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 #include "nir_deref.h"
  37
  38 struct lower_io_state {
  39    void *dead_ctx;
  40    nir_builder builder;
  41    int (*type_size)(const struct glsl_type *type);
  42    nir_variable_mode modes;
  43    nir_lower_io_options options;
  44 };
  45
  46 void
  47 nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
  48                          int (*type_size)(const struct glsl_type *))
  49 {
  50    unsigned location = 0;
  51
  52    nir_foreach_variable(var, var_list) {
  53       /*
  54        * UBOs have their own address spaces, so don't count them towards the
  55        * number of global uniforms
  56        */
  57       if ((var->data.mode == nir_var_uniform || var->data.mode == nir_var_shader_storage) &&
  58           var->interface_type != NULL)
  59          continue;
  60
  61       var->data.driver_location = location;
  62       location += type_size(var->type);
  63    }
  64
  65    *size = location;
  66 }
  67
  68 /**
  69  * Return true if the given variable is a per-vertex input/output array.
  70  * (such as geometry shader inputs).
  71  */
  72 bool
  73 nir_is_per_vertex_io(const nir_variable *var, gl_shader_stage stage)
  74 {
  75    if (var->data.patch || !glsl_type_is_array(var->type))
  76       return false;
  77
  78    if (var->data.mode == nir_var_shader_in)
  79       return stage == MESA_SHADER_GEOMETRY ||
  80              stage == MESA_SHADER_TESS_CTRL ||
  81              stage == MESA_SHADER_TESS_EVAL;
  82
  83    if (var->data.mode == nir_var_shader_out)
  84       return stage == MESA_SHADER_TESS_CTRL;
  85
  86    return false;
  87 }
  88
  89 static nir_ssa_def *
  90 get_io_offset(nir_builder *b, nir_deref_instr *deref,
  91               nir_ssa_def **vertex_index,
  92               int (*type_size)(const struct glsl_type *),
  93               unsigned *component)
  94 {
  95    nir_deref_path path;
  96    nir_deref_path_init(&path, deref, NULL);
  97
  98    assert(path.path[0]->deref_type == nir_deref_type_var);
  99    nir_deref_instr **p = &path.path[1];
 100
 101    /* For per-vertex input arrays (i.e. geometry shader inputs), keep the
 102     * outermost array index separate.  Process the rest normally.
 103     */
 104    if (vertex_index != NULL) {
 105       assert((*p)->deref_type == nir_deref_type_array);
 106       *vertex_index = nir_ssa_for_src(b, (*p)->arr.index, 1);
 107       p++;
 108    }
 109
 110    if (path.path[0]->var->data.compact) {
 111       assert((*p)->deref_type == nir_deref_type_array);
 112       assert(glsl_type_is_scalar((*p)->type));
 113
 114       /* We always lower indirect dereferences for "compact" array vars. */
 115       const unsigned index = nir_src_as_uint((*p)->arr.index);
 116       const unsigned total_offset = *component + index;
 117       const unsigned slot_offset = total_offset / 4;
 118       *component = total_offset % 4;
 119       return nir_imm_int(b, type_size(glsl_vec4_type()) * slot_offset);
 120    }
 121
 122    /* Just emit code and let constant-folding go to town */
 123    nir_ssa_def *offset = nir_imm_int(b, 0);
 124
 125    for (; *p; p++) {
 126       if ((*p)->deref_type == nir_deref_type_array) {
 127          unsigned size = type_size((*p)->type);
 128
 129          nir_ssa_def *mul =
 130             nir_imul_imm(b, nir_ssa_for_src(b, (*p)->arr.index, 1), size);
 131
 132          offset = nir_iadd(b, offset, mul);
 133       } else if ((*p)->deref_type == nir_deref_type_struct) {
 134          /* p starts at path[1], so this is safe */
 135          nir_deref_instr *parent = *(p - 1);
 136
 137          unsigned field_offset = 0;
 138          for (unsigned i = 0; i < (*p)->strct.index; i++) {
 139             field_offset += type_size(glsl_get_struct_field(parent->type, i));
 140          }
 141          offset = nir_iadd_imm(b, offset, field_offset);
 142       } else {
 143          unreachable("Unsupported deref type");
