src/amd/vulkan/radv_shader.h

   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 #ifndef RADV_SHADER_H
  29 #define RADV_SHADER_H
  30
  31 #include "ac_binary.h"
  32 #include "amd_family.h"
  33 #include "radv_constants.h"
  34
  35 #include "nir/nir.h"
  36 #include "vulkan/vulkan.h"
  37
  38 struct radv_device;
  39
  40 struct radv_shader_module {
  41         struct nir_shader *nir;
  42         unsigned char sha1[20];
  43         uint32_t size;
  44         char data[0];
  45 };
  46
  47 enum {
  48         RADV_ALPHA_ADJUST_NONE = 0,
  49         RADV_ALPHA_ADJUST_SNORM = 1,
  50         RADV_ALPHA_ADJUST_SINT = 2,
  51         RADV_ALPHA_ADJUST_SSCALED = 3,
  52 };
  53
  54 struct radv_vs_out_key {
  55         uint32_t as_es:1;
  56         uint32_t as_ls:1;
  57         uint32_t as_ngg:1;
  58         uint32_t as_ngg_passthrough:1;
  59         uint32_t export_prim_id:1;
  60         uint32_t export_layer_id:1;
  61         uint32_t export_clip_dists:1;
  62 };
  63
  64 struct radv_vs_variant_key {
  65         struct radv_vs_out_key out;
  66
  67         uint32_t instance_rate_inputs;
  68         uint32_t instance_rate_divisors[MAX_VERTEX_ATTRIBS];
  69         uint8_t vertex_attribute_formats[MAX_VERTEX_ATTRIBS];
  70         uint32_t vertex_attribute_bindings[MAX_VERTEX_ATTRIBS];
  71         uint32_t vertex_attribute_offsets[MAX_VERTEX_ATTRIBS];
  72         uint32_t vertex_attribute_strides[MAX_VERTEX_ATTRIBS];
  73
  74         /* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
  75          * so we may need to fix it up. */
  76         uint64_t alpha_adjust;
  77
  78         /* For some formats the channels have to be shuffled. */
  79         uint32_t post_shuffle;
  80
  81         /* Output primitive type. */
  82         uint8_t outprim;
  83 };
  84
  85 struct radv_tes_variant_key {
  86         struct radv_vs_out_key out;
  87
  88         uint8_t num_patches;
  89         uint8_t tcs_num_outputs;
  90 };
  91
  92 struct radv_tcs_variant_key {
  93         struct radv_vs_variant_key vs_key;
  94         unsigned primitive_mode;
  95         unsigned input_vertices;
  96         unsigned num_inputs;
  97         uint32_t tes_reads_tess_factors:1;
  98 };
  99
 100 struct radv_fs_variant_key {
 101         uint32_t col_format;
 102         uint8_t log2_ps_iter_samples;
 103         uint8_t num_samples;
 104         uint32_t is_int8;
 105         uint32_t is_int10;
 106 };
 107
 108 struct radv_cs_variant_key {
 109         uint8_t subgroup_size;
 110 };
 111
 112 struct radv_shader_variant_key {
 113         union {
 114                 struct radv_vs_variant_key vs;
 115                 struct radv_fs_variant_key fs;
 116                 struct radv_tes_variant_key tes;
 117                 struct radv_tcs_variant_key tcs;
 118                 struct radv_cs_variant_key cs;
 119
 120                 /* A common prefix of the vs and tes keys. */
 121                 struct radv_vs_out_key vs_common_out;
 122         };
 123         bool has_multiview_view_index;
 124 };
 125
 126 struct radv_nir_compiler_options {
 127         struct radv_pipeline_layout *layout;
 128         struct radv_shader_variant_key key;
 129         bool explicit_scratch_args;
 130         bool clamp_shadow_reference;
 131         bool robust_buffer_access;
 132         bool dump_shader;
 133         bool dump_preoptir;
 134         bool record_ir;
 135         bool record_stats;
 136         bool check_ir;
 137         bool has_ls_vgpr_init_bug;
 138         bool use_ngg_streamout;
 139         enum radeon_family family;
 140         enum chip_class chip_class;
 141         uint32_t tess_offchip_block_dw_size;
 142         uint32_t address32_hi;
 143 };
 144
 145 enum radv_ud_index {
 146         AC_UD_SCRATCH_RING_OFFSETS = 0,
 147         AC_UD_PUSH_CONSTANTS = 1,
 148         AC_UD_INLINE_PUSH_CONSTANTS = 2,
 149         AC_UD_INDIRECT_DESCRIPTOR_SETS = 3,
 150         AC_UD_VIEW_INDEX = 4,
 151         AC_UD_STREAMOUT_BUFFERS = 5,
 152         AC_UD_NGG_GS_STATE = 6,
 153         AC_UD_SHADER_START = 7,
 154         AC_UD_VS_VERTEX_BUFFERS = AC_UD_SHADER_START,
 155         AC_UD_VS_BASE_VERTEX_START_INSTANCE,
 156         AC_UD_VS_MAX_UD,
 157         AC_UD_PS_MAX_UD,
 158         AC_UD_CS_GRID_SIZE = AC_UD_SHADER_START,
 159         AC_UD_CS_MAX_UD,
 160         AC_UD_GS_MAX_UD,
 161         AC_UD_TCS_MAX_UD,
