src/compiler/nir/nir_lower_io.c

   1 /*
   2  * Copyright © 2014 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  * Authors:
  24  *    Connor Abbott (cwabbott0@gmail.com)
  25  *    Jason Ekstrand (jason@jlekstrand.net)
  26  *
  27  */
  28
  29 /*
  30  * This lowering pass converts references to input/output variables with
  31  * loads/stores to actual input/output intrinsics.
  32  */
  33
  34 #include "nir.h"
  35 #include "nir_builder.h"
  36
  37 struct lower_io_state {
  38    nir_builder builder;
  39    void *mem_ctx;
  40    int (*type_size)(const struct glsl_type *type);
  41    nir_variable_mode modes;
  42    nir_lower_io_options options;
  43 };
  44
  45 void
  46 nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
  47                          unsigned base_offset,
  48                          int (*type_size)(const struct glsl_type *))
  49 {
  50    unsigned location = 0;
  51
  52    /* There are 32 regular and 32 patch varyings allowed */
  53    int locations[64][2];
  54    for (unsigned i = 0; i < 64; i++) {
  55       for (unsigned j = 0; j < 2; j++)
  56          locations[i][j] = -1;
  57    }
  58
  59    nir_foreach_variable(var, var_list) {
  60       /*
  61        * UBO's have their own address spaces, so don't count them towards the
  62        * number of global uniforms
  63        */
  64       if ((var->data.mode == nir_var_uniform || var->data.mode == nir_var_shader_storage) &&
  65           var->interface_type != NULL)
  66          continue;
  67
  68       /* Make sure we give the same location to varyings packed with
  69        * ARB_enhanced_layouts.
  70        */
  71       int idx = var->data.location - base_offset;
  72       if (base_offset && idx >= 0) {
  73          assert(idx < ARRAY_SIZE(locations));
  74
  75          if (locations[idx][var->data.index] == -1) {
  76             var->data.driver_location = location;
  77             locations[idx][var->data.index] = location;
  78
  79             /* A dvec3 can be packed with a double we need special handling
  80              * for this as we are packing across two locations.
  81              */
  82             if (glsl_get_base_type(var->type) == GLSL_TYPE_DOUBLE &&
  83                 glsl_get_vector_elements(var->type) == 3) {
  84                /* Hack around type_size functions that expect vectors to be
  85                 * padded out to vec4. If a float type is the same size as a
  86                 * double then the type size is padded to vec4, otherwise
  87                 * set the offset to two doubles which offsets the location
  88                 * past the first two components in dvec3 which were stored at
  89                 * the previous location.
  90                 */
  91                unsigned dsize = type_size(glsl_double_type());
  92                unsigned offset =
  93                   dsize == type_size(glsl_float_type()) ? dsize : dsize * 2;
  94
  95                locations[idx + 1][var->data.index] = location + offset;
  96             }
  97
  98             location += type_size(var->type);
  99          } else {
 100             var->data.driver_location = locations[idx][var->data.index];
 101          }
 102       } else {
 103          var->data.driver_location = location;
 104          location += type_size(var->type);
 105       }
 106    }
 107
 108    *size = location;
 109 }
 110
 111 /**
 112  * Returns true if we're processing a stage whose inputs are arrays indexed
 113  * by a vertex number (such as geometry shader inputs).
 114  */
 115 static bool
 116 is_per_vertex_input(struct lower_io_state *state, nir_variable *var)
 117 {
 118    gl_shader_stage stage = state->builder.shader->stage;
 119
 120    return var->data.mode == nir_var_shader_in && !var->data.patch &&
 121           (stage == MESA_SHADER_TESS_CTRL ||
 122            stage == MESA_SHADER_TESS_EVAL ||
 123            stage == MESA_SHADER_GEOMETRY);
 124 }
 125
 126 static bool
 127 is_per_vertex_output(struct lower_io_state *state, nir_variable *var)
 128 {
 129    gl_shader_stage stage = state->builder.shader->stage;
 130    return var->data.mode == nir_var_shader_out && !var->data.patch &&
 131           stage == MESA_SHADER_TESS_CTRL;
 132 }
 133
 134 static nir_ssa_def *
 135 get_io_offset(nir_builder *b, nir_deref_var *deref,
 136               nir_ssa_def **vertex_index,
 137               int (*type_size)(const struct glsl_type *))
 138 {
 139    nir_deref *tail = &deref->deref;
 140
 141    /* For per-vertex input arrays (i.e. geometry shader inputs), keep the
 142     * outermost array index separate.  Process the rest normally.
