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    void *dead_ctx;
  39    nir_builder builder;
  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                          int (*type_size)(const struct glsl_type *))
  48 {
  49    unsigned location = 0;
  50
  51    nir_foreach_variable(var, var_list) {
  52       /*
  53        * UBOs have their own address spaces, so don't count them towards the
  54        * number of global uniforms
  55        */
  56       if ((var->data.mode == nir_var_uniform || var->data.mode == nir_var_shader_storage) &&
  57           var->interface_type != NULL)
  58          continue;
  59
  60       var->data.driver_location = location;
  61       location += type_size(var->type);
  62    }
  63
  64    *size = location;
  65 }
  66
  67 /**
  68  * Return true if the given variable is a per-vertex input/output array.
  69  * (such as geometry shader inputs).
  70  */
  71 bool
  72 nir_is_per_vertex_io(const nir_variable *var, gl_shader_stage stage)
  73 {
  74    if (var->data.patch || !glsl_type_is_array(var->type))
  75       return false;
  76
  77    if (var->data.mode == nir_var_shader_in)
  78       return stage == MESA_SHADER_GEOMETRY ||
  79              stage == MESA_SHADER_TESS_CTRL ||
  80              stage == MESA_SHADER_TESS_EVAL;
  81
  82    if (var->data.mode == nir_var_shader_out)
  83       return stage == MESA_SHADER_TESS_CTRL;
  84
  85    return false;
  86 }
  87
  88 static nir_ssa_def *
  89 get_io_offset(nir_builder *b, nir_deref_var *deref,
  90               nir_ssa_def **vertex_index,
  91               int (*type_size)(const struct glsl_type *),
  92               unsigned *component)
  93 {
  94    nir_deref *tail = &deref->deref;
  95
  96    /* For per-vertex input arrays (i.e. geometry shader inputs), keep the
  97     * outermost array index separate.  Process the rest normally.
  98     */
  99    if (vertex_index != NULL) {
 100       tail = tail->child;
 101       nir_deref_array *deref_array = nir_deref_as_array(tail);
 102
 103       nir_ssa_def *vtx = nir_imm_int(b, deref_array->base_offset);
 104       if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
 105          vtx = nir_iadd(b, vtx, nir_ssa_for_src(b, deref_array->indirect, 1));
 106       }
 107       *vertex_index = vtx;
 108    }
 109
 110    if (deref->var->data.compact) {
 111       assert(tail->child->deref_type == nir_deref_type_array);
 112       assert(glsl_type_is_scalar(glsl_without_array(deref->var->type)));
 113       nir_deref_array *deref_array = nir_deref_as_array(tail->child);
 114       /* We always lower indirect dereferences for "compact" array vars. */
 115       assert(deref_array->deref_array_type == nir_deref_array_type_direct);
 116
 117       const unsigned total_offset = *component + deref_array->base_offset;
 118       const unsigned slot_offset = total_offset / 4;
 119       *component = total_offset % 4;
 120       return nir_imm_int(b, type_size(glsl_vec4_type()) * slot_offset);
 121    }
 122
 123    /* Just emit code and let constant-folding go to town */
 124    nir_ssa_def *offset = nir_imm_int(b, 0);
 125
 126    while (tail->child != NULL) {
 127       const struct glsl_type *parent_type = tail->type;
 128       tail = tail->child;
 129
 130       if (tail->deref_type == nir_deref_type_array) {
 131          nir_deref_array *deref_array = nir_deref_as_array(tail);
 132          unsigned size = type_size(tail->type);
 133
 134          offset = nir_iadd(b, offset,
 135                            nir_imm_int(b, size * deref_array->base_offset));
 136
 137          if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
 138             nir_ssa_def *mul =
 139                nir_imul(b, nir_imm_int(b, size),
 140                         nir_ssa_for_src(b, deref_array->indirect, 1));
 141
 142             offset = nir_iadd(b, offset, mul);
 143          }
 144       } else if (tail->deref_type == nir_deref_type_struct) {
 145          nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
 146
 147          unsigned field_offset = 0;
 148          for (unsigned i = 0; i < deref_struct->index; i++) {
 149             field_offset += type_size(glsl_get_struct_field(parent_type, i));
