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 };
  43
  44 void
  45 nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
  46                          int (*type_size)(const struct glsl_type *))
  47 {
  48    unsigned location = 0;
  49
  50    nir_foreach_variable(var, var_list) {
  51       /*
  52        * UBO's have their own address spaces, so don't count them towards the
  53        * number of global uniforms
  54        */
  55       if ((var->data.mode == nir_var_uniform || var->data.mode == nir_var_shader_storage) &&
  56           var->interface_type != NULL)
  57          continue;
  58
  59       var->data.driver_location = location;
  60       location += type_size(var->type);
  61    }
  62
  63    *size = location;
  64 }
  65
  66 /**
  67  * Returns true if we're processing a stage whose inputs are arrays indexed
  68  * by a vertex number (such as geometry shader inputs).
  69  */
  70 static bool
  71 is_per_vertex_input(struct lower_io_state *state, nir_variable *var)
  72 {
  73    gl_shader_stage stage = state->builder.shader->stage;
  74
  75    return var->data.mode == nir_var_shader_in && !var->data.patch &&
  76           (stage == MESA_SHADER_TESS_CTRL ||
  77            stage == MESA_SHADER_TESS_EVAL ||
  78            stage == MESA_SHADER_GEOMETRY);
  79 }
  80
  81 static bool
  82 is_per_vertex_output(struct lower_io_state *state, nir_variable *var)
  83 {
  84    gl_shader_stage stage = state->builder.shader->stage;
  85    return var->data.mode == nir_var_shader_out && !var->data.patch &&
  86           stage == MESA_SHADER_TESS_CTRL;
  87 }
  88
  89 static nir_ssa_def *
  90 get_io_offset(nir_builder *b, nir_deref_var *deref,
  91               nir_ssa_def **vertex_index,
  92               int (*type_size)(const struct glsl_type *))
  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       assert(tail->deref_type == nir_deref_type_array);
 102       nir_deref_array *deref_array = nir_deref_as_array(tail);
 103
 104       nir_ssa_def *vtx = nir_imm_int(b, deref_array->base_offset);
 105       if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
 106          vtx = nir_iadd(b, vtx, nir_ssa_for_src(b, deref_array->indirect, 1));
 107       }
 108       *vertex_index = vtx;
 109    }
 110
 111    /* Just emit code and let constant-folding go to town */
 112    nir_ssa_def *offset = nir_imm_int(b, 0);
 113
 114    while (tail->child != NULL) {
 115       const struct glsl_type *parent_type = tail->type;
 116       tail = tail->child;
 117
 118       if (tail->deref_type == nir_deref_type_array) {
 119          nir_deref_array *deref_array = nir_deref_as_array(tail);
 120          unsigned size = type_size(tail->type);
 121
 122          offset = nir_iadd(b, offset,
 123                            nir_imm_int(b, size * deref_array->base_offset));
 124
 125          if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
 126             nir_ssa_def *mul =
 127                nir_imul(b, nir_imm_int(b, size),
 128                         nir_ssa_for_src(b, deref_array->indirect, 1));
 129
 130             offset = nir_iadd(b, offset, mul);
 131          }
 132       } else if (tail->deref_type == nir_deref_type_struct) {
 133          nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
 134
 135          unsigned field_offset = 0;
 136          for (unsigned i = 0; i < deref_struct->index; i++) {
 137             field_offset += type_size(glsl_get_struct_field(parent_type, i));
 138          }
 139          offset = nir_iadd(b, offset, nir_imm_int(b, field_offset));
 140       }
 141    }
 142
 143    return offset;
 144 }
 145
 146 static nir_intrinsic_op
 147 load_op(struct lower_io_state *state,
 148         nir_variable_mode mode, bool per_vertex)
 149 {
 150    nir_intrinsic_op op;
 151    switch (mode) {
 152    case nir_var_shader_in:
 153       op = per_vertex ? nir_intrinsic_load_per_vertex_input :
 154                         nir_intrinsic_load_input;
 155       break;
 156    case nir_var_shader_out:
 157       op = per_vertex ? nir_intrinsic_load_per_vertex_output :
 158                         nir_intrinsic_load_output;
 159       break;
 160    case nir_var_uniform:
 161       op = nir_intrinsic_load_uniform;
 162       break;
 163    case nir_var_shared:
 164       op = nir_intrinsic_load_shared;
 165       break;
 166    default:
 167       unreachable("Unknown variable mode");
 168    }
 169    return op;
 170 }
 171
 172 static nir_intrinsic_op
 173 store_op(struct lower_io_state *state,
 174          nir_variable_mode mode, bool per_vertex)
 175 {
 176    nir_intrinsic_op op;
 177    switch (mode) {
 178    case nir_var_shader_in:
 179    case nir_var_shader_out:
 180       op = per_vertex ? nir_intrinsic_store_per_vertex_output :
 181                         nir_intrinsic_store_output;
 182       break;
 183    case nir_var_shared:
