src/freedreno/vulkan/tu_private.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
  25  * DEALINGS IN THE SOFTWARE.
  26  */
  27
  28 #ifndef TU_PRIVATE_H
  29 #define TU_PRIVATE_H
  30
  31 #include <assert.h>
  32 #include <pthread.h>
  33 #include <stdbool.h>
  34 #include <stdint.h>
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <string.h>
  38 #ifdef HAVE_VALGRIND
  39 #include <memcheck.h>
  40 #include <valgrind.h>
  41 #define VG(x) x
  42 #else
  43 #define VG(x)
  44 #endif
  45
  46 #include "c11/threads.h"
  47 #include "compiler/shader_enums.h"
  48 #include "main/macros.h"
  49 #include "util/list.h"
  50 #include "util/macros.h"
  51 #include "vk_alloc.h"
  52 #include "vk_debug_report.h"
  53
  54 #include "drm/msm_drm.h"
  55 #include "tu_descriptor_set.h"
  56 #include "tu_extensions.h"
  57
  58 /* Pre-declarations needed for WSI entrypoints */
  59 struct wl_surface;
  60 struct wl_display;
  61 typedef struct xcb_connection_t xcb_connection_t;
  62 typedef uint32_t xcb_visualid_t;
  63 typedef uint32_t xcb_window_t;
  64
  65 #include <vulkan/vk_android_native_buffer.h>
  66 #include <vulkan/vk_icd.h>
  67 #include <vulkan/vulkan.h>
  68 #include <vulkan/vulkan_intel.h>
  69
  70 #include "tu_entrypoints.h"
  71
  72 #define MAX_VBS 32
  73 #define MAX_VERTEX_ATTRIBS 32
  74 #define MAX_RTS 8
  75 #define MAX_VIEWPORTS 16
  76 #define MAX_SCISSORS 16
  77 #define MAX_DISCARD_RECTANGLES 4
  78 #define MAX_PUSH_CONSTANTS_SIZE 128
  79 #define MAX_PUSH_DESCRIPTORS 32
  80 #define MAX_DYNAMIC_UNIFORM_BUFFERS 16
  81 #define MAX_DYNAMIC_STORAGE_BUFFERS 8
  82 #define MAX_DYNAMIC_BUFFERS                                                  \
  83    (MAX_DYNAMIC_UNIFORM_BUFFERS + MAX_DYNAMIC_STORAGE_BUFFERS)
  84 #define MAX_SAMPLES_LOG2 4
  85 #define NUM_META_FS_KEYS 13
  86 #define TU_MAX_DRM_DEVICES 8
  87 #define MAX_VIEWS 8
  88
  89 #define NUM_DEPTH_CLEAR_PIPELINES 3
  90
  91 /*
  92  * This is the point we switch from using CP to compute shader
  93  * for certain buffer operations.
  94  */
  95 #define TU_BUFFER_OPS_CS_THRESHOLD 4096
  96
  97 enum tu_mem_heap
  98 {
  99    TU_MEM_HEAP_VRAM,
 100    TU_MEM_HEAP_VRAM_CPU_ACCESS,
 101    TU_MEM_HEAP_GTT,
 102    TU_MEM_HEAP_COUNT
 103 };
 104
 105 enum tu_mem_type
 106 {
 107    TU_MEM_TYPE_VRAM,
 108    TU_MEM_TYPE_GTT_WRITE_COMBINE,
 109    TU_MEM_TYPE_VRAM_CPU_ACCESS,
 110    TU_MEM_TYPE_GTT_CACHED,
 111    TU_MEM_TYPE_COUNT
 112 };
 113
 114 #define tu_printflike(a, b) __attribute__((__format__(__printf__, a, b)))
 115
 116 static inline uint32_t
 117 align_u32(uint32_t v, uint32_t a)
 118 {
 119    assert(a != 0 && a == (a & -a));
 120    return (v + a - 1) & ~(a - 1);
 121 }
 122
 123 static inline uint32_t
 124 align_u32_npot(uint32_t v, uint32_t a)
 125 {
 126    return (v + a - 1) / a * a;
 127 }
 128
 129 static inline uint64_t
 130 align_u64(uint64_t v, uint64_t a)
 131 {
 132    assert(a != 0 && a == (a & -a));
 133    return (v + a - 1) & ~(a - 1);
 134 }
 135
 136 static inline int32_t
 137 align_i32(int32_t v, int32_t a)
 138 {
 139    assert(a != 0 && a == (a & -a));
 140    return (v + a - 1) & ~(a - 1);
 141 }
 142
 143 /** Alignment must be a power of 2. */
 144 static inline bool
 145 tu_is_aligned(uintmax_t n, uintmax_t a)
 146 {
 147    assert(a == (a & -a));
 148    return (n & (a - 1)) == 0;
 149 }
 150
 151 static inline uint32_t
 152 round_up_u32(uint32_t v, uint32_t a)
 153 {
 154    return (v + a - 1) / a;
 155 }
 156
 157 static inline uint64_t
 158 round_up_u64(uint64_t v, uint64_t a)
 159 {
 160    return (v + a - 1) / a;
 161 }
 162
 163 static inline uint32_t
 164 tu_minify(uint32_t n, uint32_t levels)
 165 {
 166    if (unlikely(n == 0))
 167       return 0;
 168    else
 169       return MAX2(n >> levels, 1);
 170 }
 171 static inline float
 172 tu_clamp_f(float f, float min, float max)
 173 {
 174    assert(min < max);
 175
 176    if (f > max)
 177       return max;
 178    else if (f < min)
 179       return min;
 180    else
 181       return f;
 182 }
 183
 184 static inline bool
 185 tu_clear_mask(uint32_t *inout_mask, uint32_t clear_mask)
 186 {
 187    if (*inout_mask & clear_mask) {
 188       *inout_mask &= ~clear_mask;
 189       return true;
 190    } else {
 191       return false;
 192    }
 193 }
 194
 195 #define for_each_bit(b, dword)                                               \
 196    for (uint32_t __dword = (dword);                                          \
 197         (b) = __builtin_ffs(__dword) - 1, __dword; __dword &= ~(1 << (b)))
 198
 199 #define typed_memcpy(dest, src, count)                                       \
 200    ({                                                                        \
 201       STATIC_ASSERT(sizeof(*src) == sizeof(*dest));                          \
 202       memcpy((dest), (src), (count) * sizeof(*(src)));                       \
 203    })
 204
 205 /* Whenever we generate an error, pass it through this function. Useful for
 206  * debugging, where we can break on it. Only call at error site, not when
 207  * propagating errors. Might be useful to plug in a stack trace here.
