src/freedreno/vulkan/tu_private.h

   1 /*
   2  * Copyright © 2016 Red Hat.
   3  * Copyright © 2016 Bas Nieuwenhuizen
   4  *
   5  * based in part on anv driver which is:
   6  * Copyright © 2015 Intel Corporation
   7  *
   8  * Permission is hereby granted, free of charge, to any person obtaining a
   9  * copy of this software and associated documentation files (the "Software"),
  10  * to deal in the Software without restriction, including without limitation
  11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  12  * and/or sell copies of the Software, and to permit persons to whom the
  13  * Software is furnished to do so, subject to the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the next
  16  * paragraph) shall be included in all copies or substantial portions of the
  17  * Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  22  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  23  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  24  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  25  * DEALINGS IN THE SOFTWARE.
  26  */
  27
  28 #ifndef TU_PRIVATE_H
  29 #define TU_PRIVATE_H
  30
  31 #include <assert.h>
  32 #include <pthread.h>
  33 #include <stdbool.h>
  34 #include <stdint.h>
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <string.h>
  38 #ifdef HAVE_VALGRIND
  39 #include <memcheck.h>
  40 #include <valgrind.h>
  41 #define VG(x) x
  42 #else
  43 #define VG(x)
  44 #endif
  45
  46 #include "c11/threads.h"
  47 #include "compiler/shader_enums.h"
  48 #include "main/macros.h"
  49 #include "util/list.h"
  50 #include "util/macros.h"
  51 #include "vk_alloc.h"
  52 #include "vk_debug_report.h"
  53
  54 #include "drm/msm_drm.h"
  55 #include "tu_descriptor_set.h"
  56 #include "tu_extensions.h"
  57
  58 /* Pre-declarations needed for WSI entrypoints */
  59 struct wl_surface;
  60 struct wl_display;
  61 typedef struct xcb_connection_t xcb_connection_t;
  62 typedef uint32_t xcb_visualid_t;
  63 typedef uint32_t xcb_window_t;
  64
  65 #include <vulkan/vk_android_native_buffer.h>
  66 #include <vulkan/vk_icd.h>
  67 #include <vulkan/vulkan.h>
  68 #include <vulkan/vulkan_intel.h>
  69
  70 #include "tu_entrypoints.h"
  71
  72 #define MAX_VBS 32
  73 #define MAX_VERTEX_ATTRIBS 32
  74 #define MAX_RTS 8
  75 #define MAX_VIEWPORTS 16
  76 #define MAX_SCISSORS 16
  77 #define MAX_DISCARD_RECTANGLES 4
  78 #define MAX_PUSH_CONSTANTS_SIZE 128
  79 #define MAX_PUSH_DESCRIPTORS 32
  80 #define MAX_DYNAMIC_UNIFORM_BUFFERS 16
  81 #define MAX_DYNAMIC_STORAGE_BUFFERS 8
  82 #define MAX_DYNAMIC_BUFFERS                                                  \
  83    (MAX_DYNAMIC_UNIFORM_BUFFERS + MAX_DYNAMIC_STORAGE_BUFFERS)
  84 #define MAX_SAMPLES_LOG2 4
  85 #define NUM_META_FS_KEYS 13
  86 #define TU_MAX_DRM_DEVICES 8
  87 #define MAX_VIEWS 8
  88
  89 #define NUM_DEPTH_CLEAR_PIPELINES 3
  90
  91 /*
  92  * This is the point we switch from using CP to compute shader
  93  * for certain buffer operations.
  94  */
  95 #define TU_BUFFER_OPS_CS_THRESHOLD 4096
  96
  97 enum tu_mem_heap
  98 {
  99    TU_MEM_HEAP_VRAM,
 100    TU_MEM_HEAP_VRAM_CPU_ACCESS,
 101    TU_MEM_HEAP_GTT,
 102    TU_MEM_HEAP_COUNT
 103 };
 104
 105 enum tu_mem_type
 106 {
 107    TU_MEM_TYPE_VRAM,
 108    TU_MEM_TYPE_GTT_WRITE_COMBINE,
 109    TU_MEM_TYPE_VRAM_CPU_ACCESS,
 110    TU_MEM_TYPE_GTT_CACHED,
 111    TU_MEM_TYPE_COUNT
 112 };
 113
 114 #define tu_printflike(a, b) __attribute__((__format__(__printf__, a, b)))
 115
 116 static inline uint32_t
 117 align_u32(uint32_t v, uint32_t a)
 118 {
 119    assert(a != 0 && a == (a & -a));
 120    return (v + a - 1) & ~(a - 1);
 121 }
 122
 123 static inline uint32_t
 124 align_u32_npot(uint32_t v, uint32_t a)
 125 {
 126    return (v + a - 1) / a * a;
 127 }
 128
 129 static inline uint64_t
 130 align_u64(uint64_t v, uint64_t a)
 131 {
 132    assert(a != 0 && a == (a & -a));
 133    return (v + a - 1) & ~(a - 1);
 134 }
 135
 136 static inline int32_t
 137 align_i32(int32_t v, int32_t a)
 138 {
 139    assert(a != 0 && a == (a & -a));
 140    return (v + a - 1) & ~(a - 1);
 141 }
 142
 143 /** Alignment must be a power of 2. */
 144 static inline bool
 145 tu_is_aligned(uintmax_t n, uintmax_t a)
 146 {
 147    assert(a == (a & -a));
 148    return (n & (a - 1)) == 0;
 149 }
 150
 151 static inline uint32_t
 152 round_up_u32(uint32_t v, uint32_t a)
 153 {
 154    return (v + a - 1) / a;
 155 }
 156
 157 static inline uint64_t
 158 round_up_u64(uint64_t v, uint64_t a)
 159 {
 160    return (v + a - 1) / a;
 161 }
 162
 163 static inline uint32_t
 164 tu_minify(uint32_t n, uint32_t levels)
 165 {
 166    if (unlikely(n == 0))
 167       return 0;
 168    else
 169       return MAX2(n >> levels, 1);
 170 }
 171 static inline float
 172 tu_clamp_f(float f, float min, float max)
