src/freedreno/vulkan/tu_device.c

   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 #include "tu_private.h"
  29
  30 #include <fcntl.h>
  31 #include <libsync.h>
  32 #include <stdbool.h>
  33 #include <string.h>
  34 #include <sys/mman.h>
  35 #include <sys/sysinfo.h>
  36 #include <unistd.h>
  37 #include <xf86drm.h>
  38
  39 #include "compiler/glsl_types.h"
  40 #include "util/debug.h"
  41 #include "util/disk_cache.h"
  42 #include "util/u_atomic.h"
  43 #include "vk_format.h"
  44 #include "vk_util.h"
  45
  46 #include "drm-uapi/msm_drm.h"
  47
  48 /* for fd_get_driver/device_uuid() */
  49 #include "freedreno/common/freedreno_uuid.h"
  50
  51 static int
  52 tu_device_get_cache_uuid(uint16_t family, void *uuid)
  53 {
  54    uint32_t mesa_timestamp;
  55    uint16_t f = family;
  56    memset(uuid, 0, VK_UUID_SIZE);
  57    if (!disk_cache_get_function_timestamp(tu_device_get_cache_uuid,
  58                                           &mesa_timestamp))
  59       return -1;
  60
  61    memcpy(uuid, &mesa_timestamp, 4);
  62    memcpy((char *) uuid + 4, &f, 2);
  63    snprintf((char *) uuid + 6, VK_UUID_SIZE - 10, "tu");
  64    return 0;
  65 }
  66
  67 static VkResult
  68 tu_bo_init(struct tu_device *dev,
  69            struct tu_bo *bo,
  70            uint32_t gem_handle,
  71            uint64_t size)
  72 {
  73    uint64_t iova = tu_gem_info_iova(dev, gem_handle);
  74    if (!iova)
  75       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
  76
  77    *bo = (struct tu_bo) {
  78       .gem_handle = gem_handle,
  79       .size = size,
  80       .iova = iova,
  81    };
  82
  83    return VK_SUCCESS;
  84 }
  85
  86 VkResult
  87 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size)
  88 {
  89    /* TODO: Choose better flags. As of 2018-11-12, freedreno/drm/msm_bo.c
  90     * always sets `flags = MSM_BO_WC`, and we copy that behavior here.
  91     */
  92    uint32_t gem_handle = tu_gem_new(dev, size, MSM_BO_WC);
  93    if (!gem_handle)
  94       return vk_error(dev->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
  95
  96    VkResult result = tu_bo_init(dev, bo, gem_handle, size);
  97    if (result != VK_SUCCESS) {
  98       tu_gem_close(dev, gem_handle);
  99       return vk_error(dev->instance, result);
 100    }
 101
 102    return VK_SUCCESS;
 103 }
 104
 105 VkResult
 106 tu_bo_init_dmabuf(struct tu_device *dev,
 107                   struct tu_bo *bo,
 108                   uint64_t size,
 109                   int fd)
 110 {
 111    uint32_t gem_handle = tu_gem_import_dmabuf(dev, fd, size);
 112    if (!gem_handle)
 113       return vk_error(dev->instance, VK_ERROR_INVALID_EXTERNAL_HANDLE);
 114
 115    VkResult result = tu_bo_init(dev, bo, gem_handle, size);
 116    if (result != VK_SUCCESS) {
 117       tu_gem_close(dev, gem_handle);
 118       return vk_error(dev->instance, result);
 119    }
 120
 121    return VK_SUCCESS;
 122 }
 123
 124 int
 125 tu_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo)
 126 {
 127    return tu_gem_export_dmabuf(dev, bo->gem_handle);
 128 }
 129
 130 VkResult
 131 tu_bo_map(struct tu_device *dev, struct tu_bo *bo)
 132 {
 133    if (bo->map)
 134       return VK_SUCCESS;
 135
 136    uint64_t offset = tu_gem_info_offset(dev, bo->gem_handle);
 137    if (!offset)
 138       return vk_error(dev->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
 139
 140    /* TODO: Should we use the wrapper os_mmap() like Freedreno does? */
 141    void *map = mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
 142                     dev->physical_device->local_fd, offset);
 143    if (map == MAP_FAILED)
 144       return vk_error(dev->instance, VK_ERROR_MEMORY_MAP_FAILED);
 145
 146    bo->map = map;
 147    return VK_SUCCESS;
 148 }
 149
 150 void
 151 tu_bo_finish(struct tu_device *dev, struct tu_bo *bo)
 152 {
 153    assert(bo->gem_handle);
 154
 155    if (bo->map)
 156       munmap(bo->map, bo->size);
 157
 158    tu_gem_close(dev, bo->gem_handle);
 159 }
 160
 161 static VkResult
 162 tu_physical_device_init(struct tu_physical_device *device,
 163                         struct tu_instance *instance,
 164                         drmDevicePtr drm_device)
 165 {
 166    const char *path = drm_device->nodes[DRM_NODE_RENDER];
 167    VkResult result = VK_SUCCESS;
 168    drmVersionPtr version;
 169    int fd;
 170    int master_fd = -1;
 171
 172    fd = open(path, O_RDWR | O_CLOEXEC);
 173    if (fd < 0) {
 174       return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 175                        "failed to open device %s", path);
 176    }
 177
 178    /* Version 1.3 added MSM_INFO_IOVA. */
 179    const int min_version_major = 1;
 180    const int min_version_minor = 3;
 181
 182    version = drmGetVersion(fd);
 183    if (!version) {
 184       close(fd);
 185       return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 186                        "failed to query kernel driver version for device %s",
 187                        path);
 188    }
 189
 190    if (strcmp(version->name, "msm")) {
 191       drmFreeVersion(version);
 192       close(fd);
 193       return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 194                        "device %s does not use the msm kernel driver", path);
 195    }
 196
 197    if (version->version_major != min_version_major ||
 198        version->version_minor < min_version_minor) {
 199       result = vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 200                          "kernel driver for device %s has version %d.%d, "
 201                          "but Vulkan requires version >= %d.%d",
 202                          path, version->version_major, version->version_minor,
 203                          min_version_major, min_version_minor);
 204       drmFreeVersion(version);
 205       close(fd);
 206       return result;
 207    }
 208
 209    drmFreeVersion(version);
 210
 211    if (instance->debug_flags & TU_DEBUG_STARTUP)
 212       tu_logi("Found compatible device '%s'.", path);
 213
 214    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 215    device->instance = instance;
 216    assert(strlen(path) < ARRAY_SIZE(device->path));
 217    strncpy(device->path, path, ARRAY_SIZE(device->path));
 218
 219    if (instance->enabled_extensions.KHR_display) {
 220       master_fd =
 221          open(drm_device->nodes[DRM_NODE_PRIMARY], O_RDWR | O_CLOEXEC);
 222       if (master_fd >= 0) {
 223          /* TODO: free master_fd is accel is not working? */
 224       }
 225    }
 226
 227    device->master_fd = master_fd;
 228    device->local_fd = fd;
 229
 230    if (tu_drm_get_gpu_id(device, &device->gpu_id)) {
 231       if (instance->debug_flags & TU_DEBUG_STARTUP)
 232          tu_logi("Could not query the GPU ID");
 233       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 234                          "could not get GPU ID");
 235       goto fail;
 236    }
 237
 238    if (tu_drm_get_gmem_size(device, &device->gmem_size)) {
 239       if (instance->debug_flags & TU_DEBUG_STARTUP)
 240          tu_logi("Could not query the GMEM size");
 241       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 242                          "could not get GMEM size");
 243       goto fail;
 244    }
 245
 246    if (tu_drm_get_gmem_base(device, &device->gmem_base)) {
 247       if (instance->debug_flags & TU_DEBUG_STARTUP)
 248          tu_logi("Could not query the GMEM size");
 249       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 250                          "could not get GMEM size");
 251       goto fail;
 252    }
 253
 254    memset(device->name, 0, sizeof(device->name));
 255    sprintf(device->name, "FD%d", device->gpu_id);
 256
 257    switch (device->gpu_id) {
 258    case 618:
 259       device->ccu_offset_gmem = 0x7c000; /* 0x7e000 in some cases? */
 260       device->ccu_offset_bypass = 0x10000;
 261       device->tile_align_w = 64;
 262       device->magic.PC_UNKNOWN_9805 = 0x0;
 263       device->magic.SP_UNKNOWN_A0F8 = 0x0;
 264       break;
 265    case 630:
 266    case 640:
 267       device->ccu_offset_gmem = 0xf8000;
 268       device->ccu_offset_bypass = 0x20000;
 269       device->tile_align_w = 64;
 270       device->magic.PC_UNKNOWN_9805 = 0x1;
 271       device->magic.SP_UNKNOWN_A0F8 = 0x1;
 272       break;
 273    case 650:
 274       device->ccu_offset_gmem = 0x114000;
 275       device->ccu_offset_bypass = 0x30000;
 276       device->tile_align_w = 96;
 277       device->magic.PC_UNKNOWN_9805 = 0x2;
 278       device->magic.SP_UNKNOWN_A0F8 = 0x2;
 279       break;
 280    default:
 281       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 282                          "device %s is unsupported", device->name);
 283       goto fail;
 284    }
 285    if (tu_device_get_cache_uuid(device->gpu_id, device->cache_uuid)) {
 286       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 287                          "cannot generate UUID");
 288       goto fail;
 289    }
 290
 291    /* The gpu id is already embedded in the uuid so we just pass "tu"
 292     * when creating the cache.
