAMD:
 337    case SpvOpGroupFAddNonUniformAMD:
 338    case SpvOpGroupFMinNonUniformAMD:
 339    case SpvOpGroupUMinNonUniformAMD:
 340    case SpvOpGroupSMinNonUniformAMD:
 341    case SpvOpGroupFMaxNonUniformAMD:
 342    case SpvOpGroupUMaxNonUniformAMD:
 343    case SpvOpGroupSMaxNonUniformAMD: {
 344       nir_op reduction_op;
 345       switch (opcode) {
 346       case SpvOpGroupNonUniformIAdd:
 347       case SpvOpGroupIAdd:
 348       case SpvOpGroupIAddNonUniformAMD:
 349          reduction_op = nir_op_iadd;
 350          break;
 351       case SpvOpGroupNonUniformFAdd:
 352       case SpvOpGroupFAdd:
 353       case SpvOpGroupFAddNonUniformAMD:
 354          reduction_op = nir_op_fadd;
 355          break;
 356       case SpvOpGroupNonUniformIMul:
 357          reduction_op = nir_op_imul;
 358          break;
 359       case SpvOpGroupNonUniformFMul:
 360          reduction_op = nir_op_fmul;
 361          break;
 362       case SpvOpGroupNonUniformSMin:
 363       case SpvOpGroupSMin:
 364       case SpvOpGroupSMinNonUniformAMD:
 365          reduction_op = nir_op_imin;
 366          break;
 367       case SpvOpGroupNonUniformUMin:
 368       case SpvOpGroupUMin:
 369       case SpvOpGroupUMinNonUniformAMD:
 370          reduction_op = nir_op_umin;
 371          break;
 372       case SpvOpGroupNonUniformFMin:
 373       case SpvOpGroupFMin:
 374       case SpvOpGroupFMinNonUniformAMD:
 375          reduction_op = nir_op_fmin;
 376          break;
 377       case SpvOpGroupNonUniformSMax:
 378       case SpvOpGroupSMax:
 379       case SpvOpGroupSMaxNonUniformAMD:
 380          reduction_op = nir_op_imax;
 381          break;
 382       case SpvOpGroupNonUniformUMax:
 383       case SpvOpGroupUMax:
 384       case SpvOpGroupUMaxNonUniformAMD:
 385          reduction_op = nir_op_umax;
 386          break;
 387       case SpvOpGroupNonUniformFMax:
 388       case SpvOpGroupFMax:
 389       case SpvOpGroupFMaxNonUniformAMD:
 390          reduction_op = nir_op_fmax;
 391          break;
 392       case SpvOpGroupNonUniformBitwiseAnd:
 393       case SpvOpGroupNonUniformLogicalAnd:
 394          reduction_op = nir_op_iand;
 395          break;
 396       case SpvOpGroupNonUniformBitwiseOr:
 397       case SpvOpGroupNonUniformLogicalOr:
 398          reduction_op = nir_op_ior;
 399          break;
 400       case SpvOpGroupNonUniformBitwiseXor:
 401       case SpvOpGroupNonUniformLogicalXor:
 402          reduction_op = nir_op_ixor;
 403          break;
 404       default:
 405          unreachable("Invalid reduction operation");
 406       }
 407
 408       nir_intrinsic_op op;
 409       unsigned cluster_size = 0;
 410       switch ((SpvGroupOperation)w[4]) {
 411       case SpvGroupOperationReduce:
 412          op = nir_intrinsic_reduce;
 413          break;
 414       case SpvGroupOperationInclusiveScan:
 415          op = nir_intrinsic_inclusive_scan;
 416          break;
 417       case SpvGroupOperationExclusiveScan:
 418          op = nir_intrinsic_exclusive_scan;
 419          break;
 420       case SpvGroupOperationClusteredReduce:
 421          op = nir_intrinsic_reduce;
 422          assert(count == 7);
 423          cluster_size = vtn_constant_uint(b, w[6]);
 424          break;
 425       default:
 426          unreachable("Invalid group operation");
 427       }
 428
 429       vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[5]),
 430                                NULL, reduction_op, cluster_size);
 431       break;
 432    }
 433
 434    default:
 435       unreachable("Invalid SPIR-V opcode");
 436    }
 437 }