src/amd/vulkan/radv_shader_info.c

   1 /*
   2  * Copyright © 2017 Red Hat
   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 #include "radv_private.h"
  24 #include "radv_shader.h"
  25 #include "nir/nir.h"
  26 #include "nir/nir_deref.h"
  27 #include "nir/nir_xfb_info.h"
  28
  29 static void mark_sampler_desc(const nir_variable *var,
  30                               struct radv_shader_info *info)
  31 {
  32         info->desc_set_used_mask |= (1 << var->data.descriptor_set);
  33 }
  34
  35 static void mark_ls_output(struct radv_shader_info *info,
  36                            uint32_t param, int num_slots)
  37 {
  38         uint64_t mask = (1ull << num_slots) - 1ull;
  39         info->vs.ls_outputs_written |= (mask << param);
  40 }
  41
  42 static void mark_tess_output(struct radv_shader_info *info,
  43                              bool is_patch, uint32_t param, int num_slots)
  44 {
  45         uint64_t mask = (1ull << num_slots) - 1ull;
  46         if (is_patch)
  47                 info->tcs.patch_outputs_written |= (mask << param);
  48         else
  49                 info->tcs.outputs_written |= (mask << param);
  50 }
  51
  52 static void
  53 get_deref_offset(nir_deref_instr *instr,
  54                  unsigned *const_out)
  55 {
  56         nir_variable *var = nir_deref_instr_get_variable(instr);
  57         nir_deref_path path;
  58         unsigned idx_lvl = 1;
  59
  60         if (var->data.compact) {
  61                 assert(instr->deref_type == nir_deref_type_array);
  62                 *const_out = nir_src_as_uint(instr->arr.index);
  63                 return;
  64         }
  65
  66         nir_deref_path_init(&path, instr, NULL);
  67
  68         uint32_t const_offset = 0;
  69
  70         for (; path.path[idx_lvl]; ++idx_lvl) {
  71                 const struct glsl_type *parent_type = path.path[idx_lvl - 1]->type;
  72                 if (path.path[idx_lvl]->deref_type == nir_deref_type_struct) {
  73                         unsigned index = path.path[idx_lvl]->strct.index;
  74
  75                         for (unsigned i = 0; i < index; i++) {
  76                                 const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
  77                                 const_offset += glsl_count_attribute_slots(ft, false);
  78                         }
  79                 } else if(path.path[idx_lvl]->deref_type == nir_deref_type_array) {
  80                         unsigned size = glsl_count_attribute_slots(path.path[idx_lvl]->type, false);
  81                         if (nir_src_is_const(path.path[idx_lvl]->arr.index))
  82                                 const_offset += nir_src_as_uint(path.path[idx_lvl]->arr.index) * size;
  83                 } else
  84                         unreachable("Uhandled deref type in get_deref_instr_offset");
  85         }
  86
  87         *const_out = const_offset;
  88
  89         nir_deref_path_finish(&path);
  90 }
  91
  92 static void
  93 gather_intrinsic_load_deref_info(const nir_shader *nir,
  94                                const nir_intrinsic_instr *instr,
  95                                struct radv_shader_info *info)
  96 {
  97         switch (nir->info.stage) {
  98         case MESA_SHADER_VERTEX: {
  99                 nir_variable *var = nir_deref_instr_get_variable(nir_instr_as_deref(instr->src[0].ssa->parent_instr));
 100
 101                 if (var && var->data.mode == nir_var_shader_in) {
 102                         unsigned idx = var->data.location;
 103                         uint8_t mask = nir_ssa_def_components_read(&instr->dest.ssa);
 104
 105                         info->vs.input_usage_mask[idx] |=
 106                                 mask << var->data.location_frac;
 107                 }
 108                 break;
 109         }
 110         default:
 111                 break;
 112         }
 113 }
 114
 115 static uint32_t
 116 widen_writemask(uint32_t wrmask)
 117 {
 118         uint32_t new_wrmask = 0;
 119         for(unsigned i = 0; i < 4; i++)
 120                 new_wrmask |= (wrmask & (1 << i) ? 0x3 : 0x0) << (i * 2);
 121         return new_wrmask;
 122 }
 123
 124 static void
 125 set_output_usage_mask(const nir_shader *nir, const nir_intrinsic_instr *instr,
 126                       uint8_t *output_usage_mask)
 127 {
 128         nir_deref_instr *deref_instr =
 129                 nir_instr_as_deref(instr->src[0].ssa->parent_instr);
 130         nir_variable *var = nir_deref_instr_get_variable(deref_instr);
 131         unsigned attrib_count = glsl_count_attribute_slots(deref_instr->type, false);
 132         unsigned idx = var->data.location;
 133         unsigned comp = var->data.location_frac;
