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 DEALINGS
  25  * 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
  71 #include "tu_entrypoints.h"
  72
  73 #define MAX_VBS 32
  74 #define MAX_VERTEX_ATTRIBS 32
  75 #define MAX_RTS 8
  76 #define MAX_VIEWPORTS 16
  77 #define MAX_SCISSORS 16
  78 #define MAX_DISCARD_RECTANGLES 4
  79 #define MAX_PUSH_CONSTANTS_SIZE 128
  80 #define MAX_PUSH_DESCRIPTORS 32
  81 #define MAX_DYNAMIC_UNIFORM_BUFFERS 16
  82 #define MAX_DYNAMIC_STORAGE_BUFFERS 8
  83 #define MAX_DYNAMIC_BUFFERS                                                    \
  84    (MAX_DYNAMIC_UNIFORM_BUFFERS + MAX_DYNAMIC_STORAGE_BUFFERS)
  85 #define MAX_SAMPLES_LOG2 4
  86 #define NUM_META_FS_KEYS 13
  87 #define TU_MAX_DRM_DEVICES 8
  88 #define MAX_VIEWS 8
  89
  90 #define NUM_DEPTH_CLEAR_PIPELINES 3
  91
  92 /*
  93  * This is the point we switch from using CP to compute shader
  94  * for certain buffer operations.
  95  */
  96 #define TU_BUFFER_OPS_CS_THRESHOLD 4096
  97
  98 enum tu_mem_heap
  99 {
 100    TU_MEM_HEAP_VRAM,
 101    TU_MEM_HEAP_VRAM_CPU_ACCESS,
 102    TU_MEM_HEAP_GTT,
 103    TU_MEM_HEAP_COUNT
 104 };
 105
 106 enum tu_mem_type
 107 {
 108    TU_MEM_TYPE_VRAM,
 109    TU_MEM_TYPE_GTT_WRITE_COMBINE,
 110    TU_MEM_TYPE_VRAM_CPU_ACCESS,
 111    TU_MEM_TYPE_GTT_CACHED,
 112    TU_MEM_TYPE_COUNT
 113 };
 114
 115 #define tu_printflike(a, b) __attribute__((__format__(__printf__, a, b)))
 116
 117 static inline uint32_t
 118 align_u32(uint32_t v, uint32_t a)
 119 {
 120    assert(a != 0 && a == (a & -a));
 121    return (v + a - 1) & ~(a - 1);
 122 }
 123
 124 static inline uint32_t
 125 align_u32_npot(uint32_t v, uint32_t a)
 126 {
 127    return (v + a - 1) / a * a;
 128 }
 129
 130 static inline uint64_t
 131 align_u64(uint64_t v, uint64_t a)
 132 {
 133    assert(a != 0 && a == (a & -a));
 134    return (v + a - 1) & ~(a - 1);
 135 }
 136
 137 static inline int32_t
 138 align_i32(int32_t v, int32_t a)
 139 {
 140    assert(a != 0 && a == (a & -a));
 141    return (v + a - 1) & ~(a - 1);
 142 }
 143
 144 /** Alignment must be a power of 2. */
 145 static inline bool
 146 tu_is_aligned(uintmax_t n, uintmax_t a)
 147 {
 148    assert(a == (a & -a));
 149    return (n & (a - 1)) == 0;
 150 }
 151
 152 static inline uint32_t
 153 round_up_u32(uint32_t v, uint32_t a)
 154 {
 155    return (v + a - 1) / a;
 156 }
 157
 158 static inline uint64_t
 159 round_up_u64(uint64_t v, uint64_t a)
 160 {
 161    return (v + a - 1) / a;
 162 }
 163
 164 static inline uint32_t
 165 tu_minify(uint32_t n, uint32_t levels)
 166 {
 167    if (unlikely(n == 0))
 168       return 0;
 169    else
 170       return MAX2(n >> levels, 1);
 171 }
 172 static inline float
 173 tu_clamp_f(float f, float min, float max)
 174 {
 175    assert(min < max);
 176
 177    if (f > max)
 178       return max;
 179    else if (f < min)
 180       return min;
 181    else
 182       return f;
 183 }
 184
 185 static inline bool
 186 tu_clear_mask(uint32_t *inout_mask, uint32_t clear_mask)
 187 {
 188    if (*inout_mask & clear_mask) {
 189       *inout_mask &= ~clear_mask;
 190       return true;
 191    } else {
 192       return false;
 193    }
 194 }
 195
 196 #define for_each_bit(b, dword)                                                 \
 197    for (uint32_t __dword = (dword); (b) = __builtin_ffs(__dword) - 1, __dword; \
 198         __dword &= ~(1 << (b)))
 199
 200 #define typed_memcpy(dest, src, count)                                         \
 201    ({                                                                          \
 202       STATIC_ASSERT(sizeof(*src) == sizeof(*dest));                            \
 203       memcpy((dest), (src), (count) * sizeof(*(src)));                         \
 204    })
 205
 206 /* Whenever we generate an error, pass it through this function. Useful for
 207  * debugging, where we can break on it. Only call at error site, not when
 208  * propagating errors. Might be useful to plug in a stack trace here.
