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 export_prim_id:1;
  59         uint32_t export_layer_id:1;
  60         uint32_t export_clip_dists:1;
  61 };
  62
  63 struct radv_vs_variant_key {
  64         struct radv_vs_out_key out;
  65
  66         uint32_t instance_rate_inputs;
  67         uint32_t instance_rate_divisors[MAX_VERTEX_ATTRIBS];
  68         uint8_t vertex_attribute_formats[MAX_VERTEX_ATTRIBS];
  69         uint32_t vertex_attribute_bindings[MAX_VERTEX_ATTRIBS];
  70         uint32_t vertex_attribute_offsets[MAX_VERTEX_ATTRIBS];
  71         uint32_t vertex_attribute_strides[MAX_VERTEX_ATTRIBS];
  72
  73         /* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
  74          * so we may need to fix it up. */
  75         uint64_t alpha_adjust;
  76
  77         /* For some formats the channels have to be shuffled. */
  78         uint32_t post_shuffle;
  79 };
  80
  81 struct radv_tes_variant_key {
  82         struct radv_vs_out_key out;
  83
  84         uint8_t num_patches;
  85         uint8_t tcs_num_outputs;
  86 };
  87
  88 struct radv_tcs_variant_key {
  89         struct radv_vs_variant_key vs_key;
  90         unsigned primitive_mode;
  91         unsigned input_vertices;
  92         unsigned num_inputs;
  93         uint32_t tes_reads_tess_factors:1;
  94 };
  95
  96 struct radv_fs_variant_key {
  97         uint32_t col_format;
  98         uint8_t log2_ps_iter_samples;
  99         uint8_t num_samples;
 100         uint32_t is_int8;
 101         uint32_t is_int10;
 102 };
 103
 104 struct radv_shader_variant_key {
 105         union {
 106                 struct radv_vs_variant_key vs;
 107                 struct radv_fs_variant_key fs;
 108                 struct radv_tes_variant_key tes;
 109                 struct radv_tcs_variant_key tcs;
 110
 111                 /* A common prefix of the vs and tes keys. */
 112                 struct radv_vs_out_key vs_common_out;
 113         };
 114         bool has_multiview_view_index;
 115 };
 116
 117 struct radv_nir_compiler_options {
 118         struct radv_pipeline_layout *layout;
 119         struct radv_shader_variant_key key;
 120         bool unsafe_math;
 121         bool supports_spill;
 122         bool clamp_shadow_reference;
 123         bool robust_buffer_access;
 124         bool dump_shader;
 125         bool dump_preoptir;
 126         bool record_llvm_ir;
 127         bool check_ir;
 128         enum radeon_family family;
 129         enum chip_class chip_class;
 130         uint32_t tess_offchip_block_dw_size;
 131         uint32_t address32_hi;
 132         uint8_t wave_size;
 133 };
 134
 135 enum radv_ud_index {
 136         AC_UD_SCRATCH_RING_OFFSETS = 0,
 137         AC_UD_PUSH_CONSTANTS = 1,
 138         AC_UD_INLINE_PUSH_CONSTANTS = 2,
 139         AC_UD_INDIRECT_DESCRIPTOR_SETS = 3,
 140         AC_UD_VIEW_INDEX = 4,
 141         AC_UD_STREAMOUT_BUFFERS = 5,
 142         AC_UD_SHADER_START = 6,
 143         AC_UD_VS_VERTEX_BUFFERS = AC_UD_SHADER_START,
 144         AC_UD_VS_BASE_VERTEX_START_INSTANCE,
 145         AC_UD_VS_MAX_UD,
 146         AC_UD_PS_MAX_UD,
 147         AC_UD_CS_GRID_SIZE = AC_UD_SHADER_START,
 148         AC_UD_CS_MAX_UD,
 149         AC_UD_GS_MAX_UD,
 150         AC_UD_TCS_MAX_UD,
 151         AC_UD_TES_MAX_UD,
 152         AC_UD_MAX_UD = AC_UD_TCS_MAX_UD,
 153 };
 154
 155 struct radv_stream_output {
 156         uint8_t location;
 157         uint8_t buffer;
 158         uint16_t offset;
 159         uint8_t component_mask;
 160         uint8_t stream;
 161 };
 162
 163 struct radv_streamout_info {
 164         uint16_t num_outputs;
 165         struct radv_stream_output outputs[MAX_SO_OUTPUTS];
 166         uint16_t strides[MAX_SO_BUFFERS];
 167         uint32_t enabled_stream_buffers_mask;
 168 };
 169
 170 struct radv_shader_info {
 171         bool loads_push_constants;
 172         bool loads_dynamic_offsets;
 173         uint8_t min_push_constant_used;
 174         uint8_t max_push_constant_used;
 175         bool has_only_32bit_push_constants;
 176         bool has_indirect_push_constants;
 177         uint8_t num_inline_push_consts;
 178         uint8_t base_inline_push_consts;
 179         uint32_t desc_set_used_mask;
 180         bool needs_multiview_view_index;
 181         bool uses_invocation_id;
 182         bool uses_prim_id;
 183         uint8_t wave_size;
 184         struct {
 185                 uint64_t ls_outputs_written;
 186                 uint8_t input_usage_mask[VERT_ATTRIB_MAX];
 187                 uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
 188                 bool has_vertex_buffers; /* needs vertex buffers and base/start */