 134         unsigned const_offset = 0;
 135
 136         get_deref_offset(deref_instr, &const_offset);
 137
 138         if (var->data.compact) {
 139                 assert(!glsl_type_is_64bit(deref_instr->type));
 140                 const_offset += comp;
 141                 output_usage_mask[idx + const_offset / 4] |= 1 << (const_offset % 4);
 142                 return;
 143         }
 144
 145         uint32_t wrmask = nir_intrinsic_write_mask(instr);
 146         if (glsl_type_is_64bit(deref_instr->type))
 147                 wrmask = widen_writemask(wrmask);
 148
 149         for (unsigned i = 0; i < attrib_count; i++)
 150                 output_usage_mask[idx + i + const_offset] |=
 151                         ((wrmask >> (i * 4)) & 0xf) << comp;
 152 }
 153
 154 static void
 155 gather_intrinsic_store_deref_info(const nir_shader *nir,
 156                                 const nir_intrinsic_instr *instr,
 157                                 struct radv_shader_info *info)
 158 {
 159         nir_variable *var = nir_deref_instr_get_variable(nir_instr_as_deref(instr->src[0].ssa->parent_instr));
 160
 161         if (var && var->data.mode == nir_var_shader_out) {
 162                 unsigned idx = var->data.location;
 163
 164                 switch (nir->info.stage) {
 165                 case MESA_SHADER_VERTEX:
 166                         set_output_usage_mask(nir, instr,
 167                                               info->vs.output_usage_mask);
 168                         break;
 169                 case MESA_SHADER_GEOMETRY:
 170                         set_output_usage_mask(nir, instr,
 171                                               info->gs.output_usage_mask);
 172                         break;
 173                 case MESA_SHADER_TESS_EVAL:
 174                         set_output_usage_mask(nir, instr,
 175                                               info->tes.output_usage_mask);
 176                         break;
 177                 case MESA_SHADER_TESS_CTRL: {
 178                         unsigned param = shader_io_get_unique_index(idx);
 179                         const struct glsl_type *type = var->type;
 180
 181                         if (!var->data.patch)
 182                                 type = glsl_get_array_element(var->type);
 183
 184                         unsigned slots =
 185                                 var->data.compact ? DIV_ROUND_UP(var->data.location_frac + glsl_get_length(type), 4)
 186                                                   : glsl_count_attribute_slots(type, false);
 187
 188                         mark_tess_output(info, var->data.patch, param, slots);
 189                         break;
 190                 }
 191                 default:
 192                         break;
 193                 }
 194         }
 195 }
 196
 197 static void
 198 gather_push_constant_info(const nir_shader *nir,
 199                           const nir_intrinsic_instr *instr,
 200                           struct radv_shader_info *info)
 201 {
 202         int base = nir_intrinsic_base(instr);
 203
 204         if (!nir_src_is_const(instr->src[0])) {
 205                 info->has_indirect_push_constants = true;
 206         } else {
 207                 uint32_t min = base + nir_src_as_uint(instr->src[0]);
 208                 uint32_t max = min + instr->num_components * 4;
 209
 210                 info->max_push_constant_used =
 211                         MAX2(max, info->max_push_constant_used);
 212                 info->min_push_constant_used =
 213                         MIN2(min, info->min_push_constant_used);
 214         }
 215
 216         if (instr->dest.ssa.bit_size != 32)
 217                 info->has_only_32bit_push_constants = false;
 218
 219         info->loads_push_constants = true;
 220 }
 221
 222 static void
 223 gather_intrinsic_info(const nir_shader *nir, const nir_intrinsic_instr *instr,
 224                       struct radv_shader_info *info)
 225 {
 226         switch (instr->intrinsic) {
 227         case nir_intrinsic_interp_deref_at_sample:
 228                 info->ps.needs_sample_positions = true;
 229                 break;
 230         case nir_intrinsic_load_draw_id:
 231                 info->vs.needs_draw_id = true;
 232                 break;
 233         case nir_intrinsic_load_instance_id:
 234                 info->vs.needs_instance_id = true;
 235                 break;
 236         case nir_intrinsic_load_num_work_groups:
 237                 info->cs.uses_grid_size = true;
 238                 break;
 239         case nir_intrinsic_load_local_invocation_id:
 240         case nir_intrinsic_load_work_group_id: {