 209  */
 210
 211 struct tu_instance;
 212
 213 VkResult
 214 __vk_errorf(struct tu_instance *instance,
 215             VkResult error,
 216             const char *file,
 217             int line,
 218             const char *format,
 219             ...);
 220
 221 #define vk_error(instance, error)                                              \
 222    __vk_errorf(instance, error, __FILE__, __LINE__, NULL);
 223 #define vk_errorf(instance, error, format, ...)                                \
 224    __vk_errorf(instance, error, __FILE__, __LINE__, format, ##__VA_ARGS__);
 225
 226 void
 227 __tu_finishme(const char *file, int line, const char *format, ...)
 228   tu_printflike(3, 4);
 229 void
 230 tu_loge(const char *format, ...) tu_printflike(1, 2);
 231 void
 232 tu_loge_v(const char *format, va_list va);
 233 void
 234 tu_logi(const char *format, ...) tu_printflike(1, 2);
 235 void
 236 tu_logi_v(const char *format, va_list va);
 237
 238 /**
 239  * Print a FINISHME message, including its source location.
 240  */
 241 #define tu_finishme(format, ...)                                              \
 242    do {                                                                        \
 243       static bool reported = false;                                            \
 244       if (!reported) {                                                         \
 245          __tu_finishme(__FILE__, __LINE__, format, ##__VA_ARGS__);            \
 246          reported = true;                                                      \
 247       }                                                                        \
 248    } while (0)
 249
 250 /* A non-fatal assert.  Useful for debugging. */
 251 #ifdef DEBUG
 252 #define tu_assert(x)                                                          \
 253    ({                                                                          \
 254       if (unlikely(!(x)))                                                      \
 255          fprintf(stderr, "%s:%d ASSERT: %s\n", __FILE__, __LINE__, #x);        \
 256    })
 257 #else
 258 #define tu_assert(x)
 259 #endif
 260
 261 #define stub_return(v)                                                         \
 262    do {                                                                        \
 263       tu_finishme("stub %s", __func__);                                       \
 264       return (v);                                                              \
 265    } while (0)
 266
 267 #define stub()                                                                 \
 268    do {                                                                        \
 269       tu_finishme("stub %s", __func__);                                       \
 270       return;                                                                  \
 271    } while (0)
 272
 273 void *
 274 tu_lookup_entrypoint_unchecked(const char *name);
 275 void *
 276 tu_lookup_entrypoint_checked(
 277    const char *name,
 278    uint32_t core_version,
 279    const struct tu_instance_extension_table *instance,
 280    const struct tu_device_extension_table *device);
 281
 282 struct tu_physical_device
 283 {
 284    VK_LOADER_DATA _loader_data;
 285
 286    struct tu_instance *instance;
 287
 288    char path[20];
 289    char name[VK_MAX_PHYSICAL_DEVICE_NAME_SIZE];
 290    uint8_t driver_uuid[VK_UUID_SIZE];
 291    uint8_t device_uuid[VK_UUID_SIZE];
 292    uint8_t cache_uuid[VK_UUID_SIZE];
 293
 294    int local_fd;
 295    int master_fd;
 296
 297    struct fd_device *drm_device;
 298    unsigned gpu_id;
 299    uint32_t gmem_size;
 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
 407 struct tu_bo_list
 408 {
 409    unsigned capacity;
 410    pthread_mutex_t mutex;
 411 };
 412
 413 struct tu_device
 414 {
 415    VK_LOADER_DATA _loader_data;
 416
 417    VkAllocationCallbacks alloc;
 418
 419    struct tu_instance *instance;
 420    struct radeon_winsys *ws;
 421
 422    struct tu_meta_state meta_state;
 423
 424    struct tu_queue *queues[TU_MAX_QUEUE_FAMILIES];
