src/glsl/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 mode;
  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    default:
 164       unreachable("Unknown variable mode");
 165    }
 166    return op;
 167 }
 168
 169 static bool
 170 nir_lower_io_block(nir_block *block, void *void_state)
 171 {
 172    struct lower_io_state *state = void_state;
 173
 174    nir_builder *b = &state->builder;
 175
 176    nir_foreach_instr_safe(block, instr) {
 177       if (instr->type != nir_instr_type_intrinsic)
 178          continue;
 179
 180       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 181
 182       if (intrin->intrinsic != nir_intrinsic_load_var &&
 183           intrin->intrinsic != nir_intrinsic_store_var)
 184          continue;
 185
 186       nir_variable_mode mode = intrin->variables[0]->var->data.mode;
 187
 188       if (state->mode != nir_var_all && state->mode != mode)
 189          continue;
 190
 191       if (mode != nir_var_shader_in &&
 192           mode != nir_var_shader_out &&
 193           mode != nir_var_uniform)
 194          continue;
 195
 196       b->cursor = nir_before_instr(instr);
 197
 198       switch (intrin->intrinsic) {
 199       case nir_intrinsic_load_var: {
 200          bool per_vertex =
 201             is_per_vertex_input(state, intrin->variables[0]->var) ||
 202             is_per_vertex_output(state, intrin->variables[0]->var);
 203
 204          nir_ssa_def *offset;
 205          nir_ssa_def *vertex_index;
 206
 207          offset = get_io_offset(b, intrin->variables[0],
 208                                 per_vertex ? &vertex_index : NULL,
 209                                 state->type_size);
 210
 211          nir_intrinsic_instr *load =
 212             nir_intrinsic_instr_create(state->mem_ctx,
 213                                        load_op(state, mode, per_vertex));
 214          load->num_components = intrin->num_components;
 215
 216          load->const_index[0] =
 217             intrin->variables[0]->var->data.driver_location;
 218
 219          if (per_vertex)
 220             load->src[0] = nir_src_for_ssa(vertex_index);
 221
 222          load->src[per_vertex ? 1 : 0] = nir_src_for_ssa(offset);
 223
 224          if (intrin->dest.is_ssa) {
 225             nir_ssa_dest_init(&load->instr, &load->dest,
 226                               intrin->num_components, NULL);
 227             nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
 228                                      nir_src_for_ssa(&load->dest.ssa));
 229          } else {
 230             nir_dest_copy(&load->dest, &intrin->dest, state->mem_ctx);
 231          }
 232
 233          nir_instr_insert_before(&intrin->instr, &load->instr);
 234          nir_instr_remove(&intrin->instr);
 235          break;
 236       }
 237
 238       case nir_intrinsic_store_var: {
 239          assert(mode == nir_var_shader_out);
 240
 241          nir_ssa_def *offset;
 242          nir_ssa_def *vertex_index;
 243
 244          bool per_vertex =
 245             is_per_vertex_output(state, intrin->variables[0]->var);
 246
 247          offset = get_io_offset(b, intrin->variables[0],
 248                                 per_vertex ? &vertex_index : NULL,
 249                                 state->type_size);
 250
 251          nir_intrinsic_op store_op =
 252             per_vertex ? nir_intrinsic_store_per_vertex_output :
 253                          nir_intrinsic_store_output;
 254
 255          nir_intrinsic_instr *store = nir_intrinsic_instr_create(state->mem_ctx,
 256                                                                  store_op);
 257          store->num_components = intrin->num_components;
 258
 259          nir_src_copy(&store->src[0], &intrin->src[0], store);
 260
 261          store->const_index[0] =
 262             intrin->variables[0]->var->data.driver_location;
 263
 264          if (per_vertex)
 265             store->src[1] = nir_src_for_ssa(vertex_index);
 266
 267          store->src[per_vertex ? 2 : 1] = nir_src_for_ssa(offset);
 268
 269          nir_instr_insert_before(&intrin->instr, &store->instr);
 270          nir_instr_remove(&intrin->instr);
 271          break;
 272       }
 273
 274       default:
 275          break;
 276       }
 277    }
 278
 279    return true;
 280 }
 281
 282 static void
 283 nir_lower_io_impl(nir_function_impl *impl,
 284                   nir_variable_mode mode,
 285                   int (*type_size)(const struct glsl_type *))
 286 {
 287    struct lower_io_state state;
 288
 289    nir_builder_init(&state.builder, impl);
 290    state.mem_ctx = ralloc_parent(impl);
 291    state.mode = mode;
 292    state.type_size = type_size;
 293
 294    nir_foreach_block(impl, nir_lower_io_block, &state);
 295
 296    nir_metadata_preserve(impl, nir_metadata_block_index |
 297                                nir_metadata_dominance);
 298 }
 299
 300 void
 301 nir_lower_io(nir_shader *shader, nir_variable_mode mode,
 302              int (*type_size)(const struct glsl_type *))
 303 {
 304    nir_foreach_overload(shader, overload) {
 305       if (overload->impl)
 306          nir_lower_io_impl(overload->impl, mode, type_size);
 307    }
 308 }
 309
 310 /**
 311  * Return the offset soruce for a load/store intrinsic.
 312  */
 313 nir_src *
 314 nir_get_io_offset_src(nir_intrinsic_instr *instr)
 315 {
 316    switch (instr->intrinsic) {
 317    case nir_intrinsic_load_input:
 318    case nir_intrinsic_load_output:
 319    case nir_intrinsic_load_uniform:
 320       return &instr->src[0];
 321    case nir_intrinsic_load_per_vertex_input:
 322    case nir_intrinsic_load_per_vertex_output:
 323    case nir_intrinsic_store_output:
 324       return &instr->src[1];
 325    case nir_intrinsic_store_per_vertex_output:
 326       return &instr->src[2];
 327    default:
 328       return NULL;
 329    }
 330 }
 331
 332 /**
 333  * Return the vertex index source for a load/store per_vertex intrinsic.
 334  */
 335 nir_src *
 336 nir_get_io_vertex_index_src(nir_intrinsic_instr *instr)
 337 {
 338    switch (instr->intrinsic) {
 339    case nir_intrinsic_load_per_vertex_input:
 340    case nir_intrinsic_load_per_vertex_output:
 341       return &instr->src[0];
 342    case nir_intrinsic_store_per_vertex_output:
 343       return &instr->src[1];
 344    default:
 345       return NULL;
 346    }
 347 }