 241                 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
 242                 while (mask) {
 243                         unsigned i = u_bit_scan(&mask);
 244
 245                         if (instr->intrinsic == nir_intrinsic_load_work_group_id)
 246                                 info->cs.uses_block_id[i] = true;
 247                         else
 248                                 info->cs.uses_thread_id[i] = true;
 249                 }
 250                 break;
 251         }
 252         case nir_intrinsic_load_local_invocation_index:
 253         case nir_intrinsic_load_subgroup_id:
 254         case nir_intrinsic_load_num_subgroups:
 255                 info->cs.uses_local_invocation_idx = true;
 256                 break;
 257         case nir_intrinsic_load_sample_id:
 258                 info->ps.force_persample = true;
 259                 break;
 260         case nir_intrinsic_load_sample_pos:
 261                 info->ps.force_persample = true;
 262                 break;
 263         case nir_intrinsic_load_view_index:
 264                 info->needs_multiview_view_index = true;
 265                 if (nir->info.stage == MESA_SHADER_FRAGMENT)
 266                         info->ps.layer_input = true;
 267                 break;
 268         case nir_intrinsic_load_invocation_id:
 269                 info->uses_invocation_id = true;
 270                 break;
 271         case nir_intrinsic_load_primitive_id:
 272                 info->uses_prim_id = true;
 273                 break;
 274         case nir_intrinsic_load_push_constant:
 275                 gather_push_constant_info(nir, instr, info);
 276                 break;
 277         case nir_intrinsic_vulkan_resource_index:
 278                 info->desc_set_used_mask |= (1 << nir_intrinsic_desc_set(instr));
 279                 break;
 280         case nir_intrinsic_image_deref_load:
 281         case nir_intrinsic_image_deref_store:
 282         case nir_intrinsic_image_deref_atomic_add:
 283         case nir_intrinsic_image_deref_atomic_min:
 284         case nir_intrinsic_image_deref_atomic_max:
 285         case nir_intrinsic_image_deref_atomic_and:
 286         case nir_intrinsic_image_deref_atomic_or:
 287         case nir_intrinsic_image_deref_atomic_xor:
 288         case nir_intrinsic_image_deref_atomic_exchange:
 289         case nir_intrinsic_image_deref_atomic_comp_swap:
 290         case nir_intrinsic_image_deref_size: {
 291                 nir_variable *var = nir_deref_instr_get_variable(nir_instr_as_deref(instr->src[0].ssa->parent_instr));
 292                 const struct glsl_type *type = glsl_without_array(var->type);
 293
 294                 enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
 295                 if (dim == GLSL_SAMPLER_DIM_SUBPASS ||
 296                     dim == GLSL_SAMPLER_DIM_SUBPASS_MS) {
 297                         info->ps.layer_input = true;
 298                         info->ps.uses_input_attachments = true;
 299                 }
 300                 mark_sampler_desc(var, info);
 301
 302                 if (instr->intrinsic == nir_intrinsic_image_deref_store ||
 303                     instr->intrinsic == nir_intrinsic_image_deref_atomic_add ||
 304                     instr->intrinsic == nir_intrinsic_image_deref_atomic_min ||
 305                     instr->intrinsic == nir_intrinsic_image_deref_atomic_max ||
 306                     instr->intrinsic == nir_intrinsic_image_deref_atomic_and ||
 307                     instr->intrinsic == nir_intrinsic_image_deref_atomic_or ||
 308                     instr->intrinsic == nir_intrinsic_image_deref_atomic_xor ||
 309                     instr->intrinsic == nir_intrinsic_image_deref_atomic_exchange ||
 310                     instr->intrinsic == nir_intrinsic_image_deref_atomic_comp_swap) {
 311                         if (nir->info.stage == MESA_SHADER_FRAGMENT)
 312                                 info->ps.writes_memory = true;
 313                 }
 314                 break;
 315         }
 316         case nir_intrinsic_store_ssbo:
 317         case nir_intrinsic_ssbo_atomic_add:
 318         case nir_intrinsic_ssbo_atomic_imin:
 319         case nir_intrinsic_ssbo_atomic_umin:
 320         case nir_intrinsic_ssbo_atomic_imax:
 321         case nir_intrinsic_ssbo_atomic_umax:
 322         case nir_intrinsic_ssbo_atomic_and:
 323         case nir_intrinsic_ssbo_atomic_or:
 324         case nir_intrinsic_ssbo_atomic_xor:
 325         case nir_intrinsic_ssbo_atomic_exchange:
 326         case nir_intrinsic_ssbo_atomic_comp_swap:
 327                 if (nir->info.stage == MESA_SHADER_FRAGMENT)
 328                         info->ps.writes_memory = true;