 189                 bool needs_draw_id;
 190                 bool needs_instance_id;
 191         } vs;
 192         struct {
 193                 uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
 194                 uint8_t num_stream_output_components[4];
 195                 uint8_t output_streams[VARYING_SLOT_VAR31 + 1];
 196                 uint8_t max_stream;
 197         } gs;
 198         struct {
 199                 uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
 200         } tes;
 201         struct {
 202                 bool force_persample;
 203                 bool needs_sample_positions;
 204                 bool writes_memory;
 205                 bool writes_z;
 206                 bool writes_stencil;
 207                 bool writes_sample_mask;
 208                 bool has_pcoord;
 209                 bool prim_id_input;
 210                 bool layer_input;
 211                 uint8_t num_input_clips_culls;
 212         } ps;
 213         struct {
 214                 bool uses_grid_size;
 215                 bool uses_block_id[3];
 216                 bool uses_thread_id[3];
 217                 bool uses_local_invocation_idx;
 218         } cs;
 219         struct {
 220                 uint64_t outputs_written;
 221                 uint64_t patch_outputs_written;
 222         } tcs;
 223
 224         struct radv_streamout_info so;
 225 };
 226
 227 struct radv_userdata_info {
 228         int8_t sgpr_idx;
 229         uint8_t num_sgprs;
 230 };
 231
 232 struct radv_userdata_locations {
 233         struct radv_userdata_info descriptor_sets[MAX_SETS];
 234         struct radv_userdata_info shader_data[AC_UD_MAX_UD];
 235         uint32_t descriptor_sets_enabled;
 236 };
 237
 238 struct radv_vs_output_info {
 239         uint8_t vs_output_param_offset[VARYING_SLOT_MAX];
 240         uint8_t clip_dist_mask;
 241         uint8_t cull_dist_mask;
 242         uint8_t param_exports;
 243         bool writes_pointsize;
 244         bool writes_layer;
 245         bool writes_viewport_index;
 246         bool export_prim_id;
 247         unsigned pos_exports;
 248 };
 249
 250 struct radv_es_output_info {
 251         uint32_t esgs_itemsize;
 252 };
 253
 254 struct radv_shader_variant_info {
 255         struct radv_userdata_locations user_sgprs_locs;
 256         struct radv_shader_info info;
 257         unsigned num_user_sgprs;
 258         unsigned num_input_sgprs;
 259         unsigned num_input_vgprs;
 260         unsigned private_mem_vgprs;
 261         bool need_indirect_descriptor_sets;
 262         bool is_ngg;
 263         struct {
 264                 struct {
 265                         struct radv_vs_output_info outinfo;
 266                         struct radv_es_output_info es_info;
 267                         bool as_es;
 268                         bool as_ls;
 269                         bool export_prim_id;
 270                 } vs;
 271                 struct {
 272                         unsigned num_interp;
 273                         uint32_t input_mask;
 274                         uint32_t flat_shaded_mask;
 275                         uint32_t float16_shaded_mask;
 276                         bool can_discard;
 277                         bool early_fragment_test;
 278                         bool post_depth_coverage;
 279                 } fs;
 280                 struct {
 281                         unsigned block_size[3];
 282                 } cs;
 283                 struct {
 284                         unsigned vertices_in;
 285                         unsigned vertices_out;
 286                         unsigned output_prim;
 287                         unsigned invocations;
 288                         unsigned gsvs_vertex_size;
 289                         unsigned max_gsvs_emit_size;
 290                         unsigned es_type; /* GFX9: VS or TES */
 291                 } gs;
 292                 struct {
 293                         unsigned tcs_vertices_out;
 294                         uint32_t num_patches;
 295                         uint32_t lds_size;
 296                 } tcs;
 297                 struct {
 298                         struct radv_vs_output_info outinfo;
 299                         struct radv_es_output_info es_info;
 300                         bool as_es;
 301                         unsigned primitive_mode;
 302                         enum gl_tess_spacing spacing;
 303                         bool ccw;
 304                         bool point_mode;
 305                         bool export_prim_id;
 306                 } tes;
 307         };
