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 "radv_shader_helper.h"
  34 #include "nir/nir.h"
  35 #include "nir/nir_builder.h"
  36 #include "spirv/nir_spirv.h"
  37
  38 #include <llvm-c/Core.h>
  39 #include <llvm-c/TargetMachine.h>
  40 #include <llvm-c/Support.h>
  41
  42 #include "sid.h"
  43 #include "gfx9d.h"
  44 #include "ac_binary.h"
  45 #include "ac_llvm_util.h"
  46 #include "ac_nir_to_llvm.h"
  47 #include "vk_format.h"
  48 #include "util/debug.h"
  49 #include "ac_exp_param.h"
  50
  51 #include "util/string_buffer.h"
  52
  53 static const struct nir_shader_compiler_options nir_options = {
  54         .vertex_id_zero_based = true,
  55         .lower_scmp = true,
  56         .lower_flrp32 = true,
  57         .lower_flrp64 = true,
  58         .lower_device_index_to_zero = true,
  59         .lower_fsat = true,
  60         .lower_fdiv = true,
  61         .lower_sub = true,
  62         .lower_pack_snorm_2x16 = true,
  63         .lower_pack_snorm_4x8 = true,
  64         .lower_pack_unorm_2x16 = true,
  65         .lower_pack_unorm_4x8 = true,
  66         .lower_unpack_snorm_2x16 = true,
  67         .lower_unpack_snorm_4x8 = true,
  68         .lower_unpack_unorm_2x16 = true,
  69         .lower_unpack_unorm_4x8 = true,
  70         .lower_extract_byte = true,
  71         .lower_extract_word = true,
  72         .lower_ffma = true,
  73         .lower_fpow = true,
  74         .max_unroll_iterations = 32
  75 };
  76
  77 VkResult radv_CreateShaderModule(
  78         VkDevice                                    _device,
  79         const VkShaderModuleCreateInfo*             pCreateInfo,
  80         const VkAllocationCallbacks*                pAllocator,
  81         VkShaderModule*                             pShaderModule)
  82 {
  83         RADV_FROM_HANDLE(radv_device, device, _device);
  84         struct radv_shader_module *module;
  85
  86         assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO);
  87         assert(pCreateInfo->flags == 0);
  88
  89         module = vk_alloc2(&device->alloc, pAllocator,
  90                              sizeof(*module) + pCreateInfo->codeSize, 8,
  91                              VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
  92         if (module == NULL)
  93                 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
  94
  95         module->nir = NULL;
  96         module->size = pCreateInfo->codeSize;
  97         memcpy(module->data, pCreateInfo->pCode, module->size);
  98
  99         _mesa_sha1_compute(module->data, module->size, module->sha1);
 100
 101         *pShaderModule = radv_shader_module_to_handle(module);
 102
 103         return VK_SUCCESS;
 104 }
 105
 106 void radv_DestroyShaderModule(
 107         VkDevice                                    _device,
 108         VkShaderModule                              _module,
 109         const VkAllocationCallbacks*                pAllocator)
 110 {
 111         RADV_FROM_HANDLE(radv_device, device, _device);
 112         RADV_FROM_HANDLE(radv_shader_module, module, _module);
 113
 114         if (!module)
 115                 return;
 116
 117         vk_free2(&device->alloc, pAllocator, module);
 118 }
 119
 120 void
 121 radv_optimize_nir(struct nir_shader *shader, bool optimize_conservatively)
 122 {
 123         bool progress;
 124
 125         do {
 126                 progress = false;
 127
 128                 NIR_PASS_V(shader, nir_lower_vars_to_ssa);
 129                 NIR_PASS_V(shader, nir_lower_pack);
 130                 NIR_PASS_V(shader, nir_lower_alu_to_scalar);
 131                 NIR_PASS_V(shader, nir_lower_phis_to_scalar);
 132
 133                 NIR_PASS(progress, shader, nir_copy_prop);
 134                 NIR_PASS(progress, shader, nir_opt_remove_phis);
 135                 NIR_PASS(progress, shader, nir_opt_dce);
 136                 if (nir_opt_trivial_continues(shader)) {
 137                         progress = true;
 138                         NIR_PASS(progress, shader, nir_copy_prop);
 139                         NIR_PASS(progress, shader, nir_opt_remove_phis);
 140                         NIR_PASS(progress, shader, nir_opt_dce);
 141                 }
 142                 NIR_PASS(progress, shader, nir_opt_if);
 143                 NIR_PASS(progress, shader, nir_opt_dead_cf);
 144                 NIR_PASS(progress, shader, nir_opt_cse);
 145                 NIR_PASS(progress, shader, nir_opt_peephole_select, 8);
 146                 NIR_PASS(progress, shader, nir_opt_algebraic);
 147                 NIR_PASS(progress, shader, nir_opt_constant_folding);
 148                 NIR_PASS(progress, shader, nir_opt_undef);
 149                 NIR_PASS(progress, shader, nir_opt_conditional_discard);
 150                 if (shader->options->max_unroll_iterations) {
 151                         NIR_PASS(progress, shader, nir_opt_loop_unroll, 0);
 152                 }
 153         } while (progress && !optimize_conservatively);
