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_fence
 400 {
 401    bool signaled;
 402    int fd;
 403 };
 404
 405 void
 406 tu_fence_init(struct tu_fence *fence, bool signaled);
 407 void
 408 tu_fence_finish(struct tu_fence *fence);
 409 void
 410 tu_fence_update_fd(struct tu_fence *fence, int fd);
 411 void
 412 tu_fence_copy(struct tu_fence *fence, const struct tu_fence *src);
 413 void
 414 tu_fence_signal(struct tu_fence *fence);
 415 void
 416 tu_fence_wait_idle(struct tu_fence *fence);
 417
 418 struct tu_queue
 419 {
 420    VK_LOADER_DATA _loader_data;
 421    struct tu_device *device;
 422    uint32_t queue_family_index;
 423    int queue_idx;
 424    VkDeviceQueueCreateFlags flags;
 425
 426    uint32_t msm_queue_id;
 427    struct tu_fence submit_fence;
 428 };
 429
 430 struct tu_device
 431 {
 432    VK_LOADER_DATA _loader_data;
 433
 434    VkAllocationCallbacks alloc;
 435
 436    struct tu_instance *instance;
 437
 438    struct tu_meta_state meta_state;
 439
 440    struct tu_queue *queues[TU_MAX_QUEUE_FAMILIES];
 441    int queue_count[TU_MAX_QUEUE_FAMILIES];
 442
 443    struct tu_physical_device *physical_device;
 444
 445    /* Backup in-memory cache to be used if the app doesn't provide one */
 446    struct tu_pipeline_cache *mem_cache;
 447
 448    struct list_head shader_slabs;
 449    mtx_t shader_slab_mutex;
 450
 451    struct tu_device_extension_table enabled_extensions;
 452 };
 453
 454 struct tu_bo
 455 {
 456    uint32_t gem_handle;
 457    uint64_t size;
 458    uint64_t iova;
 459    void *map;
 460 };
 461
 462 VkResult
 463 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size);
 464 VkResult
 465 tu_bo_init_dmabuf(struct tu_device *dev,
 466                   struct tu_bo *bo,
 467                   uint64_t size,
 468                   int fd);
 469 int
 470 tu_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo);
 471 void
 472 tu_bo_finish(struct tu_device *dev, struct tu_bo *bo);
 473 VkResult
 474 tu_bo_map(struct tu_device *dev, struct tu_bo *bo);
 475
 476 struct tu_cs_entry
 477 {
 478    /* No ownership */
 479    const struct tu_bo *bo;
 480
 481    uint32_t size;
 482    uint64_t offset;
 483 };
 484
 485 enum tu_cs_mode
 486 {
 487
 488    /*
 489     * A command stream in TU_CS_MODE_GROW mode grows automatically whenever it
 490     * is full.  tu_cs_begin must be called before command packet emission and
 491     * tu_cs_end must be called after.
 492     *
 493     * This mode may create multiple entries internally.  The entries must be
 494     * submitted together.
 495     */
 496    TU_CS_MODE_GROW,
 497
 498    /*
 499     * A command stream in TU_CS_MODE_EXTERNAL mode wraps an external,
 500     * fixed-size buffer.  tu_cs_begin and tu_cs_end are optional and have no
 501     * effect on it.
 502     *
 503     * This mode does not create any entry or any BO.
 504     */
 505    TU_CS_MODE_EXTERNAL,
 506
 507    /*
 508     * A command stream in TU_CS_MODE_SUB_STREAM mode does not support direct
 509     * command packet emission.  tu_cs_begin_sub_stream must be called to get a
 510     * sub-stream to emit comamnd packets to.  When done with the sub-stream,
 511     * tu_cs_end_sub_stream must be called.
 512     *
 513     * This mode does not create any entry internally.
