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                          unsigned base_offset,
  47                          int (*type_size)(const struct glsl_type *))
  48 {
  49    unsigned location = 0;
  50
  51    /* There are 32 regular and 32 patch varyings allowed */
  52    int locations[64][2];
  53    for (unsigned i = 0; i < 64; i++) {
  54       for (unsigned j = 0; j < 2; j++)
  55          locations[i][j] = -1;
  56    }
  57
  58    nir_foreach_variable(var, var_list) {
  59       /*
  60        * UBO's have their own address spaces, so don't count them towards the
  61        * number of global uniforms
  62        */
  63       if ((var->data.mode == nir_var_uniform || var->data.mode == nir_var_shader_storage) &&
  64           var->interface_type != NULL)
  65          continue;
  66
  67       /* Make sure we give the same location to varyings packed with
  68        * ARB_enhanced_layouts.
  69        */
  70       int idx = var->data.location - base_offset;
  71       if (base_offset && idx >= 0) {
  72          assert(idx < ARRAY_SIZE(locations));
  73
  74          if (locations[idx][var->data.index] == -1) {
  75             var->data.driver_location = location;
  76             locations[idx][var->data.index] = location;
  77             location += type_size(var->type);
  78          } else {
  79             var->data.driver_location = locations[idx][var->data.index];
  80          }
  81       } else {
  82          var->data.driver_location = location;
  83          location += type_size(var->type);
  84       }
  85    }
  86
  87    *size = location;
  88 }
  89
  90 /**
  91  * Returns true if we're processing a stage whose inputs are arrays indexed
  92  * by a vertex number (such as geometry shader inputs).
  93  */
  94 static bool
  95 is_per_vertex_input(struct lower_io_state *state, nir_variable *var)
  96 {
  97    gl_shader_stage stage = state->builder.shader->stage;
  98
  99    return var->data.mode == nir_var_shader_in && !var->data.patch &&
 100           (stage == MESA_SHADER_TESS_CTRL ||
 101            stage == MESA_SHADER_TESS_EVAL ||
 102            stage == MESA_SHADER_GEOMETRY);
 103 }
 104
 105 static bool
 106 is_per_vertex_output(struct lower_io_state *state, nir_variable *var)
 107 {
 108    gl_shader_stage stage = state->builder.shader->stage;
 109    return var->data.mode == nir_var_shader_out && !var->data.patch &&
 110           stage == MESA_SHADER_TESS_CTRL;
 111 }
 112
 113 static nir_ssa_def *
 114 get_io_offset(nir_builder *b, nir_deref_var *deref,
 115               nir_ssa_def **vertex_index,
 116               int (*type_size)(const struct glsl_type *))
 117 {
 118    nir_deref *tail = &deref->deref;
 119
 120    /* For per-vertex input arrays (i.e. geometry shader inputs), keep the
 121     * outermost array index separate.  Process the rest normally.
 122     */
 123    if (vertex_index != NULL) {
 124       tail = tail->child;
 125       assert(tail->deref_type == nir_deref_type_array);
 126       nir_deref_array *deref_array = nir_deref_as_array(tail);
 127
 128       nir_ssa_def *vtx = nir_imm_int(b, deref_array->base_offset);
 129       if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
 130          vtx = nir_iadd(b, vtx, nir_ssa_for_src(b, deref_array->indirect, 1));
 131       }
 132       *vertex_index = vtx;
 133    }
 134
 135    /* Just emit code and let constant-folding go to town */
 136    nir_ssa_def *offset = nir_imm_int(b, 0);
 137
 138    while (tail->child != NULL) {
 139       const struct glsl_type *parent_type = tail->type;
 140       tail = tail->child;
 141
 142       if (tail->deref_type == nir_deref_type_array) {
 143          nir_deref_array *deref_array = nir_deref_as_array(tail);
 144          unsigned size = type_size(tail->type);
 145
 146          offset = nir_iadd(b, offset,
 147                            nir_imm_int(b, size * deref_array->base_offset));
 148
 149          if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
 150             nir_ssa_def *mul =
 151                nir_imul(b, nir_imm_int(b, size),
 152                         nir_ssa_for_src(b, deref_array->indirect, 1));
 153
 154             offset = nir_iadd(b, offset, mul);
 155          }
 156       } else if (tail->deref_type == nir_deref_type_struct) {
 157          nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
 158
 159          unsigned field_offset = 0;
 160          for (unsigned i = 0; i < deref_struct->index; i++) {
 161             field_offset += type_size(glsl_get_struct_field(parent_type, i));
 162          }
 163          offset = nir_iadd(b, offset, nir_imm_int(b, field_offset));
 164       }
 165    }
 166
 167    return offset;
 168 }
 169
 170 static nir_intrinsic_op
 171 load_op(nir_variable_mode mode, bool per_vertex)
 172 {
 173    nir_intrinsic_op op;
 174    switch (mode) {
