src/amd/vulkan/radv_shader.c

   1 /*
   2  * Copyright © 2016 Red Hat.
   3  * Copyright © 2016 Bas Nieuwenhuizen
   4  *
   5  * based in part on anv driver which is:
   6  * Copyright © 2015 Intel Corporation
   7  *
   8  * Permission is hereby granted, free of charge, to any person obtaining a
   9  * copy of this software and associated documentation files (the "Software"),
  10  * to deal in the Software without restriction, including without limitation
  11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  12  * and/or sell copies of the Software, and to permit persons to whom the
  13  * Software is furnished to do so, subject to the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the next
  16  * paragraph) shall be included in all copies or substantial portions of the
  17  * Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  22  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  23  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  24  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  25  * IN THE SOFTWARE.
  26  */
  27
  28 #include "util/mesa-sha1.h"
  29 #include "util/u_atomic.h"
  30 #include "radv_debug.h"
  31 #include "radv_private.h"
  32 #include "radv_shader.h"
  33 #include "nir/nir.h"
  34 #include "nir/nir_builder.h"
  35 #include "spirv/nir_spirv.h"
  36
  37 #include <llvm-c/Core.h>
  38 #include <llvm-c/TargetMachine.h>
  39
  40 #include "sid.h"
  41 #include "gfx9d.h"
  42 #include "ac_binary.h"
  43 #include "ac_llvm_util.h"
  44 #include "ac_nir_to_llvm.h"
  45 #include "vk_format.h"
  46 #include "util/debug.h"
  47 #include "ac_exp_param.h"
  48
  49 #include "util/string_buffer.h"
  50
  51 static const struct nir_shader_compiler_options nir_options = {
  52         .vertex_id_zero_based = true,
  53         .lower_scmp = true,
  54         .lower_flrp32 = true,
  55         .lower_flrp64 = true,
  56         .lower_device_index_to_zero = true,
  57         .lower_fsat = true,
  58         .lower_fdiv = true,
  59         .lower_sub = true,
  60         .lower_pack_snorm_2x16 = true,
  61         .lower_pack_snorm_4x8 = true,
  62         .lower_pack_unorm_2x16 = true,
  63         .lower_pack_unorm_4x8 = true,
  64         .lower_unpack_snorm_2x16 = true,
  65         .lower_unpack_snorm_4x8 = true,
  66         .lower_unpack_unorm_2x16 = true,
  67         .lower_unpack_unorm_4x8 = true,
  68         .lower_extract_byte = true,
  69         .lower_extract_word = true,
  70         .lower_ffma = true,
  71         .lower_fpow = true,
  72         .vs_inputs_dual_locations = true,
  73         .max_unroll_iterations = 32
  74 };
  75
  76 VkResult radv_CreateShaderModule(
  77         VkDevice                                    _device,
  78         const VkShaderModuleCreateInfo*             pCreateInfo,
  79         const VkAllocationCallbacks*                pAllocator,
  80         VkShaderModule*                             pShaderModule)
  81 {
  82         RADV_FROM_HANDLE(radv_device, device, _device);
  83         struct radv_shader_module *module;
  84
  85         assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO);
  86         assert(pCreateInfo->flags == 0);
  87
  88         module = vk_alloc2(&device->alloc, pAllocator,
  89                              sizeof(*module) + pCreateInfo->codeSize, 8,
  90                              VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
  91         if (module == NULL)
  92                 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
  93
  94         module->nir = NULL;
  95         module->size = pCreateInfo->codeSize;
  96         memcpy(module->data, pCreateInfo->pCode, module->size);
  97
  98         _mesa_sha1_compute(module->data, module->size, module->sha1);
  99
 100         *pShaderModule = radv_shader_module_to_handle(module);
 101
 102         return VK_SUCCESS;
 103 }
 104
 105 void radv_DestroyShaderModule(
 106         VkDevice                                    _device,
 107         VkShaderModule                              _module,
 108         const VkAllocationCallbacks*                pAllocator)
 109 {
 110         RADV_FROM_HANDLE(radv_device, device, _device);
 111         RADV_FROM_HANDLE(radv_shader_module, module, _module);
 112
 113         if (!module)
 114                 return;
 115
 116         vk_free2(&device->alloc, pAllocator, module);
 117 }
 118
 119 void
 120 radv_optimize_nir(struct nir_shader *shader)
 121 {
 122         bool progress;
 123
 124         do {
 125                 progress = false;
 126
 127                 NIR_PASS_V(shader, nir_lower_vars_to_ssa);
 128                 NIR_PASS_V(shader, nir_lower_64bit_pack);
 129                 NIR_PASS_V(shader, nir_lower_alu_to_scalar);
 130                 NIR_PASS_V(shader, nir_lower_phis_to_scalar);
 131
 132                 NIR_PASS(progress, shader, nir_copy_prop);
 133                 NIR_PASS(progress, shader, nir_opt_remove_phis);
 134                 NIR_PASS(progress, shader, nir_opt_dce);
 135                 if (nir_opt_trivial_continues(shader)) {
 136                         progress = true;
 137                         NIR_PASS(progress, shader, nir_copy_prop);
 138                         NIR_PASS(progress, shader, nir_opt_remove_phis);
 139                         NIR_PASS(progress, shader, nir_opt_dce);