 208  */
 209
 210 struct tu_instance;
 211
 212 VkResult
 213 __vk_errorf(struct tu_instance *instance,
 214             VkResult error,
 215             const char *file,
 216             int line,
 217             const char *format,
 218             ...);
 219
 220 #define vk_error(instance, error)                                            \
 221    __vk_errorf(instance, error, __FILE__, __LINE__, NULL);
 222 #define vk_errorf(instance, error, format, ...)                              \
 223    __vk_errorf(instance, error, __FILE__, __LINE__, format, ##__VA_ARGS__);
 224
 225 void
 226 __tu_finishme(const char *file, int line, const char *format, ...)
 227    tu_printflike(3, 4);
 228 void
 229 tu_loge(const char *format, ...) tu_printflike(1, 2);
 230 void
 231 tu_loge_v(const char *format, va_list va);
 232 void
 233 tu_logi(const char *format, ...) tu_printflike(1, 2);
 234 void
 235 tu_logi_v(const char *format, va_list va);
 236
 237 /**
 238  * Print a FINISHME message, including its source location.
 239  */
 240 #define tu_finishme(format, ...)                                             \
 241    do {                                                                      \
 242       static bool reported = false;                                          \
 243       if (!reported) {                                                       \
 244          __tu_finishme(__FILE__, __LINE__, format, ##__VA_ARGS__);           \
 245          reported = true;                                                    \
 246       }                                                                      \
 247    } while (0)
 248
 249 /* A non-fatal assert.  Useful for debugging. */
 250 #ifdef DEBUG
 251 #define tu_assert(x)                                                         \
 252    ({                                                                        \
 253       if (unlikely(!(x)))                                                    \
 254          fprintf(stderr, "%s:%d ASSERT: %s\n", __FILE__, __LINE__, #x);      \
 255    })
 256 #else
 257 #define tu_assert(x)
 258 #endif
 259
 260 /* Suppress -Wunused in stub functions */
 261 #define tu_use_args(...) __tu_use_args(0, ##__VA_ARGS__)
 262 static inline void
 263 __tu_use_args(int ignore, ...)
 264 {
 265 }
 266
 267 #define tu_stub()                                                            \
 268    do {                                                                      \
 269       tu_finishme("stub %s", __func__);                                      \
 270    } while (0)
 271
 272 void *
 273 tu_lookup_entrypoint_unchecked(const char *name);
 274 void *
 275 tu_lookup_entrypoint_checked(
 276    const char *name,
 277    uint32_t core_version,
 278    const struct tu_instance_extension_table *instance,
 279    const struct tu_device_extension_table *device);
 280
 281 struct tu_physical_device
 282 {
 283    VK_LOADER_DATA _loader_data;
 284
 285    struct tu_instance *instance;
 286
 287    char path[20];
 288    char name[VK_MAX_PHYSICAL_DEVICE_NAME_SIZE];
 289    uint8_t driver_uuid[VK_UUID_SIZE];
 290    uint8_t device_uuid[VK_UUID_SIZE];
 291    uint8_t cache_uuid[VK_UUID_SIZE];
 292
 293    int local_fd;
 294    int master_fd;
 295
 296    unsigned gpu_id;
 297    uint32_t gmem_size;
 298
 299    /* This is the drivers on-disk cache used as a fallback as opposed to
 300     * the pipeline cache defined by apps.