 173 {
 174    assert(min < max);
 175
 176    if (f > max)
 177       return max;
 178    else if (f < min)
 179       return min;
 180    else
 181       return f;
 182 }
 183
 184 static inline bool
 185 tu_clear_mask(uint32_t *inout_mask, uint32_t clear_mask)
 186 {
 187    if (*inout_mask & clear_mask) {
 188       *inout_mask &= ~clear_mask;
 189       return true;
 190    } else {
 191       return false;
 192    }
 193 }
 194
 195 #define for_each_bit(b, dword)                                               \
 196    for (uint32_t __dword = (dword);                                          \
 197         (b) = __builtin_ffs(__dword) - 1, __dword; __dword &= ~(1 << (b)))
 198
 199 #define typed_memcpy(dest, src, count)                                       \
 200    ({                                                                        \
 201       STATIC_ASSERT(sizeof(*src) == sizeof(*dest));                          \
 202       memcpy((dest), (src), (count) * sizeof(*(src)));                       \
 203    })
 204
 205 /* Whenever we generate an error, pass it through this function. Useful for
 206  * debugging, where we can break on it. Only call at error site, not when
 207  * propagating errors. Might be useful to plug in a stack trace here.
 208  */
 209
 210 struct tu_instance;
 211
 212 VkResult
 213 __vk_errorf(struct tu_instance *instance,
 214             VkResult error,
 215             const char *file,
 216             int line,
 217             const char *format,
 218             ...);
 219
 220 #define vk_error(instance, error)                                            \
 221    __vk_errorf(instance, error, __FILE__, __LINE__, NULL);
 222 #define vk_errorf(instance, error, format, ...)                              \
 223    __vk_errorf(instance, error, __FILE__, __LINE__, format, ##__VA_ARGS__);
 224
 225 void
 226 __tu_finishme(const char *file, int line, const char *format, ...)
 227    tu_printflike(3, 4);
 228 void
 229 tu_loge(const char *format, ...) tu_printflike(1, 2);
 230 void
 231 tu_loge_v(const char *format, va_list va);
 232 void
 233 tu_logi(const char *format, ...) tu_printflike(1, 2);
 234 void
 235 tu_logi_v(const char *format, va_list va);
 236
 237 /**
 238  * Print a FINISHME message, including its source location.
 239  */
 240 #define tu_finishme(format, ...)                                             \
 241    do {                                                                      \
 242       static bool reported = false;                                          \
 243       if (!reported) {                                                       \
 244          __tu_finishme(__FILE__, __LINE__, format, ##__VA_ARGS__);           \
 245          reported = true;                                                    \
 246       }                                                                      \
 247    } while (0)
 248
 249 /* A non-fatal assert.  Useful for debugging. */
 250 #ifdef DEBUG
 251 #define tu_assert(x)                                                         \
 252    ({                                                                        \
 253       if (unlikely(!(x)))                                                    \
 254          fprintf(stderr, "%s:%d ASSERT: %s\n", __FILE__, __LINE__, #x);      \
 255    })
 256 #else
 257 #define tu_assert(x)
 258 #endif
 259
 260 /* Suppress -Wunused in stub functions */
 261 #define tu_use_args(...) __tu_use_args(0, ##__VA_ARGS__)
 262 static inline void
 263 __tu_use_args(int ignore, ...)
 264 {
 265 }
 266
 267 #define tu_stub()                                                            \
 268    do {                                                                      \
 269       tu_finishme("stub %s", __func__);                                      \
 270    } while (0)
 271
 272 void *
 273 tu_lookup_entrypoint_unchecked(const char *name);
 274 void *
 275 tu_lookup_entrypoint_checked(
 276    const char *name,
 277    uint32_t core_version,
 278    const struct tu_instance_extension_table *instance,
 279    const struct tu_device_extension_table *device);
 280
 281 struct tu_physical_device
 282 {
 283    VK_LOADER_DATA _loader_data;
 284
 285    struct tu_instance *instance;
 286
 287    char path[20];
 288    char name[VK_MAX_PHYSICAL_DEVICE_NAME_SIZE];
 289    uint8_t driver_uuid[VK_UUID_SIZE];
 290    uint8_t device_uuid[VK_UUID_SIZE];
 291    uint8_t cache_uuid[VK_UUID_SIZE];
 292
 293    int local_fd;
 294    int master_fd;
 295
 296    unsigned gpu_id;
 297    uint32_t gmem_size;
 298    uint32_t tile_align_w;
 299    uint32_t tile_align_h;
 300
 301    /* This is the drivers on-disk cache used as a fallback as opposed to
 302     * the pipeline cache defined by apps.
