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_VSC_PIPES 32
  76 #define MAX_VIEWPORTS 16
  77 #define MAX_SCISSORS 16
  78 #define MAX_DISCARD_RECTANGLES 4
  79 #define MAX_PUSH_CONSTANTS_SIZE 128
  80 #define MAX_PUSH_DESCRIPTORS 32
  81 #define MAX_DYNAMIC_UNIFORM_BUFFERS 16
  82 #define MAX_DYNAMIC_STORAGE_BUFFERS 8
  83 #define MAX_DYNAMIC_BUFFERS                                                  \
  84    (MAX_DYNAMIC_UNIFORM_BUFFERS + MAX_DYNAMIC_STORAGE_BUFFERS)
  85 #define MAX_SAMPLES_LOG2 4
  86 #define NUM_META_FS_KEYS 13
  87 #define TU_MAX_DRM_DEVICES 8
  88 #define MAX_VIEWS 8
  89
  90 #define NUM_DEPTH_CLEAR_PIPELINES 3
  91
  92 /*
  93  * This is the point we switch from using CP to compute shader
  94  * for certain buffer operations.
  95  */
  96 #define TU_BUFFER_OPS_CS_THRESHOLD 4096
  97
  98 enum tu_mem_heap
  99 {
 100    TU_MEM_HEAP_VRAM,
 101    TU_MEM_HEAP_VRAM_CPU_ACCESS,
 102    TU_MEM_HEAP_GTT,
 103    TU_MEM_HEAP_COUNT
 104 };
 105
 106 enum tu_mem_type
 107 {
 108    TU_MEM_TYPE_VRAM,
 109    TU_MEM_TYPE_GTT_WRITE_COMBINE,
 110    TU_MEM_TYPE_VRAM_CPU_ACCESS,
 111    TU_MEM_TYPE_GTT_CACHED,
 112    TU_MEM_TYPE_COUNT
 113 };
 114
 115 #define tu_printflike(a, b) __attribute__((__format__(__printf__, a, b)))
 116
 117 static inline uint32_t
 118 align_u32(uint32_t v, uint32_t a)
 119 {
 120    assert(a != 0 && a == (a & -a));
 121    return (v + a - 1) & ~(a - 1);
 122 }
 123
 124 static inline uint32_t
 125 align_u32_npot(uint32_t v, uint32_t a)
 126 {
 127    return (v + a - 1) / a * a;
 128 }
 129
 130 static inline uint64_t
 131 align_u64(uint64_t v, uint64_t a)
 132 {
 133    assert(a != 0 && a == (a & -a));
 134    return (v + a - 1) & ~(a - 1);
 135 }
 136
 137 static inline int32_t
 138 align_i32(int32_t v, int32_t a)
 139 {
 140    assert(a != 0 && a == (a & -a));
 141    return (v + a - 1) & ~(a - 1);
 142 }
 143
 144 /** Alignment must be a power of 2. */
 145 static inline bool
 146 tu_is_aligned(uintmax_t n, uintmax_t a)
 147 {
 148    assert(a == (a & -a));
 149    return (n & (a - 1)) == 0;
 150 }
 151
 152 static inline uint32_t
 153 round_up_u32(uint32_t v, uint32_t a)
 154 {
 155    return (v + a - 1) / a;
 156 }
 157
 158 static inline uint64_t
 159 round_up_u64(uint64_t v, uint64_t a)
 160 {
 161    return (v + a - 1) / a;
 162 }
 163
 164 static inline uint32_t
 165 tu_minify(uint32_t n, uint32_t levels)
 166 {
 167    if (unlikely(n == 0))
 168       return 0;
 169    else
 170       return MAX2(n >> levels, 1);
 171 }
 172 static inline float
 173 tu_clamp_f(float f, float min, float max)
 174 {
 175    assert(min < max);
 176
 177    if (f > max)
 178       return max;
 179    else if (f < min)
 180       return min;
 181    else
 182       return f;
 183 }
 184
 185 static inline bool
 186 tu_clear_mask(uint32_t *inout_mask, uint32_t clear_mask)
 187 {
 188    if (*inout_mask & clear_mask) {
 189       *inout_mask &= ~clear_mask;
 190       return true;
 191    } else {
 192       return false;
 193    }
 194 }
 195
 196 #define for_each_bit(b, dword)                                               \
 197    for (uint32_t __dword = (dword);                                          \
 198         (b) = __builtin_ffs(__dword) - 1, __dword; __dword &= ~(1 << (b)))
 199
 200 #define typed_memcpy(dest, src, count)                                       \
 201    ({                                                                        \
 202       STATIC_ASSERT(sizeof(*src) == sizeof(*dest));                          \
 203       memcpy((dest), (src), (count) * sizeof(*(src)));                       \
 204    })
 205
 206 /* Whenever we generate an error, pass it through this function. Useful for
 207  * debugging, where we can break on it. Only call at error site, not when
 208  * propagating errors. Might be useful to plug in a stack trace here.