 293     */
 294    char buf[VK_UUID_SIZE * 2 + 1];
 295    disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
 296    device->disk_cache = disk_cache_create(device->name, buf, 0);
 297
 298    fprintf(stderr, "WARNING: tu is not a conformant vulkan implementation, "
 299                    "testing use only.\n");
 300
 301    fd_get_driver_uuid(device->driver_uuid);
 302    fd_get_device_uuid(device->device_uuid, device->gpu_id);
 303
 304    tu_physical_device_get_supported_extensions(device, &device->supported_extensions);
 305
 306    if (result != VK_SUCCESS) {
 307       vk_error(instance, result);
 308       goto fail;
 309    }
 310
 311    result = tu_wsi_init(device);
 312    if (result != VK_SUCCESS) {
 313       vk_error(instance, result);
 314       goto fail;
 315    }
 316
 317    return VK_SUCCESS;
 318
 319 fail:
 320    close(fd);
 321    if (master_fd != -1)
 322       close(master_fd);
 323    return result;
 324 }
 325
 326 static void
 327 tu_physical_device_finish(struct tu_physical_device *device)
 328 {
 329    tu_wsi_finish(device);
 330
 331    disk_cache_destroy(device->disk_cache);
 332    close(device->local_fd);
 333    if (device->master_fd != -1)
 334       close(device->master_fd);
 335 }
 336
 337 static VKAPI_ATTR void *
 338 default_alloc_func(void *pUserData,
 339                    size_t size,
 340                    size_t align,
 341                    VkSystemAllocationScope allocationScope)
 342 {
 343    return malloc(size);
 344 }
 345
 346 static VKAPI_ATTR void *
 347 default_realloc_func(void *pUserData,
 348                      void *pOriginal,
 349                      size_t size,
 350                      size_t align,
 351                      VkSystemAllocationScope allocationScope)
 352 {
 353    return realloc(pOriginal, size);
 354 }
 355
 356 static VKAPI_ATTR void
 357 default_free_func(void *pUserData, void *pMemory)
 358 {
 359    free(pMemory);
 360 }
 361
 362 static const VkAllocationCallbacks default_alloc = {
 363    .pUserData = NULL,
 364    .pfnAllocation = default_alloc_func,
 365    .pfnReallocation = default_realloc_func,
 366    .pfnFree = default_free_func,
 367 };
 368
 369 static const struct debug_control tu_debug_options[] = {
 370    { "startup", TU_DEBUG_STARTUP },
 371    { "nir", TU_DEBUG_NIR },
 372    { "ir3", TU_DEBUG_IR3 },
 373    { "nobin", TU_DEBUG_NOBIN },
 374    { "sysmem", TU_DEBUG_SYSMEM },
 375    { "forcebin", TU_DEBUG_FORCEBIN },
 376    { "noubwc", TU_DEBUG_NOUBWC },
 377    { NULL, 0 }
 378 };
 379
 380 const char *
 381 tu_get_debug_option_name(int id)
 382 {
 383    assert(id < ARRAY_SIZE(tu_debug_options) - 1);
 384    return tu_debug_options[id].string;
 385 }
 386
 387 static int
 388 tu_get_instance_extension_index(const char *name)
 389 {
 390    for (unsigned i = 0; i < TU_INSTANCE_EXTENSION_COUNT; ++i) {
 391       if (strcmp(name, tu_instance_extensions[i].extensionName) == 0)
 392          return i;
 393    }
 394    return -1;
 395 }
 396
 397 VkResult
 398 tu_CreateInstance(const VkInstanceCreateInfo *pCreateInfo,
 399                   const VkAllocationCallbacks *pAllocator,
 400                   VkInstance *pInstance)
 401 {
 402    struct tu_instance *instance;
 403    VkResult result;
 404
 405    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 406
 407    uint32_t client_version;
 408    if (pCreateInfo->pApplicationInfo &&
 409        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 410       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 411    } else {
 412       tu_EnumerateInstanceVersion(&client_version);
 413    }
 414
 415    instance = vk_zalloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 416                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 417    if (!instance)
 418       return vk_error(NULL, VK_ERROR_OUT_OF_HOST_MEMORY);
 419
 420    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 421
 422    if (pAllocator)
 423       instance->alloc = *pAllocator;
 424    else
 425       instance->alloc = default_alloc;
 426
 427    instance->api_version = client_version;
 428    instance->physical_device_count = -1;
 429
 430    instance->debug_flags =
 431       parse_debug_string(getenv("TU_DEBUG"), tu_debug_options);
 432
 433    if (instance->debug_flags & TU_DEBUG_STARTUP)
 434       tu_logi("Created an instance");
 435
 436    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 437       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
 438       int index = tu_get_instance_extension_index(ext_name);
 439
 440       if (index < 0 || !tu_instance_extensions_supported.extensions[index]) {
 441          vk_free2(&default_alloc, pAllocator, instance);
 442          return vk_error(instance, VK_ERROR_EXTENSION_NOT_PRESENT);
 443       }
 444
 445       instance->enabled_extensions.extensions[index] = true;
 446    }
 447
 448    result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
 449    if (result != VK_SUCCESS) {
 450       vk_free2(&default_alloc, pAllocator, instance);
 451       return vk_error(instance, result);
 452    }
 453
 454    glsl_type_singleton_init_or_ref();
 455
 456    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 457
 458    *pInstance = tu_instance_to_handle(instance);
 459
 460    return VK_SUCCESS;
 461 }
 462
 463 void
 464 tu_DestroyInstance(VkInstance _instance,
 465                    const VkAllocationCallbacks *pAllocator)
 466 {
 467    TU_FROM_HANDLE(tu_instance, instance, _instance);
 468
 469    if (!instance)
 470       return;
 471
 472    for (int i = 0; i < instance->physical_device_count; ++i) {
 473       tu_physical_device_finish(instance->physical_devices + i);
 474    }
 475
 476    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 477
 478    glsl_type_singleton_decref();
 479
 480    vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
 481
 482    vk_free(&instance->alloc, instance);
 483 }
 484
 485 static VkResult
 486 tu_enumerate_devices(struct tu_instance *instance)
 487 {
 488    /* TODO: Check for more devices ? */
 489    drmDevicePtr devices[8];
 490    VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
 491    int max_devices;
 492
 493    instance->physical_device_count = 0;
 494
 495    max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
 496
 497    if (instance->debug_flags & TU_DEBUG_STARTUP)
 498       tu_logi("Found %d drm nodes", max_devices);
 499
 500    if (max_devices < 1)
 501       return vk_error(instance, VK_ERROR_INCOMPATIBLE_DRIVER);
 502
 503    for (unsigned i = 0; i < (unsigned) max_devices; i++) {
 504       if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
 505           devices[i]->bustype == DRM_BUS_PLATFORM) {
 506
 507          result = tu_physical_device_init(
 508             instance->physical_devices + instance->physical_device_count,
 509             instance, devices[i]);
 510          if (result == VK_SUCCESS)
 511             ++instance->physical_device_count;
 512          else if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 513             break;
 514       }
 515    }
 516    drmFreeDevices(devices, max_devices);
 517
 518    return result;
 519 }
 520
 521 VkResult
 522 tu_EnumeratePhysicalDevices(VkInstance _instance,
 523                             uint32_t *pPhysicalDeviceCount,