 144       }
 145    }
 146
 147    nir_deref_path_finish(&path);
 148
 149    return offset;
 150 }
 151
 152 static nir_intrinsic_instr *
 153 lower_load(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 154            nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
 155            unsigned component)
 156 {
 157    const nir_shader *nir = state->builder.shader;
 158    nir_variable_mode mode = var->data.mode;
 159    nir_ssa_def *barycentric = NULL;
 160
 161    nir_intrinsic_op op;
 162    switch (mode) {
 163    case nir_var_shader_in:
 164       if (nir->info.stage == MESA_SHADER_FRAGMENT &&
 165           nir->options->use_interpolated_input_intrinsics &&
 166           var->data.interpolation != INTERP_MODE_FLAT) {
 167          assert(vertex_index == NULL);
 168
 169          nir_intrinsic_op bary_op;
 170          if (var->data.sample ||
 171              (state->options & nir_lower_io_force_sample_interpolation))
 172             bary_op = nir_intrinsic_load_barycentric_sample;
 173          else if (var->data.centroid)
 174             bary_op = nir_intrinsic_load_barycentric_centroid;
 175          else
 176             bary_op = nir_intrinsic_load_barycentric_pixel;
 177
 178          barycentric = nir_load_barycentric(&state->builder, bary_op,
 179                                             var->data.interpolation);
 180          op = nir_intrinsic_load_interpolated_input;
 181       } else {
 182          op = vertex_index ? nir_intrinsic_load_per_vertex_input :
 183                              nir_intrinsic_load_input;
 184       }
 185       break;
 186    case nir_var_shader_out:
 187       op = vertex_index ? nir_intrinsic_load_per_vertex_output :
 188                           nir_intrinsic_load_output;
 189       break;
 190    case nir_var_uniform:
 191       op = nir_intrinsic_load_uniform;
 192       break;
 193    case nir_var_shared:
 194       op = nir_intrinsic_load_shared;
 195       break;
 196    default:
 197       unreachable("Unknown variable mode");
 198    }
 199
 200    nir_intrinsic_instr *load =
 201       nir_intrinsic_instr_create(state->builder.shader, op);
 202    load->num_components = intrin->num_components;
 203
 204    nir_intrinsic_set_base(load, var->data.driver_location);
 205    if (mode == nir_var_shader_in || mode == nir_var_shader_out)
 206       nir_intrinsic_set_component(load, component);
 207
 208    if (load->intrinsic == nir_intrinsic_load_uniform)
 209       nir_intrinsic_set_range(load, state->type_size(var->type));
 210
 211    if (vertex_index) {
 212       load->src[0] = nir_src_for_ssa(vertex_index);
 213       load->src[1] = nir_src_for_ssa(offset);
 214    } else if (barycentric) {
 215       load->src[0] = nir_src_for_ssa(barycentric);
 216       load->src[1] = nir_src_for_ssa(offset);
 217    } else {
 218       load->src[0] = nir_src_for_ssa(offset);
 219    }
 220
 221    return load;
 222 }
 223
 224 static nir_intrinsic_instr *
 225 lower_store(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 226             nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
 227             unsigned component)
 228 {
 229    nir_variable_mode mode = var->data.mode;
 230
 231    nir_intrinsic_op op;
 232    if (mode == nir_var_shared) {
 233       op = nir_intrinsic_store_shared;
 234    } else {
 235       assert(mode == nir_var_shader_out);
 236       op = vertex_index ? nir_intrinsic_store_per_vertex_output :
 237                           nir_intrinsic_store_output;
 238    }
 239
 240    nir_intrinsic_instr *store =
 241       nir_intrinsic_instr_create(state->builder.shader, op);
 242    store->num_components = intrin->num_components;
 243
 244    nir_src_copy(&store->src[0], &intrin->src[1], store);
 245
 246    nir_intrinsic_set_base(store, var->data.driver_location);
 247
 248    if (mode == nir_var_shader_out)
 249       nir_intrinsic_set_component(store, component);
 250
 251    nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