 514     */
 515    TU_CS_MODE_SUB_STREAM,
 516 };
 517
 518 struct tu_cs
 519 {
 520    uint32_t *start;
 521    uint32_t *cur;
 522    uint32_t *reserved_end;
 523    uint32_t *end;
 524
 525    enum tu_cs_mode mode;
 526    uint32_t next_bo_size;
 527
 528    struct tu_cs_entry *entries;
 529    uint32_t entry_count;
 530    uint32_t entry_capacity;
 531
 532    struct tu_bo **bos;
 533    uint32_t bo_count;
 534    uint32_t bo_capacity;
 535 };
 536
 537 struct tu_device_memory
 538 {
 539    struct tu_bo bo;
 540    VkDeviceSize size;
 541
 542    /* for dedicated allocations */
 543    struct tu_image *image;
 544    struct tu_buffer *buffer;
 545
 546    uint32_t type_index;
 547    void *map;
 548    void *user_ptr;
 549 };
 550
 551 struct tu_descriptor_range
 552 {
 553    uint64_t va;
 554    uint32_t size;
 555 };
 556
 557 struct tu_descriptor_set
 558 {
 559    const struct tu_descriptor_set_layout *layout;
 560    uint32_t size;
 561
 562    uint64_t va;
 563    uint32_t *mapped_ptr;
 564    struct tu_descriptor_range *dynamic_descriptors;
 565 };
 566
 567 struct tu_push_descriptor_set
 568 {
 569    struct tu_descriptor_set set;
 570    uint32_t capacity;
 571 };
 572
 573 struct tu_descriptor_pool_entry
 574 {
 575    uint32_t offset;
 576    uint32_t size;
 577    struct tu_descriptor_set *set;
 578 };
 579
 580 struct tu_descriptor_pool
 581 {
 582    uint8_t *mapped_ptr;
 583    uint64_t current_offset;
 584    uint64_t size;
 585
 586    uint8_t *host_memory_base;
 587    uint8_t *host_memory_ptr;
 588    uint8_t *host_memory_end;
 589
 590    uint32_t entry_count;
 591    uint32_t max_entry_count;
 592    struct tu_descriptor_pool_entry entries[0];
 593 };
 594
 595 struct tu_descriptor_update_template_entry
 596 {
 597    VkDescriptorType descriptor_type;
 598
 599    /* The number of descriptors to update */
 600    uint32_t descriptor_count;
 601
 602    /* Into mapped_ptr or dynamic_descriptors, in units of the respective array
 603     */
 604    uint32_t dst_offset;
 605
 606    /* In dwords. Not valid/used for dynamic descriptors */
 607    uint32_t dst_stride;
 608
 609    uint32_t buffer_offset;
 610
 611    /* Only valid for combined image samplers and samplers */
 612    uint16_t has_sampler;
 613
 614    /* In bytes */
 615    size_t src_offset;
 616    size_t src_stride;
 617
 618    /* For push descriptors */
 619    const uint32_t *immutable_samplers;
 620 };
 621
 622 struct tu_descriptor_update_template
 623 {
 624    uint32_t entry_count;
 625    VkPipelineBindPoint bind_point;
 626    struct tu_descriptor_update_template_entry entry[0];
 627 };
 628
 629 struct tu_buffer
 630 {
 631    VkDeviceSize size;
 632
 633    VkBufferUsageFlags usage;
 634    VkBufferCreateFlags flags;
 635 };
 636
 637 enum tu_dynamic_state_bits
 638 {
 639    TU_DYNAMIC_VIEWPORT = 1 << 0,
 640    TU_DYNAMIC_SCISSOR = 1 << 1,
 641    TU_DYNAMIC_LINE_WIDTH = 1 << 2,
 642    TU_DYNAMIC_DEPTH_BIAS = 1 << 3,
 643    TU_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
 644    TU_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
 645    TU_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
 646    TU_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
 647    TU_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
 648    TU_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
 649    TU_DYNAMIC_ALL = (1 << 10) - 1,
 650 };
 651
 652 struct tu_vertex_binding
 653 {
 654    struct tu_buffer *buffer;
 655    VkDeviceSize offset;
 656 };
 657
 658 struct tu_viewport_state
 659 {
 660    uint32_t count;
 661    VkViewport viewports[MAX_VIEWPORTS];
 662 };
 663
 664 struct tu_scissor_state
 665 {
 666    uint32_t count;
 667    VkRect2D scissors[MAX_SCISSORS];
 668 };
 669
 670 struct tu_discard_rectangle_state
 671 {
 672    uint32_t count;
 673    VkRect2D rectangles[MAX_DISCARD_RECTANGLES];
 674 };
 675
 676 struct tu_dynamic_state
 677 {
 678    /**
 679     * Bitmask of (1 << VK_DYNAMIC_STATE_*).
