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
 406 struct tu_device
 407 {
 408    VK_LOADER_DATA _loader_data;
 409
 410    VkAllocationCallbacks alloc;
 411
 412    struct tu_instance *instance;
 413
 414    struct tu_meta_state meta_state;
 415
 416    struct tu_queue *queues[TU_MAX_QUEUE_FAMILIES];
 417    int queue_count[TU_MAX_QUEUE_FAMILIES];
 418
 419    struct tu_physical_device *physical_device;
 420
 421    /* Backup in-memory cache to be used if the app doesn't provide one */
 422    struct tu_pipeline_cache *mem_cache;
 423
 424    struct list_head shader_slabs;
 425    mtx_t shader_slab_mutex;
 426
 427    struct tu_device_extension_table enabled_extensions;
 428 };
 429
 430 struct tu_bo
 431 {
 432    uint32_t gem_handle;
 433    uint64_t size;
 434    uint64_t offset;
 435    uint64_t iova;
 436    void *map;
 437 };
 438
 439 VkResult
 440 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size);
 441 void
 442 tu_bo_finish(struct tu_device *dev, struct tu_bo *bo);
 443 VkResult
 444 tu_bo_map(struct tu_device *dev, struct tu_bo *bo);
 445
 446 struct tu_device_memory
 447 {
 448    struct tu_bo bo;
 449    VkDeviceSize size;
 450
 451    /* for dedicated allocations */
 452    struct tu_image *image;
 453    struct tu_buffer *buffer;
 454
 455    uint32_t type_index;
 456    void *map;
 457    void *user_ptr;
 458 };
 459
 460 struct tu_descriptor_range
 461 {
 462    uint64_t va;
 463    uint32_t size;
 464 };
 465
 466 struct tu_descriptor_set
 467 {
 468    const struct tu_descriptor_set_layout *layout;
 469    uint32_t size;
 470
 471    uint64_t va;
 472    uint32_t *mapped_ptr;
 473    struct tu_descriptor_range *dynamic_descriptors;
 474 };
 475
 476 struct tu_push_descriptor_set
 477 {
 478    struct tu_descriptor_set set;
 479    uint32_t capacity;
 480 };
 481
 482 struct tu_descriptor_pool_entry
 483 {
 484    uint32_t offset;
 485    uint32_t size;
 486    struct tu_descriptor_set *set;
 487 };
 488
 489 struct tu_descriptor_pool
 490 {
 491    uint8_t *mapped_ptr;
 492    uint64_t current_offset;
 493    uint64_t size;
 494
 495    uint8_t *host_memory_base;
 496    uint8_t *host_memory_ptr;
 497    uint8_t *host_memory_end;
 498
 499    uint32_t entry_count;
 500    uint32_t max_entry_count;
 501    struct tu_descriptor_pool_entry entries[0];
 502 };
 503
 504 struct tu_descriptor_update_template_entry
 505 {
 506    VkDescriptorType descriptor_type;
 507
 508    /* The number of descriptors to update */
 509    uint32_t descriptor_count;
 510
 511    /* Into mapped_ptr or dynamic_descriptors, in units of the respective array
 512     */
 513    uint32_t dst_offset;
 514
 515    /* In dwords. Not valid/used for dynamic descriptors */
 516    uint32_t dst_stride;
 517
 518    uint32_t buffer_offset;
 519
 520    /* Only valid for combined image samplers and samplers */
 521    uint16_t has_sampler;
 522
 523    /* In bytes */
 524    size_t src_offset;
 525    size_t src_stride;
 526
 527    /* For push descriptors */
 528    const uint32_t *immutable_samplers;
 529 };
 530
 531 struct tu_descriptor_update_template
 532 {
 533    uint32_t entry_count;
 534    VkPipelineBindPoint bind_point;
 535    struct tu_descriptor_update_template_entry entry[0];
 536 };
 537
 538 struct tu_buffer
 539 {
 540    VkDeviceSize size;
 541
 542    VkBufferUsageFlags usage;
 543    VkBufferCreateFlags flags;
 544 };
 545
 546 enum tu_dynamic_state_bits
 547 {
 548    TU_DYNAMIC_VIEWPORT = 1 << 0,
 549    TU_DYNAMIC_SCISSOR = 1 << 1,
 550    TU_DYNAMIC_LINE_WIDTH = 1 << 2,
 551    TU_DYNAMIC_DEPTH_BIAS = 1 << 3,
 552    TU_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
 553    TU_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
 554    TU_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
 555    TU_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
 556    TU_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
 557    TU_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
 558    TU_DYNAMIC_ALL = (1 << 10) - 1,
 559 };
 560
 561 struct tu_vertex_binding
 562 {
 563    struct tu_buffer *buffer;
 564    VkDeviceSize offset;
 565 };
 566
 567 struct tu_viewport_state
 568 {
 569    uint32_t count;
 570    VkViewport viewports[MAX_VIEWPORTS];
 571 };
 572
 573 struct tu_scissor_state
 574 {
 575    uint32_t count;
 576    VkRect2D scissors[MAX_SCISSORS];
 577 };
 578
 579 struct tu_discard_rectangle_state
 580 {
 581    uint32_t count;
 582    VkRect2D rectangles[MAX_DISCARD_RECTANGLES];
 583 };
 584
 585 struct tu_dynamic_state
 586 {
 587    /**
 588     * Bitmask of (1 << VK_DYNAMIC_STATE_*).