 425    int queue_count[TU_MAX_QUEUE_FAMILIES];
 426
 427    struct tu_physical_device *physical_device;
 428
 429    /* Backup in-memory cache to be used if the app doesn't provide one */
 430    struct tu_pipeline_cache *mem_cache;
 431
 432    struct list_head shader_slabs;
 433    mtx_t shader_slab_mutex;
 434
 435    struct tu_device_extension_table enabled_extensions;
 436
 437    /* Whether the driver uses a global BO list. */
 438    bool use_global_bo_list;
 439
 440    struct tu_bo_list bo_list;
 441 };
 442
 443 struct tu_device_memory
 444 {
 445    struct fd_bo *bo;
 446    VkDeviceSize size;
 447
 448    /* for dedicated allocations */
 449    struct tu_image *image;
 450    struct tu_buffer *buffer;
 451
 452    uint32_t type_index;
 453    void *map;
 454    void *user_ptr;
 455 };
 456
 457 struct tu_descriptor_range
 458 {
 459    uint64_t va;
 460    uint32_t size;
 461 };
 462
 463 struct tu_descriptor_set
 464 {
 465    const struct tu_descriptor_set_layout *layout;
 466    uint32_t size;
 467
 468    struct radeon_winsys_bo *bo;
 469    uint64_t va;
 470    uint32_t *mapped_ptr;
 471    struct tu_descriptor_range *dynamic_descriptors;
 472 };
 473
 474 struct tu_push_descriptor_set
 475 {
 476    struct tu_descriptor_set set;
 477    uint32_t capacity;
 478 };
 479
 480 struct tu_descriptor_pool_entry
 481 {
 482    uint32_t offset;
 483    uint32_t size;
 484    struct tu_descriptor_set *set;
 485 };
 486
 487 struct tu_descriptor_pool
 488 {
 489    struct radeon_winsys_bo *bo;
 490    uint8_t *mapped_ptr;
 491    uint64_t current_offset;
 492    uint64_t size;
 493
 494    uint8_t *host_memory_base;
 495    uint8_t *host_memory_ptr;
 496    uint8_t *host_memory_end;
 497
 498    uint32_t entry_count;
 499    uint32_t max_entry_count;
 500    struct tu_descriptor_pool_entry entries[0];
 501 };
 502
 503 struct tu_descriptor_update_template_entry
 504 {
 505    VkDescriptorType descriptor_type;
 506
 507    /* The number of descriptors to update */
 508    uint32_t descriptor_count;
 509
 510    /* Into mapped_ptr or dynamic_descriptors, in units of the respective array
 511     */
 512    uint32_t dst_offset;
 513
 514    /* In dwords. Not valid/used for dynamic descriptors */
 515    uint32_t dst_stride;
 516
 517    uint32_t buffer_offset;
 518
 519    /* Only valid for combined image samplers and samplers */
 520    uint16_t has_sampler;
 521
 522    /* In bytes */
 523    size_t src_offset;
 524    size_t src_stride;
 525
 526    /* For push descriptors */
 527    const uint32_t *immutable_samplers;
 528 };
 529
 530 struct tu_descriptor_update_template
 531 {
 532    uint32_t entry_count;
 533    VkPipelineBindPoint bind_point;
 534    struct tu_descriptor_update_template_entry entry[0];
 535 };
 536
 537 struct tu_buffer
 538 {
 539    VkDeviceSize size;
 540
 541    VkBufferUsageFlags usage;
 542    VkBufferCreateFlags flags;
 543 };
 544
 545 enum tu_dynamic_state_bits
 546 {
 547    TU_DYNAMIC_VIEWPORT = 1 << 0,
 548    TU_DYNAMIC_SCISSOR = 1 << 1,
 549    TU_DYNAMIC_LINE_WIDTH = 1 << 2,
 550    TU_DYNAMIC_DEPTH_BIAS = 1 << 3,
 551    TU_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
 552    TU_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
 553    TU_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
 554    TU_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
 555    TU_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
 556    TU_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
 557    TU_DYNAMIC_ALL = (1 << 10) - 1,
 558 };
 559
 560 struct tu_vertex_binding
 561 {
 562    struct tu_buffer *buffer;
 563    VkDeviceSize offset;
 564 };
 565
 566 struct tu_viewport_state
 567 {
 568    uint32_t count;
 569    VkViewport viewports[MAX_VIEWPORTS];
 570 };
 571
 572 struct tu_scissor_state
 573 {
 574    uint32_t count;
 575    VkRect2D scissors[MAX_SCISSORS];
 576 };
 577
 578 struct tu_discard_rectangle_state
 579 {
 580    uint32_t count;
 581    VkRect2D rectangles[MAX_DISCARD_RECTANGLES];
 582 };
 583
 584 struct tu_dynamic_state
 585 {
 586    /**
 587     * Bitmask of (1 << VK_DYNAMIC_STATE_*).