 680     * Defines the set of saved dynamic state.
 681     */
 682    uint32_t mask;
 683
 684    struct tu_viewport_state viewport;
 685
 686    struct tu_scissor_state scissor;
 687
 688    float line_width;
 689
 690    struct
 691    {
 692       float bias;
 693       float clamp;
 694       float slope;
 695    } depth_bias;
 696
 697    float blend_constants[4];
 698
 699    struct
 700    {
 701       float min;
 702       float max;
 703    } depth_bounds;
 704
 705    struct
 706    {
 707       uint32_t front;
 708       uint32_t back;
 709    } stencil_compare_mask;
 710
 711    struct
 712    {
 713       uint32_t front;
 714       uint32_t back;
 715    } stencil_write_mask;
 716
 717    struct
 718    {
 719       uint32_t front;
 720       uint32_t back;
 721    } stencil_reference;
 722
 723    struct tu_discard_rectangle_state discard_rectangle;
 724 };
 725
 726 extern const struct tu_dynamic_state default_dynamic_state;
 727
 728 const char *
 729 tu_get_debug_option_name(int id);
 730
 731 const char *
 732 tu_get_perftest_option_name(int id);
 733
 734 /**
 735  * Attachment state when recording a renderpass instance.
 736  *
 737  * The clear value is valid only if there exists a pending clear.
 738  */
 739 struct tu_attachment_state
 740 {
 741    VkImageAspectFlags pending_clear_aspects;
 742    uint32_t cleared_views;
 743    VkClearValue clear_value;
 744    VkImageLayout current_layout;
 745 };
 746
 747 struct tu_descriptor_state
 748 {
 749    struct tu_descriptor_set *sets[MAX_SETS];
 750    uint32_t dirty;
 751    uint32_t valid;
 752    struct tu_push_descriptor_set push_set;
 753    bool push_dirty;
 754    uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
 755 };
 756
 757 struct tu_tile
 758 {
 759    uint8_t pipe;
 760    uint8_t slot;
 761    VkOffset2D begin;
 762    VkOffset2D end;
 763 };
 764
 765 struct tu_tiling_config
 766 {
 767    VkRect2D render_area;
 768    uint32_t buffer_cpp[MAX_RTS + 2];
 769    uint32_t buffer_count;
 770
 771    /* position and size of the first tile */
 772    VkRect2D tile0;
 773    /* number of tiles */
 774    VkExtent2D tile_count;
 775
 776    uint32_t gmem_offsets[MAX_RTS + 2];
 777
 778    /* size of the first VSC pipe */
 779    VkExtent2D pipe0;
 780    /* number of VSC pipes */
 781    VkExtent2D pipe_count;
 782
 783    /* pipe register values */
 784    uint32_t pipe_config[MAX_VSC_PIPES];
 785    uint32_t pipe_sizes[MAX_VSC_PIPES];
 786 };
 787
 788 struct tu_cmd_state
 789 {
 790    /* Vertex descriptors */
 791    uint64_t vb_va;
 792    unsigned vb_size;
 793
 794    struct tu_dynamic_state dynamic;
 795
 796    /* Index buffer */
 797    struct tu_buffer *index_buffer;
 798    uint64_t index_offset;
 799    uint32_t index_type;
 800    uint32_t max_index_count;
 801    uint64_t index_va;
 802
 803    const struct tu_render_pass *pass;
 804    const struct tu_subpass *subpass;