 303     */
 304    struct disk_cache *disk_cache;
 305
 306    struct tu_device_extension_table supported_extensions;
 307 };
 308
 309 enum tu_debug_flags
 310 {
 311    TU_DEBUG_STARTUP = 1 << 0,
 312 };
 313
 314 struct tu_instance
 315 {
 316    VK_LOADER_DATA _loader_data;
 317
 318    VkAllocationCallbacks alloc;
 319
 320    uint32_t api_version;
 321    int physical_device_count;
 322    struct tu_physical_device physical_devices[TU_MAX_DRM_DEVICES];
 323
 324    enum tu_debug_flags debug_flags;
 325
 326    struct vk_debug_report_instance debug_report_callbacks;
 327
 328    struct tu_instance_extension_table enabled_extensions;
 329 };
 330
 331 bool
 332 tu_instance_extension_supported(const char *name);
 333 uint32_t
 334 tu_physical_device_api_version(struct tu_physical_device *dev);
 335 bool
 336 tu_physical_device_extension_supported(struct tu_physical_device *dev,
 337                                        const char *name);
 338
 339 struct cache_entry;
 340
 341 struct tu_pipeline_cache
 342 {
 343    struct tu_device *device;
 344    pthread_mutex_t mutex;
 345
 346    uint32_t total_size;
 347    uint32_t table_size;
 348    uint32_t kernel_count;
 349    struct cache_entry **hash_table;
 350    bool modified;
 351
 352    VkAllocationCallbacks alloc;
 353 };
 354
 355 struct tu_pipeline_key
 356 {
 357 };
 358
 359 void
 360 tu_pipeline_cache_init(struct tu_pipeline_cache *cache,
 361                        struct tu_device *device);
 362 void
 363 tu_pipeline_cache_finish(struct tu_pipeline_cache *cache);
 364 void
 365 tu_pipeline_cache_load(struct tu_pipeline_cache *cache,
 366                        const void *data,
 367                        size_t size);
 368
 369 struct tu_shader_variant;
 370
 371 bool
 372 tu_create_shader_variants_from_pipeline_cache(
 373    struct tu_device *device,
 374    struct tu_pipeline_cache *cache,
 375    const unsigned char *sha1,
 376    struct tu_shader_variant **variants);
 377
 378 void
 379 tu_pipeline_cache_insert_shaders(struct tu_device *device,
 380                                  struct tu_pipeline_cache *cache,
 381                                  const unsigned char *sha1,
 382                                  struct tu_shader_variant **variants,
 383                                  const void *const *codes,
 384                                  const unsigned *code_sizes);
 385
 386 struct tu_meta_state
 387 {
 388    VkAllocationCallbacks alloc;
 389
 390    struct tu_pipeline_cache cache;
 391 };
 392
 393 /* queue types */
 394 #define TU_QUEUE_GENERAL 0
 395
 396 #define TU_MAX_QUEUE_FAMILIES 1
 397
 398 struct tu_queue
 399 {
 400    VK_LOADER_DATA _loader_data;
 401    struct tu_device *device;
 402    uint32_t queue_family_index;
 403    int queue_idx;
 404    VkDeviceQueueCreateFlags flags;
 405
 406    uint32_t msm_queue_id;
 407    int submit_fence_fd;
 408 };
 409
 410 struct tu_device
 411 {
 412    VK_LOADER_DATA _loader_data;
 413
 414    VkAllocationCallbacks alloc;
 415
 416    struct tu_instance *instance;
 417
 418    struct tu_meta_state meta_state;
 419
 420    struct tu_queue *queues[TU_MAX_QUEUE_FAMILIES];
 421    int queue_count[TU_MAX_QUEUE_FAMILIES];
 422
 423    struct tu_physical_device *physical_device;
 424
 425    /* Backup in-memory cache to be used if the app doesn't provide one */
 426    struct tu_pipeline_cache *mem_cache;
 427
 428    struct list_head shader_slabs;
 429    mtx_t shader_slab_mutex;
 430
 431    struct tu_device_extension_table enabled_extensions;
 432 };
 433
 434 struct tu_bo
 435 {
 436    uint32_t gem_handle;
 437    uint64_t size;
 438    uint64_t iova;
 439    void *map;
 440 };
 441
 442 VkResult
 443 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size);
 444 void
 445 tu_bo_finish(struct tu_device *dev, struct tu_bo *bo);
 446 VkResult
 447 tu_bo_map(struct tu_device *dev, struct tu_bo *bo);
 448
 449 struct tu_device_memory
 450 {
 451    struct tu_bo bo;
 452    VkDeviceSize size;
 453
 454    /* for dedicated allocations */
 455    struct tu_image *image;
 456    struct tu_buffer *buffer;
 457
 458    uint32_t type_index;
 459    void *map;
 460    void *user_ptr;
 461 };
 462
 463 struct tu_descriptor_range
 464 {
 465    uint64_t va;
 466    uint32_t size;
 467 };
 468
 469 struct tu_descriptor_set
 470 {