 301     */
 302    struct disk_cache *disk_cache;
 303
 304    struct tu_device_extension_table supported_extensions;
 305 };
 306
 307 enum tu_debug_flags
 308 {
 309    TU_DEBUG_STARTUP = 1 << 0,
 310 };
 311
 312 struct tu_instance
 313 {
 314    VK_LOADER_DATA _loader_data;
 315
 316    VkAllocationCallbacks alloc;
 317
 318    uint32_t api_version;
 319    int physical_device_count;
 320    struct tu_physical_device physical_devices[TU_MAX_DRM_DEVICES];
 321
 322    enum tu_debug_flags debug_flags;
 323
 324    struct vk_debug_report_instance debug_report_callbacks;
 325
 326    struct tu_instance_extension_table enabled_extensions;
 327 };
 328
 329 bool
 330 tu_instance_extension_supported(const char *name);
 331 uint32_t
 332 tu_physical_device_api_version(struct tu_physical_device *dev);
 333 bool
 334 tu_physical_device_extension_supported(struct tu_physical_device *dev,
 335                                        const char *name);
 336
 337 struct cache_entry;
 338
 339 struct tu_pipeline_cache
 340 {
 341    struct tu_device *device;
 342    pthread_mutex_t mutex;
 343
 344    uint32_t total_size;
 345    uint32_t table_size;
 346    uint32_t kernel_count;
 347    struct cache_entry **hash_table;
 348    bool modified;
 349
 350    VkAllocationCallbacks alloc;
 351 };
 352
 353 struct tu_pipeline_key
 354 {
 355 };
 356
 357 void
 358 tu_pipeline_cache_init(struct tu_pipeline_cache *cache,
 359                        struct tu_device *device);
 360 void
 361 tu_pipeline_cache_finish(struct tu_pipeline_cache *cache);
 362 void
 363 tu_pipeline_cache_load(struct tu_pipeline_cache *cache,
 364                        const void *data,
 365                        size_t size);
 366
 367 struct tu_shader_variant;
 368
 369 bool
 370 tu_create_shader_variants_from_pipeline_cache(
 371    struct tu_device *device,
 372    struct tu_pipeline_cache *cache,
 373    const unsigned char *sha1,
 374    struct tu_shader_variant **variants);
 375
 376 void
 377 tu_pipeline_cache_insert_shaders(struct tu_device *device,
 378                                  struct tu_pipeline_cache *cache,
 379                                  const unsigned char *sha1,
 380                                  struct tu_shader_variant **variants,
 381                                  const void *const *codes,
 382                                  const unsigned *code_sizes);
 383
 384 struct tu_meta_state
 385 {
 386    VkAllocationCallbacks alloc;
 387
 388    struct tu_pipeline_cache cache;
 389 };
 390
 391 /* queue types */
 392 #define TU_QUEUE_GENERAL 0
 393
 394 #define TU_MAX_QUEUE_FAMILIES 1
 395
 396 struct tu_queue
 397 {
 398    VK_LOADER_DATA _loader_data;
 399    struct tu_device *device;
 400    uint32_t queue_family_index;
 401    int queue_idx;
 402    VkDeviceQueueCreateFlags flags;
 403
 404    uint32_t msm_queue_id;
 405    int submit_fence_fd;
 406 };
 407
 408 struct tu_device
 409 {
 410    VK_LOADER_DATA _loader_data;
 411
 412    VkAllocationCallbacks alloc;
 413
 414    struct tu_instance *instance;
 415
 416    struct tu_meta_state meta_state;
 417
 418    struct tu_queue *queues[TU_MAX_QUEUE_FAMILIES];
 419    int queue_count[TU_MAX_QUEUE_FAMILIES];
 420
 421    struct tu_physical_device *physical_device;
 422
 423    /* Backup in-memory cache to be used if the app doesn't provide one */
 424    struct tu_pipeline_cache *mem_cache;
 425
 426    struct list_head shader_slabs;
 427    mtx_t shader_slab_mutex;
 428
 429    struct tu_device_extension_table enabled_extensions;
 430 };
 431
 432 struct tu_bo
 433 {
 434    uint32_t gem_handle;
 435    uint64_t size;
 436    uint64_t iova;
 437    void *map;
 438 };
 439
 440 VkResult
 441 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size);
 442 void
 443 tu_bo_finish(struct tu_device *dev, struct tu_bo *bo);
 444 VkResult
 445 tu_bo_map(struct tu_device *dev, struct tu_bo *bo);
 446
 447 struct tu_device_memory
 448 {
 449    struct tu_bo bo;
 450    VkDeviceSize size;
 451
 452    /* for dedicated allocations */
 453    struct tu_image *image;
 454    struct tu_buffer *buffer;
 455
 456    uint32_t type_index;
 457    void *map;
 458    void *user_ptr;
 459 };
 460
 461 struct tu_descriptor_range
 462 {
 463    uint64_t va;
 464    uint32_t size;
 465 };
 466
 467 struct tu_descriptor_set
 468 {