 805    const struct tu_framebuffer *framebuffer;
 806    struct tu_attachment_state *attachments;
 807
 808    struct tu_tiling_config tiling_config;
 809
 810    struct tu_cs_entry tile_load_ib;
 811    struct tu_cs_entry tile_store_ib;
 812 };
 813
 814 struct tu_cmd_pool
 815 {
 816    VkAllocationCallbacks alloc;
 817    struct list_head cmd_buffers;
 818    struct list_head free_cmd_buffers;
 819    uint32_t queue_family_index;
 820 };
 821
 822 struct tu_cmd_buffer_upload
 823 {
 824    uint8_t *map;
 825    unsigned offset;
 826    uint64_t size;
 827    struct list_head list;
 828 };
 829
 830 enum tu_cmd_buffer_status
 831 {
 832    TU_CMD_BUFFER_STATUS_INVALID,
 833    TU_CMD_BUFFER_STATUS_INITIAL,
 834    TU_CMD_BUFFER_STATUS_RECORDING,
 835    TU_CMD_BUFFER_STATUS_EXECUTABLE,
 836    TU_CMD_BUFFER_STATUS_PENDING,
 837 };
 838
 839 struct tu_bo_list
 840 {
 841    uint32_t count;
 842    uint32_t capacity;
 843    struct drm_msm_gem_submit_bo *bo_infos;
 844 };
 845
 846 #define TU_BO_LIST_FAILED (~0)
 847
 848 void
 849 tu_bo_list_init(struct tu_bo_list *list);
 850 void
 851 tu_bo_list_destroy(struct tu_bo_list *list);
 852 void
 853 tu_bo_list_reset(struct tu_bo_list *list);
 854 uint32_t
 855 tu_bo_list_add(struct tu_bo_list *list,
 856                const struct tu_bo *bo,
 857                uint32_t flags);
 858 VkResult
 859 tu_bo_list_merge(struct tu_bo_list *list, const struct tu_bo_list *other);
 860
 861 struct tu_cmd_buffer
 862 {
 863    VK_LOADER_DATA _loader_data;
 864
 865    struct tu_device *device;
 866
 867    struct tu_cmd_pool *pool;
 868    struct list_head pool_link;
 869
 870    VkCommandBufferUsageFlags usage_flags;
 871    VkCommandBufferLevel level;
 872    enum tu_cmd_buffer_status status;
 873
 874    struct tu_cmd_state state;
 875    struct tu_vertex_binding vertex_bindings[MAX_VBS];
 876    uint32_t queue_family_index;
 877
 878    uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
 879    VkShaderStageFlags push_constant_stages;
 880    struct tu_descriptor_set meta_push_descriptors;
 881
 882    struct tu_descriptor_state descriptors[VK_PIPELINE_BIND_POINT_RANGE_SIZE];
 883
 884    struct tu_cmd_buffer_upload upload;
 885
 886    VkResult record_result;
 887
 888    struct tu_bo_list bo_list;
 889    struct tu_cs cs;
 890    struct tu_cs tile_cs;
 891
 892    uint16_t marker_reg;
 893    uint32_t marker_seqno;
 894
 895    struct tu_bo scratch_bo;
 896    uint32_t scratch_seqno;
 897
 898    bool wait_for_idle;
 899 };
 900
 901 bool
 902 tu_get_memory_fd(struct tu_device *device,
 903                  struct tu_device_memory *memory,
 904                  int *pFD);
 905
 906 /*
 907  * Takes x,y,z as exact numbers of invocations, instead of blocks.
 908  *
 909  * Limitations: Can't call normal dispatch functions without binding or
 910  * rebinding
 911  *              the compute pipeline.