 589     * Defines the set of saved dynamic state.
 590     */
 591    uint32_t mask;
 592
 593    struct tu_viewport_state viewport;
 594
 595    struct tu_scissor_state scissor;
 596
 597    float line_width;
 598
 599    struct
 600    {
 601       float bias;
 602       float clamp;
 603       float slope;
 604    } depth_bias;
 605
 606    float blend_constants[4];
 607
 608    struct
 609    {
 610       float min;
 611       float max;
 612    } depth_bounds;
 613
 614    struct
 615    {
 616       uint32_t front;
 617       uint32_t back;
 618    } stencil_compare_mask;
 619
 620    struct
 621    {
 622       uint32_t front;
 623       uint32_t back;
 624    } stencil_write_mask;
 625
 626    struct
 627    {
 628       uint32_t front;
 629       uint32_t back;
 630    } stencil_reference;
 631
 632    struct tu_discard_rectangle_state discard_rectangle;
 633 };
 634
 635 extern const struct tu_dynamic_state default_dynamic_state;
 636
 637 const char *
 638 tu_get_debug_option_name(int id);
 639
 640 const char *
 641 tu_get_perftest_option_name(int id);
 642
 643 /**
 644  * Attachment state when recording a renderpass instance.
 645  *
 646  * The clear value is valid only if there exists a pending clear.
 647  */
 648 struct tu_attachment_state
 649 {
 650    VkImageAspectFlags pending_clear_aspects;
 651    uint32_t cleared_views;
 652    VkClearValue clear_value;
 653    VkImageLayout current_layout;
 654 };
 655
 656 struct tu_descriptor_state
 657 {
 658    struct tu_descriptor_set *sets[MAX_SETS];
 659    uint32_t dirty;
 660    uint32_t valid;
 661    struct tu_push_descriptor_set push_set;
 662    bool push_dirty;
 663    uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
 664 };
 665
 666 struct tu_cmd_state
 667 {
 668    /* Vertex descriptors */
 669    uint64_t vb_va;
 670    unsigned vb_size;
 671
 672    struct tu_dynamic_state dynamic;
 673
 674    /* Index buffer */
 675    struct tu_buffer *index_buffer;
 676    uint64_t index_offset;
 677    uint32_t index_type;
 678    uint32_t max_index_count;
 679    uint64_t index_va;
 680 };
 681
 682 struct tu_cmd_pool
 683 {
 684    VkAllocationCallbacks alloc;
 685    struct list_head cmd_buffers;
 686    struct list_head free_cmd_buffers;
 687    uint32_t queue_family_index;
 688 };
 689
 690 struct tu_cmd_buffer_upload
 691 {
 692    uint8_t *map;
 693    unsigned offset;
 694    uint64_t size;
 695    struct list_head list;
 696 };
 697
 698 enum tu_cmd_buffer_status
 699 {
 700    TU_CMD_BUFFER_STATUS_INVALID,
 701    TU_CMD_BUFFER_STATUS_INITIAL,
 702    TU_CMD_BUFFER_STATUS_RECORDING,
 703    TU_CMD_BUFFER_STATUS_EXECUTABLE,
 704    TU_CMD_BUFFER_STATUS_PENDING,
 705 };
 706
 707 struct tu_bo_list
 708 {
 709    uint32_t count;
 710    uint32_t capacity;
 711    uint32_t *handles;
 712 };
 713
 714 struct tu_cmd_buffer
 715 {
 716    VK_LOADER_DATA _loader_data;
 717
 718    struct tu_device *device;
 719
 720    struct tu_cmd_pool *pool;
 721    struct list_head pool_link;
 722
 723    VkCommandBufferUsageFlags usage_flags;
 724    VkCommandBufferLevel level;
 725    enum tu_cmd_buffer_status status;
 726
 727    struct tu_cmd_state state;
 728    struct tu_vertex_binding vertex_bindings[MAX_VBS];
 729    uint32_t queue_family_index;
 730
 731    uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
 732    VkShaderStageFlags push_constant_stages;
 733    struct tu_descriptor_set meta_push_descriptors;
 734
 735    struct tu_descriptor_state descriptors[VK_PIPELINE_BIND_POINT_RANGE_SIZE];
 736
 737    struct tu_cmd_buffer_upload upload;
 738
 739    struct tu_bo_list bo_list;
 740
 741    VkResult record_result;
 742 };
 743
 744 bool
 745 tu_get_memory_fd(struct tu_device *device,
 746                  struct tu_device_memory *memory,
 747                  int *pFD);
 748
 749 /*
 750  * Takes x,y,z as exact numbers of invocations, instead of blocks.