 471    const struct tu_descriptor_set_layout *layout;
 472    uint32_t size;
 473
 474    uint64_t va;
 475    uint32_t *mapped_ptr;
 476    struct tu_descriptor_range *dynamic_descriptors;
 477 };
 478
 479 struct tu_push_descriptor_set
 480 {
 481    struct tu_descriptor_set set;
 482    uint32_t capacity;
 483 };
 484
 485 struct tu_descriptor_pool_entry
 486 {
 487    uint32_t offset;
 488    uint32_t size;
 489    struct tu_descriptor_set *set;
 490 };
 491
 492 struct tu_descriptor_pool
 493 {
 494    uint8_t *mapped_ptr;
 495    uint64_t current_offset;
 496    uint64_t size;
 497
 498    uint8_t *host_memory_base;
 499    uint8_t *host_memory_ptr;
 500    uint8_t *host_memory_end;
 501
 502    uint32_t entry_count;
 503    uint32_t max_entry_count;
 504    struct tu_descriptor_pool_entry entries[0];
 505 };
 506
 507 struct tu_descriptor_update_template_entry
 508 {
 509    VkDescriptorType descriptor_type;
 510
 511    /* The number of descriptors to update */
 512    uint32_t descriptor_count;
 513
 514    /* Into mapped_ptr or dynamic_descriptors, in units of the respective array
 515     */
 516    uint32_t dst_offset;
 517
 518    /* In dwords. Not valid/used for dynamic descriptors */
 519    uint32_t dst_stride;
 520
 521    uint32_t buffer_offset;
 522
 523    /* Only valid for combined image samplers and samplers */
 524    uint16_t has_sampler;
 525
 526    /* In bytes */
 527    size_t src_offset;
 528    size_t src_stride;
 529
 530    /* For push descriptors */
 531    const uint32_t *immutable_samplers;
 532 };
 533
 534 struct tu_descriptor_update_template
 535 {
 536    uint32_t entry_count;
 537    VkPipelineBindPoint bind_point;
 538    struct tu_descriptor_update_template_entry entry[0];
 539 };
 540
 541 struct tu_buffer
 542 {
 543    VkDeviceSize size;
 544
 545    VkBufferUsageFlags usage;
 546    VkBufferCreateFlags flags;
 547 };
 548
 549 enum tu_dynamic_state_bits
 550 {
 551    TU_DYNAMIC_VIEWPORT = 1 << 0,
 552    TU_DYNAMIC_SCISSOR = 1 << 1,
 553    TU_DYNAMIC_LINE_WIDTH = 1 << 2,
 554    TU_DYNAMIC_DEPTH_BIAS = 1 << 3,
 555    TU_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
 556    TU_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
 557    TU_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
 558    TU_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
 559    TU_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
 560    TU_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
 561    TU_DYNAMIC_ALL = (1 << 10) - 1,
 562 };
 563
 564 struct tu_vertex_binding
 565 {
 566    struct tu_buffer *buffer;
 567    VkDeviceSize offset;
 568 };
 569
 570 struct tu_viewport_state
 571 {
 572    uint32_t count;
 573    VkViewport viewports[MAX_VIEWPORTS];
 574 };
 575
 576 struct tu_scissor_state
 577 {
 578    uint32_t count;
 579    VkRect2D scissors[MAX_SCISSORS];
 580 };
 581
 582 struct tu_discard_rectangle_state
 583 {
 584    uint32_t count;
 585    VkRect2D rectangles[MAX_DISCARD_RECTANGLES];
 586 };
 587
 588 struct tu_dynamic_state
 589 {
 590    /**
 591     * Bitmask of (1 << VK_DYNAMIC_STATE_*).
 592     * Defines the set of saved dynamic state.
 593     */
 594    uint32_t mask;
 595
 596    struct tu_viewport_state viewport;
 597
 598    struct tu_scissor_state scissor;
 599
 600    float line_width;
 601
 602    struct
 603    {
 604       float bias;
 605       float clamp;
 606       float slope;
 607    } depth_bias;
 608
 609    float blend_constants[4];
 610
 611    struct
 612    {
 613       float min;
 614       float max;
 615    } depth_bounds;
 616
 617    struct
 618    {
 619       uint32_t front;
 620       uint32_t back;
 621    } stencil_compare_mask;
 622
 623    struct
 624    {
 625       uint32_t front;
 626       uint32_t back;
 627    } stencil_write_mask;
 628
 629    struct
 630    {
 631       uint32_t front;
 632       uint32_t back;
 633    } stencil_reference;
 634
 635    struct tu_discard_rectangle_state discard_rectangle;
 636 };
 637
 638 extern const struct tu_dynamic_state default_dynamic_state;
 639
 640 const char *
 641 tu_get_debug_option_name(int id);
 642
 643 const char *
 644 tu_get_perftest_option_name(int id);
 645
 646 /**
 647  * Attachment state when recording a renderpass instance.
 648  *
 649  * The clear value is valid only if there exists a pending clear.