 140                 }
 141                 NIR_PASS(progress, shader, nir_opt_if);
 142                 NIR_PASS(progress, shader, nir_opt_dead_cf);
 143                 NIR_PASS(progress, shader, nir_opt_cse);
 144                 NIR_PASS(progress, shader, nir_opt_peephole_select, 8);
 145                 NIR_PASS(progress, shader, nir_opt_algebraic);
 146                 NIR_PASS(progress, shader, nir_opt_constant_folding);
 147                 NIR_PASS(progress, shader, nir_opt_undef);
 148                 NIR_PASS(progress, shader, nir_opt_conditional_discard);
 149                 if (shader->options->max_unroll_iterations) {
 150                         NIR_PASS(progress, shader, nir_opt_loop_unroll, 0);
 151                 }
 152         } while (progress);
 153
 154         NIR_PASS(progress, shader, nir_opt_shrink_load);
 155         NIR_PASS(progress, shader, nir_opt_move_load_ubo);
 156 }
 157
 158 nir_shader *
 159 radv_shader_compile_to_nir(struct radv_device *device,
 160                            struct radv_shader_module *module,
 161                            const char *entrypoint_name,
 162                            gl_shader_stage stage,
 163                            const VkSpecializationInfo *spec_info)
 164 {
 165         if (strcmp(entrypoint_name, "main") != 0) {
 166                 radv_finishme("Multiple shaders per module not really supported");
 167         }
 168
 169         nir_shader *nir;
 170         nir_function *entry_point;
 171         if (module->nir) {
 172                 /* Some things such as our meta clear/blit code will give us a NIR
 173                  * shader directly.  In that case, we just ignore the SPIR-V entirely
 174                  * and just use the NIR shader */
 175                 nir = module->nir;
 176                 nir->options = &nir_options;
 177                 nir_validate_shader(nir);
 178
 179                 assert(exec_list_length(&nir->functions) == 1);
 180                 struct exec_node *node = exec_list_get_head(&nir->functions);
 181                 entry_point = exec_node_data(nir_function, node, node);
 182         } else {
 183                 uint32_t *spirv = (uint32_t *) module->data;
 184                 assert(module->size % 4 == 0);
 185
 186                 if (device->instance->debug_flags & RADV_DEBUG_DUMP_SPIRV)
 187                         radv_print_spirv(spirv, module->size, stderr);
 188
 189                 uint32_t num_spec_entries = 0;
 190                 struct nir_spirv_specialization *spec_entries = NULL;
 191                 if (spec_info && spec_info->mapEntryCount > 0) {
 192                         num_spec_entries = spec_info->mapEntryCount;
 193                         spec_entries = malloc(num_spec_entries * sizeof(*spec_entries));
 194                         for (uint32_t i = 0; i < num_spec_entries; i++) {
 195                                 VkSpecializationMapEntry entry = spec_info->pMapEntries[i];
 196                                 const void *data = spec_info->pData + entry.offset;
 197                                 assert(data + entry.size <= spec_info->pData + spec_info->dataSize);
 198
 199                                 spec_entries[i].id = spec_info->pMapEntries[i].constantID;
 200                                 if (spec_info->dataSize == 8)
 201                                         spec_entries[i].data64 = *(const uint64_t *)data;
 202                                 else
 203                                         spec_entries[i].data32 = *(const uint32_t *)data;
 204                         }
 205                 }
 206                 const struct spirv_to_nir_options spirv_options = {
 207                         .caps = {
 208                                 .device_group = true,
 209                                 .draw_parameters = true,
 210                                 .float64 = true,
 211                                 .image_read_without_format = true,
 212                                 .image_write_without_format = true,
 213                                 .tessellation = true,
 214                                 .int64 = true,
 215                                 .multiview = true,
 216                                 .subgroup_basic = true,
 217                                 .variable_pointers = true,
 218                                 .gcn_shader = true,
 219                                 .trinary_minmax = true,
 220                                 .shader_viewport_index_layer = true,
 221                         },
 222                 };
 223                 entry_point = spirv_to_nir(spirv, module->size / 4,
 224                                            spec_entries, num_spec_entries,
 225                                            stage, entrypoint_name,
 226                                            &spirv_options, &nir_options);
 227                 nir = entry_point->shader;
 228                 assert(nir->info.stage == stage);
 229                 nir_validate_shader(nir);
 230
 231                 free(spec_entries);
 232
 233                 /* We have to lower away local constant initializers right before we
 234                  * inline functions.  That way they get properly initialized at the top
 235                  * of the function and not at the top of its caller.