 252
 253    if (vertex_index)
 254       store->src[1] = nir_src_for_ssa(vertex_index);
 255
 256    store->src[vertex_index ? 2 : 1] = nir_src_for_ssa(offset);
 257
 258    return store;
 259 }
 260
 261 static nir_intrinsic_instr *
 262 lower_atomic(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 263              nir_variable *var, nir_ssa_def *offset)
 264 {
 265    assert(var->data.mode == nir_var_shared);
 266
 267    nir_intrinsic_op op;
 268    switch (intrin->intrinsic) {
 269 #define OP(O) case nir_intrinsic_deref_##O: op = nir_intrinsic_shared_##O; break;
 270    OP(atomic_exchange)
 271    OP(atomic_comp_swap)
 272    OP(atomic_add)
 273    OP(atomic_imin)
 274    OP(atomic_umin)
 275    OP(atomic_imax)
 276    OP(atomic_umax)
 277    OP(atomic_and)
 278    OP(atomic_or)
 279    OP(atomic_xor)
 280    OP(atomic_fadd)
 281    OP(atomic_fmin)
 282    OP(atomic_fmax)
 283    OP(atomic_fcomp_swap)
 284 #undef OP
 285    default:
 286       unreachable("Invalid atomic");
 287    }
 288
 289    nir_intrinsic_instr *atomic =
 290       nir_intrinsic_instr_create(state->builder.shader, op);
 291
 292    nir_intrinsic_set_base(atomic, var->data.driver_location);
 293
 294    atomic->src[0] = nir_src_for_ssa(offset);
 295    assert(nir_intrinsic_infos[intrin->intrinsic].num_srcs ==
 296           nir_intrinsic_infos[op].num_srcs);
 297    for (unsigned i = 1; i < nir_intrinsic_infos[op].num_srcs; i++) {
 298       nir_src_copy(&atomic->src[i], &intrin->src[i], atomic);
 299    }
 300
 301    return atomic;
 302 }
 303
 304 static nir_intrinsic_instr *
 305 lower_interpolate_at(nir_intrinsic_instr *intrin, struct lower_io_state *state,
 306                      nir_variable *var, nir_ssa_def *offset, unsigned component)
 307 {
 308    assert(var->data.mode == nir_var_shader_in);
 309
 310    /* Ignore interpolateAt() for flat variables - flat is flat. */
 311    if (var->data.interpolation == INTERP_MODE_FLAT)
 312       return lower_load(intrin, state, NULL, var, offset, component);
 313
 314    nir_intrinsic_op bary_op;
 315    switch (intrin->intrinsic) {
 316    case nir_intrinsic_interp_deref_at_centroid:
 317       bary_op = (state->options & nir_lower_io_force_sample_interpolation) ?
 318                 nir_intrinsic_load_barycentric_sample :
 319                 nir_intrinsic_load_barycentric_centroid;
 320       break;
 321    case nir_intrinsic_interp_deref_at_sample:
 322       bary_op = nir_intrinsic_load_barycentric_at_sample;
 323       break;
 324    case nir_intrinsic_interp_deref_at_offset:
 325       bary_op = nir_intrinsic_load_barycentric_at_offset;
 326       break;
 327    default:
 328       unreachable("Bogus interpolateAt() intrinsic.");
 329    }
 330
 331    nir_intrinsic_instr *bary_setup =
 332       nir_intrinsic_instr_create(state->builder.shader, bary_op);
 333
 334    nir_ssa_dest_init(&bary_setup->instr, &bary_setup->dest, 2, 32, NULL);
 335    nir_intrinsic_set_interp_mode(bary_setup, var->data.interpolation);
 336
 337    if (intrin->intrinsic == nir_intrinsic_interp_deref_at_sample ||
 338        intrin->intrinsic == nir_intrinsic_interp_deref_at_offset)
 339       nir_src_copy(&bary_setup->src[0], &intrin->src[1], bary_setup);
 340
 341    nir_builder_instr_insert(&state->builder, &bary_setup->instr);
 342
 343    nir_intrinsic_instr *load =
 344       nir_intrinsic_instr_create(state->builder.shader,
 345                                  nir_intrinsic_load_interpolated_input);
 346    load->num_components = intrin->num_components;
 347
 348    nir_intrinsic_set_base(load, var->data.driver_location);
 349    nir_intrinsic_set_component(load, component);
 350
 351    load->src[0] = nir_src_for_ssa(&bary_setup->dest.ssa);
 352    load->src[1] = nir_src_for_ssa(offset);
 353
 354    return load;
 355 }
 356
 357 static bool
 358 nir_lower_io_block(nir_block *block,
 359                    struct lower_io_state *state)
 360 {