 751  *
 752  * Limitations: Can't call normal dispatch functions without binding or
 753  * rebinding
 754  *              the compute pipeline.
 755  */
 756 void
 757 tu_unaligned_dispatch(struct tu_cmd_buffer *cmd_buffer,
 758                       uint32_t x,
 759                       uint32_t y,
 760                       uint32_t z);
 761
 762 struct tu_event
 763 {
 764    uint64_t *map;
 765 };
 766
 767 struct tu_shader_module;
 768
 769 #define TU_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
 770 #define TU_HASH_SHADER_SISCHED (1 << 1)
 771 #define TU_HASH_SHADER_UNSAFE_MATH (1 << 2)
 772 void
 773 tu_hash_shaders(unsigned char *hash,
 774                 const VkPipelineShaderStageCreateInfo **stages,
 775                 const struct tu_pipeline_layout *layout,
 776                 const struct tu_pipeline_key *key,
 777                 uint32_t flags);
 778
 779 static inline gl_shader_stage
 780 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
 781 {
 782    assert(__builtin_popcount(vk_stage) == 1);
 783    return ffs(vk_stage) - 1;
 784 }
 785
 786 static inline VkShaderStageFlagBits
 787 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
 788 {
 789    return (1 << mesa_stage);
 790 }
 791
 792 #define TU_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
 793
 794 #define tu_foreach_stage(stage, stage_bits)                                  \
 795    for (gl_shader_stage stage,                                               \
 796         __tmp = (gl_shader_stage)((stage_bits) &TU_STAGE_MASK);              \
 797         stage = __builtin_ffs(__tmp) - 1, __tmp; __tmp &= ~(1 << (stage)))
 798
 799 struct tu_shader_module
 800 {
 801    struct nir_shader *nir;
 802    unsigned char sha1[20];
 803    uint32_t size;
 804    char data[0];
 805 };
 806
 807 struct tu_pipeline
 808 {
 809    struct tu_device *device;
 810    struct tu_dynamic_state dynamic_state;
 811
 812    struct tu_pipeline_layout *layout;
 813
 814    bool need_indirect_descriptor_sets;
 815    VkShaderStageFlags active_stages;
 816 };
 817
 818 struct tu_userdata_info *
 819 tu_lookup_user_sgpr(struct tu_pipeline *pipeline,
 820                     gl_shader_stage stage,
 821                     int idx);
 822
 823 struct tu_shader_variant *
 824 tu_get_shader(struct tu_pipeline *pipeline, gl_shader_stage stage);
 825
 826 struct tu_graphics_pipeline_create_info
 827 {
 828    bool use_rectlist;
 829    bool db_depth_clear;
 830    bool db_stencil_clear;
 831    bool db_depth_disable_expclear;
 832    bool db_stencil_disable_expclear;
 833    bool db_flush_depth_inplace;
 834    bool db_flush_stencil_inplace;
 835    bool db_resummarize;
 836    uint32_t custom_blend_mode;
 837 };
 838
 839 VkResult
 840 tu_graphics_pipeline_create(
 841    VkDevice device,
 842    VkPipelineCache cache,
 843    const VkGraphicsPipelineCreateInfo *pCreateInfo,
 844    const struct tu_graphics_pipeline_create_info *extra,
 845    const VkAllocationCallbacks *alloc,
 846    VkPipeline *pPipeline);
 847
 848 struct vk_format_description;
 849 uint32_t
 850 tu_translate_buffer_dataformat(const struct vk_format_description *desc,
 851                                int first_non_void);
 852 uint32_t
 853 tu_translate_buffer_numformat(const struct vk_format_description *desc,
 854                               int first_non_void);