 524                             VkPhysicalDevice *pPhysicalDevices)
 525 {
 526    TU_FROM_HANDLE(tu_instance, instance, _instance);
 527    VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
 528
 529    VkResult result;
 530
 531    if (instance->physical_device_count < 0) {
 532       result = tu_enumerate_devices(instance);
 533       if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
 534          return result;
 535    }
 536
 537    for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
 538       vk_outarray_append(&out, p)
 539       {
 540          *p = tu_physical_device_to_handle(instance->physical_devices + i);
 541       }
 542    }
 543
 544    return vk_outarray_status(&out);
 545 }
 546
 547 VkResult
 548 tu_EnumeratePhysicalDeviceGroups(
 549    VkInstance _instance,
 550    uint32_t *pPhysicalDeviceGroupCount,
 551    VkPhysicalDeviceGroupProperties *pPhysicalDeviceGroupProperties)
 552 {
 553    TU_FROM_HANDLE(tu_instance, instance, _instance);
 554    VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
 555                     pPhysicalDeviceGroupCount);
 556    VkResult result;
 557
 558    if (instance->physical_device_count < 0) {
 559       result = tu_enumerate_devices(instance);
 560       if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
 561          return result;
 562    }
 563
 564    for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
 565       vk_outarray_append(&out, p)
 566       {
 567          p->physicalDeviceCount = 1;
 568          p->physicalDevices[0] =
 569             tu_physical_device_to_handle(instance->physical_devices + i);
 570          p->subsetAllocation = false;
 571       }
 572    }
 573
 574    return vk_outarray_status(&out);
 575 }
 576
 577 void
 578 tu_GetPhysicalDeviceFeatures(VkPhysicalDevice physicalDevice,
 579                              VkPhysicalDeviceFeatures *pFeatures)
 580 {
 581    memset(pFeatures, 0, sizeof(*pFeatures));
 582
 583    *pFeatures = (VkPhysicalDeviceFeatures) {
 584       .robustBufferAccess = true,
 585       .fullDrawIndexUint32 = true,
 586       .imageCubeArray = true,
 587       .independentBlend = true,
 588       .geometryShader = true,
 589       .tessellationShader = true,
 590       .sampleRateShading = true,
 591       .dualSrcBlend = true,
 592       .logicOp = true,
 593       .multiDrawIndirect = true,
 594       .drawIndirectFirstInstance = true,
 595       .depthClamp = true,
 596       .depthBiasClamp = false,
 597       .fillModeNonSolid = true,
 598       .depthBounds = true,
 599       .wideLines = false,
 600       .largePoints = false,
 601       .alphaToOne = true,
 602       .multiViewport = false,
 603       .samplerAnisotropy = true,
 604       .textureCompressionETC2 = true,
 605       .textureCompressionASTC_LDR = true,
 606       .textureCompressionBC = true,
 607       .occlusionQueryPrecise = true,
 608       .pipelineStatisticsQuery = false,
 609       .vertexPipelineStoresAndAtomics = false,
 610       .fragmentStoresAndAtomics = false,
 611       .shaderTessellationAndGeometryPointSize = false,
 612       .shaderImageGatherExtended = false,
 613       .shaderStorageImageExtendedFormats = false,
 614       .shaderStorageImageMultisample = false,
 615       .shaderUniformBufferArrayDynamicIndexing = false,
 616       .shaderSampledImageArrayDynamicIndexing = false,
 617       .shaderStorageBufferArrayDynamicIndexing = false,
 618       .shaderStorageImageArrayDynamicIndexing = false,
 619       .shaderStorageImageReadWithoutFormat = false,
 620       .shaderStorageImageWriteWithoutFormat = false,
 621       .shaderClipDistance = false,
 622       .shaderCullDistance = false,
 623       .shaderFloat64 = false,
 624       .shaderInt64 = false,
 625       .shaderInt16 = false,
 626       .sparseBinding = false,
 627       .variableMultisampleRate = false,
 628       .inheritedQueries = false,
 629    };
 630 }
 631
 632 void
 633 tu_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,
 634                               VkPhysicalDeviceFeatures2 *pFeatures)
 635 {
 636    vk_foreach_struct(ext, pFeatures->pNext)
 637    {
 638       switch (ext->sType) {
 639       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES: {
 640          VkPhysicalDeviceVulkan11Features *features = (void *) ext;
 641          features->storageBuffer16BitAccess            = false;
 642          features->uniformAndStorageBuffer16BitAccess  = false;
 643          features->storagePushConstant16               = false;
 644          features->storageInputOutput16                = false;
 645          features->multiview                           = false;
 646          features->multiviewGeometryShader             = false;
 647          features->multiviewTessellationShader         = false;
 648          features->variablePointersStorageBuffer       = false;
 649          features->variablePointers                    = false;
 650          features->protectedMemory                     = false;
 651          features->samplerYcbcrConversion              = true;
 652          features->shaderDrawParameters                = true;
 653          break;
 654       }
 655       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
 656          VkPhysicalDeviceVariablePointersFeatures *features = (void *) ext;
 657          features->variablePointersStorageBuffer = false;
 658          features->variablePointers = false;
 659          break;
 660       }
 661       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
 662          VkPhysicalDeviceMultiviewFeatures *features =
 663             (VkPhysicalDeviceMultiviewFeatures *) ext;
 664          features->multiview = false;
 665          features->multiviewGeometryShader = false;
 666          features->multiviewTessellationShader = false;
 667          break;
 668       }
 669       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
 670          VkPhysicalDeviceShaderDrawParametersFeatures *features =
 671             (VkPhysicalDeviceShaderDrawParametersFeatures *) ext;
 672          features->shaderDrawParameters = true;
 673          break;
 674       }
 675       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
 676          VkPhysicalDeviceProtectedMemoryFeatures *features =
 677             (VkPhysicalDeviceProtectedMemoryFeatures *) ext;
 678          features->protectedMemory = false;
 679          break;
 680       }
 681       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
 682          VkPhysicalDevice16BitStorageFeatures *features =
 683             (VkPhysicalDevice16BitStorageFeatures *) ext;
 684          features->storageBuffer16BitAccess = false;
 685          features->uniformAndStorageBuffer16BitAccess = false;
 686          features->storagePushConstant16 = false;
 687          features->storageInputOutput16 = false;
 688          break;
 689       }
 690       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
 691          VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
 692             (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
 693          features->samplerYcbcrConversion = true;
 694          break;
 695       }
 696       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
 697          VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