 361    nir_builder *b = &state->builder;
 362    const nir_shader_compiler_options *options = b->shader->options;
 363    bool progress = false;
 364
 365    nir_foreach_instr_safe(instr, block) {
 366       if (instr->type != nir_instr_type_intrinsic)
 367          continue;
 368
 369       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 370
 371       switch (intrin->intrinsic) {
 372       case nir_intrinsic_load_deref:
 373       case nir_intrinsic_store_deref:
 374       case nir_intrinsic_deref_atomic_add:
 375       case nir_intrinsic_deref_atomic_imin:
 376       case nir_intrinsic_deref_atomic_umin:
 377       case nir_intrinsic_deref_atomic_imax:
 378       case nir_intrinsic_deref_atomic_umax:
 379       case nir_intrinsic_deref_atomic_and:
 380       case nir_intrinsic_deref_atomic_or:
 381       case nir_intrinsic_deref_atomic_xor:
 382       case nir_intrinsic_deref_atomic_exchange:
 383       case nir_intrinsic_deref_atomic_comp_swap:
 384       case nir_intrinsic_deref_atomic_fadd:
 385       case nir_intrinsic_deref_atomic_fmin:
 386       case nir_intrinsic_deref_atomic_fmax:
 387       case nir_intrinsic_deref_atomic_fcomp_swap:
 388          /* We can lower the io for this nir instrinsic */
 389          break;
 390       case nir_intrinsic_interp_deref_at_centroid:
 391       case nir_intrinsic_interp_deref_at_sample:
 392       case nir_intrinsic_interp_deref_at_offset:
 393          /* We can optionally lower these to load_interpolated_input */
 394          if (options->use_interpolated_input_intrinsics)
 395             break;
 396       default:
 397          /* We can't lower the io for this nir instrinsic, so skip it */
 398          continue;
 399       }
 400
 401       nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
 402
 403       nir_variable *var = nir_deref_instr_get_variable(deref);
 404       nir_variable_mode mode = var->data.mode;
 405
 406       if ((state->modes & mode) == 0)
 407          continue;
 408
 409       if (mode != nir_var_shader_in &&
 410           mode != nir_var_shader_out &&
 411           mode != nir_var_shared &&
 412           mode != nir_var_uniform)
 413          continue;
 414
 415       b->cursor = nir_before_instr(instr);
 416
 417       const bool per_vertex = nir_is_per_vertex_io(var, b->shader->info.stage);
 418
 419       nir_ssa_def *offset;
 420       nir_ssa_def *vertex_index = NULL;
 421       unsigned component_offset = var->data.location_frac;
 422
 423       offset = get_io_offset(b, deref, per_vertex ? &vertex_index : NULL,
 424                              state->type_size, &component_offset);
 425
 426       nir_intrinsic_instr *replacement;
 427
 428       switch (intrin->intrinsic) {
 429       case nir_intrinsic_load_deref:
 430          replacement = lower_load(intrin, state, vertex_index, var, offset,
 431                                   component_offset);
 432          break;
 433
 434       case nir_intrinsic_store_deref:
 435          replacement = lower_store(intrin, state, vertex_index, var, offset,
 436                                    component_offset);
 437          break;
 438
 439       case nir_intrinsic_deref_atomic_add:
 440       case nir_intrinsic_deref_atomic_imin:
 441       case nir_intrinsic_deref_atomic_umin:
 442       case nir_intrinsic_deref_atomic_imax:
 443       case nir_intrinsic_deref_atomic_umax:
 444       case nir_intrinsic_deref_atomic_and:
 445       case nir_intrinsic_deref_atomic_or:
 446       case nir_intrinsic_deref_atomic_xor:
 447       case nir_intrinsic_deref_atomic_exchange:
 448       case nir_intrinsic_deref_atomic_comp_swap:
 449       case nir_intrinsic_deref_atomic_fadd:
 450       case nir_intrinsic_deref_atomic_fmin:
 451       case nir_intrinsic_deref_atomic_fmax:
 452       case nir_intrinsic_deref_atomic_fcomp_swap:
 453          assert(vertex_index == NULL);
 454          replacement = lower_atomic(intrin, state, var, offset);
 455          break;
 456