 162         AC_UD_TES_MAX_UD,
 163         AC_UD_MAX_UD = AC_UD_TCS_MAX_UD,
 164 };
 165
 166 struct radv_stream_output {
 167         uint8_t location;
 168         uint8_t buffer;
 169         uint16_t offset;
 170         uint8_t component_mask;
 171         uint8_t stream;
 172 };
 173
 174 struct radv_streamout_info {
 175         uint16_t num_outputs;
 176         struct radv_stream_output outputs[MAX_SO_OUTPUTS];
 177         uint16_t strides[MAX_SO_BUFFERS];
 178         uint32_t enabled_stream_buffers_mask;
 179 };
 180
 181 struct radv_userdata_info {
 182         int8_t sgpr_idx;
 183         uint8_t num_sgprs;
 184 };
 185
 186 struct radv_userdata_locations {
 187         struct radv_userdata_info descriptor_sets[MAX_SETS];
 188         struct radv_userdata_info shader_data[AC_UD_MAX_UD];
 189         uint32_t descriptor_sets_enabled;
 190 };
 191
 192 struct radv_vs_output_info {
 193         uint8_t vs_output_param_offset[VARYING_SLOT_MAX];
 194         uint8_t clip_dist_mask;
 195         uint8_t cull_dist_mask;
 196         uint8_t param_exports;
 197         bool writes_pointsize;
 198         bool writes_layer;
 199         bool writes_viewport_index;
 200         bool export_prim_id;
 201         unsigned pos_exports;
 202 };
 203
 204 struct radv_es_output_info {
 205        uint32_t esgs_itemsize;
 206 };
 207
 208 struct gfx9_gs_info {
 209         uint32_t vgt_gs_onchip_cntl;
 210         uint32_t vgt_gs_max_prims_per_subgroup;
 211         uint32_t vgt_esgs_ring_itemsize;
 212         uint32_t lds_size;
 213 };
 214
 215 struct gfx10_ngg_info {
 216         uint16_t ngg_emit_size; /* in dwords */
 217         uint32_t hw_max_esverts;
 218         uint32_t max_gsprims;
 219         uint32_t max_out_verts;
 220         uint32_t prim_amp_factor;
 221         uint32_t vgt_esgs_ring_itemsize;
 222         uint32_t esgs_ring_size;
 223         bool max_vert_out_per_gs_instance;
 224 };
 225
 226 struct radv_shader_info {
 227         bool loads_push_constants;
 228         bool loads_dynamic_offsets;
 229         uint8_t min_push_constant_used;
 230         uint8_t max_push_constant_used;
 231         bool has_only_32bit_push_constants;
 232         bool has_indirect_push_constants;
 233         uint8_t num_inline_push_consts;
 234         uint8_t base_inline_push_consts;
 235         uint32_t desc_set_used_mask;
 236         bool needs_multiview_view_index;
 237         bool uses_invocation_id;
 238         bool uses_prim_id;
 239         uint8_t wave_size;
 240         uint8_t ballot_bit_size;
 241         struct radv_userdata_locations user_sgprs_locs;
 242         unsigned num_user_sgprs;
 243         unsigned num_input_sgprs;
 244         unsigned num_input_vgprs;
 245         unsigned private_mem_vgprs;
 246         bool need_indirect_descriptor_sets;
 247         bool is_ngg;
 248         bool is_ngg_passthrough;
 249         struct {
 250                 uint64_t ls_outputs_written;
 251                 uint8_t input_usage_mask[VERT_ATTRIB_MAX];
 252                 uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
 253                 bool has_vertex_buffers; /* needs vertex buffers and base/start */
 254                 bool needs_draw_id;
 255                 bool needs_instance_id;
 256                 struct radv_vs_output_info outinfo;
 257                 struct radv_es_output_info es_info;
 258                 bool as_es;
 259                 bool as_ls;
 260                 bool export_prim_id;
 261         } vs;
 262         struct {
 263                 uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
 264                 uint8_t num_stream_output_components[4];
 265                 uint8_t output_streams[VARYING_SLOT_VAR31 + 1];
 266                 uint8_t max_stream;
 267                 bool writes_memory;
 268                 unsigned gsvs_vertex_size;
 269                 unsigned max_gsvs_emit_size;
 270                 unsigned vertices_in;
 271                 unsigned vertices_out;
 272                 unsigned output_prim;
 273                 unsigned invocations;
 274                 unsigned es_type; /* GFX9: VS or TES */
 275         } gs;
 276         struct {
 277                 uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
 278                 struct radv_vs_output_info outinfo;
 279                 struct radv_es_output_info es_info;