 650  */
 651 struct tu_attachment_state
 652 {
 653    VkImageAspectFlags pending_clear_aspects;
 654    uint32_t cleared_views;
 655    VkClearValue clear_value;
 656    VkImageLayout current_layout;
 657 };
 658
 659 struct tu_descriptor_state
 660 {
 661    struct tu_descriptor_set *sets[MAX_SETS];
 662    uint32_t dirty;
 663    uint32_t valid;
 664    struct tu_push_descriptor_set push_set;
 665    bool push_dirty;
 666    uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
 667 };
 668
 669 struct tu_cmd_state
 670 {
 671    /* Vertex descriptors */
 672    uint64_t vb_va;
 673    unsigned vb_size;
 674
 675    struct tu_dynamic_state dynamic;
 676
 677    /* Index buffer */
 678    struct tu_buffer *index_buffer;
 679    uint64_t index_offset;
 680    uint32_t index_type;
 681    uint32_t max_index_count;
 682    uint64_t index_va;
 683
 684    const struct tu_render_pass *pass;
 685    const struct tu_subpass *subpass;
 686    const struct tu_framebuffer *framebuffer;
 687    struct tu_attachment_state *attachments;
 688 };
 689
 690 struct tu_cmd_pool
 691 {
 692    VkAllocationCallbacks alloc;
 693    struct list_head cmd_buffers;
 694    struct list_head free_cmd_buffers;
 695    uint32_t queue_family_index;
 696 };
 697
 698 struct tu_cmd_buffer_upload
 699 {
 700    uint8_t *map;
 701    unsigned offset;
 702    uint64_t size;
 703    struct list_head list;
 704 };
 705
 706 enum tu_cmd_buffer_status
 707 {
 708    TU_CMD_BUFFER_STATUS_INVALID,
 709    TU_CMD_BUFFER_STATUS_INITIAL,
 710    TU_CMD_BUFFER_STATUS_RECORDING,
 711    TU_CMD_BUFFER_STATUS_EXECUTABLE,
 712    TU_CMD_BUFFER_STATUS_PENDING,
 713 };
 714
 715 struct tu_bo_list
 716 {
 717    uint32_t count;
 718    uint32_t capacity;
 719    struct drm_msm_gem_submit_bo *bo_infos;
 720 };
 721
 722 #define TU_BO_LIST_FAILED (~0)
 723
 724 void
 725 tu_bo_list_init(struct tu_bo_list *list);
 726 void
 727 tu_bo_list_destroy(struct tu_bo_list *list);
 728 void
 729 tu_bo_list_reset(struct tu_bo_list *list);
 730 uint32_t
 731 tu_bo_list_add(struct tu_bo_list *list,
 732                const struct tu_bo *bo,
 733                uint32_t flags);
 734 VkResult
 735 tu_bo_list_merge(struct tu_bo_list *list, const struct tu_bo_list *other);
 736
 737 struct tu_cs_entry
 738 {
 739    /* No ownership */
 740    const struct tu_bo *bo;
 741
 742    uint32_t size;
 743    uint64_t offset;
 744 };
 745
 746 struct tu_cs
 747 {
 748    uint32_t *start;
 749    uint32_t *cur;
 750    uint32_t *end;
 751
 752    /* for tu_cs_reserve_space_assert */
 753    uint32_t *reserved_end;
 754
 755    struct tu_cs_entry *entries;
 756    uint32_t entry_count;
 757    uint32_t entry_capacity;
 758
 759    struct tu_bo **bos;
 760    uint32_t bo_count;
 761    uint32_t bo_capacity;
 762 };
 763
 764 struct tu_cmd_buffer
 765 {
 766    VK_LOADER_DATA _loader_data;
 767
 768    struct tu_device *device;
 769
 770    struct tu_cmd_pool *pool;
 771    struct list_head pool_link;
 772
 773    VkCommandBufferUsageFlags usage_flags;
 774    VkCommandBufferLevel level;
 775    enum tu_cmd_buffer_status status;
 776
 777    struct tu_cmd_state state;
 778    struct tu_vertex_binding vertex_bindings[MAX_VBS];
 779    uint32_t queue_family_index;
 780
 781    uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
 782    VkShaderStageFlags push_constant_stages;
 783    struct tu_descriptor_set meta_push_descriptors;
 784
 785    struct tu_descriptor_state descriptors[VK_PIPELINE_BIND_POINT_RANGE_SIZE];
 786
 787    struct tu_cmd_buffer_upload upload;
 788
 789    VkResult record_result;
 790
 791    struct tu_bo_list bo_list;
 792    struct tu_cs cs;
 793
 794    uint16_t marker_reg;
 795    uint32_t marker_seqno;
 796
 797    struct tu_bo scratch_bo;
 798    uint32_t scratch_seqno;
 799
 800    /* current cs; command packets are always emitted to it */
 801    struct tu_cs *cur_cs;
 802 };
 803
 804 bool
 805 tu_get_memory_fd(struct tu_device *device,
 806                  struct tu_device_memory *memory,
 807                  int *pFD);
 808
 809 /*
 810  * Takes x,y,z as exact numbers of invocations, instead of blocks.
 811  *
 812  * Limitations: Can't call normal dispatch functions without binding or
 813  * rebinding
 814  *              the compute pipeline.