 329                 break;
 330         case nir_intrinsic_load_deref:
 331                 gather_intrinsic_load_deref_info(nir, instr, info);
 332                 break;
 333         case nir_intrinsic_store_deref:
 334                 gather_intrinsic_store_deref_info(nir, instr, info);
 335                 break;
 336         default:
 337                 break;
 338         }
 339 }
 340
 341 static void
 342 gather_tex_info(const nir_shader *nir, const nir_tex_instr *instr,
 343                 struct radv_shader_info *info)
 344 {
 345         for (unsigned i = 0; i < instr->num_srcs; i++) {
 346                 switch (instr->src[i].src_type) {
 347                 case nir_tex_src_texture_deref:
 348                         mark_sampler_desc(nir_deref_instr_get_variable(nir_src_as_deref(instr->src[i].src)), info);
 349                         break;
 350                 case nir_tex_src_sampler_deref:
 351                         mark_sampler_desc(nir_deref_instr_get_variable(nir_src_as_deref(instr->src[i].src)), info);
 352                         break;
 353                 default:
 354                         break;
 355                 }
 356         }
 357 }
 358
 359 static void
 360 gather_info_block(const nir_shader *nir, const nir_block *block,
 361                   struct radv_shader_info *info)
 362 {
 363         nir_foreach_instr(instr, block) {
 364                 switch (instr->type) {
 365                 case nir_instr_type_intrinsic:
 366                         gather_intrinsic_info(nir, nir_instr_as_intrinsic(instr), info);
 367                         break;
 368                 case nir_instr_type_tex:
 369                         gather_tex_info(nir, nir_instr_as_tex(instr), info);
 370                         break;
 371                 default:
 372                         break;
 373                 }
 374         }
 375 }
 376
 377 static void
 378 gather_info_input_decl_vs(const nir_shader *nir, const nir_variable *var,
 379                           struct radv_shader_info *info)
 380 {
 381         int idx = var->data.location;
 382
 383         if (idx >= VERT_ATTRIB_GENERIC0 && idx <= VERT_ATTRIB_GENERIC15)
 384                 info->vs.has_vertex_buffers = true;
 385 }
 386
 387 static void
 388 gather_info_input_decl_ps(const nir_shader *nir, const nir_variable *var,
 389                           struct radv_shader_info *info)
 390 {
 391         unsigned attrib_count = glsl_count_attribute_slots(var->type, false);
 392         const struct glsl_type *type = glsl_without_array(var->type);
 393         int idx = var->data.location;
 394
 395         switch (idx) {
 396         case VARYING_SLOT_PNTC:
 397                 info->ps.has_pcoord = true;
 398                 break;
 399         case VARYING_SLOT_PRIMITIVE_ID:
 400                 info->ps.prim_id_input = true;
 401                 break;
 402         case VARYING_SLOT_LAYER:
 403                 info->ps.layer_input = true;
 404                 break;
 405         case VARYING_SLOT_CLIP_DIST0:
 406         case VARYING_SLOT_CLIP_DIST1:
 407                 info->ps.num_input_clips_culls += attrib_count;
 408                 break;
 409         default:
 410                 break;
 411         }
 412
 413         if (glsl_get_base_type(type) == GLSL_TYPE_FLOAT) {
 414                 if (var->data.sample)
 415                         info->ps.force_persample = true;
 416         }
 417 }
 418
 419 static void
 420 gather_info_input_decl(const nir_shader *nir, const nir_variable *var,
 421                        struct radv_shader_info *info)
 422 {
 423         switch (nir->info.stage) {
 424         case MESA_SHADER_VERTEX:
 425                 gather_info_input_decl_vs(nir, var, info);
 426                 break;
 427         case MESA_SHADER_FRAGMENT:
 428                 gather_info_input_decl_ps(nir, var, info);
 429                 break;
 430         default:
 431                 break;
 432         }
 433 }
 434
 435 static void
 436 gather_info_output_decl_ls(const nir_shader *nir, const nir_variable *var,
 437                            struct radv_shader_info *info)
 438 {
 439         int idx = var->data.location;
 440         unsigned param = shader_io_get_unique_index(idx);
 441         int num_slots = glsl_count_attribute_slots(var->type, false);
 442         if (var->data.compact)
 443                 num_slots = DIV_ROUND_UP(var->data.location_frac + glsl_get_length(var->type), 4);
 444         mark_ls_output(info, param, num_slots);
 445 }
 446
 447 static void
 448 gather_info_output_decl_ps(const nir_shader *nir, const nir_variable *var,
 449                            struct radv_shader_info *info)
 450 {