 280                 bool as_es;
 281                 unsigned primitive_mode;
 282                 enum gl_tess_spacing spacing;
 283                 bool ccw;
 284                 bool point_mode;
 285                 bool export_prim_id;
 286         } tes;
 287         struct {
 288                 bool force_persample;
 289                 bool needs_sample_positions;
 290                 bool writes_memory;
 291                 bool writes_z;
 292                 bool writes_stencil;
 293                 bool writes_sample_mask;
 294                 bool has_pcoord;
 295                 bool prim_id_input;
 296                 bool layer_input;
 297                 uint8_t num_input_clips_culls;
 298                 uint32_t input_mask;
 299                 uint32_t flat_shaded_mask;
 300                 uint32_t explicit_shaded_mask;
 301                 uint32_t float16_shaded_mask;
 302                 uint32_t num_interp;
 303                 bool can_discard;
 304                 bool early_fragment_test;
 305                 bool post_depth_coverage;
 306         } ps;
 307         struct {
 308                 bool uses_grid_size;
 309                 bool uses_block_id[3];
 310                 bool uses_thread_id[3];
 311                 bool uses_local_invocation_idx;
 312                 unsigned block_size[3];
 313         } cs;
 314         struct {
 315                 uint64_t outputs_written;
 316                 uint64_t patch_outputs_written;
 317                 unsigned tcs_vertices_out;
 318                 uint32_t num_patches;
 319                 uint32_t lds_size;
 320         } tcs;
 321
 322         struct radv_streamout_info so;
 323
 324         struct gfx9_gs_info gs_ring_info;
 325         struct gfx10_ngg_info ngg_info;
 326
 327         unsigned float_controls_mode;
 328 };
 329
 330 enum radv_shader_binary_type {
 331         RADV_BINARY_TYPE_LEGACY,
 332         RADV_BINARY_TYPE_RTLD
 333 };
 334
 335 struct radv_shader_binary {
 336         enum radv_shader_binary_type type;
 337         gl_shader_stage stage;
 338         bool is_gs_copy_shader;
 339
 340         struct radv_shader_info info;
 341
 342         /* Self-referential size so we avoid consistency issues. */
 343         uint32_t total_size;
 344 };
 345
 346 struct radv_shader_binary_legacy {
 347         struct radv_shader_binary base;
 348         struct ac_shader_config config;
 349         unsigned code_size;
 350         unsigned exec_size;
 351         unsigned ir_size;
 352         unsigned disasm_size;
 353         unsigned stats_size;
 354
 355         /* data has size of stats_size + code_size + ir_size + disasm_size + 2,
 356          * where the +2 is for 0 of the ir strings. */
 357         uint8_t data[0];
 358 };
 359
 360 struct radv_shader_binary_rtld {
 361         struct radv_shader_binary base;
 362         unsigned elf_size;
 363         unsigned llvm_ir_size;
 364         uint8_t data[0];
 365 };
 366
 367 struct radv_compiler_statistic_info {
 368         char name[32];
 369         char desc[64];
 370 };
 371
 372 struct radv_compiler_statistics {
 373         unsigned count;
 374         struct radv_compiler_statistic_info *infos;
 375         uint32_t values[];
 376 };
 377
 378 struct radv_shader_variant {
 379         uint32_t ref_count;
 380
 381         struct radeon_winsys_bo *bo;
 382         uint64_t bo_offset;
 383         struct ac_shader_config config;
 384         uint32_t code_size;
 385         uint32_t exec_size;
 386         struct radv_shader_info info;
 387
 388         /* debug only */
 389         char *spirv;
 390         uint32_t spirv_size;
 391         char *nir_string;
 392         char *disasm_string;
 393         char *ir_string;
 394         struct radv_compiler_statistics *statistics;
 395
 396         struct list_head slab_list;
 397 };
 398
 399 struct radv_shader_slab {
 400         struct list_head slabs;
 401         struct list_head shaders;
 402         struct radeon_winsys_bo *bo;
 403         uint64_t size;
 404         char *ptr;
 405 };
 406
 407 void
 408 radv_optimize_nir(struct nir_shader *shader, bool optimize_conservatively,
 409                   bool allow_copies);
 410 bool
 411 radv_nir_lower_ycbcr_textures(nir_shader *shader,
 412                              const struct radv_pipeline_layout *layout);
 413
 414 nir_shader *
 415 radv_shader_compile_to_nir(struct radv_device *device,