 469    const struct tu_descriptor_set_layout *layout;
 470    uint32_t size;
 471
 472    uint64_t va;
 473    uint32_t *mapped_ptr;
 474    struct tu_descriptor_range *dynamic_descriptors;
 475 };
 476
 477 struct tu_push_descriptor_set
 478 {
 479    struct tu_descriptor_set set;
 480    uint32_t capacity;
 481 };
 482
 483 struct tu_descriptor_pool_entry
 484 {
 485    uint32_t offset;
 486    uint32_t size;
 487    struct tu_descriptor_set *set;
 488 };
 489
 490 struct tu_descriptor_pool
 491 {
 492    uint8_t *mapped_ptr;
 493    uint64_t current_offset;
 494    uint64_t size;
 495
 496    uint8_t *host_memory_base;
 497    uint8_t *host_memory_ptr;
 498    uint8_t *host_memory_end;
 499
 500    uint32_t entry_count;
 501    uint32_t max_entry_count;
 502    struct tu_descriptor_pool_entry entries[0];
 503 };
 504
 505 struct tu_descriptor_update_template_entry
 506 {
 507    VkDescriptorType descriptor_type;
 508
 509    /* The number of descriptors to update */
 510    uint32_t descriptor_count;
 511
 512    /* Into mapped_ptr or dynamic_descriptors, in units of the respective array
 513     */
 514    uint32_t dst_offset;
 515
 516    /* In dwords. Not valid/used for dynamic descriptors */
 517    uint32_t dst_stride;
 518
 519    uint32_t buffer_offset;
 520
 521    /* Only valid for combined image samplers and samplers */
 522    uint16_t has_sampler;
 523
 524    /* In bytes */
 525    size_t src_offset;
 526    size_t src_stride;
 527
 528    /* For push descriptors */
 529    const uint32_t *immutable_samplers;
 530 };
 531
 532 struct tu_descriptor_update_template
 533 {
 534    uint32_t entry_count;
 535    VkPipelineBindPoint bind_point;
 536    struct tu_descriptor_update_template_entry entry[0];
 537 };
 538
 539 struct tu_buffer
 540 {
 541    VkDeviceSize size;
 542
 543    VkBufferUsageFlags usage;
 544    VkBufferCreateFlags flags;
 545 };
 546
 547 enum tu_dynamic_state_bits
 548 {
 549    TU_DYNAMIC_VIEWPORT = 1 << 0,
 550    TU_DYNAMIC_SCISSOR = 1 << 1,
 551    TU_DYNAMIC_LINE_WIDTH = 1 << 2,
 552    TU_DYNAMIC_DEPTH_BIAS = 1 << 3,
 553    TU_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
 554    TU_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
 555    TU_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
 556    TU_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
 557    TU_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
 558    TU_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
 559    TU_DYNAMIC_ALL = (1 << 10) - 1,
 560 };
 561
 562 struct tu_vertex_binding
 563 {
 564    struct tu_buffer *buffer;
 565    VkDeviceSize offset;
 566 };
 567
 568 struct tu_viewport_state
 569 {
 570    uint32_t count;
 571    VkViewport viewports[MAX_VIEWPORTS];
 572 };
 573
 574 struct tu_scissor_state
 575 {
 576    uint32_t count;
 577    VkRect2D scissors[MAX_SCISSORS];
 578 };
 579
 580 struct tu_discard_rectangle_state
 581 {
 582    uint32_t count;
 583    VkRect2D rectangles[MAX_DISCARD_RECTANGLES];
 584 };
 585
 586 struct tu_dynamic_state
 587 {
 588    /**
 589     * Bitmask of (1 << VK_DYNAMIC_STATE_*).
 590     * Defines the set of saved dynamic state.
 591     */
 592    uint32_t mask;
 593
 594    struct tu_viewport_state viewport;
 595
 596    struct tu_scissor_state scissor;
 597
 598    float line_width;
 599
 600    struct
 601    {
 602       float bias;
 603       float clamp;
 604       float slope;
 605    } depth_bias;
 606
 607    float blend_constants[4];
 608
 609    struct
 610    {
 611       float min;
 612       float max;
 613    } depth_bounds;
 614
 615    struct
 616    {
 617       uint32_t front;
 618       uint32_t back;
 619    } stencil_compare_mask;
 620
 621    struct
 622    {
 623       uint32_t front;
 624       uint32_t back;
 625    } stencil_write_mask;
 626
 627    struct
 628    {
 629       uint32_t front;
 630       uint32_t back;
 631    } stencil_reference;
 632
 633    struct tu_discard_rectangle_state discard_rectangle;
 634 };
 635
 636 extern const struct tu_dynamic_state default_dynamic_state;
 637
 638 const char *
 639 tu_get_debug_option_name(int id);
 640
 641 const char *
 642 tu_get_perftest_option_name(int id);
 643
 644 /**
 645  * Attachment state when recording a renderpass instance.
 646  *
 647  * The clear value is valid only if there exists a pending clear.
