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 "tu_descriptor_set.h"
  55 #include "tu_extensions.h"
  56
  57 /* Pre-declarations needed for WSI entrypoints */
  58 struct wl_surface;
  59 struct wl_display;
  60 typedef struct xcb_connection_t xcb_connection_t;
  61 typedef uint32_t xcb_visualid_t;
  62 typedef uint32_t xcb_window_t;
  63
  64 #include <vulkan/vk_android_native_buffer.h>
  65 #include <vulkan/vk_icd.h>
  66 #include <vulkan/vulkan.h>
  67 #include <vulkan/vulkan_intel.h>
  68
  69 #include "drm/freedreno_ringbuffer.h"
  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    struct fd_device *drm_device;
 297    unsigned gpu_id;
 298    uint32_t gmem_size;
 299
 300    /* This is the drivers on-disk cache used as a fallback as opposed to
 301     * the pipeline cache defined by apps.
 302     */
 303    struct disk_cache *disk_cache;
 304
 305    struct tu_device_extension_table supported_extensions;
 306 };
 307
 308 enum tu_debug_flags
 309 {
 310    TU_DEBUG_STARTUP = 1 << 0,
 311 };
 312
 313 struct tu_instance
 314 {
 315    VK_LOADER_DATA _loader_data;
 316
 317    VkAllocationCallbacks alloc;
 318
 319    uint32_t api_version;
 320    int physical_device_count;
 321    struct tu_physical_device physical_devices[TU_MAX_DRM_DEVICES];
 322
 323    enum tu_debug_flags debug_flags;
 324
 325    struct vk_debug_report_instance debug_report_callbacks;
 326
 327    struct tu_instance_extension_table enabled_extensions;
 328 };
 329
 330 bool
 331 tu_instance_extension_supported(const char *name);
 332 uint32_t
 333 tu_physical_device_api_version(struct tu_physical_device *dev);
 334 bool
 335 tu_physical_device_extension_supported(struct tu_physical_device *dev,
 336                                        const char *name);
 337
 338 struct cache_entry;
 339
 340 struct tu_pipeline_cache
 341 {
 342    struct tu_device *device;
 343    pthread_mutex_t mutex;
 344
 345    uint32_t total_size;
 346    uint32_t table_size;
 347    uint32_t kernel_count;
 348    struct cache_entry **hash_table;
 349    bool modified;
 350
 351    VkAllocationCallbacks alloc;
 352 };
 353
 354 struct tu_pipeline_key
 355 {
 356 };
 357
 358 void
 359 tu_pipeline_cache_init(struct tu_pipeline_cache *cache,
 360                        struct tu_device *device);
 361 void
 362 tu_pipeline_cache_finish(struct tu_pipeline_cache *cache);
 363 void
 364 tu_pipeline_cache_load(struct tu_pipeline_cache *cache,
 365                        const void *data,
 366                        size_t size);
 367
 368 struct tu_shader_variant;
 369
 370 bool
 371 tu_create_shader_variants_from_pipeline_cache(
 372    struct tu_device *device,
 373    struct tu_pipeline_cache *cache,
 374    const unsigned char *sha1,
 375    struct tu_shader_variant **variants);
 376
 377 void
 378 tu_pipeline_cache_insert_shaders(struct tu_device *device,
 379                                  struct tu_pipeline_cache *cache,
 380                                  const unsigned char *sha1,
 381                                  struct tu_shader_variant **variants,
 382                                  const void *const *codes,
 383                                  const unsigned *code_sizes);
 384
 385 struct tu_meta_state
 386 {
 387    VkAllocationCallbacks alloc;
 388
 389    struct tu_pipeline_cache cache;
 390 };
 391
 392 /* queue types */
 393 #define TU_QUEUE_GENERAL 0
 394
 395 #define TU_MAX_QUEUE_FAMILIES 1
 396
 397 struct tu_queue
 398 {
 399    VK_LOADER_DATA _loader_data;
 400    struct tu_device *device;
 401    uint32_t queue_family_index;
 402    int queue_idx;
 403    VkDeviceQueueCreateFlags flags;
 404
 405    uint32_t msm_queue_id;
 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 offset;
 437    uint64_t iova;
 438    void *map;