 855 uint32_t
 856 tu_translate_colorformat(VkFormat format);
 857 uint32_t
 858 tu_translate_color_numformat(VkFormat format,
 859                              const struct vk_format_description *desc,
 860                              int first_non_void);
 861 uint32_t
 862 tu_colorformat_endian_swap(uint32_t colorformat);
 863 unsigned
 864 tu_translate_colorswap(VkFormat format, bool do_endian_swap);
 865 uint32_t
 866 tu_translate_dbformat(VkFormat format);
 867 uint32_t
 868 tu_translate_tex_dataformat(VkFormat format,
 869                             const struct vk_format_description *desc,
 870                             int first_non_void);
 871 uint32_t
 872 tu_translate_tex_numformat(VkFormat format,
 873                            const struct vk_format_description *desc,
 874                            int first_non_void);
 875 bool
 876 tu_format_pack_clear_color(VkFormat format,
 877                            uint32_t clear_vals[2],
 878                            VkClearColorValue *value);
 879 bool
 880 tu_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
 881 bool
 882 tu_dcc_formats_compatible(VkFormat format1, VkFormat format2);
 883
 884 struct tu_image_level
 885 {
 886    VkDeviceSize offset;
 887    VkDeviceSize size;
 888    uint32_t pitch;
 889 };
 890
 891 struct tu_image
 892 {
 893    VkImageType type;
 894    /* The original VkFormat provided by the client.  This may not match any
 895     * of the actual surface formats.
 896     */
 897    VkFormat vk_format;
 898    VkImageAspectFlags aspects;
 899    VkImageUsageFlags usage;  /**< Superset of VkImageCreateInfo::usage. */
 900    VkImageTiling tiling;     /** VkImageCreateInfo::tiling */
 901    VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
 902    VkExtent3D extent;
 903
 904    VkDeviceSize size;
 905    uint32_t alignment;
 906
 907    /* memory layout */
 908    VkDeviceSize layer_size;
 909    struct tu_image_level levels[15];
 910    unsigned tile_mode;
 911
 912    unsigned queue_family_mask;
 913    bool exclusive;
 914    bool shareable;
 915
 916    /* For VK_ANDROID_native_buffer, the WSI image owns the memory, */
 917    VkDeviceMemory owned_memory;
 918 };
 919
 920 unsigned
 921 tu_image_queue_family_mask(const struct tu_image *image,
 922                            uint32_t family,
 923                            uint32_t queue_family);
 924
 925 static inline uint32_t
 926 tu_get_layerCount(const struct tu_image *image,
 927                   const VkImageSubresourceRange *range)
 928 {
 929    abort();
 930 }
 931
 932 static inline uint32_t
 933 tu_get_levelCount(const struct tu_image *image,
 934                   const VkImageSubresourceRange *range)
 935 {
 936    abort();
 937 }
 938
 939 struct tu_image_view
 940 {
 941    struct tu_image *image; /**< VkImageViewCreateInfo::image */
 942
 943    VkImageViewType type;
 944    VkImageAspectFlags aspect_mask;
 945    VkFormat vk_format;
 946    uint32_t base_layer;
 947    uint32_t layer_count;
 948    uint32_t base_mip;
 949    uint32_t level_count;
 950    VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
 951
 952    uint32_t descriptor[16];
 953
 954    /* Descriptor for use as a storage image as opposed to a sampled image.
 955     * This has a few differences for cube maps (e.g. type).