 648  */
 649 struct tu_attachment_state
 650 {
 651    VkImageAspectFlags pending_clear_aspects;
 652    uint32_t cleared_views;
 653    VkClearValue clear_value;
 654    VkImageLayout current_layout;
 655 };
 656
 657 struct tu_descriptor_state
 658 {
 659    struct tu_descriptor_set *sets[MAX_SETS];
 660    uint32_t dirty;
 661    uint32_t valid;
 662    struct tu_push_descriptor_set push_set;
 663    bool push_dirty;
 664    uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
 665 };
 666
 667 struct tu_cmd_state
 668 {
 669    /* Vertex descriptors */
 670    uint64_t vb_va;
 671    unsigned vb_size;
 672
 673    struct tu_dynamic_state dynamic;
 674
 675    /* Index buffer */
 676    struct tu_buffer *index_buffer;
 677    uint64_t index_offset;
 678    uint32_t index_type;
 679    uint32_t max_index_count;
 680    uint64_t index_va;
 681
 682    const struct tu_render_pass *pass;
 683    const struct tu_subpass *subpass;
 684    const struct tu_framebuffer *framebuffer;
 685    struct tu_attachment_state *attachments;
 686 };
 687
 688 struct tu_cmd_pool
 689 {
 690    VkAllocationCallbacks alloc;
 691    struct list_head cmd_buffers;
 692    struct list_head free_cmd_buffers;
 693    uint32_t queue_family_index;
 694 };
 695
 696 struct tu_cmd_buffer_upload
 697 {
 698    uint8_t *map;
 699    unsigned offset;
 700    uint64_t size;
 701    struct list_head list;
 702 };
 703
 704 enum tu_cmd_buffer_status
 705 {
 706    TU_CMD_BUFFER_STATUS_INVALID,
 707    TU_CMD_BUFFER_STATUS_INITIAL,
 708    TU_CMD_BUFFER_STATUS_RECORDING,
 709    TU_CMD_BUFFER_STATUS_EXECUTABLE,
 710    TU_CMD_BUFFER_STATUS_PENDING,
 711 };
 712
 713 struct tu_bo_list
 714 {
 715    uint32_t count;
 716    uint32_t capacity;
 717    struct drm_msm_gem_submit_bo *bo_infos;
 718 };
 719
 720 #define TU_BO_LIST_FAILED (~0)
 721
 722 void
 723 tu_bo_list_init(struct tu_bo_list *list);
 724 void
 725 tu_bo_list_destroy(struct tu_bo_list *list);
 726 void
 727 tu_bo_list_reset(struct tu_bo_list *list);
 728 uint32_t
 729 tu_bo_list_add(struct tu_bo_list *list,
 730                const struct tu_bo *bo,
 731                uint32_t flags);
 732 VkResult
 733 tu_bo_list_merge(struct tu_bo_list *list, const struct tu_bo_list *other);
 734
 735 struct tu_cs_entry
 736 {
 737    /* No ownership */
 738    const struct tu_bo *bo;
 739
 740    uint32_t size;
 741    uint64_t offset;
 742 };
 743
 744 struct tu_cs
 745 {
 746    uint32_t *start;
 747    uint32_t *cur;
 748    uint32_t *end;
 749
 750    /* for tu_cs_reserve_space_assert */
 751    uint32_t *reserved_end;
 752
 753    struct tu_cs_entry *entries;
 754    uint32_t entry_count;
 755    uint32_t entry_capacity;
 756
 757    struct tu_bo **bos;
 758    uint32_t bo_count;
 759    uint32_t bo_capacity;
 760 };
 761
 762 struct tu_cmd_buffer
 763 {
 764    VK_LOADER_DATA _loader_data;
 765
 766    struct tu_device *device;
 767
 768    struct tu_cmd_pool *pool;
 769    struct list_head pool_link;
 770
 771    VkCommandBufferUsageFlags usage_flags;
 772    VkCommandBufferLevel level;
 773    enum tu_cmd_buffer_status status;
 774
 775    struct tu_cmd_state state;
 776    struct tu_vertex_binding vertex_bindings[MAX_VBS];
 777    uint32_t queue_family_index;
 778
 779    uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
 780    VkShaderStageFlags push_constant_stages;
 781    struct tu_descriptor_set meta_push_descriptors;
 782
 783    struct tu_descriptor_state descriptors[VK_PIPELINE_BIND_POINT_RANGE_SIZE];
 784
 785    struct tu_cmd_buffer_upload upload;
 786
 787    VkResult record_result;
 788
 789    struct tu_bo_list bo_list;
 790    struct tu_cs cs;
 791
 792    uint16_t marker_reg;
 793    uint32_t marker_seqno;
 794
 795    struct tu_bo scratch_bo;
 796    uint32_t scratch_seqno;
 797
 798    /* current cs; command packets are always emitted to it */
 799    struct tu_cs *cur_cs;
 800 };
 801
 802 bool
 803 tu_get_memory_fd(struct tu_device *device,
 804                  struct tu_device_memory *memory,
 805                  int *pFD);
 806
 807 /*
 808  * Takes x,y,z as exact numbers of invocations, instead of blocks.
 809  *
 810  * Limitations: Can't call normal dispatch functions without binding or
 811  * rebinding
 812  *              the compute pipeline.