 698             (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *) ext;
 699          features->shaderInputAttachmentArrayDynamicIndexing = false;
 700          features->shaderUniformTexelBufferArrayDynamicIndexing = false;
 701          features->shaderStorageTexelBufferArrayDynamicIndexing = false;
 702          features->shaderUniformBufferArrayNonUniformIndexing = false;
 703          features->shaderSampledImageArrayNonUniformIndexing = false;
 704          features->shaderStorageBufferArrayNonUniformIndexing = false;
 705          features->shaderStorageImageArrayNonUniformIndexing = false;
 706          features->shaderInputAttachmentArrayNonUniformIndexing = false;
 707          features->shaderUniformTexelBufferArrayNonUniformIndexing = false;
 708          features->shaderStorageTexelBufferArrayNonUniformIndexing = false;
 709          features->descriptorBindingUniformBufferUpdateAfterBind = false;
 710          features->descriptorBindingSampledImageUpdateAfterBind = false;
 711          features->descriptorBindingStorageImageUpdateAfterBind = false;
 712          features->descriptorBindingStorageBufferUpdateAfterBind = false;
 713          features->descriptorBindingUniformTexelBufferUpdateAfterBind = false;
 714          features->descriptorBindingStorageTexelBufferUpdateAfterBind = false;
 715          features->descriptorBindingUpdateUnusedWhilePending = false;
 716          features->descriptorBindingPartiallyBound = false;
 717          features->descriptorBindingVariableDescriptorCount = false;
 718          features->runtimeDescriptorArray = false;
 719          break;
 720       }
 721       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
 722          VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
 723             (VkPhysicalDeviceConditionalRenderingFeaturesEXT *) ext;
 724          features->conditionalRendering = false;
 725          features->inheritedConditionalRendering = false;
 726          break;
 727       }
 728       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
 729          VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
 730             (VkPhysicalDeviceTransformFeedbackFeaturesEXT *) ext;
 731          features->transformFeedback = true;
 732          features->geometryStreams = false;
 733          break;
 734       }
 735       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
 736          VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
 737             (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
 738          features->indexTypeUint8 = true;
 739          break;
 740       }
 741       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
 742          VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
 743             (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
 744          features->vertexAttributeInstanceRateDivisor = true;
 745          features->vertexAttributeInstanceRateZeroDivisor = true;
 746          break;
 747       }
 748       default:
 749          break;
 750       }
 751    }
 752    return tu_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 753 }
 754
 755 void
 756 tu_GetPhysicalDeviceProperties(VkPhysicalDevice physicalDevice,
 757                                VkPhysicalDeviceProperties *pProperties)
 758 {
 759    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 760    VkSampleCountFlags sample_counts =
 761       VK_SAMPLE_COUNT_1_BIT | VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT;
 762
 763    /* I have no idea what the maximum size is, but the hardware supports very
 764     * large numbers of descriptors (at least 2^16). This limit is based on
 765     * CP_LOAD_STATE6, which has a 28-bit field for the DWORD offset, so that
 766     * we don't have to think about what to do if that overflows, but really
 767     * nothing is likely to get close to this.
 768     */
 769    const size_t max_descriptor_set_size = (1 << 28) / A6XX_TEX_CONST_DWORDS;
 770
 771    VkPhysicalDeviceLimits limits = {
 772       .maxImageDimension1D = (1 << 14),
 773       .maxImageDimension2D = (1 << 14),
 774       .maxImageDimension3D = (1 << 11),
 775       .maxImageDimensionCube = (1 << 14),
 776       .maxImageArrayLayers = (1 << 11),
 777       .maxTexelBufferElements = 128 * 1024 * 1024,
 778       .maxUniformBufferRange = MAX_UNIFORM_BUFFER_RANGE,
 779       .maxStorageBufferRange = MAX_STORAGE_BUFFER_RANGE,
 780       .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
 781       .maxMemoryAllocationCount = UINT32_MAX,
 782       .maxSamplerAllocationCount = 64 * 1024,
 783       .bufferImageGranularity = 64,          /* A cache line */
 784       .sparseAddressSpaceSize = 0xffffffffu, /* buffer max size */
 785       .maxBoundDescriptorSets = MAX_SETS,
 786       .maxPerStageDescriptorSamplers = max_descriptor_set_size,
 787       .maxPerStageDescriptorUniformBuffers = max_descriptor_set_size,
 788       .maxPerStageDescriptorStorageBuffers = max_descriptor_set_size,
 789       .maxPerStageDescriptorSampledImages = max_descriptor_set_size,
 790       .maxPerStageDescriptorStorageImages = max_descriptor_set_size,
 791       .maxPerStageDescriptorInputAttachments = MAX_RTS,
 792       .maxPerStageResources = max_descriptor_set_size,
 793       .maxDescriptorSetSamplers = max_descriptor_set_size,
 794       .maxDescriptorSetUniformBuffers = max_descriptor_set_size,
 795       .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS,
 796       .maxDescriptorSetStorageBuffers = max_descriptor_set_size,
 797       .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS,
 798       .maxDescriptorSetSampledImages = max_descriptor_set_size,
 799       .maxDescriptorSetStorageImages = max_descriptor_set_size,
 800       .maxDescriptorSetInputAttachments = MAX_RTS,
 801       .maxVertexInputAttributes = 32,
 802       .maxVertexInputBindings = 32,
 803       .maxVertexInputAttributeOffset = 4095,
 804       .maxVertexInputBindingStride = 2048,
 805       .maxVertexOutputComponents = 128,
 806       .maxTessellationGenerationLevel = 64,
 807       .maxTessellationPatchSize = 32,
 808       .maxTessellationControlPerVertexInputComponents = 128,
 809       .maxTessellationControlPerVertexOutputComponents = 128,
 810       .maxTessellationControlPerPatchOutputComponents = 120,
 811       .maxTessellationControlTotalOutputComponents = 4096,
 812       .maxTessellationEvaluationInputComponents = 128,
 813       .maxTessellationEvaluationOutputComponents = 128,
 814       .maxGeometryShaderInvocations = 32,
 815       .maxGeometryInputComponents = 64,
 816       .maxGeometryOutputComponents = 128,
 817       .maxGeometryOutputVertices = 256,
 818       .maxGeometryTotalOutputComponents = 1024,
 819       .maxFragmentInputComponents = 124,
 820       .maxFragmentOutputAttachments = 8,
 821       .maxFragmentDualSrcAttachments = 1,
 822       .maxFragmentCombinedOutputResources = 8,
 823       .maxComputeSharedMemorySize = 32768,
 824       .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
 825       .maxComputeWorkGroupInvocations = 2048,
 826       .maxComputeWorkGroupSize = { 2048, 2048, 2048 },
 827       .subPixelPrecisionBits = 8,
 828       .subTexelPrecisionBits = 8,
 829       .mipmapPrecisionBits = 8,
 830       .maxDrawIndexedIndexValue = UINT32_MAX,