 154
 155         NIR_PASS(progress, shader, nir_opt_shrink_load);
 156         NIR_PASS(progress, shader, nir_opt_move_load_ubo);
 157 }
 158
 159 nir_shader *
 160 radv_shader_compile_to_nir(struct radv_device *device,
 161                            struct radv_shader_module *module,
 162                            const char *entrypoint_name,
 163                            gl_shader_stage stage,
 164                            const VkSpecializationInfo *spec_info,
 165                            const VkPipelineCreateFlags flags)
 166 {
 167         nir_shader *nir;
 168         nir_function *entry_point;
 169         if (module->nir) {
 170                 /* Some things such as our meta clear/blit code will give us a NIR
 171                  * shader directly.  In that case, we just ignore the SPIR-V entirely
 172                  * and just use the NIR shader */
 173                 nir = module->nir;
 174                 nir->options = &nir_options;
 175                 nir_validate_shader(nir);
 176
 177                 assert(exec_list_length(&nir->functions) == 1);
 178                 struct exec_node *node = exec_list_get_head(&nir->functions);
 179                 entry_point = exec_node_data(nir_function, node, node);
 180         } else {
 181                 uint32_t *spirv = (uint32_t *) module->data;
 182                 assert(module->size % 4 == 0);
 183
 184                 if (device->instance->debug_flags & RADV_DEBUG_DUMP_SPIRV)
 185                         radv_print_spirv(spirv, module->size, stderr);
 186
 187                 uint32_t num_spec_entries = 0;
 188                 struct nir_spirv_specialization *spec_entries = NULL;
 189                 if (spec_info && spec_info->mapEntryCount > 0) {
 190                         num_spec_entries = spec_info->mapEntryCount;
 191                         spec_entries = malloc(num_spec_entries * sizeof(*spec_entries));
 192                         for (uint32_t i = 0; i < num_spec_entries; i++) {
 193                                 VkSpecializationMapEntry entry = spec_info->pMapEntries[i];
 194                                 const void *data = spec_info->pData + entry.offset;
 195                                 assert(data + entry.size <= spec_info->pData + spec_info->dataSize);
 196
 197                                 spec_entries[i].id = spec_info->pMapEntries[i].constantID;
 198                                 if (spec_info->dataSize == 8)
 199                                         spec_entries[i].data64 = *(const uint64_t *)data;
 200                                 else
 201                                         spec_entries[i].data32 = *(const uint32_t *)data;
 202                         }
 203                 }
 204                 const struct spirv_to_nir_options spirv_options = {
 205                         .caps = {
 206                                 .device_group = true,
 207                                 .draw_parameters = true,
 208                                 .float64 = true,
 209                                 .image_read_without_format = true,
 210                                 .image_write_without_format = true,
 211                                 .tessellation = true,
 212                                 .int64 = true,
 213                                 .multiview = true,
 214                                 .subgroup_ballot = true,
 215                                 .subgroup_basic = true,
 216                                 .subgroup_quad = true,
 217                                 .subgroup_shuffle = true,
 218                                 .subgroup_vote = true,
 219                                 .variable_pointers = true,
 220                                 .gcn_shader = true,
 221                                 .trinary_minmax = true,
 222                                 .shader_viewport_index_layer = true,
 223                                 .descriptor_array_dynamic_indexing = true,
 224                                 .runtime_descriptor_array = true,
 225                                 .stencil_export = true,
 226                                 .storage_16bit = true,
 227                         },
 228                 };
 229                 entry_point = spirv_to_nir(spirv, module->size / 4,
 230                                            spec_entries, num_spec_entries,
 231                                            stage, entrypoint_name,
 232                                            &spirv_options, &nir_options);
 233                 nir = entry_point->shader;
 234                 assert(nir->info.stage == stage);
 235                 nir_validate_shader(nir);
 236
 237                 free(spec_entries);
 238
 239                 /* We have to lower away local constant initializers right before we
 240                  * inline functions.  That way they get properly initialized at the top
 241                  * of the function and not at the top of its caller.