 813  */
 814 void
 815 tu_unaligned_dispatch(struct tu_cmd_buffer *cmd_buffer,
 816                       uint32_t x,
 817                       uint32_t y,
 818                       uint32_t z);
 819
 820 struct tu_event
 821 {
 822    uint64_t *map;
 823 };
 824
 825 struct tu_shader_module;
 826
 827 #define TU_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
 828 #define TU_HASH_SHADER_SISCHED (1 << 1)
 829 #define TU_HASH_SHADER_UNSAFE_MATH (1 << 2)
 830 void
 831 tu_hash_shaders(unsigned char *hash,
 832                 const VkPipelineShaderStageCreateInfo **stages,
 833                 const struct tu_pipeline_layout *layout,
 834                 const struct tu_pipeline_key *key,
 835                 uint32_t flags);
 836
 837 static inline gl_shader_stage
 838 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
 839 {
 840    assert(__builtin_popcount(vk_stage) == 1);
 841    return ffs(vk_stage) - 1;
 842 }
 843
 844 static inline VkShaderStageFlagBits
 845 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
 846 {
 847    return (1 << mesa_stage);
 848 }
 849
 850 #define TU_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
 851
 852 #define tu_foreach_stage(stage, stage_bits)                                  \
 853    for (gl_shader_stage stage,                                               \
 854         __tmp = (gl_shader_stage)((stage_bits) &TU_STAGE_MASK);              \
 855         stage = __builtin_ffs(__tmp) - 1, __tmp; __tmp &= ~(1 << (stage)))
 856
 857 struct tu_shader_module
 858 {
 859    struct nir_shader *nir;
 860    unsigned char sha1[20];
 861    uint32_t size;
 862    char data[0];
 863 };
 864
 865 struct tu_pipeline
 866 {
 867    struct tu_device *device;
 868    struct tu_dynamic_state dynamic_state;
 869
 870    struct tu_pipeline_layout *layout;
 871
 872    bool need_indirect_descriptor_sets;
 873    VkShaderStageFlags active_stages;
 874 };
 875
 876 struct tu_userdata_info *
 877 tu_lookup_user_sgpr(struct tu_pipeline *pipeline,
 878                     gl_shader_stage stage,
 879                     int idx);
 880
 881 struct tu_shader_variant *
 882 tu_get_shader(struct tu_pipeline *pipeline, gl_shader_stage stage);
 883
 884 struct tu_graphics_pipeline_create_info
 885 {
 886    bool use_rectlist;
 887    bool db_depth_clear;
 888    bool db_stencil_clear;
 889    bool db_depth_disable_expclear;
 890    bool db_stencil_disable_expclear;
 891    bool db_flush_depth_inplace;
 892    bool db_flush_stencil_inplace;
 893    bool db_resummarize;
 894    uint32_t custom_blend_mode;
 895 };
 896
 897 VkResult
 898 tu_graphics_pipeline_create(
 899    VkDevice device,
 900    VkPipelineCache cache,
 901    const VkGraphicsPipelineCreateInfo *pCreateInfo,
 902    const struct tu_graphics_pipeline_create_info *extra,
 903    const VkAllocationCallbacks *alloc,
 904    VkPipeline *pPipeline);
 905
 906 struct vk_format_description;
 907 uint32_t
 908 tu_translate_buffer_dataformat(const struct vk_format_description *desc,
 909                                int first_non_void);
 910 uint32_t
 911 tu_translate_buffer_numformat(const struct vk_format_description *desc,
 912                               int first_non_void);
 913 uint32_t
 914 tu_translate_colorformat(VkFormat format);
 915 uint32_t
 916 tu_translate_color_numformat(VkFormat format,
 917                              const struct vk_format_description *desc,
 918                              int first_non_void);
 919 uint32_t
 920 tu_colorformat_endian_swap(uint32_t colorformat);
 921 unsigned
 922 tu_translate_colorswap(VkFormat format, bool do_endian_swap);
 923 uint32_t
 924 tu_translate_dbformat(VkFormat format);
 925 uint32_t
 926 tu_translate_tex_dataformat(VkFormat format,
 927                             const struct vk_format_description *desc,
 928                             int first_non_void);
 929 uint32_t
 930 tu_translate_tex_numformat(VkFormat format,
 931                            const struct vk_format_description *desc,
 932                            int first_non_void);
 933 bool
 934 tu_format_pack_clear_color(VkFormat format,
 935                            uint32_t clear_vals[2],
 936                            VkClearColorValue *value);
 937 bool
 938 tu_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
 939 bool
 940 tu_dcc_formats_compatible(VkFormat format1, VkFormat format2);
 941
 942 struct tu_image_level
 943 {
 944    VkDeviceSize offset;
 945    VkDeviceSize size;
 946    uint32_t pitch;
 947 };
 948
 949 struct tu_image
 950 {
 951    VkImageType type;
 952    /* The original VkFormat provided by the client.  This may not match any
 953     * of the actual surface formats.
 954     */
 955    VkFormat vk_format;
 956    VkImageAspectFlags aspects;
 957    VkImageUsageFlags usage;  /**< Superset of VkImageCreateInfo::usage. */
 958    VkImageTiling tiling;     /** VkImageCreateInfo::tiling */
 959    VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
 960    VkExtent3D extent;
 961    uint32_t level_count;
 962    uint32_t layer_count;
 963
 964    VkDeviceSize size;
 965    uint32_t alignment;
 966
 967    /* memory layout */
 968    VkDeviceSize layer_size;
 969    struct tu_image_level levels[15];
 970    unsigned tile_mode;
 971
 972    unsigned queue_family_mask;
 973    bool exclusive;
 974    bool shareable;
 975
 976    /* For VK_ANDROID_native_buffer, the WSI image owns the memory, */
 977    VkDeviceMemory owned_memory;
 978
 979    /* Set when bound */
 980    const struct tu_bo *bo;
 981    VkDeviceSize bo_offset;
 982 };
 983
 984 unsigned
 985 tu_image_queue_family_mask(const struct tu_image *image,
 986                            uint32_t family,
 987                            uint32_t queue_family);
 988
 989 static inline uint32_t
 990 tu_get_layerCount(const struct tu_image *image,
 991                   const VkImageSubresourceRange *range)
 992 {
 993    return range->layerCount == VK_REMAINING_ARRAY_LAYERS
 994              ? image->layer_count - range->baseArrayLayer
 995              : range->layerCount;
 996 }
 997
 998 static inline uint32_t
 999 tu_get_levelCount(const struct tu_image *image,
1000                   const VkImageSubresourceRange *range)
1001 {
1002    return range->levelCount == VK_REMAINING_MIP_LEVELS
1003              ? image->level_count - range->baseMipLevel
1004              : range->levelCount;
1005 }
1006
1007 struct tu_image_view
1008 {
1009    struct tu_image *image; /**< VkImageViewCreateInfo::image */
1010
1011    VkImageViewType type;
1012    VkImageAspectFlags aspect_mask;
1013    VkFormat vk_format;
1014    uint32_t base_layer;
1015    uint32_t layer_count;
1016    uint32_t base_mip;
1017    uint32_t level_count;
1018    VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
1019
1020    uint32_t descriptor[16];
1021
1022    /* Descriptor for use as a storage image as opposed to a sampled image.