 150          }
 151          offset = nir_iadd(b, offset, nir_imm_int(b, field_offset));
 152       }
 153    }
 154
 155    return offset;
 156 }
 157
 158 static nir_intrinsic_instr *
 159 lower_load(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 160            nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
 161            unsigned component)
 162 {
 163    const nir_shader *nir = state->builder.shader;
 164    nir_variable_mode mode = var->data.mode;
 165    nir_ssa_def *barycentric = NULL;
 166
 167    nir_intrinsic_op op;
 168    switch (mode) {
 169    case nir_var_shader_in:
 170       if (nir->info.stage == MESA_SHADER_FRAGMENT &&
 171           nir->options->use_interpolated_input_intrinsics &&
 172           var->data.interpolation != INTERP_MODE_FLAT) {
 173          assert(vertex_index == NULL);
 174
 175          nir_intrinsic_op bary_op;
 176          if (var->data.sample ||
 177              (state->options & nir_lower_io_force_sample_interpolation))
 178             bary_op = nir_intrinsic_load_barycentric_sample;
 179          else if (var->data.centroid)
 180             bary_op = nir_intrinsic_load_barycentric_centroid;
 181          else
 182             bary_op = nir_intrinsic_load_barycentric_pixel;
 183
 184          barycentric = nir_load_barycentric(&state->builder, bary_op,
 185                                             var->data.interpolation);
 186          op = nir_intrinsic_load_interpolated_input;
 187       } else {
 188          op = vertex_index ? nir_intrinsic_load_per_vertex_input :
 189                              nir_intrinsic_load_input;
 190       }
 191       break;
 192    case nir_var_shader_out:
 193       op = vertex_index ? nir_intrinsic_load_per_vertex_output :
 194                           nir_intrinsic_load_output;
 195       break;
 196    case nir_var_uniform:
 197       op = nir_intrinsic_load_uniform;
 198       break;
 199    case nir_var_shared:
 200       op = nir_intrinsic_load_shared;
 201       break;
 202    default:
 203       unreachable("Unknown variable mode");
 204    }
 205
 206    nir_intrinsic_instr *load =
 207       nir_intrinsic_instr_create(state->builder.shader, op);
 208    load->num_components = intrin->num_components;
 209
 210    nir_intrinsic_set_base(load, var->data.driver_location);
 211    if (mode == nir_var_shader_in || mode == nir_var_shader_out)
 212       nir_intrinsic_set_component(load, component);
 213
 214    if (load->intrinsic == nir_intrinsic_load_uniform)
 215       nir_intrinsic_set_range(load, state->type_size(var->type));
 216
 217    if (vertex_index) {
 218       load->src[0] = nir_src_for_ssa(vertex_index);
 219       load->src[1] = nir_src_for_ssa(offset);
 220    } else if (barycentric) {
 221       load->src[0] = nir_src_for_ssa(barycentric);
 222       load->src[1] = nir_src_for_ssa(offset);
 223    } else {
 224       load->src[0] = nir_src_for_ssa(offset);
 225    }
 226
 227    return load;
 228 }
 229
 230 static nir_intrinsic_instr *
 231 lower_store(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 232             nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
 233             unsigned component)
 234 {
 235    nir_variable_mode mode = var->data.mode;
 236
 237    nir_intrinsic_op op;
 238    if (mode == nir_var_shared) {
 239       op = nir_intrinsic_store_shared;
 240    } else {
 241       assert(mode == nir_var_shader_out);
 242       op = vertex_index ? nir_intrinsic_store_per_vertex_output :
 243                           nir_intrinsic_store_output;
 244    }
 245
 246    nir_intrinsic_instr *store =
 247       nir_intrinsic_instr_create(state->builder.shader, op);
 248    store->num_components = intrin->num_components;
 249
 250    if (intrin->intrinsic == nir_intrinsic_store_var)
 251       nir_src_copy(&store->src[0], &intrin->src[0], store);
 252    else
 253       nir_src_copy(&store->src[0], &intrin->src[1], store);
 254
 255    nir_intrinsic_set_base(store, var->data.driver_location);
 256
 257    if (mode == nir_var_shader_out)
 258       nir_intrinsic_set_component(store, component);
 259
 260    nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