 236                  */
 237                 NIR_PASS_V(nir, nir_lower_constant_initializers, nir_var_local);
 238                 NIR_PASS_V(nir, nir_lower_returns);
 239                 NIR_PASS_V(nir, nir_inline_functions);
 240
 241                 /* Pick off the single entrypoint that we want */
 242                 foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
 243                         if (func != entry_point)
 244                                 exec_node_remove(&func->node);
 245                 }
 246                 assert(exec_list_length(&nir->functions) == 1);
 247                 entry_point->name = ralloc_strdup(entry_point, "main");
 248
 249                 /* Make sure we lower constant initializers on output variables so that
 250                  * nir_remove_dead_variables below sees the corresponding stores
 251                  */
 252                 NIR_PASS_V(nir, nir_lower_constant_initializers, nir_var_shader_out);
 253
 254                 NIR_PASS_V(nir, nir_remove_dead_variables,
 255                            nir_var_shader_in | nir_var_shader_out | nir_var_system_value);
 256
 257                 /* Now that we've deleted all but the main function, we can go ahead and
 258                  * lower the rest of the constant initializers.
 259                  */
 260                 NIR_PASS_V(nir, nir_lower_constant_initializers, ~0);
 261                 NIR_PASS_V(nir, nir_lower_system_values);
 262                 NIR_PASS_V(nir, nir_lower_clip_cull_distance_arrays);
 263         }
 264
 265         /* Vulkan uses the separate-shader linking model */
 266         nir->info.separate_shader = true;
 267
 268         nir_shader_gather_info(nir, entry_point->impl);
 269
 270         static const nir_lower_tex_options tex_options = {
 271           .lower_txp = ~0,
 272         };
 273
 274         nir_lower_tex(nir, &tex_options);
 275
 276         nir_lower_vars_to_ssa(nir);
 277         nir_lower_var_copies(nir);
 278         nir_lower_global_vars_to_local(nir);
 279         nir_remove_dead_variables(nir, nir_var_local);
 280         nir_lower_subgroups(nir, &(struct nir_lower_subgroups_options) {
 281                         .subgroup_size = 64,
 282                         .ballot_bit_size = 64,
 283                         .lower_to_scalar = 1,
 284                         .lower_subgroup_masks = 1,
 285                         .lower_shuffle = 1,
 286                         .lower_quad =  1,
 287                 });
 288
 289         radv_optimize_nir(nir);
 290
 291         /* Indirect lowering must be called after the radv_optimize_nir() loop
 292          * has been called at least once. Otherwise indirect lowering can
 293          * bloat the instruction count of the loop and cause it to be
 294          * considered too large for unrolling.