 143     */
 144    if (vertex_index != NULL) {
 145       tail = tail->child;
 146       assert(tail->deref_type == nir_deref_type_array);
 147       nir_deref_array *deref_array = nir_deref_as_array(tail);
 148
 149       nir_ssa_def *vtx = nir_imm_int(b, deref_array->base_offset);
 150       if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
 151          vtx = nir_iadd(b, vtx, nir_ssa_for_src(b, deref_array->indirect, 1));
 152       }
 153       *vertex_index = vtx;
 154    }
 155
 156    /* Just emit code and let constant-folding go to town */
 157    nir_ssa_def *offset = nir_imm_int(b, 0);
 158
 159    while (tail->child != NULL) {
 160       const struct glsl_type *parent_type = tail->type;
 161       tail = tail->child;
 162
 163       if (tail->deref_type == nir_deref_type_array) {
 164          nir_deref_array *deref_array = nir_deref_as_array(tail);
 165          unsigned size = type_size(tail->type);
 166
 167          offset = nir_iadd(b, offset,
 168                            nir_imm_int(b, size * deref_array->base_offset));
 169
 170          if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
 171             nir_ssa_def *mul =
 172                nir_imul(b, nir_imm_int(b, size),
 173                         nir_ssa_for_src(b, deref_array->indirect, 1));
 174
 175             offset = nir_iadd(b, offset, mul);
 176          }
 177       } else if (tail->deref_type == nir_deref_type_struct) {
 178          nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
 179
 180          unsigned field_offset = 0;
 181          for (unsigned i = 0; i < deref_struct->index; i++) {
 182             field_offset += type_size(glsl_get_struct_field(parent_type, i));
 183          }
 184          offset = nir_iadd(b, offset, nir_imm_int(b, field_offset));
 185       }
 186    }
 187
 188    return offset;
 189 }
 190
 191 static nir_intrinsic_instr *
 192 lower_load(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 193            nir_ssa_def *vertex_index, nir_ssa_def *offset)
 194 {
 195    const nir_shader *nir = state->builder.shader;
 196    nir_variable *var = intrin->variables[0]->var;
 197    nir_variable_mode mode = var->data.mode;
 198    nir_ssa_def *barycentric = NULL;
 199
 200    nir_intrinsic_op op;
 201    switch (mode) {
 202    case nir_var_shader_in:
 203       if (nir->stage == MESA_SHADER_FRAGMENT &&
 204           nir->options->use_interpolated_input_intrinsics &&
 205           var->data.interpolation != INTERP_MODE_FLAT) {
 206          assert(vertex_index == NULL);
 207
 208          nir_intrinsic_op bary_op;
 209          if (var->data.sample ||
 210              (state->options & nir_lower_io_force_sample_interpolation))
 211             bary_op = nir_intrinsic_load_barycentric_sample;
 212          else if (var->data.centroid)
 213             bary_op = nir_intrinsic_load_barycentric_centroid;
 214          else
 215             bary_op = nir_intrinsic_load_barycentric_pixel;
 216
 217          barycentric = nir_load_barycentric(&state->builder, bary_op,
 218                                             var->data.interpolation);
 219          op = nir_intrinsic_load_interpolated_input;
 220       } else {
 221          op = vertex_index ? nir_intrinsic_load_per_vertex_input :
 222                              nir_intrinsic_load_input;
 223       }
 224       break;
 225    case nir_var_shader_out:
 226       op = vertex_index ? nir_intrinsic_load_per_vertex_output :
 227                           nir_intrinsic_load_output;
 228       break;
 229    case nir_var_uniform:
 230       op = nir_intrinsic_load_uniform;
 231       break;
 232    case nir_var_shared:
 233       op = nir_intrinsic_load_shared;
 234       break;
 235    default:
 236       unreachable("Unknown variable mode");
 237    }
 238
 239    nir_intrinsic_instr *load = nir_intrinsic_instr_create(state->mem_ctx, op);