 439 };
 440
 441 VkResult
 442 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size);
 443 void
 444 tu_bo_finish(struct tu_device *dev, struct tu_bo *bo);
 445 VkResult
 446 tu_bo_map(struct tu_device *dev, struct tu_bo *bo);
 447
 448 struct tu_device_memory
 449 {
 450    struct tu_bo bo;
 451    VkDeviceSize size;
 452
 453    /* for dedicated allocations */
 454    struct tu_image *image;
 455    struct tu_buffer *buffer;
 456
 457    uint32_t type_index;
 458    void *map;
 459    void *user_ptr;
 460 };
 461
 462 struct tu_descriptor_range
 463 {
 464    uint64_t va;
 465    uint32_t size;
 466 };
 467
 468 struct tu_descriptor_set
 469 {
 470    const struct tu_descriptor_set_layout *layout;
 471    uint32_t size;
 472
 473    uint64_t va;
 474    uint32_t *mapped_ptr;
 475    struct tu_descriptor_range *dynamic_descriptors;
 476 };
 477
 478 struct tu_push_descriptor_set
 479 {
 480    struct tu_descriptor_set set;
 481    uint32_t capacity;
 482 };
 483
 484 struct tu_descriptor_pool_entry
 485 {
 486    uint32_t offset;
 487    uint32_t size;
 488    struct tu_descriptor_set *set;
 489 };
 490
 491 struct tu_descriptor_pool
 492 {
 493    uint8_t *mapped_ptr;
 494    uint64_t current_offset;
 495    uint64_t size;
 496
 497    uint8_t *host_memory_base;
 498    uint8_t *host_memory_ptr;
 499    uint8_t *host_memory_end;
 500
 501    uint32_t entry_count;
 502    uint32_t max_entry_count;
 503    struct tu_descriptor_pool_entry entries[0];
 504 };
 505
 506 struct tu_descriptor_update_template_entry
 507 {
 508    VkDescriptorType descriptor_type;
 509
 510    /* The number of descriptors to update */
 511    uint32_t descriptor_count;
 512
 513    /* Into mapped_ptr or dynamic_descriptors, in units of the respective array
 514     */
 515    uint32_t dst_offset;
 516
 517    /* In dwords. Not valid/used for dynamic descriptors */
 518    uint32_t dst_stride;
 519
 520    uint32_t buffer_offset;
 521
 522    /* Only valid for combined image samplers and samplers */
 523    uint16_t has_sampler;
 524
 525    /* In bytes */
 526    size_t src_offset;
 527    size_t src_stride;
 528
 529    /* For push descriptors */
 530    const uint32_t *immutable_samplers;
 531 };
 532
 533 struct tu_descriptor_update_template
 534 {
 535    uint32_t entry_count;
 536    VkPipelineBindPoint bind_point;
 537    struct tu_descriptor_update_template_entry entry[0];
 538 };
 539
 540 struct tu_buffer
 541 {
 542    VkDeviceSize size;
 543
 544    VkBufferUsageFlags usage;
 545    VkBufferCreateFlags flags;
 546 };
 547
 548 enum tu_dynamic_state_bits
 549 {
 550    TU_DYNAMIC_VIEWPORT = 1 << 0,
 551    TU_DYNAMIC_SCISSOR = 1 << 1,
 552    TU_DYNAMIC_LINE_WIDTH = 1 << 2,
 553    TU_DYNAMIC_DEPTH_BIAS = 1 << 3,
 554    TU_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
 555    TU_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
 556    TU_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
 557    TU_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
 558    TU_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
 559    TU_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
 560    TU_DYNAMIC_ALL = (1 << 10) - 1,
 561 };
 562
 563 struct tu_vertex_binding
 564 {
 565    struct tu_buffer *buffer;
 566    VkDeviceSize offset;
 567 };
 568
 569 struct tu_viewport_state
 570 {
 571    uint32_t count;
 572    VkViewport viewports[MAX_VIEWPORTS];
 573 };
 574
 575 struct tu_scissor_state
 576 {
 577    uint32_t count;
 578    VkRect2D scissors[MAX_SCISSORS];
 579 };
 580
 581 struct tu_discard_rectangle_state
 582 {
 583    uint32_t count;
 584    VkRect2D rectangles[MAX_DISCARD_RECTANGLES];
 585 };
 586
 587 struct tu_dynamic_state
 588 {
 589    /**
 590     * Bitmask of (1 << VK_DYNAMIC_STATE_*).