 416                            struct radv_shader_module *module,
 417                            const char *entrypoint_name,
 418                            gl_shader_stage stage,
 419                            const VkSpecializationInfo *spec_info,
 420                            const VkPipelineCreateFlags flags,
 421                            const struct radv_pipeline_layout *layout,
 422                            unsigned subgroup_size, unsigned ballot_bit_size);
 423
 424 void *
 425 radv_alloc_shader_memory(struct radv_device *device,
 426                           struct radv_shader_variant *shader);
 427
 428 void
 429 radv_destroy_shader_slabs(struct radv_device *device);
 430
 431 void
 432 radv_create_shaders(struct radv_pipeline *pipeline,
 433                     struct radv_device *device,
 434                     struct radv_pipeline_cache *cache,
 435                     const struct radv_pipeline_key *key,
 436                     const VkPipelineShaderStageCreateInfo **pStages,
 437                     const VkPipelineCreateFlags flags,
 438                     VkPipelineCreationFeedbackEXT *pipeline_feedback,
 439                     VkPipelineCreationFeedbackEXT **stage_feedbacks);
 440
 441 struct radv_shader_variant *
 442 radv_shader_variant_create(struct radv_device *device,
 443                            const struct radv_shader_binary *binary,
 444                            bool keep_shader_info);
 445 struct radv_shader_variant *
 446 radv_shader_variant_compile(struct radv_device *device,
 447                             struct radv_shader_module *module,
 448                             struct nir_shader *const *shaders,
 449                             int shader_count,
 450                             struct radv_pipeline_layout *layout,
 451                             const struct radv_shader_variant_key *key,
 452                             struct radv_shader_info *info,
 453                             bool keep_shader_info, bool keep_statistic_info,
 454                             struct radv_shader_binary **binary_out);
 455
 456 struct radv_shader_variant *
 457 radv_create_gs_copy_shader(struct radv_device *device, struct nir_shader *nir,
 458                            struct radv_shader_info *info,
 459                            struct radv_shader_binary **binary_out,
 460                            bool multiview,  bool keep_shader_info,
 461                            bool keep_statistic_info);
 462
 463 void
 464 radv_shader_variant_destroy(struct radv_device *device,
 465                             struct radv_shader_variant *variant);
 466
 467
 468 unsigned
 469 radv_get_max_waves(struct radv_device *device,
 470                    struct radv_shader_variant *variant,
 471                    gl_shader_stage stage);
 472
 473 unsigned
 474 radv_get_max_workgroup_size(enum chip_class chip_class,
 475                             gl_shader_stage stage,
 476                             const unsigned *sizes);
 477
 478 const char *
 479 radv_get_shader_name(struct radv_shader_info *info,
 480                      gl_shader_stage stage);
 481
 482 void
 483 radv_shader_dump_stats(struct radv_device *device,
 484                        struct radv_shader_variant *variant,
 485                        gl_shader_stage stage,
 486                        FILE *file);
 487
 488 bool
 489 radv_can_dump_shader(struct radv_device *device,
 490                      struct radv_shader_module *module,
 491                      bool is_gs_copy_shader);
 492
 493 bool
 494 radv_can_dump_shader_stats(struct radv_device *device,
 495                            struct radv_shader_module *module);
 496
 497 static inline unsigned
 498 shader_io_get_unique_index(gl_varying_slot slot)
 499 {
 500         /* handle patch indices separate */
 501         if (slot == VARYING_SLOT_TESS_LEVEL_OUTER)
 502                 return 0;
 503         if (slot == VARYING_SLOT_TESS_LEVEL_INNER)
 504                 return 1;
 505         if (slot >= VARYING_SLOT_PATCH0 && slot <= VARYING_SLOT_TESS_MAX)
 506                 return 2 + (slot - VARYING_SLOT_PATCH0);
 507         if (slot == VARYING_SLOT_POS)
 508                 return 0;
 509         if (slot == VARYING_SLOT_PSIZ)
 510                 return 1;
 511         if (slot == VARYING_SLOT_CLIP_DIST0)
 512                 return 2;