 912  */
 913 void
 914 tu_unaligned_dispatch(struct tu_cmd_buffer *cmd_buffer,
 915                       uint32_t x,
 916                       uint32_t y,
 917                       uint32_t z);
 918
 919 struct tu_event
 920 {
 921    uint64_t *map;
 922 };
 923
 924 struct tu_shader_module;
 925
 926 #define TU_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
 927 #define TU_HASH_SHADER_SISCHED (1 << 1)
 928 #define TU_HASH_SHADER_UNSAFE_MATH (1 << 2)
 929 void
 930 tu_hash_shaders(unsigned char *hash,
 931                 const VkPipelineShaderStageCreateInfo **stages,
 932                 const struct tu_pipeline_layout *layout,
 933                 const struct tu_pipeline_key *key,
 934                 uint32_t flags);
 935
 936 static inline gl_shader_stage
 937 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
 938 {
 939    assert(__builtin_popcount(vk_stage) == 1);
 940    return ffs(vk_stage) - 1;
 941 }
 942
 943 static inline VkShaderStageFlagBits
 944 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
 945 {
 946    return (1 << mesa_stage);
 947 }
 948
 949 #define TU_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
 950
 951 #define tu_foreach_stage(stage, stage_bits)                                  \
 952    for (gl_shader_stage stage,                                               \
 953         __tmp = (gl_shader_stage)((stage_bits) &TU_STAGE_MASK);              \
 954         stage = __builtin_ffs(__tmp) - 1, __tmp; __tmp &= ~(1 << (stage)))
 955
 956 struct tu_shader_module
 957 {
 958    struct nir_shader *nir;
 959    unsigned char sha1[20];
 960    uint32_t size;
 961    char data[0];
 962 };
 963
 964 struct tu_pipeline
 965 {
 966    struct tu_device *device;
 967    struct tu_dynamic_state dynamic_state;
 968
 969    struct tu_pipeline_layout *layout;
 970
 971    bool need_indirect_descriptor_sets;
 972    VkShaderStageFlags active_stages;
 973 };
 974
 975 struct tu_userdata_info *
 976 tu_lookup_user_sgpr(struct tu_pipeline *pipeline,
 977                     gl_shader_stage stage,
 978                     int idx);
 979
 980 struct tu_shader_variant *
 981 tu_get_shader(struct tu_pipeline *pipeline, gl_shader_stage stage);
 982
 983 struct tu_graphics_pipeline_create_info
 984 {
 985    bool use_rectlist;
 986    bool db_depth_clear;
 987    bool db_stencil_clear;
 988    bool db_depth_disable_expclear;
 989    bool db_stencil_disable_expclear;
 990    bool db_flush_depth_inplace;
 991    bool db_flush_stencil_inplace;
 992    bool db_resummarize;
 993    uint32_t custom_blend_mode;
 994 };
 995
 996 VkResult
 997 tu_graphics_pipeline_create(
 998    VkDevice device,
 999    VkPipelineCache cache,
1000    const VkGraphicsPipelineCreateInfo *pCreateInfo,
1001    const struct tu_graphics_pipeline_create_info *extra,
1002    const VkAllocationCallbacks *alloc,
1003    VkPipeline *pPipeline);
1004
1005 struct tu_native_format
1006 {
1007    int vtx;      /* VFMTn_xxx or -1 */
1008    int tex;      /* TFMTn_xxx or -1 */
1009    int rb;       /* RBn_xxx or -1 */
1010    int swap;     /* enum a3xx_color_swap */
1011    bool present; /* internal only; always true to external users */
1012 };
1013
1014 const struct tu_native_format *
1015 tu6_get_native_format(VkFormat format);
1016
1017 int
1018 tu_pack_clear_value(const VkClearValue *val,
1019                     VkFormat format,
1020                     uint32_t buf[4]);
1021
1022 struct tu_image_level
1023 {
1024    VkDeviceSize offset;
1025    VkDeviceSize size;
1026    uint32_t pitch;
1027 };
1028
1029 struct tu_image
1030 {
1031    VkImageType type;
1032    /* The original VkFormat provided by the client.  This may not match any
1033     * of the actual surface formats.