 591     * Defines the set of saved dynamic state.
 592     */
 593    uint32_t mask;
 594
 595    struct tu_viewport_state viewport;
 596
 597    struct tu_scissor_state scissor;
 598
 599    float line_width;
 600
 601    struct
 602    {
 603       float bias;
 604       float clamp;
 605       float slope;
 606    } depth_bias;
 607
 608    float blend_constants[4];
 609
 610    struct
 611    {
 612       float min;
 613       float max;
 614    } depth_bounds;
 615
 616    struct
 617    {
 618       uint32_t front;
 619       uint32_t back;
 620    } stencil_compare_mask;
 621
 622    struct
 623    {
 624       uint32_t front;
 625       uint32_t back;
 626    } stencil_write_mask;
 627
 628    struct
 629    {
 630       uint32_t front;
 631       uint32_t back;
 632    } stencil_reference;
 633
 634    struct tu_discard_rectangle_state discard_rectangle;
 635 };
 636
 637 extern const struct tu_dynamic_state default_dynamic_state;
 638
 639 const char *
 640 tu_get_debug_option_name(int id);
 641
 642 const char *
 643 tu_get_perftest_option_name(int id);
 644
 645 /**
 646  * Attachment state when recording a renderpass instance.
 647  *
 648  * The clear value is valid only if there exists a pending clear.
 649  */
 650 struct tu_attachment_state
 651 {
 652    VkImageAspectFlags pending_clear_aspects;
 653    uint32_t cleared_views;
 654    VkClearValue clear_value;
 655    VkImageLayout current_layout;
 656 };
 657
 658 struct tu_descriptor_state
 659 {
 660    struct tu_descriptor_set *sets[MAX_SETS];
 661    uint32_t dirty;
 662    uint32_t valid;
 663    struct tu_push_descriptor_set push_set;
 664    bool push_dirty;
 665    uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
 666 };
 667
 668 struct tu_cmd_state
 669 {
 670    /* Vertex descriptors */
 671    uint64_t vb_va;
 672    unsigned vb_size;
 673
 674    struct tu_dynamic_state dynamic;
 675
 676    /* Index buffer */
 677    struct tu_buffer *index_buffer;
 678    uint64_t index_offset;
 679    uint32_t index_type;
 680    uint32_t max_index_count;
 681    uint64_t index_va;
 682 };
 683
 684 struct tu_cmd_pool
 685 {
 686    VkAllocationCallbacks alloc;
 687    struct list_head cmd_buffers;
 688    struct list_head free_cmd_buffers;
 689    uint32_t queue_family_index;
 690 };
 691
 692 struct tu_cmd_buffer_upload
 693 {
 694    uint8_t *map;
 695    unsigned offset;
 696    uint64_t size;