 513         if (slot == VARYING_SLOT_CLIP_DIST1)
 514                 return 3;
 515         /* 3 is reserved for clip dist as well */
 516         if (slot >= VARYING_SLOT_VAR0 && slot <= VARYING_SLOT_VAR31)
 517                 return 4 + (slot - VARYING_SLOT_VAR0);
 518         unreachable("illegal slot in get unique index\n");
 519 }
 520
 521 static inline unsigned
 522 calculate_tess_lds_size(unsigned tcs_num_input_vertices,
 523                         unsigned tcs_num_output_vertices,
 524                         unsigned tcs_num_inputs,
 525                         unsigned tcs_num_patches,
 526                         unsigned tcs_outputs_written,
 527                         unsigned tcs_per_patch_outputs_written)
 528 {
 529         unsigned num_tcs_outputs = util_last_bit64(tcs_outputs_written);
 530         unsigned num_tcs_patch_outputs = util_last_bit64(tcs_per_patch_outputs_written);
 531
 532         unsigned input_vertex_size = tcs_num_inputs * 16;
 533         unsigned output_vertex_size = num_tcs_outputs * 16;
 534
 535         unsigned input_patch_size = tcs_num_input_vertices * input_vertex_size;
 536
 537         unsigned pervertex_output_patch_size = tcs_num_output_vertices * output_vertex_size;
 538         unsigned output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
 539
 540         unsigned output_patch0_offset = input_patch_size * tcs_num_patches;
 541
 542         return output_patch0_offset + output_patch_size * tcs_num_patches;
 543 }
 544
 545 static inline unsigned
 546 get_tcs_num_patches(unsigned tcs_num_input_vertices,
 547                         unsigned tcs_num_output_vertices,
 548                         unsigned tcs_num_inputs,
 549                         unsigned tcs_outputs_written,
 550                         unsigned tcs_per_patch_outputs_written,
 551                         unsigned tess_offchip_block_dw_size,
 552                         enum chip_class chip_class,
 553                         enum radeon_family family)
 554 {
 555         uint32_t input_vertex_size = tcs_num_inputs * 16;
 556         uint32_t input_patch_size = tcs_num_input_vertices * input_vertex_size;
 557         uint32_t num_tcs_outputs = util_last_bit64(tcs_outputs_written);
 558         uint32_t num_tcs_patch_outputs = util_last_bit64(tcs_per_patch_outputs_written);
 559         uint32_t output_vertex_size = num_tcs_outputs * 16;
 560         uint32_t pervertex_output_patch_size = tcs_num_output_vertices * output_vertex_size;
 561         uint32_t output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
 562
 563         /* Ensure that we only need one wave per SIMD so we don't need to check
 564          * resource usage. Also ensures that the number of tcs in and out
 565          * vertices per threadgroup are at most 256.
 566          */
 567         unsigned num_patches = 64 / MAX2(tcs_num_input_vertices, tcs_num_output_vertices) * 4;
 568         /* Make sure that the data fits in LDS. This assumes the shaders only
 569          * use LDS for the inputs and outputs.
 570          */
 571         unsigned hardware_lds_size = 32768;
 572
 573         /* Looks like STONEY hangs if we use more than 32 KiB LDS in a single
 574          * threadgroup, even though there is more than 32 KiB LDS.
 575          *
 576          * Test: dEQP-VK.tessellation.shader_input_output.barrier
 577          */
 578         if (chip_class >= GFX7 && family != CHIP_STONEY)
 579                 hardware_lds_size = 65536;
 580
 581         num_patches = MIN2(num_patches, hardware_lds_size / (input_patch_size + output_patch_size));
 582         /* Make sure the output data fits in the offchip buffer */
 583         num_patches = MIN2(num_patches, (tess_offchip_block_dw_size * 4) / output_patch_size);
 584         /* Not necessary for correctness, but improves performance. The
 585          * specific value is taken from the proprietary driver.
 586          */
 587         num_patches = MIN2(num_patches, 40);
 588
 589         /* GFX6 bug workaround - limit LS-HS threadgroups to only one wave. */
 590         if (chip_class == GFX6) {
 591                 unsigned one_wave = 64 / MAX2(tcs_num_input_vertices, tcs_num_output_vertices);
 592                 num_patches = MIN2(num_patches, one_wave);
 593         }
 594         return num_patches;
 595 }
 596
 597 void
 598 radv_lower_fs_io(nir_shader *nir);
 599
 600 #endif