 831       .maxDrawIndirectCount = UINT32_MAX,
 832       .maxSamplerLodBias = 4095.0 / 256.0, /* [-16, 15.99609375] */
 833       .maxSamplerAnisotropy = 16,
 834       .maxViewports = MAX_VIEWPORTS,
 835       .maxViewportDimensions = { (1 << 14), (1 << 14) },
 836       .viewportBoundsRange = { INT16_MIN, INT16_MAX },
 837       .viewportSubPixelBits = 8,
 838       .minMemoryMapAlignment = 4096, /* A page */
 839       .minTexelBufferOffsetAlignment = 64,
 840       .minUniformBufferOffsetAlignment = 64,
 841       .minStorageBufferOffsetAlignment = 64,
 842       .minTexelOffset = -16,
 843       .maxTexelOffset = 15,
 844       .minTexelGatherOffset = -32,
 845       .maxTexelGatherOffset = 31,
 846       .minInterpolationOffset = -0.5,
 847       .maxInterpolationOffset = 0.4375,
 848       .subPixelInterpolationOffsetBits = 4,
 849       .maxFramebufferWidth = (1 << 14),
 850       .maxFramebufferHeight = (1 << 14),
 851       .maxFramebufferLayers = (1 << 10),
 852       .framebufferColorSampleCounts = sample_counts,
 853       .framebufferDepthSampleCounts = sample_counts,
 854       .framebufferStencilSampleCounts = sample_counts,
 855       .framebufferNoAttachmentsSampleCounts = sample_counts,
 856       .maxColorAttachments = MAX_RTS,
 857       .sampledImageColorSampleCounts = sample_counts,
 858       .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 859       .sampledImageDepthSampleCounts = sample_counts,
 860       .sampledImageStencilSampleCounts = sample_counts,
 861       .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 862       .maxSampleMaskWords = 1,
 863       .timestampComputeAndGraphics = true,
 864       .timestampPeriod = 1000000000.0 / 19200000.0, /* CP_ALWAYS_ON_COUNTER is fixed 19.2MHz */
 865       .maxClipDistances = 8,
 866       .maxCullDistances = 8,
 867       .maxCombinedClipAndCullDistances = 8,
 868       .discreteQueuePriorities = 1,
 869       .pointSizeRange = { 0.125, 255.875 },
 870       .lineWidthRange = { 0.0, 7.9921875 },
 871       .pointSizeGranularity = (1.0 / 8.0),
 872       .lineWidthGranularity = (1.0 / 128.0),
 873       .strictLines = false, /* FINISHME */
 874       .standardSampleLocations = true,
 875       .optimalBufferCopyOffsetAlignment = 128,
 876       .optimalBufferCopyRowPitchAlignment = 128,
 877       .nonCoherentAtomSize = 64,
 878    };
 879
 880    *pProperties = (VkPhysicalDeviceProperties) {
 881       .apiVersion = tu_physical_device_api_version(pdevice),
 882       .driverVersion = vk_get_driver_version(),
 883       .vendorID = 0, /* TODO */
 884       .deviceID = 0,
 885       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 886       .limits = limits,
 887       .sparseProperties = { 0 },
 888    };
 889
 890    strcpy(pProperties->deviceName, pdevice->name);
 891    memcpy(pProperties->pipelineCacheUUID, pdevice->cache_uuid, VK_UUID_SIZE);
 892 }
 893
 894 void
 895 tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
 896                                 VkPhysicalDeviceProperties2 *pProperties)
 897 {
 898    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 899    tu_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 900
 901    vk_foreach_struct(ext, pProperties->pNext)
 902    {
 903       switch (ext->sType) {
 904       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 905          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 906             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 907          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 908          break;
 909       }
 910       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
 911          VkPhysicalDeviceIDProperties *properties =
 912             (VkPhysicalDeviceIDProperties *) ext;
 913          memcpy(properties->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
 914          memcpy(properties->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
 915          properties->deviceLUIDValid = false;
 916          break;
 917       }
 918       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
 919          VkPhysicalDeviceMultiviewProperties *properties =
 920             (VkPhysicalDeviceMultiviewProperties *) ext;
 921          properties->maxMultiviewViewCount = MAX_VIEWS;
 922          properties->maxMultiviewInstanceIndex = INT_MAX;
 923          break;
 924       }
 925       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
 926          VkPhysicalDevicePointClippingProperties *properties =
 927             (VkPhysicalDevicePointClippingProperties *) ext;
 928          properties->pointClippingBehavior =
 929             VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
 930          break;
 931       }
 932       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
 933          VkPhysicalDeviceMaintenance3Properties *properties =
 934             (VkPhysicalDeviceMaintenance3Properties *) ext;
 935          /* Make sure everything is addressable by a signed 32-bit int, and
 936           * our largest descriptors are 96 bytes. */
 937          properties->maxPerSetDescriptors = (1ull << 31) / 96;
 938          /* Our buffer size fields allow only this much */
 939          properties->maxMemoryAllocationSize = 0xFFFFFFFFull;
 940          break;
 941       }
 942       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
 943          VkPhysicalDeviceTransformFeedbackPropertiesEXT *properties =
 944             (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
 945
 946          properties->maxTransformFeedbackStreams = IR3_MAX_SO_STREAMS;
 947          properties->maxTransformFeedbackBuffers = IR3_MAX_SO_BUFFERS;
 948          properties->maxTransformFeedbackBufferSize = UINT32_MAX;
 949          properties->maxTransformFeedbackStreamDataSize = 512;
 950          properties->maxTransformFeedbackBufferDataSize = 512;
 951          properties->maxTransformFeedbackBufferDataStride = 512;
 952          properties->transformFeedbackQueries = true;
 953          properties->transformFeedbackStreamsLinesTriangles = false;
 954          properties->transformFeedbackRasterizationStreamSelect = false;
 955          properties->transformFeedbackDraw = true;
 956          break;
 957       }
 958       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLE_LOCATIONS_PROPERTIES_EXT: {
 959          VkPhysicalDeviceSampleLocationsPropertiesEXT *properties =
 960             (VkPhysicalDeviceSampleLocationsPropertiesEXT *)ext;
 961          properties->sampleLocationSampleCounts = 0;
 962          if (pdevice->supported_extensions.EXT_sample_locations) {
 963             properties->sampleLocationSampleCounts =
 964                VK_SAMPLE_COUNT_1_BIT | VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT;
 965          }
 966          properties->maxSampleLocationGridSize = (VkExtent2D) { 1 , 1 };
 967          properties->sampleLocationCoordinateRange[0] = 0.0f;
 968          properties->sampleLocationCoordinateRange[1] = 0.9375f;
 969          properties->sampleLocationSubPixelBits = 4;
 970          properties->variableSampleLocations = true;
 971          break;
 972       }
 973       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES: {
 974          VkPhysicalDeviceSamplerFilterMinmaxProperties *properties =