 697    struct list_head list;
 698 };
 699
 700 enum tu_cmd_buffer_status
 701 {
 702    TU_CMD_BUFFER_STATUS_INVALID,
 703    TU_CMD_BUFFER_STATUS_INITIAL,
 704    TU_CMD_BUFFER_STATUS_RECORDING,
 705    TU_CMD_BUFFER_STATUS_EXECUTABLE,
 706    TU_CMD_BUFFER_STATUS_PENDING,
 707 };
 708
 709 struct tu_bo_list
 710 {
 711    uint32_t count;
 712    uint32_t capacity;
 713    uint32_t *handles;
 714 };
 715
 716 void tu_bo_list_init(struct tu_bo_list *list);
 717 void tu_bo_list_destroy(struct tu_bo_list *list);
 718 void tu_bo_list_reset(struct tu_bo_list *list);
 719 uint32_t tu_bo_list_add(struct tu_bo_list *list,
 720                         const struct tu_bo *bo);
 721
 722 struct tu_cmd_stream_entry
 723 {
 724     /* No ownership */
 725    struct tu_bo *bo;
 726
 727    uint32_t size;
 728    uint64_t offset;
 729 };
 730
 731 struct tu_cmd_stream
 732 {
 733    uint32_t *start;
 734    uint32_t *cur;
 735    uint32_t *end;
 736
 737    struct tu_cmd_stream_entry *entries;
 738    uint32_t entry_count;
 739    uint32_t entry_capacity;
 740
 741    struct tu_bo **bos;
 742    uint32_t bo_count;
 743    uint32_t bo_capacity;
 744 };
 745
 746 struct tu_cmd_buffer
 747 {
 748    VK_LOADER_DATA _loader_data;
 749
 750    struct tu_device *device;
 751
 752    struct tu_cmd_pool *pool;
 753    struct list_head pool_link;
 754
 755    VkCommandBufferUsageFlags usage_flags;
 756    VkCommandBufferLevel level;
 757    enum tu_cmd_buffer_status status;
 758
 759    struct tu_cmd_state state;
 760    struct tu_vertex_binding vertex_bindings[MAX_VBS];
 761    uint32_t queue_family_index;
 762
 763    uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
 764    VkShaderStageFlags push_constant_stages;
 765    struct tu_descriptor_set meta_push_descriptors;
 766
 767    struct tu_descriptor_state descriptors[VK_PIPELINE_BIND_POINT_RANGE_SIZE];
 768
 769    struct tu_cmd_buffer_upload upload;
 770
 771    struct tu_bo_list bo_list;
 772    struct tu_cmd_stream primary_cmd_stream;
 773
 774    VkResult record_result;
 775 };
 776
 777 bool
 778 tu_get_memory_fd(struct tu_device *device,
 779                  struct tu_device_memory *memory,
 780                  int *pFD);
 781
 782 /*
 783  * Takes x,y,z as exact numbers of invocations, instead of blocks.
 784  *
 785  * Limitations: Can't call normal dispatch functions without binding or
 786  * rebinding
 787  *              the compute pipeline.