 295          */
 296         ac_lower_indirect_derefs(nir, device->physical_device->rad_info.chip_class);
 297         radv_optimize_nir(nir);
 298
 299         return nir;
 300 }
 301
 302 void *
 303 radv_alloc_shader_memory(struct radv_device *device,
 304                          struct radv_shader_variant *shader)
 305 {
 306         mtx_lock(&device->shader_slab_mutex);
 307         list_for_each_entry(struct radv_shader_slab, slab, &device->shader_slabs, slabs) {
 308                 uint64_t offset = 0;
 309                 list_for_each_entry(struct radv_shader_variant, s, &slab->shaders, slab_list) {
 310                         if (s->bo_offset - offset >= shader->code_size) {
 311                                 shader->bo = slab->bo;
 312                                 shader->bo_offset = offset;
 313                                 list_addtail(&shader->slab_list, &s->slab_list);
 314                                 mtx_unlock(&device->shader_slab_mutex);
 315                                 return slab->ptr + offset;
 316                         }
 317                         offset = align_u64(s->bo_offset + s->code_size, 256);
 318                 }
 319                 if (slab->size - offset >= shader->code_size) {
 320                         shader->bo = slab->bo;
 321                         shader->bo_offset = offset;
 322                         list_addtail(&shader->slab_list, &slab->shaders);
 323                         mtx_unlock(&device->shader_slab_mutex);
 324                         return slab->ptr + offset;
 325                 }
 326         }
 327
 328         mtx_unlock(&device->shader_slab_mutex);
 329         struct radv_shader_slab *slab = calloc(1, sizeof(struct radv_shader_slab));
 330
 331         slab->size = 256 * 1024;
 332         slab->bo = device->ws->buffer_create(device->ws, slab->size, 256,
 333                                              RADEON_DOMAIN_VRAM,
 334                                              RADEON_FLAG_NO_INTERPROCESS_SHARING |
 335                                              device->physical_device->cpdma_prefetch_writes_memory ?
 336                                                      0 : RADEON_FLAG_READ_ONLY);
 337         slab->ptr = (char*)device->ws->buffer_map(slab->bo);
 338         list_inithead(&slab->shaders);
 339
 340         mtx_lock(&device->shader_slab_mutex);
 341         list_add(&slab->slabs, &device->shader_slabs);
 342
 343         shader->bo = slab->bo;
 344         shader->bo_offset = 0;
 345         list_add(&shader->slab_list, &slab->shaders);
 346         mtx_unlock(&device->shader_slab_mutex);
 347         return slab->ptr;
 348 }
 349
 350 void
 351 radv_destroy_shader_slabs(struct radv_device *device)
 352 {
 353         list_for_each_entry_safe(struct radv_shader_slab, slab, &device->shader_slabs, slabs) {
 354                 device->ws->buffer_destroy(slab->bo);
 355                 free(slab);
 356         }
 357         mtx_destroy(&device->shader_slab_mutex);
 358 }
 359
 360 static void
 361 radv_fill_shader_variant(struct radv_device *device,
 362                          struct radv_shader_variant *variant,
 363                          struct ac_shader_binary *binary,
 364                          gl_shader_stage stage)
 365 {
 366         bool scratch_enabled = variant->config.scratch_bytes_per_wave > 0;
 367         unsigned vgpr_comp_cnt = 0;
 368
 369         if (scratch_enabled && !device->llvm_supports_spill)
 370                 radv_finishme("shader scratch support only available with LLVM 4.0");
 371
 372         variant->code_size = binary->code_size;
 373         variant->rsrc2 = S_00B12C_USER_SGPR(variant->info.num_user_sgprs) |
 374                         S_00B12C_SCRATCH_EN(scratch_enabled);
 375
 376         variant->rsrc1 =  S_00B848_VGPRS((variant->config.num_vgprs - 1) / 4) |
 377                 S_00B848_SGPRS((variant->config.num_sgprs - 1) / 8) |
 378                 S_00B848_DX10_CLAMP(1) |
 379                 S_00B848_FLOAT_MODE(variant->config.float_mode);
 380
 381         switch (stage) {
 382         case MESA_SHADER_TESS_EVAL:
 383                 vgpr_comp_cnt = 3;
 384                 variant->rsrc2 |= S_00B12C_OC_LDS_EN(1);
 385                 break;
 386         case MESA_SHADER_TESS_CTRL:
 387                 if (device->physical_device->rad_info.chip_class >= GFX9)
 388                         vgpr_comp_cnt = variant->info.vs.vgpr_comp_cnt;
 389                 else
 390                         variant->rsrc2 |= S_00B12C_OC_LDS_EN(1);
 391                 break;
 392         case MESA_SHADER_VERTEX:
 393         case MESA_SHADER_GEOMETRY:
 394                 vgpr_comp_cnt = variant->info.vs.vgpr_comp_cnt;
 395                 break;
 396         case MESA_SHADER_FRAGMENT:
 397                 break;
 398         case MESA_SHADER_COMPUTE: {
 399                 struct radv_shader_info *info = &variant->info.info;
 400                 variant->rsrc2 |=
 401                         S_00B84C_TGID_X_EN(info->cs.uses_block_id[0]) |
 402                         S_00B84C_TGID_Y_EN(info->cs.uses_block_id[1]) |
 403                         S_00B84C_TGID_Z_EN(info->cs.uses_block_id[2]) |
 404                         S_00B84C_TIDIG_COMP_CNT(info->cs.uses_thread_id[2] ? 2 :
 405                                                 info->cs.uses_thread_id[1] ? 1 : 0) |
 406                         S_00B84C_TG_SIZE_EN(info->cs.uses_local_invocation_idx) |
 407                         S_00B84C_LDS_SIZE(variant->config.lds_size);
 408                 break;
 409         }
 410         default:
 411                 unreachable("unsupported shader type");
 412                 break;
 413         }
 414
 415         if (device->physical_device->rad_info.chip_class >= GFX9 &&
 416             stage == MESA_SHADER_GEOMETRY) {
 417                 struct radv_shader_info *info = &variant->info.info;
 418                 unsigned es_type = variant->info.gs.es_type;
 419                 unsigned gs_vgpr_comp_cnt, es_vgpr_comp_cnt;
 420
 421                 if (es_type == MESA_SHADER_VERTEX) {
 422                         es_vgpr_comp_cnt = variant->info.vs.vgpr_comp_cnt;
 423                 } else if (es_type == MESA_SHADER_TESS_EVAL) {
 424                         es_vgpr_comp_cnt = 3;
 425                 } else {
 426                         unreachable("invalid shader ES type");
 427                 }
 428
 429                 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
 430                  * VGPR[0:4] are always loaded.
 431                  */
 432                 if (info->uses_invocation_id)
 433                         gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID. */
 434                 else if (info->uses_prim_id)
 435                         gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
 436                 else if (variant->info.gs.vertices_in >= 3)
 437                         gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
 438                 else
 439                         gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
 440
 441                 variant->rsrc1 |= S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt);
 442                 variant->rsrc2 |= S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
 443                                   S_00B22C_OC_LDS_EN(es_type == MESA_SHADER_TESS_EVAL);
 444         } else if (device->physical_device->rad_info.chip_class >= GFX9 &&
 445             stage == MESA_SHADER_TESS_CTRL)
 446                 variant->rsrc1 |= S_00B428_LS_VGPR_COMP_CNT(vgpr_comp_cnt);
 447         else
 448                 variant->rsrc1 |= S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt);
 449
 450         void *ptr = radv_alloc_shader_memory(device, variant);
 451         memcpy(ptr, binary->code, binary->code_size);
 452 }
 453
 454 static struct radv_shader_variant *
 455 shader_variant_create(struct radv_device *device,
 456                       struct radv_shader_module *module,
 457                       struct nir_shader * const *shaders,
 458                       int shader_count,
 459                       gl_shader_stage stage,
 460                       struct radv_nir_compiler_options *options,
 461                       bool gs_copy_shader,
 462                       void **code_out,
 463                       unsigned *code_size_out)
 464 {
 465         enum radeon_family chip_family = device->physical_device->rad_info.family;
 466         enum ac_target_machine_options tm_options = 0;
 467         struct radv_shader_variant *variant;
 468         struct ac_shader_binary binary;
 469         LLVMTargetMachineRef tm;
 470
 471         variant = calloc(1, sizeof(struct radv_shader_variant));
 472         if (!variant)
 473                 return NULL;
 474
 475         options->family = chip_family;
 476         options->chip_class = device->physical_device->rad_info.chip_class;
 477         options->dump_shader = radv_can_dump_shader(device, module);
 478         options->dump_preoptir = options->dump_shader &&
 479                                  device->instance->debug_flags & RADV_DEBUG_PREOPTIR;