 242                  */
 243                 NIR_PASS_V(nir, nir_lower_constant_initializers, nir_var_local);
 244                 NIR_PASS_V(nir, nir_lower_returns);
 245                 NIR_PASS_V(nir, nir_inline_functions);
 246                 NIR_PASS_V(nir, nir_copy_prop);
 247
 248                 /* Pick off the single entrypoint that we want */
 249                 foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
 250                         if (func != entry_point)
 251                                 exec_node_remove(&func->node);
 252                 }
 253                 assert(exec_list_length(&nir->functions) == 1);
 254                 entry_point->name = ralloc_strdup(entry_point, "main");
 255
 256                 /* Make sure we lower constant initializers on output variables so that
 257                  * nir_remove_dead_variables below sees the corresponding stores
 258                  */
 259                 NIR_PASS_V(nir, nir_lower_constant_initializers, nir_var_shader_out);
 260
 261                 /* Now that we've deleted all but the main function, we can go ahead and
 262                  * lower the rest of the constant initializers.
 263                  */
 264                 NIR_PASS_V(nir, nir_lower_constant_initializers, ~0);
 265
 266                 /* Split member structs.  We do this before lower_io_to_temporaries so that
 267                  * it doesn't lower system values to temporaries by accident.
 268                  */
 269                 NIR_PASS_V(nir, nir_split_var_copies);
 270                 NIR_PASS_V(nir, nir_split_per_member_structs);
 271
 272                 NIR_PASS_V(nir, nir_remove_dead_variables,
 273                            nir_var_shader_in | nir_var_shader_out | nir_var_system_value);
 274
 275                 NIR_PASS_V(nir, nir_lower_system_values);
 276                 NIR_PASS_V(nir, nir_lower_clip_cull_distance_arrays);
 277         }
 278
 279         /* Vulkan uses the separate-shader linking model */
 280         nir->info.separate_shader = true;
 281
 282         nir_shader_gather_info(nir, entry_point->impl);
 283
 284         static const nir_lower_tex_options tex_options = {
 285           .lower_txp = ~0,
 286         };
 287
 288         nir_lower_tex(nir, &tex_options);
 289
 290         nir_lower_vars_to_ssa(nir);
 291
 292         if (nir->info.stage == MESA_SHADER_VERTEX ||
 293             nir->info.stage == MESA_SHADER_GEOMETRY) {
 294                 NIR_PASS_V(nir, nir_lower_io_to_temporaries,
 295                            nir_shader_get_entrypoint(nir), true, true);
 296         } else if (nir->info.stage == MESA_SHADER_TESS_EVAL||
 297                    nir->info.stage == MESA_SHADER_FRAGMENT) {
 298                 NIR_PASS_V(nir, nir_lower_io_to_temporaries,
 299                            nir_shader_get_entrypoint(nir), true, false);
 300         }
 301
 302         nir_split_var_copies(nir);
 303         nir_lower_var_copies(nir);
 304
 305         nir_lower_global_vars_to_local(nir);
 306         nir_remove_dead_variables(nir, nir_var_local);
 307         nir_lower_subgroups(nir, &(struct nir_lower_subgroups_options) {
 308                         .subgroup_size = 64,
 309                         .ballot_bit_size = 64,
 310                         .lower_to_scalar = 1,
 311                         .lower_subgroup_masks = 1,
 312                         .lower_shuffle = 1,
 313                         .lower_shuffle_to_32bit = 1,
 314                         .lower_vote_eq_to_ballot = 1,
 315                 });
 316
 317         if (!(flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT))
 318                 radv_optimize_nir(nir, false);
 319
 320         /* Indirect lowering must be called after the radv_optimize_nir() loop
 321          * has been called at least once. Otherwise indirect lowering can
 322          * bloat the instruction count of the loop and cause it to be
 323          * considered too large for unrolling.
 324          */
 325         ac_lower_indirect_derefs(nir, device->physical_device->rad_info.chip_class);
 326         radv_optimize_nir(nir, flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT);
 327
 328         return nir;
 329 }
 330
 331 void *
 332 radv_alloc_shader_memory(struct radv_device *device,
 333                          struct radv_shader_variant *shader)
 334 {
 335         mtx_lock(&device->shader_slab_mutex);
 336         list_for_each_entry(struct radv_shader_slab, slab, &device->shader_slabs, slabs) {
 337                 uint64_t offset = 0;
 338                 list_for_each_entry(struct radv_shader_variant, s, &slab->shaders, slab_list) {
 339                         if (s->bo_offset - offset >= shader->code_size) {
 340                                 shader->bo = slab->bo;
 341                                 shader->bo_offset = offset;
 342                                 list_addtail(&shader->slab_list, &s->slab_list);
 343                                 mtx_unlock(&device->shader_slab_mutex);
 344                                 return slab->ptr + offset;
 345                         }
 346                         offset = align_u64(s->bo_offset + s->code_size, 256);
 347                 }
 348                 if (slab->size - offset >= shader->code_size) {
 349                         shader->bo = slab->bo;
 350                         shader->bo_offset = offset;
 351                         list_addtail(&shader->slab_list, &slab->shaders);
 352                         mtx_unlock(&device->shader_slab_mutex);
 353                         return slab->ptr + offset;
 354                 }
 355         }
 356
 357         mtx_unlock(&device->shader_slab_mutex);
 358         struct radv_shader_slab *slab = calloc(1, sizeof(struct radv_shader_slab));
 359
 360         slab->size = 256 * 1024;
 361         slab->bo = device->ws->buffer_create(device->ws, slab->size, 256,
 362                                              RADEON_DOMAIN_VRAM,
 363                                              RADEON_FLAG_NO_INTERPROCESS_SHARING |
 364                                              (device->physical_device->cpdma_prefetch_writes_memory ?