 209  */
 210
 211 struct tu_instance;
 212
 213 VkResult
 214 __vk_errorf(struct tu_instance *instance,
 215             VkResult error,
 216             const char *file,
 217             int line,
 218             const char *format,
 219             ...);
 220
 221 #define vk_error(instance, error)                                            \
 222    __vk_errorf(instance, error, __FILE__, __LINE__, NULL);
 223 #define vk_errorf(instance, error, format, ...)                              \
 224    __vk_errorf(instance, error, __FILE__, __LINE__, format, ##__VA_ARGS__);
 225
 226 void
 227 __tu_finishme(const char *file, int line, const char *format, ...)
 228    tu_printflike(3, 4);
 229 void
 230 tu_loge(const char *format, ...) tu_printflike(1, 2);
 231 void
 232 tu_loge_v(const char *format, va_list va);
 233 void
 234 tu_logi(const char *format, ...) tu_printflike(1, 2);
 235 void
 236 tu_logi_v(const char *format, va_list va);
 237
 238 /**
 239  * Print a FINISHME message, including its source location.
 240  */
 241 #define tu_finishme(format, ...)                                             \
 242    do {                                                                      \
 243       static bool reported = false;                                          \
 244       if (!reported) {                                                       \
 245          __tu_finishme(__FILE__, __LINE__, format, ##__VA_ARGS__);           \
 246          reported = true;                                                    \
 247       }                                                                      \
 248    } while (0)
 249
 250 /* A non-fatal assert.  Useful for debugging. */
 251 #ifdef DEBUG
 252 #define tu_assert(x)                                                         \
 253    ({                                                                        \
 254       if (unlikely(!(x)))                                                    \
 255          fprintf(stderr, "%s:%d ASSERT: %s\n", __FILE__, __LINE__, #x);      \
 256    })
 257 #else
 258 #define tu_assert(x)
 259 #endif
 260
 261 /* Suppress -Wunused in stub functions */
 262 #define tu_use_args(...) __tu_use_args(0, ##__VA_ARGS__)
 263 static inline void
 264 __tu_use_args(int ignore, ...)
 265 {
 266 }
 267
 268 #define tu_stub()                                                            \
 269    do {                                                                      \
 270       tu_finishme("stub %s", __func__);                                      \
 271    } while (0)
 272
 273 void *
 274 tu_lookup_entrypoint_unchecked(const char *name);
 275 void *
 276 tu_lookup_entrypoint_checked(
 277    const char *name,
 278    uint32_t core_version,
 279    const struct tu_instance_extension_table *instance,
 280    const struct tu_device_extension_table *device);
 281
 282 struct tu_physical_device
 283 {
 284    VK_LOADER_DATA _loader_data;
 285
 286    struct tu_instance *instance;
 287
 288    char path[20];
 289    char name[VK_MAX_PHYSICAL_DEVICE_NAME_SIZE];
 290    uint8_t driver_uuid[VK_UUID_SIZE];
 291    uint8_t device_uuid[VK_UUID_SIZE];
 292    uint8_t cache_uuid[VK_UUID_SIZE];
 293
 294    int local_fd;
 295    int master_fd;
 296
 297    unsigned gpu_id;
 298    uint32_t gmem_size;
 299    uint32_t tile_align_w;
 300    uint32_t tile_align_h;
 301
 302    /* This is the drivers on-disk cache used as a fallback as opposed to
 303     * the pipeline cache defined by apps.
 304     */
 305    struct disk_cache *disk_cache;
 306
 307    struct tu_device_extension_table supported_extensions;
 308 };
 309
 310 enum tu_debug_flags
 311 {
 312    TU_DEBUG_STARTUP = 1 << 0,
 313 };
 314
 315 struct tu_instance
 316 {
 317    VK_LOADER_DATA _loader_data;
 318
 319    VkAllocationCallbacks alloc;
 320
 321    uint32_t api_version;
 322    int physical_device_count;
 323    struct tu_physical_device physical_devices[TU_MAX_DRM_DEVICES];
 324
 325    enum tu_debug_flags debug_flags;
 326
 327    struct vk_debug_report_instance debug_report_callbacks;
 328
 329    struct tu_instance_extension_table enabled_extensions;
 330 };
 331
 332 bool
 333 tu_instance_extension_supported(const char *name);
 334 uint32_t
 335 tu_physical_device_api_version(struct tu_physical_device *dev);
 336 bool
 337 tu_physical_device_extension_supported(struct tu_physical_device *dev,
 338                                        const char *name);
 339
 340 struct cache_entry;
 341
 342 struct tu_pipeline_cache
 343 {
 344    struct tu_device *device;
 345    pthread_mutex_t mutex;
 346
 347    uint32_t total_size;
 348    uint32_t table_size;
 349    uint32_t kernel_count;
 350    struct cache_entry **hash_table;
 351    bool modified;
 352
 353    VkAllocationCallbacks alloc;
 354 };
 355
 356 struct tu_pipeline_key
 357 {
 358 };
 359
 360 void
 361 tu_pipeline_cache_init(struct tu_pipeline_cache *cache,
 362                        struct tu_device *device);
 363 void
 364 tu_pipeline_cache_finish(struct tu_pipeline_cache *cache);
 365 void
 366 tu_pipeline_cache_load(struct tu_pipeline_cache *cache,
 367                        const void *data,
 368                        size_t size);
 369
 370 struct tu_shader_variant;
 371
 372 bool
 373 tu_create_shader_variants_from_pipeline_cache(
 374    struct tu_device *device,
 375    struct tu_pipeline_cache *cache,
 376    const unsigned char *sha1,
 377    struct tu_shader_variant **variants);