 240    load->num_components = intrin->num_components;
 241
 242    nir_intrinsic_set_base(load, var->data.driver_location);
 243    if (mode == nir_var_shader_in || mode == nir_var_shader_out)
 244       nir_intrinsic_set_component(load, var->data.location_frac);
 245
 246    if (load->intrinsic == nir_intrinsic_load_uniform)
 247       nir_intrinsic_set_range(load, state->type_size(var->type));
 248
 249    if (vertex_index) {
 250       load->src[0] = nir_src_for_ssa(vertex_index);
 251       load->src[1] = nir_src_for_ssa(offset);
 252    } else if (barycentric) {
 253       load->src[0] = nir_src_for_ssa(barycentric);
 254       load->src[1] = nir_src_for_ssa(offset);
 255    } else {
 256       load->src[0] = nir_src_for_ssa(offset);
 257    }
 258
 259    return load;
 260 }
 261
 262 static nir_intrinsic_instr *
 263 lower_store(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 264             nir_ssa_def *vertex_index, nir_ssa_def *offset)
 265 {
 266    nir_variable *var = intrin->variables[0]->var;
 267    nir_variable_mode mode = var->data.mode;
 268
 269    nir_intrinsic_op op;
 270    if (mode == nir_var_shared) {
 271       op = nir_intrinsic_store_shared;
 272    } else {
 273       assert(mode == nir_var_shader_out);
 274       op = vertex_index ? nir_intrinsic_store_per_vertex_output :
 275                           nir_intrinsic_store_output;
 276    }
 277
 278    nir_intrinsic_instr *store = nir_intrinsic_instr_create(state->mem_ctx, op);
 279    store->num_components = intrin->num_components;
 280
 281    nir_src_copy(&store->src[0], &intrin->src[0], store);
 282
 283    nir_intrinsic_set_base(store, var->data.driver_location);
 284
 285    if (mode == nir_var_shader_out)
 286       nir_intrinsic_set_component(store, var->data.location_frac);
 287
 288    nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
 289
 290    if (vertex_index)
 291       store->src[1] = nir_src_for_ssa(vertex_index);
 292
 293    store->src[vertex_index ? 2 : 1] = nir_src_for_ssa(offset);
 294
 295    return store;
 296 }
 297
 298 static nir_intrinsic_instr *
 299 lower_atomic(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 300              nir_ssa_def *offset)
 301 {
 302    nir_variable *var = intrin->variables[0]->var;
 303
 304    assert(var->data.mode == nir_var_shared);
 305
 306    nir_intrinsic_op op;
 307    switch (intrin->intrinsic) {
 308 #define OP(O) case nir_intrinsic_var_##O: op = nir_intrinsic_shared_##O; break;
 309    OP(atomic_exchange)
 310    OP(atomic_comp_swap)
 311    OP(atomic_add)
 312    OP(atomic_imin)
 313    OP(atomic_umin)
 314    OP(atomic_imax)
 315    OP(atomic_umax)
 316    OP(atomic_and)
 317    OP(atomic_or)
 318    OP(atomic_xor)
 319 #undef OP
 320    default:
 321       unreachable("Invalid atomic");
 322    }
 323
 324    nir_intrinsic_instr *atomic =
 325       nir_intrinsic_instr_create(state->mem_ctx, op);
 326
 327    nir_intrinsic_set_base(atomic, var->data.driver_location);
 328
 329    atomic->src[0] = nir_src_for_ssa(offset);
 330    for (unsigned i = 0; i < nir_op_infos[intrin->intrinsic].num_inputs; i++) {
 331       nir_src_copy(&atomic->src[i+1], &intrin->src[i], atomic);
 332    }
 333
 334    return atomic;
 335 }
 336
 337 static nir_intrinsic_instr *
 338 lower_interpolate_at(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 339                      nir_ssa_def *offset)
 340 {
 341    nir_variable *var = intrin->variables[0]->var;
 342
 343    assert(var->data.mode == nir_var_shader_in);
 344
 345    /* Ignore interpolateAt() for flat variables - flat is flat. */
 346    if (var->data.interpolation == INTERP_MODE_FLAT)
 347       return lower_load(intrin, state, NULL, offset);
 348
 349    nir_intrinsic_op bary_op;
 350    switch (intrin->intrinsic) {
 351    case nir_intrinsic_interp_var_at_centroid:
 352       bary_op = (state->options & nir_lower_io_force_sample_interpolation) ?