 261
 262    if (vertex_index)
 263       store->src[1] = nir_src_for_ssa(vertex_index);
 264
 265    store->src[vertex_index ? 2 : 1] = nir_src_for_ssa(offset);
 266
 267    return store;
 268 }
 269
 270 static nir_intrinsic_instr *
 271 lower_atomic(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 272              nir_variable *var, nir_ssa_def *offset)
 273 {
 274    assert(var->data.mode == nir_var_shared);
 275
 276    nir_intrinsic_op op;
 277    switch (intrin->intrinsic) {
 278 #define OP(O) case nir_intrinsic_deref_##O: op = nir_intrinsic_shared_##O; break;
 279    OP(atomic_exchange)
 280    OP(atomic_comp_swap)
 281    OP(atomic_add)
 282    OP(atomic_imin)
 283    OP(atomic_umin)
 284    OP(atomic_imax)
 285    OP(atomic_umax)
 286    OP(atomic_and)
 287    OP(atomic_or)
 288    OP(atomic_xor)
 289 #undef OP
 290    default:
 291       unreachable("Invalid atomic");
 292    }
 293
 294    nir_intrinsic_instr *atomic =
 295       nir_intrinsic_instr_create(state->builder.shader, op);
 296
 297    nir_intrinsic_set_base(atomic, var->data.driver_location);
 298
 299    atomic->src[0] = nir_src_for_ssa(offset);
 300    assert(nir_intrinsic_infos[intrin->intrinsic].num_srcs ==
 301           nir_intrinsic_infos[op].num_srcs);
 302    for (unsigned i = 1; i < nir_intrinsic_infos[op].num_srcs; i++) {
 303       nir_src_copy(&atomic->src[i], &intrin->src[i], atomic);
 304    }
 305
 306    return atomic;
 307 }
 308
 309 static nir_intrinsic_instr *
 310 lower_interpolate_at(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 311                      nir_variable *var, nir_ssa_def *offset, unsigned component)
 312 {
 313    assert(var->data.mode == nir_var_shader_in);
 314
 315    /* Ignore interpolateAt() for flat variables - flat is flat. */
 316    if (var->data.interpolation == INTERP_MODE_FLAT)
 317       return lower_load(intrin, state, NULL, var, offset, component);
 318
 319    nir_intrinsic_op bary_op;
 320    switch (intrin->intrinsic) {
 321    case nir_intrinsic_interp_var_at_centroid:
 322    case nir_intrinsic_interp_deref_at_centroid:
 323       bary_op = (state->options & nir_lower_io_force_sample_interpolation) ?
 324                 nir_intrinsic_load_barycentric_sample :
 325                 nir_intrinsic_load_barycentric_centroid;
 326       break;
 327    case nir_intrinsic_interp_var_at_sample:
 328    case nir_intrinsic_interp_deref_at_sample:
 329       bary_op = nir_intrinsic_load_barycentric_at_sample;
 330       break;
 331    case nir_intrinsic_interp_var_at_offset:
 332    case nir_intrinsic_interp_deref_at_offset:
 333       bary_op = nir_intrinsic_load_barycentric_at_offset;
 334       break;
 335    default:
 336       unreachable("Bogus interpolateAt() intrinsic.");
 337    }
 338
 339    nir_intrinsic_instr *bary_setup =
 340       nir_intrinsic_instr_create(state->builder.shader, bary_op);
 341
 342    nir_ssa_dest_init(&bary_setup->instr, &bary_setup->dest, 2, 32, NULL);
 343    nir_intrinsic_set_interp_mode(bary_setup, var->data.interpolation);
 344
 345    if (intrin->intrinsic == nir_intrinsic_interp_var_at_sample ||
 346        intrin->intrinsic == nir_intrinsic_interp_var_at_offset)
 347       nir_src_copy(&bary_setup->src[0], &intrin->src[0], bary_setup);
 348
 349    if (intrin->intrinsic == nir_intrinsic_interp_deref_at_sample ||
 350        intrin->intrinsic == nir_intrinsic_interp_deref_at_offset)
 351       nir_src_copy(&bary_setup->src[0], &intrin->src[1], bary_setup);
 352
 353    nir_builder_instr_insert(&state->builder, &bary_setup->instr);
 354
 355    nir_intrinsic_instr *load =
 356       nir_intrinsic_instr_create(state->builder.shader,
 357                                  nir_intrinsic_load_interpolated_input);
 358    load->num_components = intrin->num_components;
 359
 360    nir_intrinsic_set_base(load, var->data.driver_location);
 361    nir_intrinsic_set_component(load, component);
 362
 363    load->src[0] = nir_src_for_ssa(&bary_setup->dest.ssa);
 364    load->src[1] = nir_src_for_ssa(offset);
 365
 366    return load;
 367 }
 368
 369 static bool
 370 nir_lower_io_block(nir_block *block,
 371                    struct lower_io_state *state)
 372 {