 175    case nir_var_shader_in:
 176       op = per_vertex ? nir_intrinsic_load_per_vertex_input :
 177                         nir_intrinsic_load_input;
 178       break;
 179    case nir_var_shader_out:
 180       op = per_vertex ? nir_intrinsic_load_per_vertex_output :
 181                         nir_intrinsic_load_output;
 182       break;
 183    case nir_var_uniform:
 184       op = nir_intrinsic_load_uniform;
 185       break;
 186    case nir_var_shared:
 187       op = nir_intrinsic_load_shared;
 188       break;
 189    default:
 190       unreachable("Unknown variable mode");
 191    }
 192    return op;
 193 }
 194
 195 static nir_intrinsic_op
 196 store_op(struct lower_io_state *state,
 197          nir_variable_mode mode, bool per_vertex)
 198 {
 199    nir_intrinsic_op op;
 200    switch (mode) {
 201    case nir_var_shader_out:
 202       op = per_vertex ? nir_intrinsic_store_per_vertex_output :
 203                         nir_intrinsic_store_output;
 204       break;
 205    case nir_var_shared:
 206       op = nir_intrinsic_store_shared;
 207       break;
 208    default:
 209       unreachable("Unknown variable mode");
 210    }
 211    return op;
 212 }
 213
 214 static nir_intrinsic_op
 215 atomic_op(nir_intrinsic_op opcode)
 216 {
 217    switch (opcode) {
 218 #define OP(O) case nir_intrinsic_var_##O: return nir_intrinsic_shared_##O;
 219    OP(atomic_exchange)
 220    OP(atomic_comp_swap)
 221    OP(atomic_add)
 222    OP(atomic_imin)
 223    OP(atomic_umin)
 224    OP(atomic_imax)
 225    OP(atomic_umax)
 226    OP(atomic_and)
 227    OP(atomic_or)
 228    OP(atomic_xor)
 229 #undef OP
 230    default:
 231       unreachable("Invalid atomic");
 232    }
 233 }
 234
 235 static bool
 236 nir_lower_io_block(nir_block *block,
 237                    struct lower_io_state *state)
 238 {
 239    nir_builder *b = &state->builder;
 240
 241    nir_foreach_instr_safe(instr, block) {
 242       if (instr->type != nir_instr_type_intrinsic)
 243          continue;
 244
 245       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 246
 247       switch (intrin->intrinsic) {
 248       case nir_intrinsic_load_var:
 249       case nir_intrinsic_store_var:
 250       case nir_intrinsic_var_atomic_add:
 251       case nir_intrinsic_var_atomic_imin:
 252       case nir_intrinsic_var_atomic_umin:
 253       case nir_intrinsic_var_atomic_imax:
 254       case nir_intrinsic_var_atomic_umax:
 255       case nir_intrinsic_var_atomic_and:
 256       case nir_intrinsic_var_atomic_or:
 257       case nir_intrinsic_var_atomic_xor:
 258       case nir_intrinsic_var_atomic_exchange:
 259       case nir_intrinsic_var_atomic_comp_swap:
 260          /* We can lower the io for this nir instrinsic */
 261          break;
 262       default:
 263          /* We can't lower the io for this nir instrinsic, so skip it */
 264          continue;
 265       }
 266
 267       nir_variable *var = intrin->variables[0]->var;
 268       nir_variable_mode mode = var->data.mode;
 269
 270       if ((state->modes & mode) == 0)
 271          continue;
 272
 273       if (mode != nir_var_shader_in &&
 274           mode != nir_var_shader_out &&
 275           mode != nir_var_shared &&
 276           mode != nir_var_uniform)
 277          continue;
 278
 279       b->cursor = nir_before_instr(instr);
 280
 281       const bool per_vertex =
 282          is_per_vertex_input(state, var) || is_per_vertex_output(state, var);
 283
 284       nir_ssa_def *offset;
 285       nir_ssa_def *vertex_index;
 286
 287       offset = get_io_offset(b, intrin->variables[0],
 288                              per_vertex ? &vertex_index : NULL,
 289                              state->type_size);
 290
 291       nir_intrinsic_instr *replacement;
 292
 293       switch (intrin->intrinsic) {
 294       case nir_intrinsic_load_var: {
 295          nir_intrinsic_instr *load =
 296             nir_intrinsic_instr_create(state->mem_ctx,
 297                                        load_op(mode, per_vertex));
 298          load->num_components = intrin->num_components;
 299
 300          nir_intrinsic_set_base(load,
 301             var->data.driver_location);
 302          if (mode == nir_var_shader_in || mode == nir_var_shader_out) {
 303             nir_intrinsic_set_component(load, var->data.location_frac);
 304          }
 305
 306          if (load->intrinsic == nir_intrinsic_load_uniform) {
 307             nir_intrinsic_set_range(load, state->type_size(var->type));
 308          }
 309
 310          if (per_vertex)
 311             load->src[0] = nir_src_for_ssa(vertex_index);
 312
 313          load->src[per_vertex ? 1 : 0] = nir_src_for_ssa(offset);
 314
 315          replacement = load;
 316          break;
 317       }
 318
 319       case nir_intrinsic_store_var: {
 320          assert(mode == nir_var_shader_out || mode == nir_var_shared);