 588     * Defines the set of saved dynamic state.
 589     */
 590    uint32_t mask;
 591
 592    struct tu_viewport_state viewport;
 593
 594    struct tu_scissor_state scissor;
 595
 596    float line_width;
 597
 598    struct
 599    {
 600       float bias;
 601       float clamp;
 602       float slope;
 603    } depth_bias;
 604
 605    float blend_constants[4];
 606
 607    struct
 608    {
 609       float min;
 610       float max;
 611    } depth_bounds;
 612
 613    struct
 614    {
 615       uint32_t front;
 616       uint32_t back;
 617    } stencil_compare_mask;
 618
 619    struct
 620    {
 621       uint32_t front;
 622       uint32_t back;
 623    } stencil_write_mask;
 624
 625    struct
 626    {
 627       uint32_t front;
 628       uint32_t back;
 629    } stencil_reference;
 630
 631    struct tu_discard_rectangle_state discard_rectangle;
 632 };
 633
 634 extern const struct tu_dynamic_state default_dynamic_state;
 635
 636 const char *
 637 tu_get_debug_option_name(int id);
 638
 639 const char *
 640 tu_get_perftest_option_name(int id);
 641
 642 /**
 643  * Attachment state when recording a renderpass instance.
 644  *
 645  * The clear value is valid only if there exists a pending clear.
 646  */
 647 struct tu_attachment_state
 648 {
 649    VkImageAspectFlags pending_clear_aspects;
 650    uint32_t cleared_views;
 651    VkClearValue clear_value;
 652    VkImageLayout current_layout;
 653 };
 654
 655 struct tu_descriptor_state
 656 {
 657    struct tu_descriptor_set *sets[MAX_SETS];
 658    uint32_t dirty;
 659    uint32_t valid;
 660    struct tu_push_descriptor_set push_set;
 661    bool push_dirty;
 662    uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
 663 };
 664
 665 struct tu_cmd_state
 666 {
 667    /* Vertex descriptors */
 668    uint64_t vb_va;
 669    unsigned vb_size;
 670
 671    struct tu_dynamic_state dynamic;
 672
 673    /* Index buffer */
 674    struct tu_buffer *index_buffer;
 675    uint64_t index_offset;
 676    uint32_t index_type;
 677    uint32_t max_index_count;
 678    uint64_t index_va;
 679 };
 680
 681 struct tu_cmd_pool
 682 {
 683    VkAllocationCallbacks alloc;
 684    struct list_head cmd_buffers;
 685    struct list_head free_cmd_buffers;
 686    uint32_t queue_family_index;
 687 };
 688
 689 struct tu_cmd_buffer_upload
 690 {
 691    uint8_t *map;
 692    unsigned offset;
 693    uint64_t size;
 694    struct radeon_winsys_bo *upload_bo;
 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_cmd_buffer
 708 {
 709    VK_LOADER_DATA _loader_data;
 710
 711    struct tu_device *device;
 712
 713    struct tu_cmd_pool *pool;
 714    struct list_head pool_link;
 715
 716    VkCommandBufferUsageFlags usage_flags;
 717    VkCommandBufferLevel level;
 718    enum tu_cmd_buffer_status status;
 719    struct radeon_cmdbuf *cs;
 720    struct tu_cmd_state state;
 721    struct tu_vertex_binding vertex_bindings[MAX_VBS];
 722    uint32_t queue_family_index;
 723
 724    uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
 725    VkShaderStageFlags push_constant_stages;
 726    struct tu_descriptor_set meta_push_descriptors;
 727
 728    struct tu_descriptor_state descriptors[VK_PIPELINE_BIND_POINT_RANGE_SIZE];
 729
 730    struct tu_cmd_buffer_upload upload;
 731
 732    uint32_t scratch_size_needed;
 733    uint32_t compute_scratch_size_needed;
 734    uint32_t esgs_ring_size_needed;
 735    uint32_t gsvs_ring_size_needed;
 736    bool tess_rings_needed;
 737    bool sample_positions_needed;
 738
 739    VkResult record_result;
 740
 741    uint32_t gfx9_fence_offset;
 742    struct radeon_winsys_bo *gfx9_fence_bo;
 743    uint32_t gfx9_fence_idx;
 744    uint64_t gfx9_eop_bug_va;
 745
 746    /**
 747     * Whether a query pool has been resetted and we have to flush caches.