 956     */
 957    uint32_t storage_descriptor[16];
 958 };
 959
 960 struct tu_sampler
 961 {
 962 };
 963
 964 struct tu_image_create_info
 965 {
 966    const VkImageCreateInfo *vk_info;
 967    bool scanout;
 968    bool no_metadata_planes;
 969 };
 970
 971 VkResult
 972 tu_image_create(VkDevice _device,
 973                 const struct tu_image_create_info *info,
 974                 const VkAllocationCallbacks *alloc,
 975                 VkImage *pImage);
 976
 977 VkResult
 978 tu_image_from_gralloc(VkDevice device_h,
 979                       const VkImageCreateInfo *base_info,
 980                       const VkNativeBufferANDROID *gralloc_info,
 981                       const VkAllocationCallbacks *alloc,
 982                       VkImage *out_image_h);
 983
 984 void
 985 tu_image_view_init(struct tu_image_view *view,
 986                    struct tu_device *device,
 987                    const VkImageViewCreateInfo *pCreateInfo);
 988
 989 struct tu_buffer_view
 990 {
 991    VkFormat vk_format;
 992    uint64_t range; /**< VkBufferViewCreateInfo::range */
 993    uint32_t state[4];
 994 };
 995 void
 996 tu_buffer_view_init(struct tu_buffer_view *view,
 997                     struct tu_device *device,
 998                     const VkBufferViewCreateInfo *pCreateInfo);
 999
1000 static inline struct VkExtent3D
1001 tu_sanitize_image_extent(const VkImageType imageType,
1002                          const struct VkExtent3D imageExtent)
1003 {
1004    switch (imageType) {
1005    case VK_IMAGE_TYPE_1D:
1006       return (VkExtent3D) { imageExtent.width, 1, 1 };
1007    case VK_IMAGE_TYPE_2D:
1008       return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
1009    case VK_IMAGE_TYPE_3D:
1010       return imageExtent;
1011    default:
1012       unreachable("invalid image type");
1013    }
1014 }
1015
1016 static inline struct VkOffset3D
1017 tu_sanitize_image_offset(const VkImageType imageType,
1018                          const struct VkOffset3D imageOffset)
1019 {
1020    switch (imageType) {
1021    case VK_IMAGE_TYPE_1D:
1022       return (VkOffset3D) { imageOffset.x, 0, 0 };
1023    case VK_IMAGE_TYPE_2D:
1024       return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
1025    case VK_IMAGE_TYPE_3D:
1026       return imageOffset;
1027    default:
1028       unreachable("invalid image type");
1029    }
1030 }
1031
1032 struct tu_attachment_info
1033 {
1034    struct tu_image_view *attachment;
1035 };
1036
1037 struct tu_framebuffer
1038 {
1039    uint32_t width;
1040    uint32_t height;
1041    uint32_t layers;
1042
1043    uint32_t attachment_count;
1044    struct tu_attachment_info attachments[0];
1045 };
1046
1047 struct tu_subpass_barrier
1048 {
1049    VkPipelineStageFlags src_stage_mask;
1050    VkAccessFlags src_access_mask;
1051    VkAccessFlags dst_access_mask;
1052 };
1053
1054 void
1055 tu_subpass_barrier(struct tu_cmd_buffer *cmd_buffer,
1056                    const struct tu_subpass_barrier *barrier);
1057
1058 struct tu_subpass_attachment
1059 {
1060    uint32_t attachment;
1061    VkImageLayout layout;
1062 };
1063
1064 struct tu_subpass
1065 {
1066    uint32_t input_count;
1067    uint32_t color_count;
1068    struct tu_subpass_attachment *input_attachments;
1069    struct tu_subpass_attachment *color_attachments;
1070    struct tu_subpass_attachment *resolve_attachments;
1071    struct tu_subpass_attachment depth_stencil_attachment;
1072
1073    /** Subpass has at least one resolve attachment */
1074    bool has_resolve;
1075
1076    struct tu_subpass_barrier start_barrier;
1077
1078    uint32_t view_mask;
1079    VkSampleCountFlagBits max_sample_count;
1080 };
1081
1082 struct tu_render_pass_attachment
1083 {
1084    VkFormat format;
1085    uint32_t samples;
1086    VkAttachmentLoadOp load_op;
1087    VkAttachmentLoadOp stencil_load_op;
1088    VkImageLayout initial_layout;
1089    VkImageLayout final_layout;
1090    uint32_t view_mask;
1091 };
1092
1093 struct tu_render_pass
1094 {
1095    uint32_t attachment_count;
1096    uint32_t subpass_count;
1097    struct tu_subpass_attachment *subpass_attachments;
1098    struct tu_render_pass_attachment *attachments;
1099    struct tu_subpass_barrier end_barrier;
1100    struct tu_subpass subpasses[0];
1101 };
1102
1103 VkResult
1104 tu_device_init_meta(struct tu_device *device);
1105 void
1106 tu_device_finish_meta(struct tu_device *device);
1107
1108 struct tu_query_pool
1109 {
1110    uint32_t stride;
1111    uint32_t availability_offset;
1112    uint64_t size;
1113    char *ptr;
1114    VkQueryType type;
1115    uint32_t pipeline_stats_mask;
1116 };
1117
1118 struct tu_semaphore
1119 {
1120    uint32_t syncobj;
1121    uint32_t temp_syncobj;
1122 };
1123
1124 void
1125 tu_set_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1126                       VkPipelineBindPoint bind_point,
1127                       struct tu_descriptor_set *set,
1128                       unsigned idx);
1129
1130 void
1131 tu_update_descriptor_sets(struct tu_device *device,
1132                           struct tu_cmd_buffer *cmd_buffer,
1133                           VkDescriptorSet overrideSet,
1134                           uint32_t descriptorWriteCount,
1135                           const VkWriteDescriptorSet *pDescriptorWrites,