 184       op = nir_intrinsic_store_shared;
 185       break;
 186    default:
 187       unreachable("Unknown variable mode");
 188    }
 189    return op;
 190 }
 191
 192 static nir_intrinsic_op
 193 atomic_op(nir_intrinsic_op opcode)
 194 {
 195    switch (opcode) {
 196 #define OP(O) case nir_intrinsic_var_##O: return nir_intrinsic_shared_##O;
 197    OP(atomic_exchange)
 198    OP(atomic_comp_swap)
 199    OP(atomic_add)
 200    OP(atomic_imin)
 201    OP(atomic_umin)
 202    OP(atomic_imax)
 203    OP(atomic_umax)
 204    OP(atomic_and)
 205    OP(atomic_or)
 206    OP(atomic_xor)
 207 #undef OP
 208    default:
 209       unreachable("Invalid atomic");
 210    }
 211 }
 212
 213 static bool
 214 nir_lower_io_block(nir_block *block, void *void_state)
 215 {
 216    struct lower_io_state *state = void_state;
 217
 218    nir_builder *b = &state->builder;
 219
 220    nir_foreach_instr_safe(block, instr) {
 221       if (instr->type != nir_instr_type_intrinsic)
 222          continue;
 223
 224       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 225
 226       switch (intrin->intrinsic) {
 227       case nir_intrinsic_load_var:
 228       case nir_intrinsic_store_var:
 229       case nir_intrinsic_var_atomic_add:
 230       case nir_intrinsic_var_atomic_imin:
 231       case nir_intrinsic_var_atomic_umin:
 232       case nir_intrinsic_var_atomic_imax:
 233       case nir_intrinsic_var_atomic_umax:
 234       case nir_intrinsic_var_atomic_and:
 235       case nir_intrinsic_var_atomic_or:
 236       case nir_intrinsic_var_atomic_xor:
 237       case nir_intrinsic_var_atomic_exchange:
 238       case nir_intrinsic_var_atomic_comp_swap:
 239          /* We can lower the io for this nir instrinsic */
 240          break;
 241       default:
 242          /* We can't lower the io for this nir instrinsic, so skip it */
 243          continue;
 244       }
 245
 246       nir_variable_mode mode = intrin->variables[0]->var->data.mode;
 247
 248       if ((state->modes & mode) == 0)
 249          continue;
 250
 251       if (mode != nir_var_shader_in &&
 252           mode != nir_var_shader_out &&
 253           mode != nir_var_shared &&
 254           mode != nir_var_uniform)
 255          continue;
 256
 257       b->cursor = nir_before_instr(instr);
 258
 259       switch (intrin->intrinsic) {
 260       case nir_intrinsic_load_var: {
 261          bool per_vertex =
 262             is_per_vertex_input(state, intrin->variables[0]->var) ||
 263             is_per_vertex_output(state, intrin->variables[0]->var);
 264
 265          nir_ssa_def *offset;
 266          nir_ssa_def *vertex_index;
 267
 268          offset = get_io_offset(b, intrin->variables[0],
 269                                 per_vertex ? &vertex_index : NULL,
 270                                 state->type_size);
 271
 272          nir_intrinsic_instr *load =
 273             nir_intrinsic_instr_create(state->mem_ctx,
 274                                        load_op(state, mode, per_vertex));
 275          load->num_components = intrin->num_components;
 276
 277          nir_intrinsic_set_base(load,
 278             intrin->variables[0]->var->data.driver_location);
 279
 280          if (load->intrinsic == nir_intrinsic_load_uniform) {
 281             nir_intrinsic_set_range(load,
 282                state->type_size(intrin->variables[0]->var->type));
 283          }
 284
 285          if (per_vertex)
 286             load->src[0] = nir_src_for_ssa(vertex_index);
 287
 288          load->src[per_vertex ? 1 : 0] = nir_src_for_ssa(offset);
 289
 290          if (intrin->dest.is_ssa) {
 291             nir_ssa_dest_init(&load->instr, &load->dest,
 292                               intrin->num_components,
 293                               intrin->dest.ssa.bit_size, NULL);
 294             nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
 295                                      nir_src_for_ssa(&load->dest.ssa));
 296          } else {
 297             nir_dest_copy(&load->dest, &intrin->dest, state->mem_ctx);
 298          }
 299
 300          nir_instr_insert_before(&intrin->instr, &load->instr);
 301          nir_instr_remove(&intrin->instr);
 302          break;
 303       }
 304
 305       case nir_intrinsic_store_var: {
 306          assert(mode == nir_var_shader_out || mode == nir_var_shared);
 307
 308          nir_ssa_def *offset;
 309          nir_ssa_def *vertex_index;
 310
 311          bool per_vertex =
 312             is_per_vertex_output(state, intrin->variables[0]->var);
 313
 314          offset = get_io_offset(b, intrin->variables[0],
 315                                 per_vertex ? &vertex_index : NULL,
 316                                 state->type_size);
 317
 318          nir_intrinsic_instr *store =
 319             nir_intrinsic_instr_create(state->mem_ctx,
 320                                        store_op(state, mode, per_vertex));