 788  */
 789 void
 790 tu_unaligned_dispatch(struct tu_cmd_buffer *cmd_buffer,
 791                       uint32_t x,
 792                       uint32_t y,
 793                       uint32_t z);
 794
 795 struct tu_event
 796 {
 797    uint64_t *map;
 798 };
 799
 800 struct tu_shader_module;
 801
 802 #define TU_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
 803 #define TU_HASH_SHADER_SISCHED (1 << 1)
 804 #define TU_HASH_SHADER_UNSAFE_MATH (1 << 2)
 805 void
 806 tu_hash_shaders(unsigned char *hash,
 807                 const VkPipelineShaderStageCreateInfo **stages,
 808                 const struct tu_pipeline_layout *layout,
 809                 const struct tu_pipeline_key *key,
 810                 uint32_t flags);
 811
 812 static inline gl_shader_stage
 813 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
 814 {
 815    assert(__builtin_popcount(vk_stage) == 1);
 816    return ffs(vk_stage) - 1;
 817 }
 818
 819 static inline VkShaderStageFlagBits
 820 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
 821 {
 822    return (1 << mesa_stage);
 823 }
 824
 825 #define TU_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
 826
 827 #define tu_foreach_stage(stage, stage_bits)                                  \
 828    for (gl_shader_stage stage,                                               \
 829         __tmp = (gl_shader_stage)((stage_bits) &TU_STAGE_MASK);              \
 830         stage = __builtin_ffs(__tmp) - 1, __tmp; __tmp &= ~(1 << (stage)))
 831
 832 struct tu_shader_module
 833 {
 834    struct nir_shader *nir;
 835    unsigned char sha1[20];
 836    uint32_t size;
 837    char data[0];
 838 };
 839
 840 struct tu_pipeline
 841 {
 842    struct tu_device *device;
 843    struct tu_dynamic_state dynamic_state;
 844
 845    struct tu_pipeline_layout *layout;
 846
 847    bool need_indirect_descriptor_sets;
 848    VkShaderStageFlags active_stages;
 849 };
 850
 851 struct tu_userdata_info *
 852 tu_lookup_user_sgpr(struct tu_pipeline *pipeline,
 853                     gl_shader_stage stage,
 854                     int idx);
 855
 856 struct tu_shader_variant *
 857 tu_get_shader(struct tu_pipeline *pipeline, gl_shader_stage stage);
 858
 859 struct tu_graphics_pipeline_create_info
 860 {
 861    bool use_rectlist;
 862    bool db_depth_clear;
 863    bool db_stencil_clear;
 864    bool db_depth_disable_expclear;
 865    bool db_stencil_disable_expclear;
 866    bool db_flush_depth_inplace;
 867    bool db_flush_stencil_inplace;
 868    bool db_resummarize;
 869    uint32_t custom_blend_mode;
 870 };
 871
 872 VkResult
 873 tu_graphics_pipeline_create(
 874    VkDevice device,
 875    VkPipelineCache cache,
 876    const VkGraphicsPipelineCreateInfo *pCreateInfo,
 877    const struct tu_graphics_pipeline_create_info *extra,
 878    const VkAllocationCallbacks *alloc,
 879    VkPipeline *pPipeline);
 880
 881 struct vk_format_description;
 882 uint32_t
 883 tu_translate_buffer_dataformat(const struct vk_format_description *desc,
 884                                int first_non_void);
 885 uint32_t
 886 tu_translate_buffer_numformat(const struct vk_format_description *desc,
 887                               int first_non_void);
 888 uint32_t
 889 tu_translate_colorformat(VkFormat format);
 890 uint32_t
 891 tu_translate_color_numformat(VkFormat format,
 892                              const struct vk_format_description *desc,
 893                              int first_non_void);
 894 uint32_t
 895 tu_colorformat_endian_swap(uint32_t colorformat);
 896 unsigned
 897 tu_translate_colorswap(VkFormat format, bool do_endian_swap);
 898 uint32_t
 899 tu_translate_dbformat(VkFormat format);
 900 uint32_t
 901 tu_translate_tex_dataformat(VkFormat format,
 902                             const struct vk_format_description *desc,
 903                             int first_non_void);
 904 uint32_t
 905 tu_translate_tex_numformat(VkFormat format,
 906                            const struct vk_format_description *desc,
 907                            int first_non_void);
 908 bool
 909 tu_format_pack_clear_color(VkFormat format,
 910                            uint32_t clear_vals[2],
 911                            VkClearColorValue *value);
 912 bool
 913 tu_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
 914 bool
 915 tu_dcc_formats_compatible(VkFormat format1, VkFormat format2);
 916
 917 struct tu_image_level
 918 {
 919    VkDeviceSize offset;
 920    VkDeviceSize size;
 921    uint32_t pitch;
 922 };
 923
 924 struct tu_image
 925 {
 926    VkImageType type;
 927    /* The original VkFormat provided by the client.  This may not match any
 928     * of the actual surface formats.