 748     */
 749    bool pending_reset_query;
 750 };
 751
 752 bool
 753 tu_get_memory_fd(struct tu_device *device,
 754                  struct tu_device_memory *memory,
 755                  int *pFD);
 756
 757 /*
 758  * Takes x,y,z as exact numbers of invocations, instead of blocks.
 759  *
 760  * Limitations: Can't call normal dispatch functions without binding or
 761  * rebinding
 762  *              the compute pipeline.
 763  */
 764 void
 765 tu_unaligned_dispatch(struct tu_cmd_buffer *cmd_buffer,
 766                       uint32_t x,
 767                       uint32_t y,
 768                       uint32_t z);
 769
 770 struct tu_event
 771 {
 772    struct radeon_winsys_bo *bo;
 773    uint64_t *map;
 774 };
 775
 776 struct tu_shader_module;
 777
 778 #define TU_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
 779 #define TU_HASH_SHADER_SISCHED (1 << 1)
 780 #define TU_HASH_SHADER_UNSAFE_MATH (1 << 2)
 781 void
 782 tu_hash_shaders(unsigned char *hash,
 783                 const VkPipelineShaderStageCreateInfo **stages,
 784                 const struct tu_pipeline_layout *layout,
 785                 const struct tu_pipeline_key *key,
 786                 uint32_t flags);
 787
 788 static inline gl_shader_stage
 789 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
 790 {
 791    assert(__builtin_popcount(vk_stage) == 1);
 792    return ffs(vk_stage) - 1;
 793 }
 794
 795 static inline VkShaderStageFlagBits
 796 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
 797 {
 798    return (1 << mesa_stage);
 799 }
 800
 801 #define TU_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
 802
 803 #define tu_foreach_stage(stage, stage_bits)                                   \
 804    for (gl_shader_stage stage,                                                 \
 805         __tmp = (gl_shader_stage)((stage_bits)&TU_STAGE_MASK);                \
 806         stage = __builtin_ffs(__tmp) - 1, __tmp;                               \
 807         __tmp &= ~(1 << (stage)))
 808
 809 struct tu_shader_module
 810 {
 811    struct nir_shader *nir;
 812    unsigned char sha1[20];
 813    uint32_t size;
 814    char data[0];
 815 };
 816
 817 struct tu_pipeline
 818 {
 819    struct tu_device *device;
 820    struct tu_dynamic_state dynamic_state;
 821
 822    struct tu_pipeline_layout *layout;
 823
 824    bool need_indirect_descriptor_sets;
 825    VkShaderStageFlags active_stages;
 826 };
 827
 828 struct tu_userdata_info *
 829 tu_lookup_user_sgpr(struct tu_pipeline *pipeline,
 830                     gl_shader_stage stage,
 831                     int idx);
 832
 833 struct tu_shader_variant *
 834 tu_get_shader(struct tu_pipeline *pipeline, gl_shader_stage stage);
 835
 836 struct tu_graphics_pipeline_create_info
 837 {
 838    bool use_rectlist;
 839    bool db_depth_clear;
 840    bool db_stencil_clear;
 841    bool db_depth_disable_expclear;
 842    bool db_stencil_disable_expclear;
 843    bool db_flush_depth_inplace;
 844    bool db_flush_stencil_inplace;
 845    bool db_resummarize;
 846    uint32_t custom_blend_mode;
 847 };
 848
 849 VkResult
 850 tu_graphics_pipeline_create(
 851    VkDevice device,
 852    VkPipelineCache cache,
 853    const VkGraphicsPipelineCreateInfo *pCreateInfo,
 854    const struct tu_graphics_pipeline_create_info *extra,
 855    const VkAllocationCallbacks *alloc,