 321
 322          nir_intrinsic_instr *store =
 323             nir_intrinsic_instr_create(state->mem_ctx,
 324                                        store_op(state, mode, per_vertex));
 325          store->num_components = intrin->num_components;
 326
 327          nir_src_copy(&store->src[0], &intrin->src[0], store);
 328
 329          nir_intrinsic_set_base(store,
 330             var->data.driver_location);
 331          if (mode == nir_var_shader_out) {
 332             nir_intrinsic_set_component(store, var->data.location_frac);
 333          }
 334          nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
 335
 336          if (per_vertex)
 337             store->src[1] = nir_src_for_ssa(vertex_index);
 338
 339          store->src[per_vertex ? 2 : 1] = nir_src_for_ssa(offset);
 340
 341          replacement = store;
 342          break;
 343       }
 344
 345       case nir_intrinsic_var_atomic_add:
 346       case nir_intrinsic_var_atomic_imin:
 347       case nir_intrinsic_var_atomic_umin:
 348       case nir_intrinsic_var_atomic_imax:
 349       case nir_intrinsic_var_atomic_umax:
 350       case nir_intrinsic_var_atomic_and:
 351       case nir_intrinsic_var_atomic_or:
 352       case nir_intrinsic_var_atomic_xor:
 353       case nir_intrinsic_var_atomic_exchange:
 354       case nir_intrinsic_var_atomic_comp_swap: {
 355          assert(mode == nir_var_shared);
 356
 357          nir_intrinsic_instr *atomic =
 358             nir_intrinsic_instr_create(state->mem_ctx,
 359                                        atomic_op(intrin->intrinsic));
 360
 361          atomic->src[0] = nir_src_for_ssa(offset);
 362
 363          atomic->const_index[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          replacement = atomic;
 372          break;
 373       }
 374
 375       default:
 376          break;
 377       }
 378
 379       if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
 380          if (intrin->dest.is_ssa) {
 381             nir_ssa_dest_init(&replacement->instr, &replacement->dest,
 382                               intrin->num_components,
 383                               intrin->dest.ssa.bit_size, NULL);
 384             nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
 385                                      nir_src_for_ssa(&replacement->dest.ssa));
 386          } else {
 387             nir_dest_copy(&replacement->dest, &intrin->dest, state->mem_ctx);
 388          }
 389       }
 390
 391       nir_instr_insert_before(&intrin->instr, &replacement->instr);
 392       nir_instr_remove(&intrin->instr);
 393    }
 394
 395    return true;
 396 }
 397
 398 static void
 399 nir_lower_io_impl(nir_function_impl *impl,
 400                   nir_variable_mode modes,
 401                   int (*type_size)(const struct glsl_type *))
 402 {
 403    struct lower_io_state state;
 404
 405    nir_builder_init(&state.builder, impl);
 406    state.mem_ctx = ralloc_parent(impl);
 407    state.modes = modes;
 408    state.type_size = type_size;
 409
 410    nir_foreach_block(block, impl) {
 411       nir_lower_io_block(block, &state);
 412    }
 413
 414    nir_metadata_preserve(impl, nir_metadata_block_index |
 415                                nir_metadata_dominance);
 416 }
 417
 418 void
 419 nir_lower_io(nir_shader *shader, nir_variable_mode modes,
 420              int (*type_size)(const struct glsl_type *))
 421 {
 422    nir_foreach_function(function, shader) {
 423       if (function->impl)
 424          nir_lower_io_impl(function->impl, modes, type_size);
 425    }
 426 }
 427
 428 /**
 429  * Return the offset soruce for a load/store intrinsic.
 430  */
 431 nir_src *
 432 nir_get_io_offset_src(nir_intrinsic_instr *instr)
 433 {
 434    switch (instr->intrinsic) {
 435    case nir_intrinsic_load_input:
 436    case nir_intrinsic_load_output:
 437    case nir_intrinsic_load_uniform:
 438       return &instr->src[0];
 439    case nir_intrinsic_load_ubo:
 440    case nir_intrinsic_load_ssbo:
 441    case nir_intrinsic_load_per_vertex_input:
 442    case nir_intrinsic_load_per_vertex_output:
 443    case nir_intrinsic_store_output:
 444       return &instr->src[1];
 445    case nir_intrinsic_store_ssbo:
 446    case nir_intrinsic_store_per_vertex_output:
 447       return &instr->src[2];
 448    default:
 449       return NULL;
 450    }
 451 }
 452
 453 /**
 454  * Return the vertex index source for a load/store per_vertex intrinsic.
 455  */
 456 nir_src *
 457 nir_get_io_vertex_index_src(nir_intrinsic_instr *instr)
 458 {
 459    switch (instr->intrinsic) {
 460    case nir_intrinsic_load_per_vertex_input:
 461    case nir_intrinsic_load_per_vertex_output:
 462       return &instr->src[0];
 463    case nir_intrinsic_store_per_vertex_output:
 464       return &instr->src[1];
 465    default:
 466       return NULL;
 467    }
 468 }