 929     */
 930    VkFormat vk_format;
 931    VkImageAspectFlags aspects;
 932    VkImageUsageFlags usage;  /**< Superset of VkImageCreateInfo::usage. */
 933    VkImageTiling tiling;     /** VkImageCreateInfo::tiling */
 934    VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
 935    VkExtent3D extent;
 936
 937    VkDeviceSize size;
 938    uint32_t alignment;
 939
 940    /* memory layout */
 941    VkDeviceSize layer_size;
 942    struct tu_image_level levels[15];
 943    unsigned tile_mode;
 944
 945    unsigned queue_family_mask;
 946    bool exclusive;
 947    bool shareable;
 948
 949    /* For VK_ANDROID_native_buffer, the WSI image owns the memory, */
 950    VkDeviceMemory owned_memory;
 951 };
 952
 953 unsigned
 954 tu_image_queue_family_mask(const struct tu_image *image,
 955                            uint32_t family,
 956                            uint32_t queue_family);
 957
 958 static inline uint32_t
 959 tu_get_layerCount(const struct tu_image *image,
 960                   const VkImageSubresourceRange *range)
 961 {
 962    abort();
 963 }
 964
 965 static inline uint32_t
 966 tu_get_levelCount(const struct tu_image *image,
 967                   const VkImageSubresourceRange *range)
 968 {
 969    abort();
 970 }
 971
 972 struct tu_image_view
 973 {
 974    struct tu_image *image; /**< VkImageViewCreateInfo::image */
 975
 976    VkImageViewType type;
 977    VkImageAspectFlags aspect_mask;
 978    VkFormat vk_format;
 979    uint32_t base_layer;
 980    uint32_t layer_count;
 981    uint32_t base_mip;
 982    uint32_t level_count;
 983    VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
 984
 985    uint32_t descriptor[16];
 986
 987    /* Descriptor for use as a storage image as opposed to a sampled image.
 988     * This has a few differences for cube maps (e.g. type).
 989     */
 990    uint32_t storage_descriptor[16];
 991 };
 992
 993 struct tu_sampler
 994 {
 995 };
 996
 997 struct tu_image_create_info
 998 {
 999    const VkImageCreateInfo *vk_info;
1000    bool scanout;
1001    bool no_metadata_planes;
1002 };
1003
1004 VkResult
1005 tu_image_create(VkDevice _device,
1006                 const struct tu_image_create_info *info,
1007                 const VkAllocationCallbacks *alloc,
1008                 VkImage *pImage);
1009
1010 VkResult
1011 tu_image_from_gralloc(VkDevice device_h,
1012                       const VkImageCreateInfo *base_info,
1013                       const VkNativeBufferANDROID *gralloc_info,
1014                       const VkAllocationCallbacks *alloc,
1015                       VkImage *out_image_h);
1016
1017 void
1018 tu_image_view_init(struct tu_image_view *view,
1019                    struct tu_device *device,
1020                    const VkImageViewCreateInfo *pCreateInfo);
1021
1022 struct tu_buffer_view
1023 {
1024    VkFormat vk_format;
1025    uint64_t range; /**< VkBufferViewCreateInfo::range */
1026    uint32_t state[4];
1027 };
1028 void
1029 tu_buffer_view_init(struct tu_buffer_view *view,
1030                     struct tu_device *device,
1031                     const VkBufferViewCreateInfo *pCreateInfo);
1032
1033 static inline struct VkExtent3D
1034 tu_sanitize_image_extent(const VkImageType imageType,
1035                          const struct VkExtent3D imageExtent)
1036 {
1037    switch (imageType) {
1038    case VK_IMAGE_TYPE_1D:
1039       return (VkExtent3D) { imageExtent.width, 1, 1 };
1040    case VK_IMAGE_TYPE_2D:
1041       return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
1042    case VK_IMAGE_TYPE_3D:
1043       return imageExtent;
1044    default:
1045       unreachable("invalid image type");
1046    }
1047 }
1048
1049 static inline struct VkOffset3D
1050 tu_sanitize_image_offset(const VkImageType imageType,
1051                          const struct VkOffset3D imageOffset)
1052 {
1053    switch (imageType) {
1054    case VK_IMAGE_TYPE_1D:
1055       return (VkOffset3D) { imageOffset.x, 0, 0 };
1056    case VK_IMAGE_TYPE_2D:
1057       return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
1058    case VK_IMAGE_TYPE_3D:
1059       return imageOffset;