 480         options->record_llvm_ir = device->keep_shader_info;
 481         options->tess_offchip_block_dw_size = device->tess_offchip_block_dw_size;
 482
 483         if (options->supports_spill)
 484                 tm_options |= AC_TM_SUPPORTS_SPILL;
 485         if (device->instance->perftest_flags & RADV_PERFTEST_SISCHED)
 486                 tm_options |= AC_TM_SISCHED;
 487         tm = ac_create_target_machine(chip_family, tm_options);
 488
 489         if (gs_copy_shader) {
 490                 assert(shader_count == 1);
 491                 radv_compile_gs_copy_shader(tm, *shaders, &binary,
 492                                             &variant->config, &variant->info,
 493                                             options);
 494         } else {
 495                 radv_compile_nir_shader(tm, &binary, &variant->config,
 496                                         &variant->info, shaders, shader_count,
 497                                         options);
 498         }
 499
 500         LLVMDisposeTargetMachine(tm);
 501
 502         radv_fill_shader_variant(device, variant, &binary, stage);
 503
 504         if (code_out) {
 505                 *code_out = binary.code;
 506                 *code_size_out = binary.code_size;
 507         } else
 508                 free(binary.code);
 509         free(binary.config);
 510         free(binary.rodata);
 511         free(binary.global_symbol_offsets);
 512         free(binary.relocs);
 513         variant->ref_count = 1;
 514
 515         if (device->keep_shader_info) {
 516                 variant->disasm_string = binary.disasm_string;
 517                 variant->llvm_ir_string = binary.llvm_ir_string;
 518                 if (!gs_copy_shader && !module->nir) {
 519                         variant->nir = *shaders;
 520                         variant->spirv = (uint32_t *)module->data;
 521                         variant->spirv_size = module->size;
 522                 }
 523         } else {
 524                 free(binary.disasm_string);
 525         }
 526
 527         return variant;
 528 }
 529
 530 struct radv_shader_variant *
 531 radv_shader_variant_create(struct radv_device *device,
 532                            struct radv_shader_module *module,
 533                            struct nir_shader *const *shaders,
 534                            int shader_count,
 535                            struct radv_pipeline_layout *layout,
 536                            const struct radv_shader_variant_key *key,
 537                            void **code_out,
 538                            unsigned *code_size_out)
 539 {
 540         struct radv_nir_compiler_options options = {0};
 541
 542         options.layout = layout;
 543         if (key)
 544                 options.key = *key;
 545
 546         options.unsafe_math = !!(device->instance->debug_flags & RADV_DEBUG_UNSAFE_MATH);
 547         options.supports_spill = device->llvm_supports_spill;
 548
 549         return shader_variant_create(device, module, shaders, shader_count, shaders[shader_count - 1]->info.stage,
 550                                      &options, false, code_out, code_size_out);
 551 }
 552
 553 struct radv_shader_variant *
 554 radv_create_gs_copy_shader(struct radv_device *device,
 555                            struct nir_shader *shader,
 556                            void **code_out,
 557                            unsigned *code_size_out,
 558                            bool multiview)
 559 {
 560         struct radv_nir_compiler_options options = {0};
 561
 562         options.key.has_multiview_view_index = multiview;
 563
 564         return shader_variant_create(device, NULL, &shader, 1, MESA_SHADER_VERTEX,
 565                                      &options, true, code_out, code_size_out);
 566 }
 567
 568 void
 569 radv_shader_variant_destroy(struct radv_device *device,
 570                             struct radv_shader_variant *variant)
 571 {
 572         if (!p_atomic_dec_zero(&variant->ref_count))
 573                 return;
 574
 575         mtx_lock(&device->shader_slab_mutex);
 576         list_del(&variant->slab_list);
 577         mtx_unlock(&device->shader_slab_mutex);
 578
 579         ralloc_free(variant->nir);
 580         free(variant->disasm_string);
 581         free(variant->llvm_ir_string);
 582         free(variant);
 583 }
 584
 585 const char *
 586 radv_get_shader_name(struct radv_shader_variant *var, gl_shader_stage stage)
 587 {
 588         switch (stage) {
 589         case MESA_SHADER_VERTEX: return var->info.vs.as_ls ? "Vertex Shader as LS" : var->info.vs.as_es ? "Vertex Shader as ES" : "Vertex Shader as VS";