 378
 379 void
 380 tu_pipeline_cache_insert_shaders(struct tu_device *device,
 381                                  struct tu_pipeline_cache *cache,
 382                                  const unsigned char *sha1,
 383                                  struct tu_shader_variant **variants,
 384                                  const void *const *codes,
 385                                  const unsigned *code_sizes);
 386
 387 struct tu_meta_state
 388 {
 389    VkAllocationCallbacks alloc;
 390
 391    struct tu_pipeline_cache cache;
 392 };
 393
 394 /* queue types */
 395 #define TU_QUEUE_GENERAL 0
 396
 397 #define TU_MAX_QUEUE_FAMILIES 1
 398
 399 struct tu_queue
 400 {
 401    VK_LOADER_DATA _loader_data;
 402    struct tu_device *device;
 403    uint32_t queue_family_index;
 404    int queue_idx;
 405    VkDeviceQueueCreateFlags flags;
 406
 407    uint32_t msm_queue_id;
 408    int submit_fence_fd;
 409 };
 410
 411 struct tu_device
 412 {
 413    VK_LOADER_DATA _loader_data;
 414
 415    VkAllocationCallbacks alloc;
 416
 417    struct tu_instance *instance;
 418
 419    struct tu_meta_state meta_state;
 420
 421    struct tu_queue *queues[TU_MAX_QUEUE_FAMILIES];
 422    int queue_count[TU_MAX_QUEUE_FAMILIES];
 423
 424    struct tu_physical_device *physical_device;
 425
 426    /* Backup in-memory cache to be used if the app doesn't provide one */
 427    struct tu_pipeline_cache *mem_cache;
 428
 429    struct list_head shader_slabs;
 430    mtx_t shader_slab_mutex;
 431
 432    struct tu_device_extension_table enabled_extensions;
 433 };
 434
 435 struct tu_bo
 436 {
 437    uint32_t gem_handle;
 438    uint64_t size;
 439    uint64_t iova;
 440    void *map;
 441 };
 442
 443 VkResult
 444 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size);
 445 void
 446 tu_bo_finish(struct tu_device *dev, struct tu_bo *bo);
 447 VkResult
 448 tu_bo_map(struct tu_device *dev, struct tu_bo *bo);
 449
 450 struct tu_device_memory
 451 {
 452    struct tu_bo bo;
 453    VkDeviceSize size;
 454
 455    /* for dedicated allocations */
 456    struct tu_image *image;
 457    struct tu_buffer *buffer;
 458
 459    uint32_t type_index;
 460    void *map;
 461    void *user_ptr;
 462 };
 463
 464 struct tu_descriptor_range
 465 {
 466    uint64_t va;
 467    uint32_t size;
 468 };
 469
 470 struct tu_descriptor_set
 471 {
 472    const struct tu_descriptor_set_layout *layout;
 473    uint32_t size;
 474
 475    uint64_t va;
 476    uint32_t *mapped_ptr;
 477    struct tu_descriptor_range *dynamic_descriptors;
 478 };
 479
 480 struct tu_push_descriptor_set
 481 {
 482    struct tu_descriptor_set set;
 483    uint32_t capacity;
 484 };
 485
 486 struct tu_descriptor_pool_entry
 487 {
 488    uint32_t offset;
 489    uint32_t size;
 490    struct tu_descriptor_set *set;
 491 };
 492
 493 struct tu_descriptor_pool
 494 {
 495    uint8_t *mapped_ptr;
 496    uint64_t current_offset;
 497    uint64_t size;
 498
 499    uint8_t *host_memory_base;
 500    uint8_t *host_memory_ptr;
 501    uint8_t *host_memory_end;
 502
 503    uint32_t entry_count;
 504    uint32_t max_entry_count;
 505    struct tu_descriptor_pool_entry entries[0];
 506 };
 507
 508 struct tu_descriptor_update_template_entry
 509 {
 510    VkDescriptorType descriptor_type;
 511
 512    /* The number of descriptors to update */
 513    uint32_t descriptor_count;
 514
 515    /* Into mapped_ptr or dynamic_descriptors, in units of the respective array
 516     */
 517    uint32_t dst_offset;
 518
 519    /* In dwords. Not valid/used for dynamic descriptors */
 520    uint32_t dst_stride;
 521
 522    uint32_t buffer_offset;
 523
 524    /* Only valid for combined image samplers and samplers */
 525    uint16_t has_sampler;
 526
 527    /* In bytes */
 528    size_t src_offset;
 529    size_t src_stride;
 530
 531    /* For push descriptors */
 532    const uint32_t *immutable_samplers;
 533 };
 534
 535 struct tu_descriptor_update_template
 536 {
 537    uint32_t entry_count;
 538    VkPipelineBindPoint bind_point;
 539    struct tu_descriptor_update_template_entry entry[0];
 540 };
 541
 542 struct tu_buffer
 543 {
 544    VkDeviceSize size;
 545
 546    VkBufferUsageFlags usage;
 547    VkBufferCreateFlags flags;
 548 };
 549
 550 enum tu_dynamic_state_bits
 551 {
 552    TU_DYNAMIC_VIEWPORT = 1 << 0,
 553    TU_DYNAMIC_SCISSOR = 1 << 1,
 554    TU_DYNAMIC_LINE_WIDTH = 1 << 2,
 555    TU_DYNAMIC_DEPTH_BIAS = 1 << 3,
 556    TU_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
 557    TU_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
 558    TU_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
 559    TU_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
 560    TU_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
 561    TU_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
 562    TU_DYNAMIC_ALL = (1 << 10) - 1,
 563 };
 564
 565 struct tu_vertex_binding
 566 {
 567    struct tu_buffer *buffer;
 568    VkDeviceSize offset;
 569 };
 570
 571 struct tu_viewport_state
 572 {
 573    uint32_t count;
 574    VkViewport viewports[MAX_VIEWPORTS];
 575 };
 576
 577 struct tu_scissor_state
 578 {
 579    uint32_t count;
 580    VkRect2D scissors[MAX_SCISSORS];
 581 };
 582
 583 struct tu_discard_rectangle_state
 584 {
 585    uint32_t count;
 586    VkRect2D rectangles[MAX_DISCARD_RECTANGLES];
 587 };
 588
 589 struct tu_dynamic_state
 590 {
 591    /**
 592     * Bitmask of (1 << VK_DYNAMIC_STATE_*).