1023     * This has a few differences for cube maps (e.g. type).
1024     */
1025    uint32_t storage_descriptor[16];
1026 };
1027
1028 struct tu_sampler
1029 {
1030 };
1031
1032 struct tu_image_create_info
1033 {
1034    const VkImageCreateInfo *vk_info;
1035    bool scanout;
1036    bool no_metadata_planes;
1037 };
1038
1039 VkResult
1040 tu_image_create(VkDevice _device,
1041                 const struct tu_image_create_info *info,
1042                 const VkAllocationCallbacks *alloc,
1043                 VkImage *pImage);
1044
1045 VkResult
1046 tu_image_from_gralloc(VkDevice device_h,
1047                       const VkImageCreateInfo *base_info,
1048                       const VkNativeBufferANDROID *gralloc_info,
1049                       const VkAllocationCallbacks *alloc,
1050                       VkImage *out_image_h);
1051
1052 void
1053 tu_image_view_init(struct tu_image_view *view,
1054                    struct tu_device *device,
1055                    const VkImageViewCreateInfo *pCreateInfo);
1056
1057 struct tu_buffer_view
1058 {
1059    VkFormat vk_format;
1060    uint64_t range; /**< VkBufferViewCreateInfo::range */
1061    uint32_t state[4];
1062 };
1063 void
1064 tu_buffer_view_init(struct tu_buffer_view *view,
1065                     struct tu_device *device,
1066                     const VkBufferViewCreateInfo *pCreateInfo);
1067
1068 static inline struct VkExtent3D
1069 tu_sanitize_image_extent(const VkImageType imageType,
1070                          const struct VkExtent3D imageExtent)
1071 {
1072    switch (imageType) {
1073    case VK_IMAGE_TYPE_1D:
1074       return (VkExtent3D) { imageExtent.width, 1, 1 };
1075    case VK_IMAGE_TYPE_2D:
1076       return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
1077    case VK_IMAGE_TYPE_3D:
1078       return imageExtent;
1079    default:
1080       unreachable("invalid image type");
1081    }
1082 }
1083
1084 static inline struct VkOffset3D
1085 tu_sanitize_image_offset(const VkImageType imageType,
1086                          const struct VkOffset3D imageOffset)
1087 {
1088    switch (imageType) {
1089    case VK_IMAGE_TYPE_1D:
1090       return (VkOffset3D) { imageOffset.x, 0, 0 };
1091    case VK_IMAGE_TYPE_2D:
1092       return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
1093    case VK_IMAGE_TYPE_3D:
1094       return imageOffset;
1095    default:
1096       unreachable("invalid image type");
1097    }
1098 }
1099
1100 struct tu_attachment_info
1101 {
1102    struct tu_image_view *attachment;
1103 };
1104
1105 struct tu_framebuffer
1106 {
1107    uint32_t width;
1108    uint32_t height;
1109    uint32_t layers;
1110
1111    uint32_t attachment_count;
1112    struct tu_attachment_info attachments[0];
1113 };
1114
1115 struct tu_subpass_barrier
1116 {
1117    VkPipelineStageFlags src_stage_mask;
1118    VkAccessFlags src_access_mask;
1119    VkAccessFlags dst_access_mask;
1120 };
1121
1122 void
1123 tu_subpass_barrier(struct tu_cmd_buffer *cmd_buffer,
1124                    const struct tu_subpass_barrier *barrier);
1125
1126 struct tu_subpass_attachment
1127 {
1128    uint32_t attachment;
1129    VkImageLayout layout;
1130 };
1131
1132 struct tu_subpass
1133 {
1134    uint32_t input_count;
1135    uint32_t color_count;
1136    struct tu_subpass_attachment *input_attachments;
1137    struct tu_subpass_attachment *color_attachments;
1138    struct tu_subpass_attachment *resolve_attachments;
1139    struct tu_subpass_attachment depth_stencil_attachment;
1140
1141    /** Subpass has at least one resolve attachment */
1142    bool has_resolve;
1143
1144    struct tu_subpass_barrier start_barrier;
1145
1146    uint32_t view_mask;
1147    VkSampleCountFlagBits max_sample_count;
1148 };
1149
1150 struct tu_render_pass_attachment
1151 {
1152    VkFormat format;
1153    uint32_t samples;
1154    VkAttachmentLoadOp load_op;
1155    VkAttachmentLoadOp stencil_load_op;
1156    VkImageLayout initial_layout;
1157    VkImageLayout final_layout;
1158    uint32_t view_mask;
1159 };
1160
1161 struct tu_render_pass
1162 {
1163    uint32_t attachment_count;
1164    uint32_t subpass_count;
1165    struct tu_subpass_attachment *subpass_attachments;
1166    struct tu_render_pass_attachment *attachments;
1167    struct tu_subpass_barrier end_barrier;
1168    struct tu_subpass subpasses[0];
1169 };
1170
1171 VkResult
1172 tu_device_init_meta(struct tu_device *device);
1173 void
1174 tu_device_finish_meta(struct tu_device *device);
1175
1176 struct tu_query_pool
1177 {
1178    uint32_t stride;
1179    uint32_t availability_offset;
1180    uint64_t size;
1181    char *ptr;
1182    VkQueryType type;
1183    uint32_t pipeline_stats_mask;
1184 };
1185
1186 struct tu_semaphore
1187 {
1188    uint32_t syncobj;
1189    uint32_t temp_syncobj;
1190 };
1191
1192 void
1193 tu_set_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1194                       VkPipelineBindPoint bind_point,
1195                       struct tu_descriptor_set *set,
1196                       unsigned idx);
1197
1198 void
1199 tu_update_descriptor_sets(struct tu_device *device,
1200                           struct tu_cmd_buffer *cmd_buffer,
1201                           VkDescriptorSet overrideSet,
1202                           uint32_t descriptorWriteCount,
1203                           const VkWriteDescriptorSet *pDescriptorWrites,
1204                           uint32_t descriptorCopyCount,
1205                           const VkCopyDescriptorSet *pDescriptorCopies);