 590         case MESA_SHADER_GEOMETRY: return "Geometry Shader";
 591         case MESA_SHADER_FRAGMENT: return "Pixel Shader";
 592         case MESA_SHADER_COMPUTE: return "Compute Shader";
 593         case MESA_SHADER_TESS_CTRL: return "Tessellation Control Shader";
 594         case MESA_SHADER_TESS_EVAL: return var->info.tes.as_es ? "Tessellation Evaluation Shader as ES" : "Tessellation Evaluation Shader as VS";
 595         default:
 596                 return "Unknown shader";
 597         };
 598 }
 599
 600 static void
 601 generate_shader_stats(struct radv_device *device,
 602                       struct radv_shader_variant *variant,
 603                       gl_shader_stage stage,
 604                       struct _mesa_string_buffer *buf)
 605 {
 606         unsigned lds_increment = device->physical_device->rad_info.chip_class >= CIK ? 512 : 256;
 607         struct ac_shader_config *conf;
 608         unsigned max_simd_waves;
 609         unsigned lds_per_wave = 0;
 610
 611         switch (device->physical_device->rad_info.family) {
 612         /* These always have 8 waves: */
 613         case CHIP_POLARIS10:
 614         case CHIP_POLARIS11:
 615         case CHIP_POLARIS12:
 616                 max_simd_waves = 8;
 617                 break;
 618         default:
 619                 max_simd_waves = 10;
 620         }
 621
 622         conf = &variant->config;
 623
 624         if (stage == MESA_SHADER_FRAGMENT) {
 625                 lds_per_wave = conf->lds_size * lds_increment +
 626                                align(variant->info.fs.num_interp * 48,
 627                                      lds_increment);
 628         }
 629
 630         if (conf->num_sgprs)
 631                 max_simd_waves =
 632                         MIN2(max_simd_waves,
 633                              radv_get_num_physical_sgprs(device->physical_device) / conf->num_sgprs);
 634
 635         if (conf->num_vgprs)
 636                 max_simd_waves = MIN2(max_simd_waves, 256 / conf->num_vgprs);
 637
 638         /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
 639          * that PS can use.
 640          */
 641         if (lds_per_wave)
 642                 max_simd_waves = MIN2(max_simd_waves, 16384 / lds_per_wave);
 643
 644         if (stage == MESA_SHADER_FRAGMENT) {
 645                 _mesa_string_buffer_printf(buf, "*** SHADER CONFIG ***\n"
 646                                            "SPI_PS_INPUT_ADDR = 0x%04x\n"
 647                                            "SPI_PS_INPUT_ENA  = 0x%04x\n",
 648                                            conf->spi_ps_input_addr, conf->spi_ps_input_ena);
 649         }
 650
 651         _mesa_string_buffer_printf(buf, "*** SHADER STATS ***\n"
 652                                    "SGPRS: %d\n"
 653                                    "VGPRS: %d\n"
 654                                    "Spilled SGPRs: %d\n"
 655                                    "Spilled VGPRs: %d\n"
 656                                    "PrivMem VGPRS: %d\n"
 657                                    "Code Size: %d bytes\n"
 658                                    "LDS: %d blocks\n"
 659                                    "Scratch: %d bytes per wave\n"
 660                                    "Max Waves: %d\n"
 661                                    "********************\n\n\n",
 662                                    conf->num_sgprs, conf->num_vgprs,
 663                                    conf->spilled_sgprs, conf->spilled_vgprs,
 664                                    variant->info.private_mem_vgprs, variant->code_size,
 665                                    conf->lds_size, conf->scratch_bytes_per_wave,
 666                                    max_simd_waves);
 667 }
 668
 669 void
 670 radv_shader_dump_stats(struct radv_device *device,
 671                        struct radv_shader_variant *variant,
 672                        gl_shader_stage stage,
 673                        FILE *file)
 674 {
 675         struct _mesa_string_buffer *buf = _mesa_string_buffer_create(NULL, 256);
 676
 677         generate_shader_stats(device, variant, stage, buf);
 678
 679         fprintf(file, "\n%s:\n", radv_get_shader_name(variant, stage));
 680         fprintf(file, "%s", buf->buf);
 681
 682         _mesa_string_buffer_destroy(buf);
 683 }
 684
 685 VkResult
 686 radv_GetShaderInfoAMD(VkDevice _device,
 687                       VkPipeline _pipeline,
 688                       VkShaderStageFlagBits shaderStage,
 689                       VkShaderInfoTypeAMD infoType,