 975             (VkPhysicalDeviceSamplerFilterMinmaxProperties *)ext;
 976          properties->filterMinmaxImageComponentMapping = true;
 977          properties->filterMinmaxSingleComponentFormats = true;
 978          break;
 979       }
 980       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
 981          VkPhysicalDeviceSubgroupProperties *properties =
 982             (VkPhysicalDeviceSubgroupProperties *)ext;
 983          properties->subgroupSize = 64;
 984          properties->supportedStages = VK_SHADER_STAGE_COMPUTE_BIT;
 985          properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
 986                                            VK_SUBGROUP_FEATURE_VOTE_BIT;
 987          properties->quadOperationsInAllStages = false;
 988          break;
 989       }
 990       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
 991          VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
 992             (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
 993          props->maxVertexAttribDivisor = UINT32_MAX;
 994          break;
 995       }
 996       default:
 997          break;
 998       }
 999    }
1000 }
1001
1002 static const VkQueueFamilyProperties tu_queue_family_properties = {
1003    .queueFlags =
1004       VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT,
1005    .queueCount = 1,
1006    .timestampValidBits = 48,
1007    .minImageTransferGranularity = { 1, 1, 1 },
1008 };
1009
1010 void
1011 tu_GetPhysicalDeviceQueueFamilyProperties(
1012    VkPhysicalDevice physicalDevice,
1013    uint32_t *pQueueFamilyPropertyCount,
1014    VkQueueFamilyProperties *pQueueFamilyProperties)
1015 {
1016    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1017
1018    vk_outarray_append(&out, p) { *p = tu_queue_family_properties; }
1019 }
1020
1021 void
1022 tu_GetPhysicalDeviceQueueFamilyProperties2(
1023    VkPhysicalDevice physicalDevice,
1024    uint32_t *pQueueFamilyPropertyCount,
1025    VkQueueFamilyProperties2 *pQueueFamilyProperties)
1026 {
1027    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1028
1029    vk_outarray_append(&out, p)
1030    {
1031       p->queueFamilyProperties = tu_queue_family_properties;
1032    }
1033 }
1034
1035 static uint64_t
1036 tu_get_system_heap_size()
1037 {
1038    struct sysinfo info;
1039    sysinfo(&info);
1040
1041    uint64_t total_ram = (uint64_t) info.totalram * (uint64_t) info.mem_unit;
1042
1043    /* We don't want to burn too much ram with the GPU.  If the user has 4GiB
1044     * or less, we use at most half.  If they have more than 4GiB, we use 3/4.
1045     */
1046    uint64_t available_ram;
1047    if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
1048       available_ram = total_ram / 2;
1049    else
1050       available_ram = total_ram * 3 / 4;
1051
1052    return available_ram;
1053 }
1054
1055 void
1056 tu_GetPhysicalDeviceMemoryProperties(
1057    VkPhysicalDevice physicalDevice,
1058    VkPhysicalDeviceMemoryProperties *pMemoryProperties)
1059 {
1060    pMemoryProperties->memoryHeapCount = 1;
1061    pMemoryProperties->memoryHeaps[0].size = tu_get_system_heap_size();
1062    pMemoryProperties->memoryHeaps[0].flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT;
1063
1064    pMemoryProperties->memoryTypeCount = 1;
1065    pMemoryProperties->memoryTypes[0].propertyFlags =
1066       VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
1067       VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
1068       VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
1069    pMemoryProperties->memoryTypes[0].heapIndex = 0;
1070 }
1071
1072 void
1073 tu_GetPhysicalDeviceMemoryProperties2(
1074    VkPhysicalDevice physicalDevice,
1075    VkPhysicalDeviceMemoryProperties2 *pMemoryProperties)
1076 {
1077    return tu_GetPhysicalDeviceMemoryProperties(
1078       physicalDevice, &pMemoryProperties->memoryProperties);
1079 }
1080
1081 static VkResult
1082 tu_queue_init(struct tu_device *device,
1083               struct tu_queue *queue,
1084               uint32_t queue_family_index,
1085               int idx,
1086               VkDeviceQueueCreateFlags flags)
1087 {
1088    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1089    queue->device = device;
1090    queue->queue_family_index = queue_family_index;
1091    queue->queue_idx = idx;
1092    queue->flags = flags;
1093
1094    int ret = tu_drm_submitqueue_new(device, 0, &queue->msm_queue_id);
1095    if (ret)
1096       return VK_ERROR_INITIALIZATION_FAILED;
1097
1098    tu_fence_init(&queue->submit_fence, false);
1099
1100    return VK_SUCCESS;
1101 }
1102
1103 static void
1104 tu_queue_finish(struct tu_queue *queue)
1105 {
1106    tu_fence_finish(&queue->submit_fence);
1107    tu_drm_submitqueue_close(queue->device, queue->msm_queue_id);
1108 }
1109
1110 static int
1111 tu_get_device_extension_index(const char *name)
1112 {
1113    for (unsigned i = 0; i < TU_DEVICE_EXTENSION_COUNT; ++i) {
1114       if (strcmp(name, tu_device_extensions[i].extensionName) == 0)
1115          return i;
1116    }
1117    return -1;
1118 }
1119
1120 struct PACKED bcolor_entry {
1121    uint32_t fp32[4];
1122    uint16_t ui16[4];
1123    int16_t  si16[4];
1124    uint16_t fp16[4];
1125    uint16_t rgb565;
1126    uint16_t rgb5a1;
1127    uint16_t rgba4;
1128    uint8_t __pad0[2];
1129    uint8_t  ui8[4];
1130    int8_t   si8[4];
1131    uint32_t rgb10a2;
1132    uint32_t z24; /* also s8? */
1133    uint16_t srgb[4];      /* appears to duplicate fp16[], but clamped, used for srgb */
1134    uint8_t  __pad1[56];
1135 } border_color[] = {
1136    [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = {},
1137    [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = {},
1138    [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = {
1139       .fp32[3] = 0x3f800000,
1140       .ui16[3] = 0xffff,
1141       .si16[3] = 0x7fff,
1142       .fp16[3] = 0x3c00,
1143       .rgb5a1 = 0x8000,
1144       .rgba4 = 0xf000,
1145       .ui8[3] = 0xff,
1146       .si8[3] = 0x7f,
1147       .rgb10a2 = 0xc0000000,
1148       .srgb[3] = 0x3c00,
1149    },
1150    [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = {
1151       .fp32[3] = 1,
1152       .fp16[3] = 1,
1153    },
1154    [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = {
1155       .fp32[0 ... 3] = 0x3f800000,
1156       .ui16[0 ... 3] = 0xffff,
1157       .si16[0 ... 3] = 0x7fff,
1158       .fp16[0 ... 3] = 0x3c00,
1159       .rgb565 = 0xffff,
1160       .rgb5a1 = 0xffff,
1161       .rgba4 = 0xffff,
1162       .ui8[0 ... 3] = 0xff,
1163       .si8[0 ... 3] = 0x7f,
1164       .rgb10a2 = 0xffffffff,
1165       .z24 = 0xffffff,
1166       .srgb[0 ... 3] = 0x3c00,
1167    },
1168    [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = {
1169       .fp32[0 ... 3] = 1,
1170       .fp16[0 ... 3] = 1,
1171    },
1172 };
1173
1174
1175 VkResult
1176 tu_CreateDevice(VkPhysicalDevice physicalDevice,
1177                 const VkDeviceCreateInfo *pCreateInfo,
1178                 const VkAllocationCallbacks *pAllocator,
1179                 VkDevice *pDevice)
1180 {
1181    TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
1182    VkResult result;
1183    struct tu_device *device;
1184
1185    /* Check enabled features */
1186    if (pCreateInfo->pEnabledFeatures) {
1187       VkPhysicalDeviceFeatures supported_features;