1060    default:
1061       unreachable("invalid image type");
1062    }
1063 }
1064
1065 struct tu_attachment_info
1066 {
1067    struct tu_image_view *attachment;
1068 };
1069
1070 struct tu_framebuffer
1071 {
1072    uint32_t width;
1073    uint32_t height;
1074    uint32_t layers;
1075
1076    uint32_t attachment_count;
1077    struct tu_attachment_info attachments[0];
1078 };
1079
1080 struct tu_subpass_barrier
1081 {
1082    VkPipelineStageFlags src_stage_mask;
1083    VkAccessFlags src_access_mask;
1084    VkAccessFlags dst_access_mask;
1085 };
1086
1087 void
1088 tu_subpass_barrier(struct tu_cmd_buffer *cmd_buffer,
1089                    const struct tu_subpass_barrier *barrier);
1090
1091 struct tu_subpass_attachment
1092 {
1093    uint32_t attachment;
1094    VkImageLayout layout;
1095 };
1096
1097 struct tu_subpass
1098 {
1099    uint32_t input_count;
1100    uint32_t color_count;
1101    struct tu_subpass_attachment *input_attachments;
1102    struct tu_subpass_attachment *color_attachments;
1103    struct tu_subpass_attachment *resolve_attachments;
1104    struct tu_subpass_attachment depth_stencil_attachment;
1105
1106    /** Subpass has at least one resolve attachment */
1107    bool has_resolve;
1108
1109    struct tu_subpass_barrier start_barrier;
1110
1111    uint32_t view_mask;
1112    VkSampleCountFlagBits max_sample_count;
1113 };
1114
1115 struct tu_render_pass_attachment
1116 {
1117    VkFormat format;
1118    uint32_t samples;
1119    VkAttachmentLoadOp load_op;
1120    VkAttachmentLoadOp stencil_load_op;
1121    VkImageLayout initial_layout;
1122    VkImageLayout final_layout;
1123    uint32_t view_mask;
1124 };
1125
1126 struct tu_render_pass
1127 {
1128    uint32_t attachment_count;
1129    uint32_t subpass_count;
1130    struct tu_subpass_attachment *subpass_attachments;
1131    struct tu_render_pass_attachment *attachments;
1132    struct tu_subpass_barrier end_barrier;
1133    struct tu_subpass subpasses[0];
1134 };
1135
1136 VkResult
1137 tu_device_init_meta(struct tu_device *device);
1138 void
1139 tu_device_finish_meta(struct tu_device *device);
1140
1141 struct tu_query_pool
1142 {
1143    uint32_t stride;
1144    uint32_t availability_offset;
1145    uint64_t size;
1146    char *ptr;
1147    VkQueryType type;
1148    uint32_t pipeline_stats_mask;
1149 };
1150
1151 struct tu_semaphore
1152 {
1153    uint32_t syncobj;
1154    uint32_t temp_syncobj;
1155 };
1156
1157 void
1158 tu_set_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1159                       VkPipelineBindPoint bind_point,
1160                       struct tu_descriptor_set *set,
1161                       unsigned idx);
1162
1163 void
1164 tu_update_descriptor_sets(struct tu_device *device,
1165                           struct tu_cmd_buffer *cmd_buffer,
1166                           VkDescriptorSet overrideSet,
1167                           uint32_t descriptorWriteCount,
1168                           const VkWriteDescriptorSet *pDescriptorWrites,
1169                           uint32_t descriptorCopyCount,
1170                           const VkCopyDescriptorSet *pDescriptorCopies);
1171
1172 void
1173 tu_update_descriptor_set_with_template(
1174    struct tu_device *device,
1175    struct tu_cmd_buffer *cmd_buffer,
1176    struct tu_descriptor_set *set,
1177    VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1178    const void *pData);
1179
1180 void
1181 tu_meta_push_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1182                             VkPipelineBindPoint pipelineBindPoint,
1183                             VkPipelineLayout _layout,
1184                             uint32_t set,
1185                             uint32_t descriptorWriteCount,
1186                             const VkWriteDescriptorSet *pDescriptorWrites);
1187
1188 struct tu_fence
1189 {
1190    uint32_t syncobj;
1191    uint32_t temp_syncobj;
1192 };
1193
1194 uint32_t
1195 tu_gem_new(struct tu_device *dev, uint64_t size, uint32_t flags);
1196 void
1197 tu_gem_close(struct tu_device *dev, uint32_t gem_handle);
1198 uint64_t
1199 tu_gem_info_offset(struct tu_device *dev, uint32_t gem_handle);