 353                 nir_intrinsic_load_barycentric_sample :
 354                 nir_intrinsic_load_barycentric_centroid;
 355       break;
 356    case nir_intrinsic_interp_var_at_sample:
 357       bary_op = nir_intrinsic_load_barycentric_at_sample;
 358       break;
 359    case nir_intrinsic_interp_var_at_offset:
 360       bary_op = nir_intrinsic_load_barycentric_at_offset;
 361       break;
 362    default:
 363       unreachable("Bogus interpolateAt() intrinsic.");
 364    }
 365
 366    nir_intrinsic_instr *bary_setup =
 367       nir_intrinsic_instr_create(state->mem_ctx, bary_op);
 368
 369    nir_ssa_dest_init(&bary_setup->instr, &bary_setup->dest, 2, 32, NULL);
 370    nir_intrinsic_set_interp_mode(bary_setup, var->data.interpolation);
 371
 372    if (intrin->intrinsic != nir_intrinsic_interp_var_at_centroid)
 373       nir_src_copy(&bary_setup->src[0], &intrin->src[0], bary_setup);
 374
 375    nir_builder_instr_insert(&state->builder, &bary_setup->instr);
 376
 377    nir_intrinsic_instr *load =
 378       nir_intrinsic_instr_create(state->mem_ctx,
 379                                  nir_intrinsic_load_interpolated_input);
 380    load->num_components = intrin->num_components;
 381
 382    nir_intrinsic_set_base(load, var->data.driver_location);
 383    nir_intrinsic_set_component(load, var->data.location_frac);
 384
 385    load->src[0] = nir_src_for_ssa(&bary_setup->dest.ssa);
 386    load->src[1] = nir_src_for_ssa(offset);
 387
 388    return load;
 389 }
 390
 391 static bool
 392 nir_lower_io_block(nir_block *block,
 393                    struct lower_io_state *state)
 394 {
 395    nir_builder *b = &state->builder;
 396    const nir_shader_compiler_options *options = b->shader->options;
 397
 398    nir_foreach_instr_safe(instr, block) {
 399       if (instr->type != nir_instr_type_intrinsic)
 400          continue;
 401
 402       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 403
 404       switch (intrin->intrinsic) {
 405       case nir_intrinsic_load_var:
 406       case nir_intrinsic_store_var:
 407       case nir_intrinsic_var_atomic_add:
 408       case nir_intrinsic_var_atomic_imin:
 409       case nir_intrinsic_var_atomic_umin:
 410       case nir_intrinsic_var_atomic_imax:
 411       case nir_intrinsic_var_atomic_umax:
 412       case nir_intrinsic_var_atomic_and:
 413       case nir_intrinsic_var_atomic_or:
 414       case nir_intrinsic_var_atomic_xor:
 415       case nir_intrinsic_var_atomic_exchange:
 416       case nir_intrinsic_var_atomic_comp_swap:
 417          /* We can lower the io for this nir instrinsic */
 418          break;
 419       case nir_intrinsic_interp_var_at_centroid:
 420       case nir_intrinsic_interp_var_at_sample:
 421       case nir_intrinsic_interp_var_at_offset:
 422          /* We can optionally lower these to load_interpolated_input */
 423          if (options->use_interpolated_input_intrinsics)
 424             break;
 425       default:
 426          /* We can't lower the io for this nir instrinsic, so skip it */
 427          continue;
 428       }
 429
 430       nir_variable *var = intrin->variables[0]->var;
 431       nir_variable_mode mode = var->data.mode;
 432
 433       if ((state->modes & mode) == 0)
 434          continue;
 435
 436       if (mode != nir_var_shader_in &&
 437           mode != nir_var_shader_out &&
 438           mode != nir_var_shared &&
 439           mode != nir_var_uniform)
 440          continue;
 441
 442       b->cursor = nir_before_instr(instr);
 443
 444       const bool per_vertex =
 445          is_per_vertex_input(state, var) || is_per_vertex_output(state, var);
 446
 447       nir_ssa_def *offset;
 448       nir_ssa_def *vertex_index = NULL;
 449
 450       offset = get_io_offset(b, intrin->variables[0],
 451                              per_vertex ? &vertex_index : NULL,
 452                              state->type_size);
 453
 454       nir_intrinsic_instr *replacement;
 455
 456       switch (intrin->intrinsic) {
 457       case nir_intrinsic_load_var:
 458          replacement = lower_load(intrin, state, vertex_index, offset);
 459          break;
 460
 461       case nir_intrinsic_store_var:
 462          replacement = lower_store(intrin, state, vertex_index, offset);