1136                           uint32_t descriptorCopyCount,
1137                           const VkCopyDescriptorSet *pDescriptorCopies);
1138
1139 void
1140 tu_update_descriptor_set_with_template(
1141    struct tu_device *device,
1142    struct tu_cmd_buffer *cmd_buffer,
1143    struct tu_descriptor_set *set,
1144    VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1145    const void *pData);
1146
1147 void
1148 tu_meta_push_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1149                             VkPipelineBindPoint pipelineBindPoint,
1150                             VkPipelineLayout _layout,
1151                             uint32_t set,
1152                             uint32_t descriptorWriteCount,
1153                             const VkWriteDescriptorSet *pDescriptorWrites);
1154
1155 struct tu_fence
1156 {
1157    uint32_t syncobj;
1158    uint32_t temp_syncobj;
1159 };
1160
1161 uint32_t
1162 tu_gem_new(struct tu_device *dev, uint64_t size, uint32_t flags);
1163 void
1164 tu_gem_close(struct tu_device *dev, uint32_t gem_handle);
1165 uint64_t
1166 tu_gem_info_offset(struct tu_device *dev, uint32_t gem_handle);
1167 uint64_t
1168 tu_gem_info_iova(struct tu_device *dev, uint32_t gem_handle);
1169 int
1170 tu_drm_query_param(struct tu_physical_device *dev,
1171                    uint32_t param,
1172                    uint64_t *value);
1173
1174 #define TU_DEFINE_HANDLE_CASTS(__tu_type, __VkType)                          \
1175                                                                              \
1176    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1177    {                                                                         \
1178       return (struct __tu_type *) _handle;                                   \
1179    }                                                                         \
1180                                                                              \
1181    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1182    {                                                                         \
1183       return (__VkType) _obj;                                                \
1184    }
1185
1186 #define TU_DEFINE_NONDISP_HANDLE_CASTS(__tu_type, __VkType)                  \
1187                                                                              \
1188    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1189    {                                                                         \
1190       return (struct __tu_type *) (uintptr_t) _handle;                       \
1191    }                                                                         \
1192                                                                              \
1193    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1194    {                                                                         \
1195       return (__VkType)(uintptr_t) _obj;                                     \
1196    }
1197
1198 #define TU_FROM_HANDLE(__tu_type, __name, __handle)                          \
1199    struct __tu_type *__name = __tu_type##_from_handle(__handle)
1200
1201 TU_DEFINE_HANDLE_CASTS(tu_cmd_buffer, VkCommandBuffer)
1202 TU_DEFINE_HANDLE_CASTS(tu_device, VkDevice)
1203 TU_DEFINE_HANDLE_CASTS(tu_instance, VkInstance)
1204 TU_DEFINE_HANDLE_CASTS(tu_physical_device, VkPhysicalDevice)
1205 TU_DEFINE_HANDLE_CASTS(tu_queue, VkQueue)
1206
1207 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_cmd_pool, VkCommandPool)
1208 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer, VkBuffer)
1209 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer_view, VkBufferView)
1210 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_pool, VkDescriptorPool)
1211 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set, VkDescriptorSet)
1212 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set_layout,
1213                                VkDescriptorSetLayout)
1214 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_update_template,
1215                                VkDescriptorUpdateTemplateKHR)
1216 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_device_memory, VkDeviceMemory)
1217 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_fence, VkFence)
1218 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_event, VkEvent)
1219 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_framebuffer, VkFramebuffer)
1220 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image, VkImage)
1221 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image_view, VkImageView);
1222 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_cache, VkPipelineCache)
1223 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline, VkPipeline)
1224 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_layout, VkPipelineLayout)
1225 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_query_pool, VkQueryPool)
1226 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_render_pass, VkRenderPass)
1227 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_sampler, VkSampler)
1228 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_shader_module, VkShaderModule)
1229 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_semaphore, VkSemaphore)
1230
1231 #endif /* TU_PRIVATE_H */