1200 uint64_t
1201 tu_gem_info_iova(struct tu_device *dev, uint32_t gem_handle);
1202 int
1203 tu_drm_query_param(struct tu_physical_device *dev,
1204                    uint32_t param,
1205                    uint64_t *value);
1206
1207 #define TU_DEFINE_HANDLE_CASTS(__tu_type, __VkType)                          \
1208                                                                              \
1209    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1210    {                                                                         \
1211       return (struct __tu_type *) _handle;                                   \
1212    }                                                                         \
1213                                                                              \
1214    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1215    {                                                                         \
1216       return (__VkType) _obj;                                                \
1217    }
1218
1219 #define TU_DEFINE_NONDISP_HANDLE_CASTS(__tu_type, __VkType)                  \
1220                                                                              \
1221    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1222    {                                                                         \
1223       return (struct __tu_type *) (uintptr_t) _handle;                       \
1224    }                                                                         \
1225                                                                              \
1226    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1227    {                                                                         \
1228       return (__VkType)(uintptr_t) _obj;                                     \
1229    }
1230
1231 #define TU_FROM_HANDLE(__tu_type, __name, __handle)                          \
1232    struct __tu_type *__name = __tu_type##_from_handle(__handle)
1233
1234 TU_DEFINE_HANDLE_CASTS(tu_cmd_buffer, VkCommandBuffer)
1235 TU_DEFINE_HANDLE_CASTS(tu_device, VkDevice)
1236 TU_DEFINE_HANDLE_CASTS(tu_instance, VkInstance)
1237 TU_DEFINE_HANDLE_CASTS(tu_physical_device, VkPhysicalDevice)
1238 TU_DEFINE_HANDLE_CASTS(tu_queue, VkQueue)
1239
1240 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_cmd_pool, VkCommandPool)
1241 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer, VkBuffer)
1242 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer_view, VkBufferView)
1243 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_pool, VkDescriptorPool)
1244 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set, VkDescriptorSet)
1245 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set_layout,
1246                                VkDescriptorSetLayout)
1247 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_update_template,
1248                                VkDescriptorUpdateTemplateKHR)
1249 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_device_memory, VkDeviceMemory)
1250 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_fence, VkFence)
1251 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_event, VkEvent)
1252 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_framebuffer, VkFramebuffer)
1253 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image, VkImage)
1254 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image_view, VkImageView);
1255 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_cache, VkPipelineCache)
1256 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline, VkPipeline)
1257 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_layout, VkPipelineLayout)
1258 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_query_pool, VkQueryPool)
1259 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_render_pass, VkRenderPass)
1260 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_sampler, VkSampler)
1261 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_shader_module, VkShaderModule)
1262 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_semaphore, VkSemaphore)
1263
1264 #endif /* TU_PRIVATE_H */