 690                       size_t* pInfoSize,
 691                       void* pInfo)
 692 {
 693         RADV_FROM_HANDLE(radv_device, device, _device);
 694         RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline);
 695         gl_shader_stage stage = vk_to_mesa_shader_stage(shaderStage);
 696         struct radv_shader_variant *variant = pipeline->shaders[stage];
 697         struct _mesa_string_buffer *buf;
 698         VkResult result = VK_SUCCESS;
 699
 700         /* Spec doesn't indicate what to do if the stage is invalid, so just
 701          * return no info for this. */
 702         if (!variant)
 703                 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
 704
 705         switch (infoType) {
 706         case VK_SHADER_INFO_TYPE_STATISTICS_AMD:
 707                 if (!pInfo) {
 708                         *pInfoSize = sizeof(VkShaderStatisticsInfoAMD);
 709                 } else {
 710                         unsigned lds_multiplier = device->physical_device->rad_info.chip_class >= CIK ? 512 : 256;
 711                         struct ac_shader_config *conf = &variant->config;
 712
 713                         VkShaderStatisticsInfoAMD statistics = {};
 714                         statistics.shaderStageMask = shaderStage;
 715                         statistics.numPhysicalVgprs = 256;
 716                         statistics.numPhysicalSgprs = radv_get_num_physical_sgprs(device->physical_device);
 717                         statistics.numAvailableSgprs = statistics.numPhysicalSgprs;
 718
 719                         if (stage == MESA_SHADER_COMPUTE) {
 720                                 unsigned *local_size = variant->nir->info.cs.local_size;
 721                                 unsigned workgroup_size = local_size[0] * local_size[1] * local_size[2];
 722
 723                                 statistics.numAvailableVgprs = statistics.numPhysicalVgprs /
 724                                                                ceil(workgroup_size / statistics.numPhysicalVgprs);
 725
 726                                 statistics.computeWorkGroupSize[0] = local_size[0];
 727                                 statistics.computeWorkGroupSize[1] = local_size[1];
 728                                 statistics.computeWorkGroupSize[2] = local_size[2];
 729                         } else {
 730                                 statistics.numAvailableVgprs = statistics.numPhysicalVgprs;
 731                         }
 732
 733                         statistics.resourceUsage.numUsedVgprs = conf->num_vgprs;
 734                         statistics.resourceUsage.numUsedSgprs = conf->num_sgprs;
 735                         statistics.resourceUsage.ldsSizePerLocalWorkGroup = 32768;
 736                         statistics.resourceUsage.ldsUsageSizeInBytes = conf->lds_size * lds_multiplier;
 737                         statistics.resourceUsage.scratchMemUsageInBytes = conf->scratch_bytes_per_wave;
 738
 739                         size_t size = *pInfoSize;
 740                         *pInfoSize = sizeof(statistics);
 741
 742                         memcpy(pInfo, &statistics, MIN2(size, *pInfoSize));
 743
 744                         if (size < *pInfoSize)
 745                                 result = VK_INCOMPLETE;
 746                 }
 747
 748                 break;
 749         case VK_SHADER_INFO_TYPE_DISASSEMBLY_AMD:
 750                 buf = _mesa_string_buffer_create(NULL, 1024);
 751
 752                 _mesa_string_buffer_printf(buf, "%s:\n", radv_get_shader_name(variant, stage));
 753                 _mesa_string_buffer_printf(buf, "%s\n\n", variant->disasm_string);
 754                 generate_shader_stats(device, variant, stage, buf);
 755
 756                 /* Need to include the null terminator. */
 757                 size_t length = buf->length + 1;
 758
 759                 if (!pInfo) {
 760                         *pInfoSize = length;
 761                 } else {
 762                         size_t size = *pInfoSize;
 763                         *pInfoSize = length;
 764
 765                         memcpy(pInfo, buf->buf, MIN2(size, length));
 766
 767                         if (size < length)
 768                                 result = VK_INCOMPLETE;
 769                 }
 770
 771                 _mesa_string_buffer_destroy(buf);
 772                 break;
 773         default:
 774                 /* VK_SHADER_INFO_TYPE_BINARY_AMD unimplemented for now. */
 775                 result = VK_ERROR_FEATURE_NOT_PRESENT;
 776                 break;
 777         }
 778
 779         return result;
 780 }