1206
1207 void
1208 tu_update_descriptor_set_with_template(
1209    struct tu_device *device,
1210    struct tu_cmd_buffer *cmd_buffer,
1211    struct tu_descriptor_set *set,
1212    VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1213    const void *pData);
1214
1215 void
1216 tu_meta_push_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1217                             VkPipelineBindPoint pipelineBindPoint,
1218                             VkPipelineLayout _layout,
1219                             uint32_t set,
1220                             uint32_t descriptorWriteCount,
1221                             const VkWriteDescriptorSet *pDescriptorWrites);
1222
1223 struct tu_fence
1224 {
1225    uint32_t syncobj;
1226    uint32_t temp_syncobj;
1227 };
1228
1229 int
1230 tu_drm_get_gpu_id(const struct tu_physical_device *dev, uint32_t *id);
1231
1232 int
1233 tu_drm_get_gmem_size(const struct tu_physical_device *dev, uint32_t *size);
1234
1235 int
1236 tu_drm_submitqueue_new(const struct tu_device *dev,
1237                        int priority,
1238                        uint32_t *queue_id);
1239
1240 void
1241 tu_drm_submitqueue_close(const struct tu_device *dev, uint32_t queue_id);
1242
1243 uint32_t
1244 tu_gem_new(const struct tu_device *dev, uint64_t size, uint32_t flags);
1245 void
1246 tu_gem_close(const struct tu_device *dev, uint32_t gem_handle);
1247 uint64_t
1248 tu_gem_info_offset(const struct tu_device *dev, uint32_t gem_handle);
1249 uint64_t
1250 tu_gem_info_iova(const struct tu_device *dev, uint32_t gem_handle);
1251
1252 #define TU_DEFINE_HANDLE_CASTS(__tu_type, __VkType)                          \
1253                                                                              \
1254    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1255    {                                                                         \
1256       return (struct __tu_type *) _handle;                                   \
1257    }                                                                         \
1258                                                                              \
1259    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1260    {                                                                         \
1261       return (__VkType) _obj;                                                \
1262    }
1263
1264 #define TU_DEFINE_NONDISP_HANDLE_CASTS(__tu_type, __VkType)                  \
1265                                                                              \
1266    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1267    {                                                                         \
1268       return (struct __tu_type *) (uintptr_t) _handle;                       \
1269    }                                                                         \
1270                                                                              \
1271    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1272    {                                                                         \
1273       return (__VkType)(uintptr_t) _obj;                                     \
1274    }
1275
1276 #define TU_FROM_HANDLE(__tu_type, __name, __handle)                          \
1277    struct __tu_type *__name = __tu_type##_from_handle(__handle)
1278
1279 TU_DEFINE_HANDLE_CASTS(tu_cmd_buffer, VkCommandBuffer)
1280 TU_DEFINE_HANDLE_CASTS(tu_device, VkDevice)
1281 TU_DEFINE_HANDLE_CASTS(tu_instance, VkInstance)
1282 TU_DEFINE_HANDLE_CASTS(tu_physical_device, VkPhysicalDevice)
1283 TU_DEFINE_HANDLE_CASTS(tu_queue, VkQueue)
1284
1285 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_cmd_pool, VkCommandPool)
1286 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer, VkBuffer)
1287 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer_view, VkBufferView)
1288 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_pool, VkDescriptorPool)
1289 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set, VkDescriptorSet)
1290 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set_layout,
1291                                VkDescriptorSetLayout)
1292 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_update_template,
1293                                VkDescriptorUpdateTemplateKHR)
1294 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_device_memory, VkDeviceMemory)
1295 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_fence, VkFence)
1296 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_event, VkEvent)
1297 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_framebuffer, VkFramebuffer)
1298 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image, VkImage)
1299 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image_view, VkImageView);
1300 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_cache, VkPipelineCache)
1301 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline, VkPipeline)
1302 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_layout, VkPipelineLayout)
1303 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_query_pool, VkQueryPool)
1304 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_render_pass, VkRenderPass)
1305 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_sampler, VkSampler)
1306 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_shader_module, VkShaderModule)
1307 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_semaphore, VkSemaphore)
1308
1309 #endif /* TU_PRIVATE_H */