 365                                                      0 : RADEON_FLAG_READ_ONLY));
 366         slab->ptr = (char*)device->ws->buffer_map(slab->bo);
 367         list_inithead(&slab->shaders);
 368
 369         mtx_lock(&device->shader_slab_mutex);
 370         list_add(&slab->slabs, &device->shader_slabs);
 371
 372         shader->bo = slab->bo;
 373         shader->bo_offset = 0;
 374         list_add(&shader->slab_list, &slab->shaders);
 375         mtx_unlock(&device->shader_slab_mutex);
 376         return slab->ptr;
 377 }
 378
 379 void
 380 radv_destroy_shader_slabs(struct radv_device *device)
 381 {
 382         list_for_each_entry_safe(struct radv_shader_slab, slab, &device->shader_slabs, slabs) {
 383                 device->ws->buffer_destroy(slab->bo);
 384                 free(slab);
 385         }
 386         mtx_destroy(&device->shader_slab_mutex);
 387 }
 388
 389 /* For the UMR disassembler. */
 390 #define DEBUGGER_END_OF_CODE_MARKER    0xbf9f0000 /* invalid instruction */
 391 #define DEBUGGER_NUM_MARKERS           5
 392
 393 static unsigned
 394 radv_get_shader_binary_size(struct ac_shader_binary *binary)
 395 {
 396         return binary->code_size + DEBUGGER_NUM_MARKERS * 4;
 397 }
 398
 399 static void
 400 radv_fill_shader_variant(struct radv_device *device,
 401                          struct radv_shader_variant *variant,
 402                          struct ac_shader_binary *binary,
 403                          gl_shader_stage stage)
 404 {
 405         bool scratch_enabled = variant->config.scratch_bytes_per_wave > 0;
 406         struct radv_shader_info *info = &variant->info.info;
 407         unsigned vgpr_comp_cnt = 0;
 408
 409         variant->code_size = radv_get_shader_binary_size(binary);
 410         variant->rsrc2 = S_00B12C_USER_SGPR(variant->info.num_user_sgprs) |
 411                          S_00B12C_SCRATCH_EN(scratch_enabled);
 412
 413         variant->rsrc1 = S_00B848_VGPRS((variant->config.num_vgprs - 1) / 4) |
 414                 S_00B848_SGPRS((variant->config.num_sgprs - 1) / 8) |
 415                 S_00B848_DX10_CLAMP(1) |
 416                 S_00B848_FLOAT_MODE(variant->config.float_mode);
 417
 418         switch (stage) {
 419         case MESA_SHADER_TESS_EVAL:
 420                 vgpr_comp_cnt = 3;
 421                 variant->rsrc2 |= S_00B12C_OC_LDS_EN(1);
 422                 break;
 423         case MESA_SHADER_TESS_CTRL:
 424                 if (device->physical_device->rad_info.chip_class >= GFX9) {
 425                         vgpr_comp_cnt = variant->info.vs.vgpr_comp_cnt;
 426                 } else {
 427                         variant->rsrc2 |= S_00B12C_OC_LDS_EN(1);
 428                 }
 429                 break;
 430         case MESA_SHADER_VERTEX:
 431         case MESA_SHADER_GEOMETRY:
 432                 vgpr_comp_cnt = variant->info.vs.vgpr_comp_cnt;
 433                 break;
 434         case MESA_SHADER_FRAGMENT:
 435                 break;
 436         case MESA_SHADER_COMPUTE:
 437                 variant->rsrc2 |=
 438                         S_00B84C_TGID_X_EN(info->cs.uses_block_id[0]) |
 439                         S_00B84C_TGID_Y_EN(info->cs.uses_block_id[1]) |
 440                         S_00B84C_TGID_Z_EN(info->cs.uses_block_id[2]) |
 441                         S_00B84C_TIDIG_COMP_CNT(info->cs.uses_thread_id[2] ? 2 :
 442                                                 info->cs.uses_thread_id[1] ? 1 : 0) |
 443                         S_00B84C_TG_SIZE_EN(info->cs.uses_local_invocation_idx) |
 444                         S_00B84C_LDS_SIZE(variant->config.lds_size);
 445                 break;
 446         default:
 447                 unreachable("unsupported shader type");
 448                 break;
 449         }
 450
 451         if (device->physical_device->rad_info.chip_class >= GFX9 &&
 452             stage == MESA_SHADER_GEOMETRY) {
 453                 unsigned es_type = variant->info.gs.es_type;
 454                 unsigned gs_vgpr_comp_cnt, es_vgpr_comp_cnt;
 455
 456                 if (es_type == MESA_SHADER_VERTEX) {
 457                         es_vgpr_comp_cnt = variant->info.vs.vgpr_comp_cnt;
 458                 } else if (es_type == MESA_SHADER_TESS_EVAL) {
 459                         es_vgpr_comp_cnt = 3;
 460                 } else {
 461                         unreachable("invalid shader ES type");
 462                 }
 463
 464                 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
 465                  * VGPR[0:4] are always loaded.