1034     */
1035    VkFormat vk_format;
1036    VkImageAspectFlags aspects;
1037    VkImageUsageFlags usage;  /**< Superset of VkImageCreateInfo::usage. */
1038    VkImageTiling tiling;     /** VkImageCreateInfo::tiling */
1039    VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
1040    VkExtent3D extent;
1041    uint32_t level_count;
1042    uint32_t layer_count;
1043
1044    VkDeviceSize size;
1045    uint32_t alignment;
1046
1047    /* memory layout */
1048    VkDeviceSize layer_size;
1049    struct tu_image_level levels[15];
1050    unsigned tile_mode;
1051
1052    unsigned queue_family_mask;
1053    bool exclusive;
1054    bool shareable;
1055
1056    /* For VK_ANDROID_native_buffer, the WSI image owns the memory, */
1057    VkDeviceMemory owned_memory;
1058
1059    /* Set when bound */
1060    const struct tu_bo *bo;
1061    VkDeviceSize bo_offset;
1062 };
1063
1064 unsigned
1065 tu_image_queue_family_mask(const struct tu_image *image,
1066                            uint32_t family,
1067                            uint32_t queue_family);
1068
1069 static inline uint32_t
1070 tu_get_layerCount(const struct tu_image *image,
1071                   const VkImageSubresourceRange *range)
1072 {
1073    return range->layerCount == VK_REMAINING_ARRAY_LAYERS
1074              ? image->layer_count - range->baseArrayLayer
1075              : range->layerCount;
1076 }
1077
1078 static inline uint32_t
1079 tu_get_levelCount(const struct tu_image *image,
1080                   const VkImageSubresourceRange *range)
1081 {
1082    return range->levelCount == VK_REMAINING_MIP_LEVELS
1083              ? image->level_count - range->baseMipLevel
1084              : range->levelCount;
1085 }
1086
1087 struct tu_image_view
1088 {
1089    struct tu_image *image; /**< VkImageViewCreateInfo::image */
1090
1091    VkImageViewType type;
1092    VkImageAspectFlags aspect_mask;
1093    VkFormat vk_format;
1094    uint32_t base_layer;
1095    uint32_t layer_count;
1096    uint32_t base_mip;
1097    uint32_t level_count;
1098    VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
1099
1100    uint32_t descriptor[16];
1101
1102    /* Descriptor for use as a storage image as opposed to a sampled image.
1103     * This has a few differences for cube maps (e.g. type).
1104     */
1105    uint32_t storage_descriptor[16];
1106 };
1107
1108 struct tu_sampler
1109 {
1110 };
1111
1112 struct tu_image_create_info
1113 {
1114    const VkImageCreateInfo *vk_info;
1115    bool scanout;
1116    bool no_metadata_planes;
1117 };
1118
1119 VkResult
1120 tu_image_create(VkDevice _device,
1121                 const struct tu_image_create_info *info,
1122                 const VkAllocationCallbacks *alloc,
1123                 VkImage *pImage);
1124
1125 VkResult
1126 tu_image_from_gralloc(VkDevice device_h,
1127                       const VkImageCreateInfo *base_info,
1128                       const VkNativeBufferANDROID *gralloc_info,
1129                       const VkAllocationCallbacks *alloc,
1130                       VkImage *out_image_h);
1131
1132 void
1133 tu_image_view_init(struct tu_image_view *view,
1134                    struct tu_device *device,
1135                    const VkImageViewCreateInfo *pCreateInfo);
1136
1137 struct tu_buffer_view
1138 {
1139    VkFormat vk_format;
1140    uint64_t range; /**< VkBufferViewCreateInfo::range */
1141    uint32_t state[4];
1142 };
1143 void
1144 tu_buffer_view_init(struct tu_buffer_view *view,
1145                     struct tu_device *device,
1146                     const VkBufferViewCreateInfo *pCreateInfo);
1147
1148 static inline struct VkExtent3D
1149 tu_sanitize_image_extent(const VkImageType imageType,
1150                          const struct VkExtent3D imageExtent)
1151 {
1152    switch (imageType) {
1153    case VK_IMAGE_TYPE_1D:
1154       return (VkExtent3D) { imageExtent.width, 1, 1 };
1155    case VK_IMAGE_TYPE_2D:
1156       return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
1157    case VK_IMAGE_TYPE_3D:
1158       return imageExtent;
1159    default:
1160       unreachable("invalid image type");
1161    }
1162 }
1163
1164 static inline struct VkOffset3D
1165 tu_sanitize_image_offset(const VkImageType imageType,
1166                          const struct VkOffset3D imageOffset)
1167 {
1168    switch (imageType) {
1169    case VK_IMAGE_TYPE_1D:
1170       return (VkOffset3D) { imageOffset.x, 0, 0 };
1171    case VK_IMAGE_TYPE_2D:
1172       return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
1173    case VK_IMAGE_TYPE_3D:
1174       return imageOffset;
1175    default:
1176       unreachable("invalid image type");
1177    }
1178 }
1179
1180 struct tu_attachment_info
1181 {
1182    struct tu_image_view *attachment;
1183 };
1184
1185 struct tu_framebuffer
1186 {
1187    uint32_t width;
1188    uint32_t height;
1189    uint32_t layers;
1190
1191    uint32_t attachment_count;
1192    struct tu_attachment_info attachments[0];
1193 };
1194
1195 struct tu_subpass_barrier
1196 {
1197    VkPipelineStageFlags src_stage_mask;
1198    VkAccessFlags src_access_mask;
1199    VkAccessFlags dst_access_mask;
1200 };
1201
1202 void
1203 tu_subpass_barrier(struct tu_cmd_buffer *cmd_buffer,
1204                    const struct tu_subpass_barrier *barrier);
1205
1206 struct tu_subpass_attachment
1207 {
1208    uint32_t attachment;
1209    VkImageLayout layout;
1210 };
1211
1212 struct tu_subpass
1213 {
1214    uint32_t input_count;
1215    uint32_t color_count;
1216    struct tu_subpass_attachment *input_attachments;
1217    struct tu_subpass_attachment *color_attachments;
1218    struct tu_subpass_attachment *resolve_attachments;
1219    struct tu_subpass_attachment depth_stencil_attachment;
1220
1221    /** Subpass has at least one resolve attachment */
1222    bool has_resolve;
1223
1224    struct tu_subpass_barrier start_barrier;
1225
1226    uint32_t view_mask;
1227    VkSampleCountFlagBits max_sample_count;
1228 };
1229
1230 struct tu_render_pass_attachment
1231 {
1232    VkFormat format;
1233    uint32_t samples;
1234    VkAttachmentLoadOp load_op;
1235    VkAttachmentLoadOp stencil_load_op;
1236    VkImageLayout initial_layout;
1237    VkImageLayout final_layout;
1238    uint32_t view_mask;
1239 };
1240
1241 struct tu_render_pass
1242 {
1243    uint32_t attachment_count;
1244    uint32_t subpass_count;
1245    struct tu_subpass_attachment *subpass_attachments;
1246    struct tu_render_pass_attachment *attachments;
1247    struct tu_subpass_barrier end_barrier;
1248    struct tu_subpass subpasses[0];
1249 };
1250
1251 VkResult
1252 tu_device_init_meta(struct tu_device *device);
1253 void
1254 tu_device_finish_meta(struct tu_device *device);
1255
1256 struct tu_query_pool
1257 {
1258    uint32_t stride;
1259    uint32_t availability_offset;
1260    uint64_t size;
1261    char *ptr;
1262    VkQueryType type;
1263    uint32_t pipeline_stats_mask;
1264 };
1265
1266 struct tu_semaphore
1267 {
1268    uint32_t syncobj;
1269    uint32_t temp_syncobj;
1270 };
1271
1272 void
1273 tu_set_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1274                       VkPipelineBindPoint bind_point,
1275                       struct tu_descriptor_set *set,
1276                       unsigned idx);
1277
1278 void
1279 tu_update_descriptor_sets(struct tu_device *device,
1280                           struct tu_cmd_buffer *cmd_buffer,
1281                           VkDescriptorSet overrideSet,
1282                           uint32_t descriptorWriteCount,
1283                           const VkWriteDescriptorSet *pDescriptorWrites,
1284                           uint32_t descriptorCopyCount,
1285                           const VkCopyDescriptorSet *pDescriptorCopies);
1286
1287 void
1288 tu_update_descriptor_set_with_template(
1289    struct tu_device *device,
1290    struct tu_cmd_buffer *cmd_buffer,
1291    struct tu_descriptor_set *set,
1292    VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1293    const void *pData);
1294
1295 void
1296 tu_meta_push_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1297                             VkPipelineBindPoint pipelineBindPoint,
1298                             VkPipelineLayout _layout,
1299                             uint32_t set,
1300                             uint32_t descriptorWriteCount,
1301                             const VkWriteDescriptorSet *pDescriptorWrites);
1302
1303 int
1304 tu_drm_get_gpu_id(const struct tu_physical_device *dev, uint32_t *id);
1305
1306 int
1307 tu_drm_get_gmem_size(const struct tu_physical_device *dev, uint32_t *size);
1308
1309 int
1310 tu_drm_submitqueue_new(const struct tu_device *dev,
1311                        int priority,
1312                        uint32_t *queue_id);
1313
1314 void
1315 tu_drm_submitqueue_close(const struct tu_device *dev, uint32_t queue_id);
1316
1317 uint32_t