 463          break;
 464
 465       case nir_intrinsic_var_atomic_add:
 466       case nir_intrinsic_var_atomic_imin:
 467       case nir_intrinsic_var_atomic_umin:
 468       case nir_intrinsic_var_atomic_imax:
 469       case nir_intrinsic_var_atomic_umax:
 470       case nir_intrinsic_var_atomic_and:
 471       case nir_intrinsic_var_atomic_or:
 472       case nir_intrinsic_var_atomic_xor:
 473       case nir_intrinsic_var_atomic_exchange:
 474       case nir_intrinsic_var_atomic_comp_swap:
 475          assert(vertex_index == NULL);
 476          replacement = lower_atomic(intrin, state, offset);
 477          break;
 478
 479       case nir_intrinsic_interp_var_at_centroid:
 480       case nir_intrinsic_interp_var_at_sample:
 481       case nir_intrinsic_interp_var_at_offset:
 482          assert(vertex_index == NULL);
 483          replacement = lower_interpolate_at(intrin, state, offset);
 484          break;
 485
 486       default:
 487          continue;
 488       }
 489
 490       if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
 491          if (intrin->dest.is_ssa) {
 492             nir_ssa_dest_init(&replacement->instr, &replacement->dest,
 493                               intrin->dest.ssa.num_components,
 494                               intrin->dest.ssa.bit_size, NULL);
 495             nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
 496                                      nir_src_for_ssa(&replacement->dest.ssa));
 497          } else {
 498             nir_dest_copy(&replacement->dest, &intrin->dest, state->mem_ctx);
 499          }
 500       }
 501
 502       nir_instr_insert_before(&intrin->instr, &replacement->instr);
 503       nir_instr_remove(&intrin->instr);
 504    }
 505
 506    return true;
 507 }
 508
 509 static void
 510 nir_lower_io_impl(nir_function_impl *impl,
 511                   nir_variable_mode modes,
 512                   int (*type_size)(const struct glsl_type *),
 513                   nir_lower_io_options options)
 514 {
 515    struct lower_io_state state;
 516
 517    nir_builder_init(&state.builder, impl);
 518    state.mem_ctx = ralloc_parent(impl);
 519    state.modes = modes;
 520    state.type_size = type_size;
 521    state.options = options;
 522
 523    nir_foreach_block(block, impl) {
 524       nir_lower_io_block(block, &state);
 525    }
 526
 527    nir_metadata_preserve(impl, nir_metadata_block_index |
 528                                nir_metadata_dominance);
 529 }
 530
 531 void
 532 nir_lower_io(nir_shader *shader, nir_variable_mode modes,
 533              int (*type_size)(const struct glsl_type *),
 534              nir_lower_io_options options)
 535 {
 536    nir_foreach_function(function, shader) {
 537       if (function->impl) {
 538          nir_lower_io_impl(function->impl, modes, type_size, options);
 539       }
 540    }
 541 }
 542
 543 /**
 544  * Return the offset soruce for a load/store intrinsic.
 545  */
 546 nir_src *
 547 nir_get_io_offset_src(nir_intrinsic_instr *instr)
 548 {
 549    switch (instr->intrinsic) {
 550    case nir_intrinsic_load_input:
 551    case nir_intrinsic_load_output:
 552    case nir_intrinsic_load_uniform:
 553       return &instr->src[0];
 554    case nir_intrinsic_load_ubo:
 555    case nir_intrinsic_load_ssbo:
 556    case nir_intrinsic_load_per_vertex_input:
 557    case nir_intrinsic_load_per_vertex_output:
 558    case nir_intrinsic_load_interpolated_input:
 559    case nir_intrinsic_store_output:
 560       return &instr->src[1];
 561    case nir_intrinsic_store_ssbo:
 562    case nir_intrinsic_store_per_vertex_output:
 563       return &instr->src[2];
 564    default:
 565       return NULL;
 566    }
 567 }
 568
 569 /**
 570  * Return the vertex index source for a load/store per_vertex intrinsic.
 571  */
 572 nir_src *
 573 nir_get_io_vertex_index_src(nir_intrinsic_instr *instr)
 574 {
 575    switch (instr->intrinsic) {
 576    case nir_intrinsic_load_per_vertex_input:
 577    case nir_intrinsic_load_per_vertex_output:
 578       return &instr->src[0];
 579    case nir_intrinsic_store_per_vertex_output:
 580       return &instr->src[1];
 581    default:
 582       return NULL;
 583    }
 584 }