 815  */
 816 void
 817 tu_unaligned_dispatch(struct tu_cmd_buffer *cmd_buffer,
 818                       uint32_t x,
 819                       uint32_t y,
 820                       uint32_t z);
 821
 822 struct tu_event
 823 {
 824    uint64_t *map;
 825 };
 826
 827 struct tu_shader_module;
 828
 829 #define TU_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
 830 #define TU_HASH_SHADER_SISCHED (1 << 1)
 831 #define TU_HASH_SHADER_UNSAFE_MATH (1 << 2)
 832 void
 833 tu_hash_shaders(unsigned char *hash,
 834                 const VkPipelineShaderStageCreateInfo **stages,
 835                 const struct tu_pipeline_layout *layout,
 836                 const struct tu_pipeline_key *key,
 837                 uint32_t flags);
 838
 839 static inline gl_shader_stage
 840 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
 841 {
 842    assert(__builtin_popcount(vk_stage) == 1);
 843    return ffs(vk_stage) - 1;
 844 }
 845
 846 static inline VkShaderStageFlagBits
 847 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
 848 {
 849    return (1 << mesa_stage);
 850 }
 851
 852 #define TU_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
 853
 854 #define tu_foreach_stage(stage, stage_bits)                                  \
 855    for (gl_shader_stage stage,                                               \
 856         __tmp = (gl_shader_stage)((stage_bits) &TU_STAGE_MASK);              \
 857         stage = __builtin_ffs(__tmp) - 1, __tmp; __tmp &= ~(1 << (stage)))
 858
 859 struct tu_shader_module
 860 {
 861    struct nir_shader *nir;
 862    unsigned char sha1[20];
 863    uint32_t size;
 864    char data[0];
 865 };
 866
 867 struct tu_pipeline
 868 {
 869    struct tu_device *device;
 870    struct tu_dynamic_state dynamic_state;
 871
 872    struct tu_pipeline_layout *layout;
 873
 874    bool need_indirect_descriptor_sets;
 875    VkShaderStageFlags active_stages;
 876 };
 877
 878 struct tu_userdata_info *
 879 tu_lookup_user_sgpr(struct tu_pipeline *pipeline,
 880                     gl_shader_stage stage,
 881                     int idx);
 882
 883 struct tu_shader_variant *
 884 tu_get_shader(struct tu_pipeline *pipeline, gl_shader_stage stage);
 885
 886 struct tu_graphics_pipeline_create_info
 887 {
 888    bool use_rectlist;
 889    bool db_depth_clear;
 890    bool db_stencil_clear;
 891    bool db_depth_disable_expclear;
 892    bool db_stencil_disable_expclear;
 893    bool db_flush_depth_inplace;
 894    bool db_flush_stencil_inplace;
 895    bool db_resummarize;
 896    uint32_t custom_blend_mode;
 897 };
 898
 899 VkResult
 900 tu_graphics_pipeline_create(
 901    VkDevice device,
 902    VkPipelineCache cache,
 903    const VkGraphicsPipelineCreateInfo *pCreateInfo,
 904    const struct tu_graphics_pipeline_create_info *extra,
 905    const VkAllocationCallbacks *alloc,
 906    VkPipeline *pPipeline);
 907
 908 struct vk_format_description;
 909 uint32_t
 910 tu_translate_buffer_dataformat(const struct vk_format_description *desc,
 911                                int first_non_void);
 912 uint32_t
 913 tu_translate_buffer_numformat(const struct vk_format_description *desc,
 914                               int first_non_void);
 915 uint32_t
 916 tu_translate_colorformat(VkFormat format);
 917 uint32_t
 918 tu_translate_color_numformat(VkFormat format,
 919                              const struct vk_format_description *desc,
 920                              int first_non_void);
 921 uint32_t
 922 tu_colorformat_endian_swap(uint32_t colorformat);
 923 unsigned
 924 tu_translate_colorswap(VkFormat format, bool do_endian_swap);
 925 uint32_t
 926 tu_translate_dbformat(VkFormat format);
 927 uint32_t
 928 tu_translate_tex_dataformat(VkFormat format,
 929                             const struct vk_format_description *desc,
 930                             int first_non_void);
 931 uint32_t
 932 tu_translate_tex_numformat(VkFormat format,
 933                            const struct vk_format_description *desc,
 934                            int first_non_void);
 935 bool
 936 tu_format_pack_clear_color(VkFormat format,
 937                            uint32_t clear_vals[2],
 938                            VkClearColorValue *value);
 939 bool
 940 tu_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
 941 bool
 942 tu_dcc_formats_compatible(VkFormat format1, VkFormat format2);
 943
 944 struct tu_image_level
 945 {
 946    VkDeviceSize offset;
 947    VkDeviceSize size;
 948    uint32_t pitch;
 949 };
 950
 951 struct tu_image
 952 {
 953    VkImageType type;
 954    /* The original VkFormat provided by the client.  This may not match any
 955     * of the actual surface formats.
 956     */
 957    VkFormat vk_format;
 958    VkImageAspectFlags aspects;
 959    VkImageUsageFlags usage;  /**< Superset of VkImageCreateInfo::usage. */
 960    VkImageTiling tiling;     /** VkImageCreateInfo::tiling */
 961    VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
 962    VkExtent3D extent;
 963    uint32_t level_count;
 964    uint32_t layer_count;
 965
 966    VkDeviceSize size;
 967    uint32_t alignment;
 968
 969    /* memory layout */
 970    VkDeviceSize layer_size;
 971    struct tu_image_level levels[15];
 972    unsigned tile_mode;
 973
 974    unsigned queue_family_mask;
 975    bool exclusive;
 976    bool shareable;
 977
 978    /* For VK_ANDROID_native_buffer, the WSI image owns the memory, */
 979    VkDeviceMemory owned_memory;
 980
 981    /* Set when bound */
 982    const struct tu_bo *bo;
 983    VkDeviceSize bo_offset;
 984 };
 985
 986 unsigned
 987 tu_image_queue_family_mask(const struct tu_image *image,
 988                            uint32_t family,
 989                            uint32_t queue_family);
 990
 991 static inline uint32_t
 992 tu_get_layerCount(const struct tu_image *image,
 993                   const VkImageSubresourceRange *range)
 994 {
 995    return range->layerCount == VK_REMAINING_ARRAY_LAYERS
 996              ? image->layer_count - range->baseArrayLayer
 997              : range->layerCount;
 998 }
 999
1000 static inline uint32_t
1001 tu_get_levelCount(const struct tu_image *image,
1002                   const VkImageSubresourceRange *range)
1003 {
1004    return range->levelCount == VK_REMAINING_MIP_LEVELS
1005              ? image->level_count - range->baseMipLevel
1006              : range->levelCount;
1007 }
1008
1009 struct tu_image_view
1010 {
1011    struct tu_image *image; /**< VkImageViewCreateInfo::image */
1012
1013    VkImageViewType type;
1014    VkImageAspectFlags aspect_mask;
1015    VkFormat vk_format;
1016    uint32_t base_layer;
1017    uint32_t layer_count;
1018    uint32_t base_mip;
1019    uint32_t level_count;
1020    VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
1021
1022    uint32_t descriptor[16];
1023
1024    /* Descriptor for use as a storage image as opposed to a sampled image.