 593     * Defines the set of saved dynamic state.
 594     */
 595    uint32_t mask;
 596
 597    struct tu_viewport_state viewport;
 598
 599    struct tu_scissor_state scissor;
 600
 601    float line_width;
 602
 603    struct
 604    {
 605       float bias;
 606       float clamp;
 607       float slope;
 608    } depth_bias;
 609
 610    float blend_constants[4];
 611
 612    struct
 613    {
 614       float min;
 615       float max;
 616    } depth_bounds;
 617
 618    struct
 619    {
 620       uint32_t front;
 621       uint32_t back;
 622    } stencil_compare_mask;
 623
 624    struct
 625    {
 626       uint32_t front;
 627       uint32_t back;
 628    } stencil_write_mask;
 629
 630    struct
 631    {
 632       uint32_t front;
 633       uint32_t back;
 634    } stencil_reference;
 635
 636    struct tu_discard_rectangle_state discard_rectangle;
 637 };
 638
 639 extern const struct tu_dynamic_state default_dynamic_state;
 640
 641 const char *
 642 tu_get_debug_option_name(int id);
 643
 644 const char *
 645 tu_get_perftest_option_name(int id);
 646
 647 /**
 648  * Attachment state when recording a renderpass instance.
 649  *
 650  * The clear value is valid only if there exists a pending clear.
 651  */
 652 struct tu_attachment_state
 653 {
 654    VkImageAspectFlags pending_clear_aspects;
 655    uint32_t cleared_views;
 656    VkClearValue clear_value;
 657    VkImageLayout current_layout;
 658 };
 659
 660 struct tu_descriptor_state
 661 {
 662    struct tu_descriptor_set *sets[MAX_SETS];
 663    uint32_t dirty;
 664    uint32_t valid;
 665    struct tu_push_descriptor_set push_set;
 666    bool push_dirty;
 667    uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
 668 };
 669
 670 struct tu_tile
 671 {
 672    uint8_t pipe;
 673    uint8_t slot;
 674    VkOffset2D begin;
 675    VkOffset2D end;
 676 };
 677
 678 struct tu_tiling_config
 679 {
 680    VkRect2D render_area;
 681    uint32_t buffer_cpp[MAX_RTS + 2];
 682    uint32_t buffer_count;
 683
 684    /* position and size of the first tile */
 685    VkRect2D tile0;
 686    /* number of tiles */
 687    VkExtent2D tile_count;
 688
 689    uint32_t gmem_offsets[MAX_RTS + 2];
 690
 691    /* size of the first VSC pipe */
 692    VkExtent2D pipe0;
 693    /* number of VSC pipes */
 694    VkExtent2D pipe_count;
 695
 696    /* pipe register values */
 697    uint32_t pipe_config[MAX_VSC_PIPES];
 698    uint32_t pipe_sizes[MAX_VSC_PIPES];
 699 };
 700
 701 struct tu_cmd_state
 702 {
 703    /* Vertex descriptors */
 704    uint64_t vb_va;
 705    unsigned vb_size;
 706
 707    struct tu_dynamic_state dynamic;
 708
 709    /* Index buffer */
 710    struct tu_buffer *index_buffer;
 711    uint64_t index_offset;
 712    uint32_t index_type;
 713    uint32_t max_index_count;
 714    uint64_t index_va;
 715
 716    const struct tu_render_pass *pass;
 717    const struct tu_subpass *subpass;
 718    const struct tu_framebuffer *framebuffer;
 719    struct tu_attachment_state *attachments;
 720
 721    struct tu_tiling_config tiling_config;
 722 };
 723
 724 struct tu_cmd_pool
 725 {
 726    VkAllocationCallbacks alloc;
 727    struct list_head cmd_buffers;
 728    struct list_head free_cmd_buffers;
 729    uint32_t queue_family_index;
 730 };
 731
 732 struct tu_cmd_buffer_upload
 733 {
 734    uint8_t *map;
 735    unsigned offset;
 736    uint64_t size;
 737    struct list_head list;
 738 };
 739
 740 enum tu_cmd_buffer_status
 741 {
 742    TU_CMD_BUFFER_STATUS_INVALID,
 743    TU_CMD_BUFFER_STATUS_INITIAL,
 744    TU_CMD_BUFFER_STATUS_RECORDING,
 745    TU_CMD_BUFFER_STATUS_EXECUTABLE,
 746    TU_CMD_BUFFER_STATUS_PENDING,
 747 };
 748
 749 struct tu_bo_list
 750 {
 751    uint32_t count;
 752    uint32_t capacity;
 753    struct drm_msm_gem_submit_bo *bo_infos;
 754 };
 755
 756 #define TU_BO_LIST_FAILED (~0)
 757
 758 void
 759 tu_bo_list_init(struct tu_bo_list *list);
 760 void
 761 tu_bo_list_destroy(struct tu_bo_list *list);
 762 void
 763 tu_bo_list_reset(struct tu_bo_list *list);
 764 uint32_t
 765 tu_bo_list_add(struct tu_bo_list *list,
 766                const struct tu_bo *bo,
 767                uint32_t flags);
 768 VkResult
 769 tu_bo_list_merge(struct tu_bo_list *list, const struct tu_bo_list *other);
 770
 771 struct tu_cs_entry
 772 {
 773    /* No ownership */
 774    const struct tu_bo *bo;
 775
 776    uint32_t size;