 466                  */
 467                 if (info->uses_invocation_id) {
 468                         gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID. */
 469                 } else if (info->uses_prim_id) {
 470                         gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
 471                 } else if (variant->info.gs.vertices_in >= 3) {
 472                         gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
 473                 } else {
 474                         gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
 475                 }
 476
 477                 variant->rsrc1 |= S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt);
 478                 variant->rsrc2 |= S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
 479                                   S_00B22C_OC_LDS_EN(es_type == MESA_SHADER_TESS_EVAL);
 480         } else if (device->physical_device->rad_info.chip_class >= GFX9 &&
 481                    stage == MESA_SHADER_TESS_CTRL) {
 482                 variant->rsrc1 |= S_00B428_LS_VGPR_COMP_CNT(vgpr_comp_cnt);
 483         } else {
 484                 variant->rsrc1 |= S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt);
 485         }
 486
 487         void *ptr = radv_alloc_shader_memory(device, variant);
 488         memcpy(ptr, binary->code, binary->code_size);
 489
 490         /* Add end-of-code markers for the UMR disassembler. */
 491        uint32_t *ptr32 = (uint32_t *)ptr + binary->code_size / 4;
 492        for (unsigned i = 0; i < DEBUGGER_NUM_MARKERS; i++)
 493                 ptr32[i] = DEBUGGER_END_OF_CODE_MARKER;
 494
 495 }
 496
 497 static void radv_init_llvm_target()
 498 {
 499         LLVMInitializeAMDGPUTargetInfo();
 500         LLVMInitializeAMDGPUTarget();
 501         LLVMInitializeAMDGPUTargetMC();
 502         LLVMInitializeAMDGPUAsmPrinter();
 503
 504         /* For inline assembly. */
 505         LLVMInitializeAMDGPUAsmParser();
 506
 507         /* Workaround for bug in llvm 4.0 that causes image intrinsics
 508          * to disappear.
 509          * https://reviews.llvm.org/D26348
 510          *
 511          * Workaround for bug in llvm that causes the GPU to hang in presence
 512          * of nested loops because there is an exec mask issue. The proper
 513          * solution is to fix LLVM but this might require a bunch of work.
 514          * https://bugs.llvm.org/show_bug.cgi?id=37744
 515          *
 516          * "mesa" is the prefix for error messages.