 321          store->num_components = intrin->num_components;
 322
 323          nir_src_copy(&store->src[0], &intrin->src[0], store);
 324
 325          nir_intrinsic_set_base(store,
 326             intrin->variables[0]->var->data.driver_location);
 327          nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
 328
 329          if (per_vertex)
 330             store->src[1] = nir_src_for_ssa(vertex_index);
 331
 332          store->src[per_vertex ? 2 : 1] = nir_src_for_ssa(offset);
 333
 334          nir_instr_insert_before(&intrin->instr, &store->instr);
 335          nir_instr_remove(&intrin->instr);
 336          break;
 337       }
 338
 339       case nir_intrinsic_var_atomic_add:
 340       case nir_intrinsic_var_atomic_imin:
 341       case nir_intrinsic_var_atomic_umin:
 342       case nir_intrinsic_var_atomic_imax:
 343       case nir_intrinsic_var_atomic_umax:
 344       case nir_intrinsic_var_atomic_and:
 345       case nir_intrinsic_var_atomic_or:
 346       case nir_intrinsic_var_atomic_xor:
 347       case nir_intrinsic_var_atomic_exchange:
 348       case nir_intrinsic_var_atomic_comp_swap: {
 349          assert(mode == nir_var_shared);
 350
 351          nir_ssa_def *offset;
 352
 353          offset = get_io_offset(b, intrin->variables[0],
 354                                 NULL, state->type_size);
 355
 356          nir_intrinsic_instr *atomic =
 357             nir_intrinsic_instr_create(state->mem_ctx,
 358                                        atomic_op(intrin->intrinsic));
 359
 360          atomic->src[0] = nir_src_for_ssa(offset);
 361
 362          atomic->const_index[0] =
 363             intrin->variables[0]->var->data.driver_location;
 364
 365          for (unsigned i = 0;
 366               i < nir_op_infos[intrin->intrinsic].num_inputs;
 367               i++) {
 368             nir_src_copy(&atomic->src[i+1], &intrin->src[i], atomic);
 369          }
 370
 371          if (intrin->dest.is_ssa) {
 372             nir_ssa_dest_init(&atomic->instr, &atomic->dest,
 373                               intrin->dest.ssa.num_components,
 374                               intrin->dest.ssa.bit_size, NULL);
 375             nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
 376                                      nir_src_for_ssa(&atomic->dest.ssa));
 377          } else {
 378             nir_dest_copy(&atomic->dest, &intrin->dest, state->mem_ctx);
 379          }
 380
 381          nir_instr_insert_before(&intrin->instr, &atomic->instr);
 382          nir_instr_remove(&intrin->instr);
 383          break;
 384       }
 385
 386       default:
 387          break;
 388       }
 389    }
 390
 391    return true;
 392 }
 393
 394 static void
 395 nir_lower_io_impl(nir_function_impl *impl,
 396                   nir_variable_mode modes,
 397                   int (*type_size)(const struct glsl_type *))
 398 {
 399    struct lower_io_state state;
 400
 401    nir_builder_init(&state.builder, impl);
 402    state.mem_ctx = ralloc_parent(impl);
 403    state.modes = modes;
 404    state.type_size = type_size;
 405
 406    nir_foreach_block(impl, nir_lower_io_block, &state);
 407
 408    nir_metadata_preserve(impl, nir_metadata_block_index |
 409                                nir_metadata_dominance);
 410 }
 411
 412 void
 413 nir_lower_io(nir_shader *shader, nir_variable_mode modes,
 414              int (*type_size)(const struct glsl_type *))
 415 {
 416    nir_foreach_function(shader, function) {
 417       if (function->impl)
 418          nir_lower_io_impl(function->impl, modes, type_size);
 419    }
 420 }
 421
 422 /**
 423  * Return the offset soruce for a load/store intrinsic.
 424  */
 425 nir_src *
 426 nir_get_io_offset_src(nir_intrinsic_instr *instr)
 427 {
 428    switch (instr->intrinsic) {
 429    case nir_intrinsic_load_input:
 430    case nir_intrinsic_load_output:
 431    case nir_intrinsic_load_uniform:
 432       return &instr->src[0];
 433    case nir_intrinsic_load_ubo:
 434    case nir_intrinsic_load_ssbo:
 435    case nir_intrinsic_load_per_vertex_input:
 436    case nir_intrinsic_load_per_vertex_output:
 437    case nir_intrinsic_store_output:
 438       return &instr->src[1];
 439    case nir_intrinsic_store_ssbo:
 440    case nir_intrinsic_store_per_vertex_output:
 441       return &instr->src[2];
 442    default:
 443       return NULL;
 444    }
 445 }
 446
 447 /**
 448  * Return the vertex index source for a load/store per_vertex intrinsic.
 449  */
 450 nir_src *
 451 nir_get_io_vertex_index_src(nir_intrinsic_instr *instr)
 452 {
 453    switch (instr->intrinsic) {
 454    case nir_intrinsic_load_per_vertex_input:
 455    case nir_intrinsic_load_per_vertex_output:
 456       return &instr->src[0];
 457    case nir_intrinsic_store_per_vertex_output:
 458       return &instr->src[1];
 459    default:
 460       return NULL;
 461    }
 462 }