1188       tu_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1189       VkBool32 *supported_feature = (VkBool32 *) &supported_features;
1190       VkBool32 *enabled_feature = (VkBool32 *) pCreateInfo->pEnabledFeatures;
1191       unsigned num_features =
1192          sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1193       for (uint32_t i = 0; i < num_features; i++) {
1194          if (enabled_feature[i] && !supported_feature[i])
1195             return vk_error(physical_device->instance,
1196                             VK_ERROR_FEATURE_NOT_PRESENT);
1197       }
1198    }
1199
1200    device = vk_zalloc2(&physical_device->instance->alloc, pAllocator,
1201                        sizeof(*device), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1202    if (!device)
1203       return vk_error(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1204
1205    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1206    device->instance = physical_device->instance;
1207    device->physical_device = physical_device;
1208
1209    if (pAllocator)
1210       device->alloc = *pAllocator;
1211    else
1212       device->alloc = physical_device->instance->alloc;
1213
1214    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1215       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
1216       int index = tu_get_device_extension_index(ext_name);
1217       if (index < 0 ||
1218           !physical_device->supported_extensions.extensions[index]) {
1219          vk_free(&device->alloc, device);
1220          return vk_error(physical_device->instance,
1221                          VK_ERROR_EXTENSION_NOT_PRESENT);
1222       }
1223
1224       device->enabled_extensions.extensions[index] = true;
1225    }
1226
1227    for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1228       const VkDeviceQueueCreateInfo *queue_create =
1229          &pCreateInfo->pQueueCreateInfos[i];
1230       uint32_t qfi = queue_create->queueFamilyIndex;
1231       device->queues[qfi] = vk_alloc(
1232          &device->alloc, queue_create->queueCount * sizeof(struct tu_queue),
1233          8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1234       if (!device->queues[qfi]) {
1235          result = VK_ERROR_OUT_OF_HOST_MEMORY;
1236          goto fail_queues;
1237       }
1238
1239       memset(device->queues[qfi], 0,
1240              queue_create->queueCount * sizeof(struct tu_queue));
1241
1242       device->queue_count[qfi] = queue_create->queueCount;
1243
1244       for (unsigned q = 0; q < queue_create->queueCount; q++) {
1245          result = tu_queue_init(device, &device->queues[qfi][q], qfi, q,
1246                                 queue_create->flags);
1247          if (result != VK_SUCCESS)
1248             goto fail_queues;
1249       }
1250    }
1251
1252    device->compiler = ir3_compiler_create(NULL, physical_device->gpu_id);
1253    if (!device->compiler)
1254       goto fail_queues;
1255
1256 #define VSC_DRAW_STRM_SIZE(pitch)  ((pitch) * 32 + 0x100)  /* extra size to store VSC_SIZE */
1257 #define VSC_PRIM_STRM_SIZE(pitch) ((pitch) * 32)
1258
1259    device->vsc_draw_strm_pitch = 0x440 * 4;
1260    device->vsc_prim_strm_pitch = 0x1040 * 4;
1261
1262    result = tu_bo_init_new(device, &device->vsc_draw_strm, VSC_DRAW_STRM_SIZE(device->vsc_draw_strm_pitch));
1263    if (result != VK_SUCCESS)
1264       goto fail_vsc_data;
1265
1266    result = tu_bo_init_new(device, &device->vsc_prim_strm, VSC_PRIM_STRM_SIZE(device->vsc_prim_strm_pitch));
1267    if (result != VK_SUCCESS)
1268       goto fail_vsc_data2;
1269
1270    STATIC_ASSERT(sizeof(struct bcolor_entry) == 128);
1271    result = tu_bo_init_new(device, &device->border_color, sizeof(border_color));
1272    if (result != VK_SUCCESS)
1273       goto fail_border_color;
1274
1275    result = tu_bo_map(device, &device->border_color);
1276    if (result != VK_SUCCESS)
1277       goto fail_border_color_map;
1278
1279    memcpy(device->border_color.map, border_color, sizeof(border_color));
1280
1281    VkPipelineCacheCreateInfo ci;
1282    ci.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
1283    ci.pNext = NULL;
1284    ci.flags = 0;
1285    ci.pInitialData = NULL;
1286    ci.initialDataSize = 0;
1287    VkPipelineCache pc;
1288    result =
1289       tu_CreatePipelineCache(tu_device_to_handle(device), &ci, NULL, &pc);
1290    if (result != VK_SUCCESS)
1291       goto fail_pipeline_cache;
1292
1293    device->mem_cache = tu_pipeline_cache_from_handle(pc);
1294
1295    for (unsigned i = 0; i < ARRAY_SIZE(device->scratch_bos); i++)
1296       mtx_init(&device->scratch_bos[i].construct_mtx, mtx_plain);
1297
1298    *pDevice = tu_device_to_handle(device);
1299    return VK_SUCCESS;
1300
1301 fail_pipeline_cache:
1302 fail_border_color_map:
1303    tu_bo_finish(device, &device->border_color);
1304
1305 fail_border_color:
1306    tu_bo_finish(device, &device->vsc_prim_strm);
1307
1308 fail_vsc_data2:
1309    tu_bo_finish(device, &device->vsc_draw_strm);
1310
1311 fail_vsc_data:
1312    ralloc_free(device->compiler);
1313
1314 fail_queues:
1315    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1316       for (unsigned q = 0; q < device->queue_count[i]; q++)
1317          tu_queue_finish(&device->queues[i][q]);
1318       if (device->queue_count[i])
1319          vk_free(&device->alloc, device->queues[i]);
1320    }
1321
1322    vk_free(&device->alloc, device);
1323    return result;
1324 }
1325
1326 void
1327 tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
1328 {
1329    TU_FROM_HANDLE(tu_device, device, _device);
1330
1331    if (!device)
1332       return;
1333
1334    tu_bo_finish(device, &device->vsc_draw_strm);
1335    tu_bo_finish(device, &device->vsc_prim_strm);
1336
1337    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1338       for (unsigned q = 0; q < device->queue_count[i]; q++)
1339          tu_queue_finish(&device->queues[i][q]);
1340       if (device->queue_count[i])
1341          vk_free(&device->alloc, device->queues[i]);
1342    }
1343
1344    for (unsigned i = 0; i < ARRAY_SIZE(device->scratch_bos); i++) {
1345       if (device->scratch_bos[i].initialized)
1346          tu_bo_finish(device, &device->scratch_bos[i].bo);
1347    }
1348
1349    ir3_compiler_destroy(device->compiler);
1350
1351    VkPipelineCache pc = tu_pipeline_cache_to_handle(device->mem_cache);
1352    tu_DestroyPipelineCache(tu_device_to_handle(device), pc, NULL);
1353
1354    vk_free(&device->alloc, device);
1355 }
1356
1357 VkResult
1358 tu_get_scratch_bo(struct tu_device *dev, uint64_t size, struct tu_bo **bo)
1359 {
1360    unsigned size_log2 = MAX2(util_logbase2_ceil64(size), MIN_SCRATCH_BO_SIZE_LOG2);
1361    unsigned index = size_log2 - MIN_SCRATCH_BO_SIZE_LOG2;
1362    assert(index < ARRAY_SIZE(dev->scratch_bos));
1363
1364    for (unsigned i = index; i < ARRAY_SIZE(dev->scratch_bos); i++) {
1365       if (p_atomic_read(&dev->scratch_bos[i].initialized)) {
1366          /* Fast path: just return the already-allocated BO. */
1367          *bo = &dev->scratch_bos[i].bo;
1368          return VK_SUCCESS;
1369       }
1370    }
1371
1372    /* Slow path: actually allocate the BO. We take a lock because the process
1373     * of allocating it is slow, and we don't want to block the CPU while it
1374     * finishes.