1025     * This has a few differences for cube maps (e.g. type).
1026     */
1027    uint32_t storage_descriptor[16];
1028 };
1029
1030 struct tu_sampler
1031 {
1032 };
1033
1034 struct tu_image_create_info
1035 {
1036    const VkImageCreateInfo *vk_info;
1037    bool scanout;
1038    bool no_metadata_planes;
1039 };
1040
1041 VkResult
1042 tu_image_create(VkDevice _device,
1043                 const struct tu_image_create_info *info,
1044                 const VkAllocationCallbacks *alloc,
1045                 VkImage *pImage);
1046
1047 VkResult
1048 tu_image_from_gralloc(VkDevice device_h,
1049                       const VkImageCreateInfo *base_info,
1050                       const VkNativeBufferANDROID *gralloc_info,
1051                       const VkAllocationCallbacks *alloc,
1052                       VkImage *out_image_h);
1053
1054 void
1055 tu_image_view_init(struct tu_image_view *view,
1056                    struct tu_device *device,
1057                    const VkImageViewCreateInfo *pCreateInfo);
1058
1059 struct tu_buffer_view
1060 {
1061    VkFormat vk_format;
1062    uint64_t range; /**< VkBufferViewCreateInfo::range */
1063    uint32_t state[4];
1064 };
1065 void
1066 tu_buffer_view_init(struct tu_buffer_view *view,
1067                     struct tu_device *device,
1068                     const VkBufferViewCreateInfo *pCreateInfo);
1069
1070 static inline struct VkExtent3D
1071 tu_sanitize_image_extent(const VkImageType imageType,
1072                          const struct VkExtent3D imageExtent)
1073 {
1074    switch (imageType) {
1075    case VK_IMAGE_TYPE_1D:
1076       return (VkExtent3D) { imageExtent.width, 1, 1 };
1077    case VK_IMAGE_TYPE_2D:
1078       return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
1079    case VK_IMAGE_TYPE_3D:
1080       return imageExtent;
1081    default:
1082       unreachable("invalid image type");
1083    }
1084 }
1085
1086 static inline struct VkOffset3D
1087 tu_sanitize_image_offset(const VkImageType imageType,
1088                          const struct VkOffset3D imageOffset)
1089 {
1090    switch (imageType) {
1091    case VK_IMAGE_TYPE_1D:
1092       return (VkOffset3D) { imageOffset.x, 0, 0 };
1093    case VK_IMAGE_TYPE_2D:
1094       return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
1095    case VK_IMAGE_TYPE_3D:
1096       return imageOffset;
1097    default:
1098       unreachable("invalid image type");
1099    }
1100 }
1101
1102 struct tu_attachment_info
1103 {
1104    struct tu_image_view *attachment;
1105 };
1106
1107 struct tu_framebuffer
1108 {
1109    uint32_t width;
1110    uint32_t height;
1111    uint32_t layers;
1112
1113    uint32_t attachment_count;
1114    struct tu_attachment_info attachments[0];
1115 };
1116
1117 struct tu_subpass_barrier
1118 {
1119    VkPipelineStageFlags src_stage_mask;
1120    VkAccessFlags src_access_mask;
1121    VkAccessFlags dst_access_mask;
1122 };
1123
1124 void
1125 tu_subpass_barrier(struct tu_cmd_buffer *cmd_buffer,
1126                    const struct tu_subpass_barrier *barrier);
1127
1128 struct tu_subpass_attachment
1129 {
1130    uint32_t attachment;
1131    VkImageLayout layout;
1132 };
1133
1134 struct tu_subpass
1135 {
1136    uint32_t input_count;
1137    uint32_t color_count;
1138    struct tu_subpass_attachment *input_attachments;
1139    struct tu_subpass_attachment *color_attachments;
1140    struct tu_subpass_attachment *resolve_attachments;
1141    struct tu_subpass_attachment depth_stencil_attachment;
1142
1143    /** Subpass has at least one resolve attachment */
1144    bool has_resolve;
1145
1146    struct tu_subpass_barrier start_barrier;
1147
1148    uint32_t view_mask;
1149    VkSampleCountFlagBits max_sample_count;
1150 };
1151
1152 struct tu_render_pass_attachment
1153 {
1154    VkFormat format;
1155    uint32_t samples;
1156    VkAttachmentLoadOp load_op;
1157    VkAttachmentLoadOp stencil_load_op;
1158    VkImageLayout initial_layout;
1159    VkImageLayout final_layout;
1160    uint32_t view_mask;
1161 };
1162
1163 struct tu_render_pass
1164 {
1165    uint32_t attachment_count;
1166    uint32_t subpass_count;
1167    struct tu_subpass_attachment *subpass_attachments;
1168    struct tu_render_pass_attachment *attachments;
1169    struct tu_subpass_barrier end_barrier;
1170    struct tu_subpass subpasses[0];
1171 };
1172
1173 VkResult
1174 tu_device_init_meta(struct tu_device *device);
1175 void
1176 tu_device_finish_meta(struct tu_device *device);
1177
1178 struct tu_query_pool
1179 {
1180    uint32_t stride;
1181    uint32_t availability_offset;
1182    uint64_t size;
1183    char *ptr;
1184    VkQueryType type;
1185    uint32_t pipeline_stats_mask;
1186 };
1187
1188 struct tu_semaphore
1189 {
1190    uint32_t syncobj;
1191    uint32_t temp_syncobj;
1192 };
1193
1194 void
1195 tu_set_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1196                       VkPipelineBindPoint bind_point,
1197                       struct tu_descriptor_set *set,
1198                       unsigned idx);
1199
1200 void
1201 tu_update_descriptor_sets(struct tu_device *device,
1202                           struct tu_cmd_buffer *cmd_buffer,
1203                           VkDescriptorSet overrideSet,
1204                           uint32_t descriptorWriteCount,
1205                           const VkWriteDescriptorSet *pDescriptorWrites,
1206                           uint32_t descriptorCopyCount,
1207                           const VkCopyDescriptorSet *pDescriptorCopies);