 777    uint64_t offset;
 778 };
 779
 780 struct tu_cs
 781 {
 782    uint32_t *start;
 783    uint32_t *cur;
 784    uint32_t *reserved_end;
 785    uint32_t *end;
 786
 787    uint32_t next_bo_size;
 788
 789    struct tu_cs_entry *entries;
 790    uint32_t entry_count;
 791    uint32_t entry_capacity;
 792
 793    struct tu_bo **bos;
 794    uint32_t bo_count;
 795    uint32_t bo_capacity;
 796 };
 797
 798 struct tu_cmd_buffer
 799 {
 800    VK_LOADER_DATA _loader_data;
 801
 802    struct tu_device *device;
 803
 804    struct tu_cmd_pool *pool;
 805    struct list_head pool_link;
 806
 807    VkCommandBufferUsageFlags usage_flags;
 808    VkCommandBufferLevel level;
 809    enum tu_cmd_buffer_status status;
 810
 811    struct tu_cmd_state state;
 812    struct tu_vertex_binding vertex_bindings[MAX_VBS];
 813    uint32_t queue_family_index;
 814
 815    uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
 816    VkShaderStageFlags push_constant_stages;
 817    struct tu_descriptor_set meta_push_descriptors;
 818
 819    struct tu_descriptor_state descriptors[VK_PIPELINE_BIND_POINT_RANGE_SIZE];
 820
 821    struct tu_cmd_buffer_upload upload;
 822
 823    VkResult record_result;
 824
 825    struct tu_bo_list bo_list;
 826    struct tu_cs cs;
 827
 828    uint16_t marker_reg;
 829    uint32_t marker_seqno;
 830
 831    struct tu_bo scratch_bo;
 832    uint32_t scratch_seqno;
 833
 834    /* current cs; command packets are always emitted to it */
 835    struct tu_cs *cur_cs;
 836 };
 837
 838 bool
 839 tu_get_memory_fd(struct tu_device *device,
 840                  struct tu_device_memory *memory,
 841                  int *pFD);
 842
 843 /*
 844  * Takes x,y,z as exact numbers of invocations, instead of blocks.
 845  *
 846  * Limitations: Can't call normal dispatch functions without binding or
 847  * rebinding
 848  *              the compute pipeline.
 849  */
 850 void
 851 tu_unaligned_dispatch(struct tu_cmd_buffer *cmd_buffer,
 852                       uint32_t x,
 853                       uint32_t y,
 854                       uint32_t z);
 855
 856 struct tu_event
 857 {
 858    uint64_t *map;
 859 };
 860
 861 struct tu_shader_module;
 862
 863 #define TU_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
 864 #define TU_HASH_SHADER_SISCHED (1 << 1)
 865 #define TU_HASH_SHADER_UNSAFE_MATH (1 << 2)
 866 void
 867 tu_hash_shaders(unsigned char *hash,
 868                 const VkPipelineShaderStageCreateInfo **stages,
 869                 const struct tu_pipeline_layout *layout,
 870                 const struct tu_pipeline_key *key,
 871                 uint32_t flags);
 872
 873 static inline gl_shader_stage
 874 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
 875 {
 876    assert(__builtin_popcount(vk_stage) == 1);
 877    return ffs(vk_stage) - 1;
 878 }
 879
 880 static inline VkShaderStageFlagBits
 881 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
 882 {
 883    return (1 << mesa_stage);
 884 }
 885
 886 #define TU_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
 887
 888 #define tu_foreach_stage(stage, stage_bits)                                  \
 889    for (gl_shader_stage stage,                                               \
 890         __tmp = (gl_shader_stage)((stage_bits) &TU_STAGE_MASK);              \
 891         stage = __builtin_ffs(__tmp) - 1, __tmp; __tmp &= ~(1 << (stage)))
 892
 893 struct tu_shader_module
 894 {
 895    struct nir_shader *nir;
 896    unsigned char sha1[20];
 897    uint32_t size;
 898    char data[0];
 899 };
 900
 901 struct tu_pipeline
 902 {
 903    struct tu_device *device;
 904    struct tu_dynamic_state dynamic_state;
 905
 906    struct tu_pipeline_layout *layout;
 907
 908    bool need_indirect_descriptor_sets;
 909    VkShaderStageFlags active_stages;
 910 };
 911
 912 struct tu_userdata_info *
 913 tu_lookup_user_sgpr(struct tu_pipeline *pipeline,
 914                     gl_shader_stage stage,
 915                     int idx);
 916
 917 struct tu_shader_variant *
 918 tu_get_shader(struct tu_pipeline *pipeline, gl_shader_stage stage);
 919
 920 struct tu_graphics_pipeline_create_info
 921 {
 922    bool use_rectlist;
 923    bool db_depth_clear;
 924    bool db_stencil_clear;
 925    bool db_depth_disable_expclear;
 926    bool db_stencil_disable_expclear;
 927    bool db_flush_depth_inplace;
 928    bool db_flush_stencil_inplace;