 517          */
 518         const char *argv[3] = { "mesa", "-simplifycfg-sink-common=false",
 519                                 "-amdgpu-skip-threshold=1" };
 520         LLVMParseCommandLineOptions(3, argv, NULL);
 521 }
 522
 523 static once_flag radv_init_llvm_target_once_flag = ONCE_FLAG_INIT;
 524
 525 static void radv_init_llvm_once(void)
 526 {
 527         call_once(&radv_init_llvm_target_once_flag, radv_init_llvm_target);
 528 }
 529
 530 static struct radv_shader_variant *
 531 shader_variant_create(struct radv_device *device,
 532                       struct radv_shader_module *module,
 533                       struct nir_shader * const *shaders,
 534                       int shader_count,
 535                       gl_shader_stage stage,
 536                       struct radv_nir_compiler_options *options,
 537                       bool gs_copy_shader,
 538                       void **code_out,
 539                       unsigned *code_size_out)
 540 {
 541         enum radeon_family chip_family = device->physical_device->rad_info.family;
 542         enum ac_target_machine_options tm_options = 0;
 543         struct radv_shader_variant *variant;
 544         struct ac_shader_binary binary;
 545         struct ac_llvm_compiler ac_llvm;
 546         bool thread_compiler;
 547         variant = calloc(1, sizeof(struct radv_shader_variant));
 548         if (!variant)
 549                 return NULL;
 550
 551         options->family = chip_family;
 552         options->chip_class = device->physical_device->rad_info.chip_class;
 553         options->dump_shader = radv_can_dump_shader(device, module, gs_copy_shader);
 554         options->dump_preoptir = options->dump_shader &&
 555                                  device->instance->debug_flags & RADV_DEBUG_PREOPTIR;
 556         options->record_llvm_ir = device->keep_shader_info;
 557         options->check_ir = device->instance->debug_flags & RADV_DEBUG_CHECKIR;
 558         options->tess_offchip_block_dw_size = device->tess_offchip_block_dw_size;
 559         options->address32_hi = device->physical_device->rad_info.address32_hi;
 560
 561         if (options->supports_spill)
 562                 tm_options |= AC_TM_SUPPORTS_SPILL;
 563         if (device->instance->perftest_flags & RADV_PERFTEST_SISCHED)
 564                 tm_options |= AC_TM_SISCHED;
 565         if (options->check_ir)
 566                 tm_options |= AC_TM_CHECK_IR;
 567
 568         thread_compiler = !(device->instance->debug_flags & RADV_DEBUG_NOTHREADLLVM);
 569         radv_init_llvm_once();
 570         radv_init_llvm_compiler(&ac_llvm, false,
 571                                 thread_compiler,
 572                                 chip_family, tm_options);
 573         if (gs_copy_shader) {
 574                 assert(shader_count == 1);
 575                 radv_compile_gs_copy_shader(&ac_llvm, *shaders, &binary,
 576                                             &variant->config, &variant->info,
 577                                             options);
 578         } else {
 579                 radv_compile_nir_shader(&ac_llvm, &binary, &variant->config,
 580                                         &variant->info, shaders, shader_count,
 581                                         options);
 582         }
 583
 584         radv_destroy_llvm_compiler(&ac_llvm, thread_compiler);
 585
 586         radv_fill_shader_variant(device, variant, &binary, stage);
 587
 588         if (code_out) {
 589                 *code_out = binary.code;
 590                 *code_size_out = binary.code_size;
 591         } else
 592                 free(binary.code);
 593         free(binary.config);
 594         free(binary.rodata);
 595         free(binary.global_symbol_offsets);
 596         free(binary.relocs);
 597         variant->ref_count = 1;
 598
 599         if (device->keep_shader_info) {
 600                 variant->disasm_string = binary.disasm_string;
 601                 variant->llvm_ir_string = binary.llvm_ir_string;
 602                 if (!gs_copy_shader && !module->nir) {
 603                         variant->nir = *shaders;
 604                         variant->spirv = (uint32_t *)module->data;
 605                         variant->spirv_size = module->size;
 606                 }
 607         } else {
 608                 free(binary.disasm_string);
 609         }
 610
 611         return variant;
 612 }
 613
 614 struct radv_shader_variant *
 615 radv_shader_variant_create(struct radv_device *device,
 616                            struct radv_shader_module *module,
 617                            struct nir_shader *const *shaders,
 618                            int shader_count,
 619                            struct radv_pipeline_layout *layout,
 620                            const struct radv_shader_variant_key *key,
 621                            void **code_out,
 622                            unsigned *code_size_out)
 623 {
 624         struct radv_nir_compiler_options options = {0};
 625
 626         options.layout = layout;
 627         if (key)
 628                 options.key = *key;
 629
 630         options.unsafe_math = !!(device->instance->debug_flags & RADV_DEBUG_UNSAFE_MATH);
 631         options.supports_spill = true;
 632