 856    VkPipeline *pPipeline);
 857
 858 struct vk_format_description;
 859 uint32_t
 860 tu_translate_buffer_dataformat(const struct vk_format_description *desc,
 861                                int first_non_void);
 862 uint32_t
 863 tu_translate_buffer_numformat(const struct vk_format_description *desc,
 864                               int first_non_void);
 865 uint32_t
 866 tu_translate_colorformat(VkFormat format);
 867 uint32_t
 868 tu_translate_color_numformat(VkFormat format,
 869                              const struct vk_format_description *desc,
 870                              int first_non_void);
 871 uint32_t
 872 tu_colorformat_endian_swap(uint32_t colorformat);
 873 unsigned
 874 tu_translate_colorswap(VkFormat format, bool do_endian_swap);
 875 uint32_t
 876 tu_translate_dbformat(VkFormat format);
 877 uint32_t
 878 tu_translate_tex_dataformat(VkFormat format,
 879                             const struct vk_format_description *desc,
 880                             int first_non_void);
 881 uint32_t
 882 tu_translate_tex_numformat(VkFormat format,
 883                            const struct vk_format_description *desc,
 884                            int first_non_void);
 885 bool
 886 tu_format_pack_clear_color(VkFormat format,
 887                            uint32_t clear_vals[2],
 888                            VkClearColorValue *value);
 889 bool
 890 tu_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
 891 bool
 892 tu_dcc_formats_compatible(VkFormat format1, VkFormat format2);
 893
 894 struct tu_image
 895 {
 896    VkImageType type;
 897    /* The original VkFormat provided by the client.  This may not match any
 898     * of the actual surface formats.
 899     */
 900    VkFormat vk_format;
 901    VkImageAspectFlags aspects;
 902    VkImageUsageFlags usage;  /**< Superset of VkImageCreateInfo::usage. */
 903    VkImageTiling tiling;     /** VkImageCreateInfo::tiling */
 904    VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
 905
 906    VkDeviceSize size;
 907    uint32_t alignment;
 908
 909    unsigned queue_family_mask;
 910    bool exclusive;
 911    bool shareable;
 912
 913    /* For VK_ANDROID_native_buffer, the WSI image owns the memory, */
 914    VkDeviceMemory owned_memory;
 915 };
 916
 917 unsigned
 918 tu_image_queue_family_mask(const struct tu_image *image,
 919                            uint32_t family,
 920                            uint32_t queue_family);
 921
 922 static inline uint32_t
 923 tu_get_layerCount(const struct tu_image *image,
 924                   const VkImageSubresourceRange *range)
 925 {
 926    abort();
 927 }
 928
 929 static inline uint32_t
 930 tu_get_levelCount(const struct tu_image *image,
 931                   const VkImageSubresourceRange *range)
 932 {
 933    abort();
 934 }
 935
 936 struct tu_image_view
 937 {
 938    struct tu_image *image; /**< VkImageViewCreateInfo::image */
 939
 940    VkImageViewType type;
 941    VkImageAspectFlags aspect_mask;
 942    VkFormat vk_format;
 943    uint32_t base_layer;
 944    uint32_t layer_count;
 945    uint32_t base_mip;
 946    uint32_t level_count;
 947    VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
 948
 949    uint32_t descriptor[16];
 950
 951    /* Descriptor for use as a storage image as opposed to a sampled image.
 952     * This has a few differences for cube maps (e.g. type).