 929    bool db_resummarize;
 930    uint32_t custom_blend_mode;
 931 };
 932
 933 VkResult
 934 tu_graphics_pipeline_create(
 935    VkDevice device,
 936    VkPipelineCache cache,
 937    const VkGraphicsPipelineCreateInfo *pCreateInfo,
 938    const struct tu_graphics_pipeline_create_info *extra,
 939    const VkAllocationCallbacks *alloc,
 940    VkPipeline *pPipeline);
 941
 942 struct vk_format_description;
 943 uint32_t
 944 tu_translate_buffer_dataformat(const struct vk_format_description *desc,
 945                                int first_non_void);
 946 uint32_t
 947 tu_translate_buffer_numformat(const struct vk_format_description *desc,
 948                               int first_non_void);
 949 uint32_t
 950 tu_translate_colorformat(VkFormat format);
 951 uint32_t
 952 tu_translate_color_numformat(VkFormat format,
 953                              const struct vk_format_description *desc,
 954                              int first_non_void);
 955 uint32_t
 956 tu_colorformat_endian_swap(uint32_t colorformat);
 957 unsigned
 958 tu_translate_colorswap(VkFormat format, bool do_endian_swap);
 959 uint32_t
 960 tu_translate_dbformat(VkFormat format);
 961 uint32_t
 962 tu_translate_tex_dataformat(VkFormat format,
 963                             const struct vk_format_description *desc,
 964                             int first_non_void);
 965 uint32_t
 966 tu_translate_tex_numformat(VkFormat format,
 967                            const struct vk_format_description *desc,
 968                            int first_non_void);
 969 bool
 970 tu_format_pack_clear_color(VkFormat format,
 971                            uint32_t clear_vals[2],
 972                            VkClearColorValue *value);
 973 bool
 974 tu_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
 975 bool
 976 tu_dcc_formats_compatible(VkFormat format1, VkFormat format2);
 977
 978 struct tu_image_level
 979 {
 980    VkDeviceSize offset;
 981    VkDeviceSize size;
 982    uint32_t pitch;
 983 };
 984
 985 struct tu_image
 986 {
 987    VkImageType type;
 988    /* The original VkFormat provided by the client.  This may not match any
 989     * of the actual surface formats.
 990     */
 991    VkFormat vk_format;
 992    VkImageAspectFlags aspects;
 993    VkImageUsageFlags usage;  /**< Superset of VkImageCreateInfo::usage. */
 994    VkImageTiling tiling;     /** VkImageCreateInfo::tiling */
 995    VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
 996    VkExtent3D extent;
 997    uint32_t level_count;
 998    uint32_t layer_count;
 999
1000    VkDeviceSize size;
1001    uint32_t alignment;
1002
1003    /* memory layout */
1004    VkDeviceSize layer_size;
1005    struct tu_image_level levels[15];
1006    unsigned tile_mode;
1007
1008    unsigned queue_family_mask;
1009    bool exclusive;
1010    bool shareable;
1011
1012    /* For VK_ANDROID_native_buffer, the WSI image owns the memory, */
1013    VkDeviceMemory owned_memory;
1014
1015    /* Set when bound */
1016    const struct tu_bo *bo;
1017    VkDeviceSize bo_offset;
1018 };
1019
1020 unsigned
1021 tu_image_queue_family_mask(const struct tu_image *image,
1022                            uint32_t family,
1023                            uint32_t queue_family);
1024
1025 static inline uint32_t
1026 tu_get_layerCount(const struct tu_image *image,
1027                   const VkImageSubresourceRange *range)
1028 {
1029    return range->layerCount == VK_REMAINING_ARRAY_LAYERS
1030              ? image->layer_count - range->baseArrayLayer
1031              : range->layerCount;
1032 }
1033
1034 static inline uint32_t
1035 tu_get_levelCount(const struct tu_image *image,
1036                   const VkImageSubresourceRange *range)
1037 {
1038    return range->levelCount == VK_REMAINING_MIP_LEVELS
1039              ? image->level_count - range->baseMipLevel
1040              : range->levelCount;
1041 }
1042
1043 struct tu_image_view
1044 {
1045    struct tu_image *image; /**< VkImageViewCreateInfo::image */
1046
1047    VkImageViewType type;
1048    VkImageAspectFlags aspect_mask;
1049    VkFormat vk_format;
1050    uint32_t base_layer;
1051    uint32_t layer_count;
1052    uint32_t base_mip;
1053    uint32_t level_count;
1054    VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
1055
1056    uint32_t descriptor[16];
1057
1058    /* Descriptor for use as a storage image as opposed to a sampled image.