1318 tu_gem_new(const struct tu_device *dev, uint64_t size, uint32_t flags);
1319 uint32_t
1320 tu_gem_import_dmabuf(const struct tu_device *dev,
1321                      int prime_fd,
1322                      uint64_t size);
1323 int
1324 tu_gem_export_dmabuf(const struct tu_device *dev, uint32_t gem_handle);
1325 void
1326 tu_gem_close(const struct tu_device *dev, uint32_t gem_handle);
1327 uint64_t
1328 tu_gem_info_offset(const struct tu_device *dev, uint32_t gem_handle);
1329 uint64_t
1330 tu_gem_info_iova(const struct tu_device *dev, uint32_t gem_handle);
1331
1332 #define TU_DEFINE_HANDLE_CASTS(__tu_type, __VkType)                          \
1333                                                                              \
1334    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1335    {                                                                         \
1336       return (struct __tu_type *) _handle;                                   \
1337    }                                                                         \
1338                                                                              \
1339    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1340    {                                                                         \
1341       return (__VkType) _obj;                                                \
1342    }
1343
1344 #define TU_DEFINE_NONDISP_HANDLE_CASTS(__tu_type, __VkType)                  \
1345                                                                              \
1346    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1347    {                                                                         \
1348       return (struct __tu_type *) (uintptr_t) _handle;                       \
1349    }                                                                         \
1350                                                                              \
1351    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1352    {                                                                         \
1353       return (__VkType)(uintptr_t) _obj;                                     \
1354    }
1355
1356 #define TU_FROM_HANDLE(__tu_type, __name, __handle)                          \
1357    struct __tu_type *__name = __tu_type##_from_handle(__handle)
1358
1359 TU_DEFINE_HANDLE_CASTS(tu_cmd_buffer, VkCommandBuffer)
1360 TU_DEFINE_HANDLE_CASTS(tu_device, VkDevice)
1361 TU_DEFINE_HANDLE_CASTS(tu_instance, VkInstance)
1362 TU_DEFINE_HANDLE_CASTS(tu_physical_device, VkPhysicalDevice)
1363 TU_DEFINE_HANDLE_CASTS(tu_queue, VkQueue)
1364
1365 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_cmd_pool, VkCommandPool)
1366 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer, VkBuffer)
1367 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer_view, VkBufferView)
1368 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_pool, VkDescriptorPool)
1369 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set, VkDescriptorSet)
1370 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set_layout,
1371                                VkDescriptorSetLayout)
1372 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_update_template,
1373                                VkDescriptorUpdateTemplateKHR)
1374 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_device_memory, VkDeviceMemory)
1375 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_fence, VkFence)
1376 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_event, VkEvent)
1377 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_framebuffer, VkFramebuffer)
1378 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image, VkImage)
1379 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image_view, VkImageView);
1380 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_cache, VkPipelineCache)
1381 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline, VkPipeline)
1382 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_layout, VkPipelineLayout)
1383 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_query_pool, VkQueryPool)
1384 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_render_pass, VkRenderPass)
1385 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_sampler, VkSampler)
1386 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_shader_module, VkShaderModule)
1387 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_semaphore, VkSemaphore)
1388
1389 #endif /* TU_PRIVATE_H */