 308 };
 309
 310 enum radv_shader_binary_type {
 311         RADV_BINARY_TYPE_LEGACY,
 312         RADV_BINARY_TYPE_RTLD
 313 };
 314
 315 struct radv_shader_binary {
 316         enum radv_shader_binary_type type;
 317         gl_shader_stage stage;
 318         bool is_gs_copy_shader;
 319
 320         struct radv_shader_variant_info variant_info;
 321
 322         /* Self-referential size so we avoid consistency issues. */
 323         uint32_t total_size;
 324 };
 325
 326 struct radv_shader_binary_legacy {
 327         struct radv_shader_binary base;
 328         struct ac_shader_config config;
 329         unsigned code_size;
 330         unsigned llvm_ir_size;
 331         unsigned disasm_size;
 332
 333         /* data has size of code_size + llvm_ir_size + disasm_size + 2, where
 334          * the +2 is for 0 of the ir strings. */
 335         uint8_t data[0];
 336 };
 337
 338 struct radv_shader_binary_rtld {
 339         struct radv_shader_binary base;
 340         unsigned elf_size;
 341         unsigned llvm_ir_size;
 342         uint8_t data[0];
 343 };
 344
 345 struct radv_shader_variant {
 346         uint32_t ref_count;
 347
 348         struct radeon_winsys_bo *bo;
 349         uint64_t bo_offset;
 350         struct ac_shader_config config;
 351         uint32_t code_size;
 352         struct radv_shader_variant_info info;
 353
 354         /* debug only */
 355         uint32_t *spirv;
 356         uint32_t spirv_size;
 357         struct nir_shader *nir;
 358         char *disasm_string;
 359         char *llvm_ir_string;
 360
 361         struct list_head slab_list;
 362 };
 363
 364 struct radv_shader_slab {
 365         struct list_head slabs;
 366         struct list_head shaders;
 367         struct radeon_winsys_bo *bo;
 368         uint64_t size;
 369         char *ptr;
 370 };
 371
 372 void
 373 radv_optimize_nir(struct nir_shader *shader, bool optimize_conservatively,
 374                   bool allow_copies);
 375 bool
 376 radv_nir_lower_ycbcr_textures(nir_shader *shader,
 377                              const struct radv_pipeline_layout *layout);
 378
 379 nir_shader *
 380 radv_shader_compile_to_nir(struct radv_device *device,
 381                            struct radv_shader_module *module,
 382                            const char *entrypoint_name,
 383                            gl_shader_stage stage,
 384                            const VkSpecializationInfo *spec_info,
 385                            const VkPipelineCreateFlags flags,
 386                            const struct radv_pipeline_layout *layout);
 387
 388 void *
 389 radv_alloc_shader_memory(struct radv_device *device,
 390                           struct radv_shader_variant *shader);
 391
 392 void
 393 radv_destroy_shader_slabs(struct radv_device *device);
 394
 395 struct radv_shader_variant *
 396 radv_shader_variant_create(struct radv_device *device,
 397                            const struct radv_shader_binary *binary);
 398 struct radv_shader_variant *
 399 radv_shader_variant_compile(struct radv_device *device,
 400                             struct radv_shader_module *module,
 401                             struct nir_shader *const *shaders,
 402                             int shader_count,
 403                             struct radv_pipeline_layout *layout,
 404                             const struct radv_shader_variant_key *key,
 405                             struct radv_shader_binary **binary_out);
 406
 407 struct radv_shader_variant *
 408 radv_create_gs_copy_shader(struct radv_device *device, struct nir_shader *nir,
 409                            struct radv_shader_binary **binary_out,
 410                            bool multiview);
 411
 412 void
 413 radv_shader_variant_destroy(struct radv_device *device,
 414                             struct radv_shader_variant *variant);
 415
 416 const char *
 417 radv_get_shader_name(struct radv_shader_variant_info *info,
 418                      gl_shader_stage stage);
 419
 420 void
 421 radv_shader_dump_stats(struct radv_device *device,
 422                        struct radv_shader_variant *variant,
 423                        gl_shader_stage stage,
 424                        FILE *file);
 425
 426 bool
 427 radv_can_dump_shader(struct radv_device *device,
 428                      struct radv_shader_module *module,
 429                      bool is_gs_copy_shader);
 430
 431 bool
 432 radv_can_dump_shader_stats(struct radv_device *device,
 433                            struct radv_shader_module *module);
 434
 435 unsigned
 436 shader_io_get_unique_index(gl_varying_slot slot);
 437
 438 #endif