1059     * This has a few differences for cube maps (e.g. type).
1060     */
1061    uint32_t storage_descriptor[16];
1062 };
1063
1064 struct tu_sampler
1065 {
1066 };
1067
1068 struct tu_image_create_info
1069 {
1070    const VkImageCreateInfo *vk_info;
1071    bool scanout;
1072    bool no_metadata_planes;
1073 };
1074
1075 VkResult
1076 tu_image_create(VkDevice _device,
1077                 const struct tu_image_create_info *info,
1078                 const VkAllocationCallbacks *alloc,
1079                 VkImage *pImage);
1080
1081 VkResult
1082 tu_image_from_gralloc(VkDevice device_h,
1083                       const VkImageCreateInfo *base_info,
1084                       const VkNativeBufferANDROID *gralloc_info,
1085                       const VkAllocationCallbacks *alloc,
1086                       VkImage *out_image_h);
1087
1088 void
1089 tu_image_view_init(struct tu_image_view *view,
1090                    struct tu_device *device,
1091                    const VkImageViewCreateInfo *pCreateInfo);
1092
1093 struct tu_buffer_view
1094 {
1095    VkFormat vk_format;
1096    uint64_t range; /**< VkBufferViewCreateInfo::range */
1097    uint32_t state[4];
1098 };
1099 void
1100 tu_buffer_view_init(struct tu_buffer_view *view,
1101                     struct tu_device *device,
1102                     const VkBufferViewCreateInfo *pCreateInfo);
1103
1104 static inline struct VkExtent3D
1105 tu_sanitize_image_extent(const VkImageType imageType,
1106                          const struct VkExtent3D imageExtent)
1107 {
1108    switch (imageType) {
1109    case VK_IMAGE_TYPE_1D:
1110       return (VkExtent3D) { imageExtent.width, 1, 1 };
1111    case VK_IMAGE_TYPE_2D:
1112       return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
1113    case VK_IMAGE_TYPE_3D:
1114       return imageExtent;
1115    default:
1116       unreachable("invalid image type");
1117    }
1118 }
1119
1120 static inline struct VkOffset3D
1121 tu_sanitize_image_offset(const VkImageType imageType,
1122                          const struct VkOffset3D imageOffset)
1123 {
1124    switch (imageType) {
1125    case VK_IMAGE_TYPE_1D:
1126       return (VkOffset3D) { imageOffset.x, 0, 0 };
1127    case VK_IMAGE_TYPE_2D:
1128       return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
1129    case VK_IMAGE_TYPE_3D:
1130       return imageOffset;
1131    default:
1132       unreachable("invalid image type");
1133    }
1134 }
1135
1136 struct tu_attachment_info
1137 {
1138    struct tu_image_view *attachment;
1139 };
1140
1141 struct tu_framebuffer
1142 {
1143    uint32_t width;
1144    uint32_t height;
1145    uint32_t layers;
1146
1147    uint32_t attachment_count;
1148    struct tu_attachment_info attachments[0];
1149 };
1150
1151 struct tu_subpass_barrier
1152 {
1153    VkPipelineStageFlags src_stage_mask;
1154    VkAccessFlags src_access_mask;
1155    VkAccessFlags dst_access_mask;
1156 };
1157
1158 void
1159 tu_subpass_barrier(struct tu_cmd_buffer *cmd_buffer,
1160                    const struct tu_subpass_barrier *barrier);
1161
1162 struct tu_subpass_attachment
1163 {
1164    uint32_t attachment;
1165    VkImageLayout layout;
1166 };
1167
1168 struct tu_subpass
1169 {
1170    uint32_t input_count;
1171    uint32_t color_count;
1172    struct tu_subpass_attachment *input_attachments;
1173    struct tu_subpass_attachment *color_attachments;
1174    struct tu_subpass_attachment *resolve_attachments;
1175    struct tu_subpass_attachment depth_stencil_attachment;
1176
1177    /** Subpass has at least one resolve attachment */
1178    bool has_resolve;
1179
1180    struct tu_subpass_barrier start_barrier;
1181
1182    uint32_t view_mask;
1183    VkSampleCountFlagBits max_sample_count;
1184 };
1185
1186 struct tu_render_pass_attachment
1187 {
1188    VkFormat format;
1189    uint32_t samples;
1190    VkAttachmentLoadOp load_op;
1191    VkAttachmentLoadOp stencil_load_op;
1192    VkImageLayout initial_layout;
1193    VkImageLayout final_layout;
1194    uint32_t view_mask;
1195 };
1196
1197 struct tu_render_pass
1198 {
1199    uint32_t attachment_count;
1200    uint32_t subpass_count;
1201    struct tu_subpass_attachment *subpass_attachments;
1202    struct tu_render_pass_attachment *attachments;
1203    struct tu_subpass_barrier end_barrier;
1204    struct tu_subpass subpasses[0];
1205 };
1206
1207 VkResult
1208 tu_device_init_meta(struct tu_device *device);
1209 void
1210 tu_device_finish_meta(struct tu_device *device);
1211
1212 struct tu_query_pool
1213 {
1214    uint32_t stride;
1215    uint32_t availability_offset;
1216    uint64_t size;
1217    char *ptr;
1218    VkQueryType type;
1219    uint32_t pipeline_stats_mask;
1220 };
1221
1222 struct tu_semaphore
1223 {
1224    uint32_t syncobj;
1225    uint32_t temp_syncobj;
1226 };
1227
1228 void
1229 tu_set_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1230                       VkPipelineBindPoint bind_point,
1231                       struct tu_descriptor_set *set,
1232                       unsigned idx);
1233
1234 void
1235 tu_update_descriptor_sets(struct tu_device *device,
1236                           struct tu_cmd_buffer *cmd_buffer,
1237                           VkDescriptorSet overrideSet,
1238                           uint32_t descriptorWriteCount,
1239                           const VkWriteDescriptorSet *pDescriptorWrites,
1240                           uint32_t descriptorCopyCount,
1241                           const VkCopyDescriptorSet *pDescriptorCopies);