 953     */
 954    uint32_t storage_descriptor[16];
 955 };
 956
 957 struct tu_sampler
 958 {
 959 };
 960
 961 struct tu_image_create_info
 962 {
 963    const VkImageCreateInfo *vk_info;
 964    bool scanout;
 965    bool no_metadata_planes;
 966 };
 967
 968 VkResult
 969 tu_image_create(VkDevice _device,
 970                 const struct tu_image_create_info *info,
 971                 const VkAllocationCallbacks *alloc,
 972                 VkImage *pImage);
 973
 974 VkResult
 975 tu_image_from_gralloc(VkDevice device_h,
 976                       const VkImageCreateInfo *base_info,
 977                       const VkNativeBufferANDROID *gralloc_info,
 978                       const VkAllocationCallbacks *alloc,
 979                       VkImage *out_image_h);
 980
 981 void
 982 tu_image_view_init(struct tu_image_view *view,
 983                    struct tu_device *device,
 984                    const VkImageViewCreateInfo *pCreateInfo);
 985
 986 struct tu_buffer_view
 987 {
 988    struct radeon_winsys_bo *bo;
 989    VkFormat vk_format;
 990    uint64_t range; /**< VkBufferViewCreateInfo::range */
 991    uint32_t state[4];
 992 };
 993 void
 994 tu_buffer_view_init(struct tu_buffer_view *view,
 995                     struct tu_device *device,
 996                     const VkBufferViewCreateInfo *pCreateInfo);
 997
 998 static inline struct VkExtent3D
 999 tu_sanitize_image_extent(const VkImageType imageType,
1000                          const struct VkExtent3D imageExtent)
1001 {
1002    switch (imageType) {
1003       case VK_IMAGE_TYPE_1D:
1004          return (VkExtent3D){ imageExtent.width, 1, 1 };
1005       case VK_IMAGE_TYPE_2D:
1006          return (VkExtent3D){ imageExtent.width, imageExtent.height, 1 };
1007       case VK_IMAGE_TYPE_3D:
1008          return imageExtent;
1009       default:
1010          unreachable("invalid image type");
1011    }
1012 }
1013
1014 static inline struct VkOffset3D
1015 tu_sanitize_image_offset(const VkImageType imageType,
1016                          const struct VkOffset3D imageOffset)
1017 {
1018    switch (imageType) {
1019       case VK_IMAGE_TYPE_1D:
1020          return (VkOffset3D){ imageOffset.x, 0, 0 };
1021       case VK_IMAGE_TYPE_2D:
1022          return (VkOffset3D){ imageOffset.x, imageOffset.y, 0 };
1023       case VK_IMAGE_TYPE_3D:
1024          return imageOffset;
1025       default:
1026          unreachable("invalid image type");
1027    }
1028 }
1029
1030 struct tu_attachment_info
1031 {
1032    struct tu_image_view *attachment;
1033 };
1034
1035 struct tu_framebuffer
1036 {
1037    uint32_t width;
1038    uint32_t height;
1039    uint32_t layers;
1040
1041    uint32_t attachment_count;
1042    struct tu_attachment_info attachments[0];
1043 };
1044
1045 struct tu_subpass_barrier
1046 {
1047    VkPipelineStageFlags src_stage_mask;
1048    VkAccessFlags src_access_mask;
1049    VkAccessFlags dst_access_mask;
1050 };
1051
1052 void
1053 tu_subpass_barrier(struct tu_cmd_buffer *cmd_buffer,
1054                    const struct tu_subpass_barrier *barrier);
1055
1056 struct tu_subpass_attachment
1057 {
1058    uint32_t attachment;
1059    VkImageLayout layout;
1060 };
1061
1062 struct tu_subpass
1063 {
1064    uint32_t input_count;
1065    uint32_t color_count;
1066    struct tu_subpass_attachment *input_attachments;
1067    struct tu_subpass_attachment *color_attachments;
1068    struct tu_subpass_attachment *resolve_attachments;
1069    struct tu_subpass_attachment depth_stencil_attachment;
1070
1071    /** Subpass has at least one resolve attachment */
1072    bool has_resolve;
1073
1074    struct tu_subpass_barrier start_barrier;
1075
1076    uint32_t view_mask;
1077    VkSampleCountFlagBits max_sample_count;
1078 };
1079
1080 struct tu_render_pass_attachment
1081 {
1082    VkFormat format;
1083    uint32_t samples;
1084    VkAttachmentLoadOp load_op;
1085    VkAttachmentLoadOp stencil_load_op;
1086    VkImageLayout initial_layout;
1087    VkImageLayout final_layout;
1088    uint32_t view_mask;
1089 };
1090
1091 struct tu_render_pass
1092 {
1093    uint32_t attachment_count;
1094    uint32_t subpass_count;
1095    struct tu_subpass_attachment *subpass_attachments;
1096    struct tu_render_pass_attachment *attachments;
1097    struct tu_subpass_barrier end_barrier;
1098    struct tu_subpass subpasses[0];
1099 };
1100
1101 VkResult
1102 tu_device_init_meta(struct tu_device *device);
1103 void
1104 tu_device_finish_meta(struct tu_device *device);
1105
1106 struct tu_query_pool
1107 {
1108    struct radeon_winsys_bo *bo;
1109    uint32_t stride;
1110    uint32_t availability_offset;
1111    uint64_t size;
1112    char *ptr;
1113    VkQueryType type;
1114    uint32_t pipeline_stats_mask;
1115 };
1116
1117 struct tu_semaphore
1118 {
1119    /* use a winsys sem for non-exportable */
1120    struct radeon_winsys_sem *sem;
1121    uint32_t syncobj;
1122    uint32_t temp_syncobj;
1123 };
1124
1125 void
1126 tu_set_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1127                       VkPipelineBindPoint bind_point,
1128                       struct tu_descriptor_set *set,
1129                       unsigned idx);
1130
1131 void
1132 tu_update_descriptor_sets(struct tu_device *device,
1133                           struct tu_cmd_buffer *cmd_buffer,