 451         int idx = var->data.location;
 452
 453         switch (idx) {
 454         case FRAG_RESULT_DEPTH:
 455                 info->ps.writes_z = true;
 456                 break;
 457         case FRAG_RESULT_STENCIL:
 458                 info->ps.writes_stencil = true;
 459                 break;
 460         case FRAG_RESULT_SAMPLE_MASK:
 461                 info->ps.writes_sample_mask = true;
 462                 break;
 463         default:
 464                 break;
 465         }
 466 }
 467
 468 static void
 469 gather_info_output_decl_gs(const nir_shader *nir, const nir_variable *var,
 470                            struct radv_shader_info *info)
 471 {
 472         unsigned num_components = glsl_get_component_slots(var->type);
 473         unsigned stream = var->data.stream;
 474         unsigned idx = var->data.location;
 475
 476         assert(stream < 4);
 477
 478         info->gs.max_stream = MAX2(info->gs.max_stream, stream);
 479         info->gs.num_stream_output_components[stream] += num_components;
 480         info->gs.output_streams[idx] = stream;
 481 }
 482
 483 static void
 484 gather_info_output_decl(const nir_shader *nir, const nir_variable *var,
 485                         struct radv_shader_info *info,
 486                         const struct radv_nir_compiler_options *options)
 487 {
 488         switch (nir->info.stage) {
 489         case MESA_SHADER_FRAGMENT:
 490                 gather_info_output_decl_ps(nir, var, info);
 491                 break;
 492         case MESA_SHADER_VERTEX:
 493                 if (options->key.vs.as_ls)
 494                         gather_info_output_decl_ls(nir, var, info);
 495                 break;
 496         case MESA_SHADER_GEOMETRY:
 497                 gather_info_output_decl_gs(nir, var, info);
 498                 break;
 499         default:
 500                 break;
 501         }
 502 }
 503
 504 static void
 505 gather_xfb_info(const nir_shader *nir, struct radv_shader_info *info)
 506 {
 507         nir_xfb_info *xfb = nir_gather_xfb_info(nir, NULL);
 508         struct radv_streamout_info *so = &info->so;
 509
 510         if (!xfb)
 511                 return;
 512
 513         assert(xfb->output_count < MAX_SO_OUTPUTS);
 514         so->num_outputs = xfb->output_count;
 515
 516         for (unsigned i = 0; i < xfb->output_count; i++) {
 517                 struct radv_stream_output *output = &so->outputs[i];
 518
 519                 output->buffer = xfb->outputs[i].buffer;
 520                 output->stream = xfb->buffer_to_stream[xfb->outputs[i].buffer];
 521                 output->offset = xfb->outputs[i].offset;
 522                 output->location = xfb->outputs[i].location;
 523                 output->component_mask = xfb->outputs[i].component_mask;
 524
 525                 so->enabled_stream_buffers_mask |=
 526                         (1 << output->buffer) << (output->stream * 4);
 527
 528         }
 529
 530         for (unsigned i = 0; i < NIR_MAX_XFB_BUFFERS; i++) {
 531                 so->strides[i] = xfb->buffers[i].stride / 4;
 532         }
 533
 534         ralloc_free(xfb);
 535 }
 536
 537 void
 538 radv_nir_shader_info_init(struct radv_shader_info *info)
 539 {
 540         /* Assume that shaders only have 32-bit push constants by default. */
 541         info->min_push_constant_used = UINT8_MAX;
 542         info->has_only_32bit_push_constants = true;
 543 }
 544
 545 void
 546 radv_nir_shader_info_pass(const struct nir_shader *nir,
 547                           const struct radv_nir_compiler_options *options,
 548                           struct radv_shader_info *info)
 549 {
 550         struct nir_function *func =
 551                 (struct nir_function *)exec_list_get_head_const(&nir->functions);
 552
 553         if (options->layout && options->layout->dynamic_offset_count &&
 554             (options->layout->dynamic_shader_stages & mesa_to_vk_shader_stage(nir->info.stage))) {
 555                 info->loads_push_constants = true;
 556                 info->loads_dynamic_offsets = true;
 557         }
 558
 559         nir_foreach_variable(variable, &nir->inputs)
 560                 gather_info_input_decl(nir, variable, info);
 561
 562         nir_foreach_block(block, func->impl) {
 563                 gather_info_block(nir, block, info);
 564         }
 565
 566         nir_foreach_variable(variable, &nir->outputs)
 567                 gather_info_output_decl(nir, variable, info, options);
 568
 569         if (nir->info.stage == MESA_SHADER_VERTEX ||
 570             nir->info.stage == MESA_SHADER_TESS_EVAL ||
 571             nir->info.stage == MESA_SHADER_GEOMETRY)
 572                 gather_xfb_info(nir, info);
 573 }