 633         return shader_variant_create(device, module, shaders, shader_count, shaders[shader_count - 1]->info.stage,
 634                                      &options, false, code_out, code_size_out);
 635 }
 636
 637 struct radv_shader_variant *
 638 radv_create_gs_copy_shader(struct radv_device *device,
 639                            struct nir_shader *shader,
 640                            void **code_out,
 641                            unsigned *code_size_out,
 642                            bool multiview)
 643 {
 644         struct radv_nir_compiler_options options = {0};
 645
 646         options.key.has_multiview_view_index = multiview;
 647
 648         return shader_variant_create(device, NULL, &shader, 1, MESA_SHADER_VERTEX,
 649                                      &options, true, code_out, code_size_out);
 650 }
 651
 652 void
 653 radv_shader_variant_destroy(struct radv_device *device,
 654                             struct radv_shader_variant *variant)
 655 {
 656         if (!p_atomic_dec_zero(&variant->ref_count))
 657                 return;
 658
 659         mtx_lock(&device->shader_slab_mutex);
 660         list_del(&variant->slab_list);
 661         mtx_unlock(&device->shader_slab_mutex);
 662
 663         ralloc_free(variant->nir);
 664         free(variant->disasm_string);
 665         free(variant->llvm_ir_string);
 666         free(variant);
 667 }
 668
 669 const char *
 670 radv_get_shader_name(struct radv_shader_variant *var, gl_shader_stage stage)
 671 {
 672         switch (stage) {
 673         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";
 674         case MESA_SHADER_GEOMETRY: return "Geometry Shader";
 675         case MESA_SHADER_FRAGMENT: return "Pixel Shader";
 676         case MESA_SHADER_COMPUTE: return "Compute Shader";
 677         case MESA_SHADER_TESS_CTRL: return "Tessellation Control Shader";
 678         case MESA_SHADER_TESS_EVAL: return var->info.tes.as_es ? "Tessellation Evaluation Shader as ES" : "Tessellation Evaluation Shader as VS";
 679         default:
 680                 return "Unknown shader";
 681         };
 682 }
 683
 684 static void
 685 generate_shader_stats(struct radv_device *device,
 686                       struct radv_shader_variant *variant,
 687                       gl_shader_stage stage,
 688                       struct _mesa_string_buffer *buf)
 689 {
 690         unsigned lds_increment = device->physical_device->rad_info.chip_class >= CIK ? 512 : 256;
 691         struct ac_shader_config *conf;
 692         unsigned max_simd_waves;
 693         unsigned lds_per_wave = 0;
 694
 695         max_simd_waves = ac_get_max_simd_waves(device->physical_device->rad_info.family);
 696
 697         conf = &variant->config;
 698
 699         if (stage == MESA_SHADER_FRAGMENT) {
 700                 lds_per_wave = conf->lds_size * lds_increment +
 701                                align(variant->info.fs.num_interp * 48,
 702                                      lds_increment);
 703         }
 704
 705         if (conf->num_sgprs)
 706                 max_simd_waves =
 707                         MIN2(max_simd_waves,
 708                              radv_get_num_physical_sgprs(device->physical_device) / conf->num_sgprs);
 709
 710         if (conf->num_vgprs)
 711                 max_simd_waves =
 712                         MIN2(max_simd_waves,
 713                              RADV_NUM_PHYSICAL_VGPRS / conf->num_vgprs);
 714
 715         /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
 716          * that PS can use.
 717          */
 718         if (lds_per_wave)
 719                 max_simd_waves = MIN2(max_simd_waves, 16384 / lds_per_wave);
 720
 721         if (stage == MESA_SHADER_FRAGMENT) {
 722                 _mesa_string_buffer_printf(buf, "*** SHADER CONFIG ***\n"
 723                                            "SPI_PS_INPUT_ADDR = 0x%04x\n"
 724                                            "SPI_PS_INPUT_ENA  = 0x%04x\n",
 725                                            conf->spi_ps_input_addr, conf->spi_ps_input_ena);
 726         }
 727
 728         _mesa_string_buffer_printf(buf, "*** SHADER STATS ***\n"
 729                                    "SGPRS: %d\n"
 730                                    "VGPRS: %d\n"
 731                                    "Spilled SGPRs: %d\n"
 732                                    "Spilled VGPRs: %d\n"
 733                                    "PrivMem VGPRS: %d\n"
 734                                    "Code Size: %d bytes\n"
 735                                    "LDS: %d blocks\n"
 736                                    "Scratch: %d bytes per wave\n"
 737                                    "Max Waves: %d\n"
 738                                    "********************\n\n\n",
 739                                    conf->num_sgprs, conf->num_vgprs,
 740                                    conf->spilled_sgprs, conf->spilled_vgprs,
 741                                    variant->info.private_mem_vgprs, variant->code_size,
 742                                    conf->lds_size, conf->scratch_bytes_per_wave,
 743                                    max_simd_waves);
 744 }
 745
 746 void
 747 radv_shader_dump_stats(struct radv_device *device,
 748                        struct radv_shader_variant *variant,