1208
1209 void
1210 tu_update_descriptor_set_with_template(
1211    struct tu_device *device,
1212    struct tu_cmd_buffer *cmd_buffer,
1213    struct tu_descriptor_set *set,
1214    VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1215    const void *pData);
1216
1217 void
1218 tu_meta_push_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1219                             VkPipelineBindPoint pipelineBindPoint,
1220                             VkPipelineLayout _layout,
1221                             uint32_t set,
1222                             uint32_t descriptorWriteCount,
1223                             const VkWriteDescriptorSet *pDescriptorWrites);
1224
1225 struct tu_fence
1226 {
1227    uint32_t syncobj;
1228    uint32_t temp_syncobj;
1229 };
1230
1231 int
1232 tu_drm_get_gpu_id(const struct tu_physical_device *dev, uint32_t *id);
1233
1234 int
1235 tu_drm_get_gmem_size(const struct tu_physical_device *dev, uint32_t *size);
1236
1237 int
1238 tu_drm_submitqueue_new(const struct tu_device *dev,
1239                        int priority,
1240                        uint32_t *queue_id);
1241
1242 void
1243 tu_drm_submitqueue_close(const struct tu_device *dev, uint32_t queue_id);
1244
1245 uint32_t
1246 tu_gem_new(const struct tu_device *dev, uint64_t size, uint32_t flags);
1247 void
1248 tu_gem_close(const struct tu_device *dev, uint32_t gem_handle);
1249 uint64_t
1250 tu_gem_info_offset(const struct tu_device *dev, uint32_t gem_handle);
1251 uint64_t
1252 tu_gem_info_iova(const struct tu_device *dev, uint32_t gem_handle);
1253
1254 #define TU_DEFINE_HANDLE_CASTS(__tu_type, __VkType)                          \
1255                                                                              \
1256    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1257    {                                                                         \
1258       return (struct __tu_type *) _handle;                                   \
1259    }                                                                         \
1260                                                                              \
1261    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1262    {                                                                         \
1263       return (__VkType) _obj;                                                \
1264    }
1265
1266 #define TU_DEFINE_NONDISP_HANDLE_CASTS(__tu_type, __VkType)                  \
1267                                                                              \
1268    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1269    {                                                                         \
1270       return (struct __tu_type *) (uintptr_t) _handle;                       \
1271    }                                                                         \
1272                                                                              \
1273    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1274    {                                                                         \
1275       return (__VkType)(uintptr_t) _obj;                                     \
1276    }
1277
1278 #define TU_FROM_HANDLE(__tu_type, __name, __handle)                          \
1279    struct __tu_type *__name = __tu_type##_from_handle(__handle)
1280
1281 TU_DEFINE_HANDLE_CASTS(tu_cmd_buffer, VkCommandBuffer)
1282 TU_DEFINE_HANDLE_CASTS(tu_device, VkDevice)
1283 TU_DEFINE_HANDLE_CASTS(tu_instance, VkInstance)
1284 TU_DEFINE_HANDLE_CASTS(tu_physical_device, VkPhysicalDevice)
1285 TU_DEFINE_HANDLE_CASTS(tu_queue, VkQueue)
1286
1287 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_cmd_pool, VkCommandPool)
1288 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer, VkBuffer)
1289 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer_view, VkBufferView)
1290 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_pool, VkDescriptorPool)
1291 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set, VkDescriptorSet)
1292 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set_layout,
1293                                VkDescriptorSetLayout)
1294 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_update_template,
1295                                VkDescriptorUpdateTemplateKHR)
1296 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_device_memory, VkDeviceMemory)
1297 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_fence, VkFence)
1298 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_event, VkEvent)
1299 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_framebuffer, VkFramebuffer)
1300 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image, VkImage)
1301 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image_view, VkImageView);
1302 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_cache, VkPipelineCache)
1303 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline, VkPipeline)
1304 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_layout, VkPipelineLayout)
1305 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_query_pool, VkQueryPool)
1306 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_render_pass, VkRenderPass)
1307 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_sampler, VkSampler)
1308 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_shader_module, VkShaderModule)
1309 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_semaphore, VkSemaphore)
1310
1311 #endif /* TU_PRIVATE_H */