1134                           VkDescriptorSet overrideSet,
1135                           uint32_t descriptorWriteCount,
1136                           const VkWriteDescriptorSet *pDescriptorWrites,
1137                           uint32_t descriptorCopyCount,
1138                           const VkCopyDescriptorSet *pDescriptorCopies);
1139
1140 void
1141 tu_update_descriptor_set_with_template(
1142    struct tu_device *device,
1143    struct tu_cmd_buffer *cmd_buffer,
1144    struct tu_descriptor_set *set,
1145    VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1146    const void *pData);
1147
1148 void
1149 tu_meta_push_descriptor_set(struct tu_cmd_buffer *cmd_buffer,
1150                             VkPipelineBindPoint pipelineBindPoint,
1151                             VkPipelineLayout _layout,
1152                             uint32_t set,
1153                             uint32_t descriptorWriteCount,
1154                             const VkWriteDescriptorSet *pDescriptorWrites);
1155
1156 struct tu_fence
1157 {
1158    struct radeon_winsys_fence *fence;
1159    bool submitted;
1160    bool signalled;
1161
1162    uint32_t syncobj;
1163    uint32_t temp_syncobj;
1164 };
1165
1166 /* tu_nir_to_llvm.c */
1167 struct tu_shader_variant_info;
1168 struct tu_nir_compiler_options;
1169
1170 struct radeon_winsys_sem;
1171
1172 #define TU_DEFINE_HANDLE_CASTS(__tu_type, __VkType)                          \
1173                                                                                \
1174    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1175    {                                                                           \
1176       return (struct __tu_type *)_handle;                                     \
1177    }                                                                           \
1178                                                                                \
1179    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1180    {                                                                           \
1181       return (__VkType)_obj;                                                   \
1182    }
1183
1184 #define TU_DEFINE_NONDISP_HANDLE_CASTS(__tu_type, __VkType)                  \
1185                                                                                \
1186    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \
1187    {                                                                           \
1188       return (struct __tu_type *)(uintptr_t)_handle;                          \
1189    }                                                                           \
1190                                                                                \
1191    static inline __VkType __tu_type##_to_handle(struct __tu_type *_obj)      \
1192    {                                                                           \
1193       return (__VkType)(uintptr_t)_obj;                                        \
1194    }
1195
1196 #define TU_FROM_HANDLE(__tu_type, __name, __handle)                          \
1197    struct __tu_type *__name = __tu_type##_from_handle(__handle)
1198
1199 TU_DEFINE_HANDLE_CASTS(tu_cmd_buffer, VkCommandBuffer)
1200 TU_DEFINE_HANDLE_CASTS(tu_device, VkDevice)
1201 TU_DEFINE_HANDLE_CASTS(tu_instance, VkInstance)
1202 TU_DEFINE_HANDLE_CASTS(tu_physical_device, VkPhysicalDevice)
1203 TU_DEFINE_HANDLE_CASTS(tu_queue, VkQueue)
1204
1205 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_cmd_pool, VkCommandPool)
1206 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer, VkBuffer)
1207 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_buffer_view, VkBufferView)
1208 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_pool, VkDescriptorPool)
1209 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set, VkDescriptorSet)
1210 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_set_layout,
1211                                 VkDescriptorSetLayout)
1212 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_descriptor_update_template,
1213                                 VkDescriptorUpdateTemplateKHR)
1214 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_device_memory, VkDeviceMemory)
1215 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_fence, VkFence)
1216 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_event, VkEvent)
1217 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_framebuffer, VkFramebuffer)
1218 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image, VkImage)
1219 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_image_view, VkImageView);
1220 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_cache, VkPipelineCache)
1221 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline, VkPipeline)
1222 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_pipeline_layout, VkPipelineLayout)
1223 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_query_pool, VkQueryPool)
1224 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_render_pass, VkRenderPass)
1225 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_sampler, VkSampler)
1226 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_shader_module, VkShaderModule)
1227 TU_DEFINE_NONDISP_HANDLE_CASTS(tu_semaphore, VkSemaphore)
1228
1229 #endif /* TU_PRIVATE_H */