 373    nir_builder *b = &state->builder;
 374    const nir_shader_compiler_options *options = b->shader->options;
 375    bool progress = false;
 376
 377    nir_foreach_instr_safe(instr, block) {
 378       if (instr->type != nir_instr_type_intrinsic)
 379          continue;
 380
 381       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 382
 383       switch (intrin->intrinsic) {
 384       case nir_intrinsic_load_var:
 385       case nir_intrinsic_load_deref:
 386       case nir_intrinsic_store_var:
 387       case nir_intrinsic_store_deref:
 388       case nir_intrinsic_deref_atomic_add:
 389       case nir_intrinsic_deref_atomic_imin:
 390       case nir_intrinsic_deref_atomic_umin:
 391       case nir_intrinsic_deref_atomic_imax:
 392       case nir_intrinsic_deref_atomic_umax:
 393       case nir_intrinsic_deref_atomic_and:
 394       case nir_intrinsic_deref_atomic_or:
 395       case nir_intrinsic_deref_atomic_xor:
 396       case nir_intrinsic_deref_atomic_exchange:
 397       case nir_intrinsic_deref_atomic_comp_swap:
 398          /* We can lower the io for this nir instrinsic */
 399          break;
 400       case nir_intrinsic_interp_var_at_centroid:
 401       case nir_intrinsic_interp_var_at_sample:
 402       case nir_intrinsic_interp_var_at_offset:
 403       case nir_intrinsic_interp_deref_at_centroid:
 404       case nir_intrinsic_interp_deref_at_sample:
 405       case nir_intrinsic_interp_deref_at_offset:
 406          /* We can optionally lower these to load_interpolated_input */
 407          if (options->use_interpolated_input_intrinsics)
 408             break;
 409       default:
 410          /* We can't lower the io for this nir instrinsic, so skip it */
 411          continue;
 412       }
 413
 414       nir_deref_var *deref;
 415       nir_deref_instr *deref_instr = NULL;
 416       if (nir_intrinsic_infos[intrin->intrinsic].num_variables > 0) {
 417          assert(nir_intrinsic_infos[intrin->intrinsic].num_variables == 1);
 418          deref = intrin->variables[0];
 419       } else {
 420          deref_instr = nir_src_as_deref(intrin->src[0]);
 421          deref = nir_deref_instr_to_deref(deref_instr, state->dead_ctx);
 422       }
 423
 424       nir_variable *var = deref->var;
 425       nir_variable_mode mode = var->data.mode;
 426
 427       if ((state->modes & mode) == 0)
 428          continue;
 429
 430       if (mode != nir_var_shader_in &&
 431           mode != nir_var_shader_out &&
 432           mode != nir_var_shared &&
 433           mode != nir_var_uniform)
 434          continue;
 435
 436       b->cursor = nir_before_instr(instr);
 437
 438       const bool per_vertex = nir_is_per_vertex_io(var, b->shader->info.stage);
 439
 440       nir_ssa_def *offset;
 441       nir_ssa_def *vertex_index = NULL;
 442       unsigned component_offset = var->data.location_frac;
 443
 444       offset = get_io_offset(b, deref, per_vertex ? &vertex_index : NULL,
 445                              state->type_size, &component_offset);
 446
 447       nir_intrinsic_instr *replacement;
 448
 449       switch (intrin->intrinsic) {
 450       case nir_intrinsic_load_var:
 451       case nir_intrinsic_load_deref:
 452          replacement = lower_load(intrin, state, vertex_index, var, offset,
 453                                   component_offset);
 454          break;
 455
 456       case nir_intrinsic_store_var:
 457       case nir_intrinsic_store_deref:
 458          replacement = lower_store(intrin, state, vertex_index, var, offset,
 459                                    component_offset);
 460          break;
 461
 462       case nir_intrinsic_deref_atomic_add:
 463       case nir_intrinsic_deref_atomic_imin:
 464       case nir_intrinsic_deref_atomic_umin:
 465       case nir_intrinsic_deref_atomic_imax:
 466       case nir_intrinsic_deref_atomic_umax:
 467       case nir_intrinsic_deref_atomic_and:
 468       case nir_intrinsic_deref_atomic_or:
 469       case nir_intrinsic_deref_atomic_xor:
 470       case nir_intrinsic_deref_atomic_exchange:
 471       case nir_intrinsic_deref_atomic_comp_swap:
 472          assert(vertex_index == NULL);
 473          replacement = lower_atomic(intrin, state, var, offset);