1242
1243 void
1244 tu_update_descriptor_set_with_template(
1245    struct tu_device *device,
1246    struct tu_cmd_buffer *cmd_buffer,
1247    struct tu_descriptor_set *set,
1248    VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1249    const void *pData);
1250
1251 void
1252 tu_meta_push_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1253                             VkPipelineBindPoint pipelineBindPoint,
1254                             VkPipelineLayout _layout,
1255                             uint32_t set,
1256                             uint32_t descriptorWriteCount,
1257                             const VkWriteDescriptorSet *pDescriptorWrites);
1258
1259 struct tu_fence
1260 {
1261    uint32_t syncobj;
1262    uint32_t temp_syncobj;
1263 };
1264
1265 int
1266 tu_drm_get_gpu_id(const struct tu_physical_device *dev, uint32_t *id);
1267
1268 int
1269 tu_drm_get_gmem_size(const struct tu_physical_device *dev, uint32_t *size);
1270
1271 int
1272 tu_drm_submitqueue_new(const struct tu_device *dev,
1273                        int priority,
1274                        uint32_t *queue_id);
1275
1276 void
1277 tu_drm_submitqueue_close(const struct tu_device *dev, uint32_t queue_id);
1278
1279 uint32_t
1280 tu_gem_new(const struct tu_device *dev, uint64_t size, uint32_t flags);
1281 void
1282 tu_gem_close(const struct tu_device *dev, uint32_t gem_handle);
1283 uint64_t
1284 tu_gem_info_offset(const struct tu_device *dev, uint32_t gem_handle);
1285 uint64_t
1286 tu_gem_info_iova(const struct tu_device *dev, uint32_t gem_handle);
1287
1288 #define TU_DEFINE_HANDLE_CASTS(__tu_type, __VkType)                          \
1289                                                                              \
1290    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1291    {                                                                         \
1292       return (struct __tu_type *) _handle;                                   \
1293    }                                                                         \
1294                                                                              \
1295    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1296    {                                                                         \
1297       return (__VkType) _obj;                                                \
1298    }
1299
1300 #define TU_DEFINE_NONDISP_HANDLE_CASTS(__tu_type, __VkType)                  \
1301                                                                              \
1302    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1303    {                                                                         \
1304       return (struct __tu_type *) (uintptr_t) _handle;                       \
1305    }                                                                         \
1306                                                                              \
1307    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1308    {                                                                         \
1309       return (__VkType)(uintptr_t) _obj;                                     \
1310    }
1311
1312 #define TU_FROM_HANDLE(__tu_type, __name, __handle)                          \
1313    struct __tu_type *__name = __tu_type##_from_handle(__handle)
1314
1315 TU_DEFINE_HANDLE_CASTS(tu_cmd_buffer, VkCommandBuffer)
1316 TU_DEFINE_HANDLE_CASTS(tu_device, VkDevice)
1317 TU_DEFINE_HANDLE_CASTS(tu_instance, VkInstance)
1318 TU_DEFINE_HANDLE_CASTS(tu_physical_device, VkPhysicalDevice)
1319 TU_DEFINE_HANDLE_CASTS(tu_queue, VkQueue)
1320
1321 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_cmd_pool, VkCommandPool)
1322 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer, VkBuffer)
1323 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer_view, VkBufferView)
1324 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_pool, VkDescriptorPool)
1325 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set, VkDescriptorSet)
1326 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set_layout,
1327                                VkDescriptorSetLayout)
1328 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_update_template,
1329                                VkDescriptorUpdateTemplateKHR)
1330 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_device_memory, VkDeviceMemory)
1331 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_fence, VkFence)
1332 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_event, VkEvent)
1333 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_framebuffer, VkFramebuffer)
1334 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image, VkImage)
1335 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image_view, VkImageView);
1336 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_cache, VkPipelineCache)
1337 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline, VkPipeline)
1338 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_layout, VkPipelineLayout)
1339 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_query_pool, VkQueryPool)
1340 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_render_pass, VkRenderPass)
1341 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_sampler, VkSampler)
1342 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_shader_module, VkShaderModule)
1343 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_semaphore, VkSemaphore)
1344
1345 #endif /* TU_PRIVATE_H */