 749                        gl_shader_stage stage,
 750                        FILE *file)
 751 {
 752         struct _mesa_string_buffer *buf = _mesa_string_buffer_create(NULL, 256);
 753
 754         generate_shader_stats(device, variant, stage, buf);
 755
 756         fprintf(file, "\n%s:\n", radv_get_shader_name(variant, stage));
 757         fprintf(file, "%s", buf->buf);
 758
 759         _mesa_string_buffer_destroy(buf);
 760 }
 761
 762 VkResult
 763 radv_GetShaderInfoAMD(VkDevice _device,
 764                       VkPipeline _pipeline,
 765                       VkShaderStageFlagBits shaderStage,
 766                       VkShaderInfoTypeAMD infoType,
 767                       size_t* pInfoSize,
 768                       void* pInfo)
 769 {
 770         RADV_FROM_HANDLE(radv_device, device, _device);
 771         RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline);
 772         gl_shader_stage stage = vk_to_mesa_shader_stage(shaderStage);
 773         struct radv_shader_variant *variant = pipeline->shaders[stage];
 774         struct _mesa_string_buffer *buf;
 775         VkResult result = VK_SUCCESS;
 776
 777         /* Spec doesn't indicate what to do if the stage is invalid, so just
 778          * return no info for this. */
 779         if (!variant)
 780                 return vk_error(device->instance, VK_ERROR_FEATURE_NOT_PRESENT);
 781
 782         switch (infoType) {
 783         case VK_SHADER_INFO_TYPE_STATISTICS_AMD:
 784                 if (!pInfo) {
 785                         *pInfoSize = sizeof(VkShaderStatisticsInfoAMD);
 786                 } else {
 787                         unsigned lds_multiplier = device->physical_device->rad_info.chip_class >= CIK ? 512 : 256;
 788                         struct ac_shader_config *conf = &variant->config;
 789
 790                         VkShaderStatisticsInfoAMD statistics = {};
 791                         statistics.shaderStageMask = shaderStage;
 792                         statistics.numPhysicalVgprs = RADV_NUM_PHYSICAL_VGPRS;
 793                         statistics.numPhysicalSgprs = radv_get_num_physical_sgprs(device->physical_device);
 794                         statistics.numAvailableSgprs = statistics.numPhysicalSgprs;
 795
 796                         if (stage == MESA_SHADER_COMPUTE) {
 797                                 unsigned *local_size = variant->nir->info.cs.local_size;
 798                                 unsigned workgroup_size = local_size[0] * local_size[1] * local_size[2];
 799
 800                                 statistics.numAvailableVgprs = statistics.numPhysicalVgprs /
 801                                                                ceil((double)workgroup_size / statistics.numPhysicalVgprs);
 802
 803                                 statistics.computeWorkGroupSize[0] = local_size[0];
 804                                 statistics.computeWorkGroupSize[1] = local_size[1];
 805                                 statistics.computeWorkGroupSize[2] = local_size[2];
 806                         } else {
 807                                 statistics.numAvailableVgprs = statistics.numPhysicalVgprs;
 808                         }
 809
 810                         statistics.resourceUsage.numUsedVgprs = conf->num_vgprs;
 811                         statistics.resourceUsage.numUsedSgprs = conf->num_sgprs;
 812                         statistics.resourceUsage.ldsSizePerLocalWorkGroup = 32768;
 813                         statistics.resourceUsage.ldsUsageSizeInBytes = conf->lds_size * lds_multiplier;
 814                         statistics.resourceUsage.scratchMemUsageInBytes = conf->scratch_bytes_per_wave;
 815
 816                         size_t size = *pInfoSize;
 817                         *pInfoSize = sizeof(statistics);
 818
 819                         memcpy(pInfo, &statistics, MIN2(size, *pInfoSize));
 820
 821                         if (size < *pInfoSize)
 822                                 result = VK_INCOMPLETE;
 823                 }
 824
 825                 break;
 826         case VK_SHADER_INFO_TYPE_DISASSEMBLY_AMD:
 827                 buf = _mesa_string_buffer_create(NULL, 1024);
 828
 829                 _mesa_string_buffer_printf(buf, "%s:\n", radv_get_shader_name(variant, stage));
 830                 _mesa_string_buffer_printf(buf, "%s\n\n", variant->disasm_string);
 831                 generate_shader_stats(device, variant, stage, buf);
 832
 833                 /* Need to include the null terminator. */
 834                 size_t length = buf->length + 1;
 835
 836                 if (!pInfo) {
 837                         *pInfoSize = length;
 838                 } else {
 839                         size_t size = *pInfoSize;
 840                         *pInfoSize = length;
 841
 842                         memcpy(pInfo, buf->buf, MIN2(size, length));
 843
 844                         if (size < length)
 845                                 result = VK_INCOMPLETE;
 846                 }
 847
 848                 _mesa_string_buffer_destroy(buf);
 849                 break;
 850         default:
 851                 /* VK_SHADER_INFO_TYPE_BINARY_AMD unimplemented for now. */
 852                 result = VK_ERROR_FEATURE_NOT_PRESENT;
 853                 break;
 854         }
 855
 856         return result;
 857 }