 474          break;
 475
 476       case nir_intrinsic_interp_var_at_centroid:
 477       case nir_intrinsic_interp_var_at_sample:
 478       case nir_intrinsic_interp_var_at_offset:
 479       case nir_intrinsic_interp_deref_at_centroid:
 480       case nir_intrinsic_interp_deref_at_sample:
 481       case nir_intrinsic_interp_deref_at_offset:
 482          assert(vertex_index == NULL);
 483          replacement = lower_interpolate_at(intrin, state, var, offset,
 484                                             component_offset);
 485          break;
 486
 487       default:
 488          continue;
 489       }
 490
 491       if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
 492          if (intrin->dest.is_ssa) {
 493             nir_ssa_dest_init(&replacement->instr, &replacement->dest,
 494                               intrin->dest.ssa.num_components,
 495                               intrin->dest.ssa.bit_size, NULL);
 496             nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
 497                                      nir_src_for_ssa(&replacement->dest.ssa));
 498          } else {
 499             nir_dest_copy(&replacement->dest, &intrin->dest, &intrin->instr);
 500          }
 501       }
 502
 503       nir_instr_insert_before(&intrin->instr, &replacement->instr);
 504       nir_instr_remove(&intrin->instr);
 505       progress = true;
 506    }
 507
 508    return progress;
 509 }
 510
 511 static bool
 512 nir_lower_io_impl(nir_function_impl *impl,
 513                   nir_variable_mode modes,
 514                   int (*type_size)(const struct glsl_type *),
 515                   nir_lower_io_options options)
 516 {
 517    struct lower_io_state state;
 518    bool progress = false;
 519
 520    nir_builder_init(&state.builder, impl);
 521    state.dead_ctx = ralloc_context(NULL);
 522    state.modes = modes;
 523    state.type_size = type_size;
 524    state.options = options;
 525
 526    nir_foreach_block(block, impl) {
 527       progress |= nir_lower_io_block(block, &state);
 528    }
 529
 530    ralloc_free(state.dead_ctx);
 531
 532    nir_metadata_preserve(impl, nir_metadata_block_index |
 533                                nir_metadata_dominance);
 534    return progress;
 535 }
 536
 537 bool
 538 nir_lower_io(nir_shader *shader, nir_variable_mode modes,
 539              int (*type_size)(const struct glsl_type *),
 540              nir_lower_io_options options)
 541 {
 542    bool progress = false;
 543
 544    nir_assert_unlowered_derefs(shader, nir_lower_atomic_derefs);
 545
 546    nir_foreach_function(function, shader) {
 547       if (function->impl) {
 548          progress |= nir_lower_io_impl(function->impl, modes,
 549                                        type_size, options);
 550       }
 551    }
 552
 553    return progress;
 554 }
 555
 556 /**
 557  * Return the offset source for a load/store intrinsic.
 558  */
 559 nir_src *
 560 nir_get_io_offset_src(nir_intrinsic_instr *instr)
 561 {
 562    switch (instr->intrinsic) {
 563    case nir_intrinsic_load_input:
 564    case nir_intrinsic_load_output:
 565    case nir_intrinsic_load_uniform:
 566       return &instr->src[0];
 567    case nir_intrinsic_load_ubo:
 568    case nir_intrinsic_load_ssbo:
 569    case nir_intrinsic_load_per_vertex_input:
 570    case nir_intrinsic_load_per_vertex_output:
 571    case nir_intrinsic_load_interpolated_input:
 572    case nir_intrinsic_store_output:
 573       return &instr->src[1];
 574    case nir_intrinsic_store_ssbo:
 575    case nir_intrinsic_store_per_vertex_output:
 576       return &instr->src[2];
 577    default:
 578       return NULL;
 579    }
 580 }
 581
 582 /**
 583  * Return the vertex index source for a load/store per_vertex intrinsic.
 584  */
 585 nir_src *
 586 nir_get_io_vertex_index_src(nir_intrinsic_instr *instr)
 587 {
 588    switch (instr->intrinsic) {
 589    case nir_intrinsic_load_per_vertex_input:
 590    case nir_intrinsic_load_per_vertex_output:
 591       return &instr->src[0];
 592    case nir_intrinsic_store_per_vertex_output:
 593       return &instr->src[1];
 594    default:
 595       return NULL;
 596    }
 597 }