 457       case nir_intrinsic_interp_deref_at_centroid:
 458       case nir_intrinsic_interp_deref_at_sample:
 459       case nir_intrinsic_interp_deref_at_offset:
 460          assert(vertex_index == NULL);
 461          replacement = lower_interpolate_at(intrin, state, var, offset,
 462                                             component_offset);
 463          break;
 464
 465       default:
 466          continue;
 467       }
 468
 469       if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
 470          if (intrin->dest.is_ssa) {
 471             nir_ssa_dest_init(&replacement->instr, &replacement->dest,
 472                               intrin->dest.ssa.num_components,
 473                               intrin->dest.ssa.bit_size, NULL);
 474             nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
 475                                      nir_src_for_ssa(&replacement->dest.ssa));
 476          } else {
 477             nir_dest_copy(&replacement->dest, &intrin->dest, &intrin->instr);
 478          }
 479       }
 480
 481       nir_instr_insert_before(&intrin->instr, &replacement->instr);
 482       nir_instr_remove(&intrin->instr);
 483       progress = true;
 484    }
 485
 486    return progress;
 487 }
 488
 489 static bool
 490 nir_lower_io_impl(nir_function_impl *impl,
 491                   nir_variable_mode modes,
 492                   int (*type_size)(const struct glsl_type *),
 493                   nir_lower_io_options options)
 494 {
 495    struct lower_io_state state;
 496    bool progress = false;
 497
 498    nir_builder_init(&state.builder, impl);
 499    state.dead_ctx = ralloc_context(NULL);
 500    state.modes = modes;
 501    state.type_size = type_size;
 502    state.options = options;
 503
 504    nir_foreach_block(block, impl) {
 505       progress |= nir_lower_io_block(block, &state);
 506    }
 507
 508    ralloc_free(state.dead_ctx);
 509
 510    nir_metadata_preserve(impl, nir_metadata_block_index |
 511                                nir_metadata_dominance);
 512    return progress;
 513 }
 514
 515 bool
 516 nir_lower_io(nir_shader *shader, nir_variable_mode modes,
 517              int (*type_size)(const struct glsl_type *),
 518              nir_lower_io_options options)
 519 {
 520    bool progress = false;
 521
 522    nir_foreach_function(function, shader) {
 523       if (function->impl) {
 524          progress |= nir_lower_io_impl(function->impl, modes,
 525                                        type_size, options);
 526       }
 527    }
 528
 529    return progress;
 530 }
 531
 532 /**
 533  * Return the offset source for a load/store intrinsic.
 534  */
 535 nir_src *
 536 nir_get_io_offset_src(nir_intrinsic_instr *instr)
 537 {
 538    switch (instr->intrinsic) {
 539    case nir_intrinsic_load_input:
 540    case nir_intrinsic_load_output:
 541    case nir_intrinsic_load_shared:
 542    case nir_intrinsic_load_uniform:
 543       return &instr->src[0];
 544    case nir_intrinsic_load_ubo:
 545    case nir_intrinsic_load_ssbo:
 546    case nir_intrinsic_load_per_vertex_input:
 547    case nir_intrinsic_load_per_vertex_output:
 548    case nir_intrinsic_load_interpolated_input:
 549    case nir_intrinsic_store_output:
 550    case nir_intrinsic_store_shared:
 551       return &instr->src[1];
 552    case nir_intrinsic_store_ssbo:
 553    case nir_intrinsic_store_per_vertex_output:
 554       return &instr->src[2];
 555    default:
 556       return NULL;
 557    }
 558 }
 559
 560 /**
 561  * Return the vertex index source for a load/store per_vertex intrinsic.
 562  */
 563 nir_src *
 564 nir_get_io_vertex_index_src(nir_intrinsic_instr *instr)
 565 {
 566    switch (instr->intrinsic) {
 567    case nir_intrinsic_load_per_vertex_input:
 568    case nir_intrinsic_load_per_vertex_output:
 569       return &instr->src[0];
 570    case nir_intrinsic_store_per_vertex_output:
 571       return &instr->src[1];
 572    default:
 573       return NULL;
 574    }
 575 }