1375    */
1376    mtx_lock(&dev->scratch_bos[index].construct_mtx);
1377
1378    /* Another thread may have allocated it already while we were waiting on
1379     * the lock. We need to check this in order to avoid double-allocating.
1380     */
1381    if (dev->scratch_bos[index].initialized) {
1382       mtx_unlock(&dev->scratch_bos[index].construct_mtx);
1383       *bo = &dev->scratch_bos[index].bo;
1384       return VK_SUCCESS;
1385    }
1386
1387    unsigned bo_size = 1ull << size_log2;
1388    VkResult result = tu_bo_init_new(dev, &dev->scratch_bos[index].bo, bo_size);
1389    if (result != VK_SUCCESS) {
1390       mtx_unlock(&dev->scratch_bos[index].construct_mtx);
1391       return result;
1392    }
1393
1394    p_atomic_set(&dev->scratch_bos[index].initialized, true);
1395
1396    mtx_unlock(&dev->scratch_bos[index].construct_mtx);
1397
1398    *bo = &dev->scratch_bos[index].bo;
1399    return VK_SUCCESS;
1400 }
1401
1402 VkResult
1403 tu_EnumerateInstanceLayerProperties(uint32_t *pPropertyCount,
1404                                     VkLayerProperties *pProperties)
1405 {
1406    *pPropertyCount = 0;
1407    return VK_SUCCESS;
1408 }
1409
1410 VkResult
1411 tu_EnumerateDeviceLayerProperties(VkPhysicalDevice physicalDevice,
1412                                   uint32_t *pPropertyCount,
1413                                   VkLayerProperties *pProperties)
1414 {
1415    *pPropertyCount = 0;
1416    return VK_SUCCESS;
1417 }
1418
1419 void
1420 tu_GetDeviceQueue2(VkDevice _device,
1421                    const VkDeviceQueueInfo2 *pQueueInfo,
1422                    VkQueue *pQueue)
1423 {
1424    TU_FROM_HANDLE(tu_device, device, _device);
1425    struct tu_queue *queue;
1426
1427    queue =
1428       &device->queues[pQueueInfo->queueFamilyIndex][pQueueInfo->queueIndex];
1429    if (pQueueInfo->flags != queue->flags) {
1430       /* From the Vulkan 1.1.70 spec:
1431        *
1432        * "The queue returned by vkGetDeviceQueue2 must have the same
1433        * flags value from this structure as that used at device
1434        * creation time in a VkDeviceQueueCreateInfo instance. If no
1435        * matching flags were specified at device creation time then
1436        * pQueue will return VK_NULL_HANDLE."
1437        */
1438       *pQueue = VK_NULL_HANDLE;
1439       return;
1440    }
1441
1442    *pQueue = tu_queue_to_handle(queue);
1443 }
1444
1445 void
1446 tu_GetDeviceQueue(VkDevice _device,
1447                   uint32_t queueFamilyIndex,
1448                   uint32_t queueIndex,
1449                   VkQueue *pQueue)
1450 {
1451    const VkDeviceQueueInfo2 info =
1452       (VkDeviceQueueInfo2) { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
1453                              .queueFamilyIndex = queueFamilyIndex,
1454                              .queueIndex = queueIndex };
1455
1456    tu_GetDeviceQueue2(_device, &info, pQueue);
1457 }
1458
1459 VkResult
1460 tu_QueueSubmit(VkQueue _queue,
1461                uint32_t submitCount,
1462                const VkSubmitInfo *pSubmits,
1463                VkFence _fence)
1464 {
1465    TU_FROM_HANDLE(tu_queue, queue, _queue);
1466
1467    for (uint32_t i = 0; i < submitCount; ++i) {
1468       const VkSubmitInfo *submit = pSubmits + i;
1469       const bool last_submit = (i == submitCount - 1);
1470       struct tu_bo_list bo_list;
1471       tu_bo_list_init(&bo_list);
1472
1473       uint32_t entry_count = 0;
1474       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1475          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1476          entry_count += cmdbuf->cs.entry_count;
1477       }
1478
1479       struct drm_msm_gem_submit_cmd cmds[entry_count];
1480       uint32_t entry_idx = 0;
1481       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1482          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1483          struct tu_cs *cs = &cmdbuf->cs;
1484          for (unsigned i = 0; i < cs->entry_count; ++i, ++entry_idx) {
1485             cmds[entry_idx].type = MSM_SUBMIT_CMD_BUF;
1486             cmds[entry_idx].submit_idx =
1487                tu_bo_list_add(&bo_list, cs->entries[i].bo,
1488                               MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
1489             cmds[entry_idx].submit_offset = cs->entries[i].offset;
1490             cmds[entry_idx].size = cs->entries[i].size;
1491             cmds[entry_idx].pad = 0;
1492             cmds[entry_idx].nr_relocs = 0;
1493             cmds[entry_idx].relocs = 0;
1494          }
1495
1496          tu_bo_list_merge(&bo_list, &cmdbuf->bo_list);
1497       }
1498
1499       uint32_t flags = MSM_PIPE_3D0;
1500       if (last_submit) {
1501          flags |= MSM_SUBMIT_FENCE_FD_OUT;
1502       }
1503
1504       struct drm_msm_gem_submit req = {
1505          .flags = flags,
1506          .queueid = queue->msm_queue_id,
1507          .bos = (uint64_t)(uintptr_t) bo_list.bo_infos,
1508          .nr_bos = bo_list.count,
1509          .cmds = (uint64_t)(uintptr_t)cmds,
1510          .nr_cmds = entry_count,
1511       };
1512
1513       int ret = drmCommandWriteRead(queue->device->physical_device->local_fd,
1514                                     DRM_MSM_GEM_SUBMIT,
1515                                     &req, sizeof(req));
1516       if (ret) {
1517          fprintf(stderr, "submit failed: %s\n", strerror(errno));
1518          abort();
1519       }
1520
1521       tu_bo_list_destroy(&bo_list);
1522
1523       if (last_submit) {
1524          /* no need to merge fences as queue execution is serialized */
1525          tu_fence_update_fd(&queue->submit_fence, req.fence_fd);
1526       }
1527    }
1528
1529    if (_fence != VK_NULL_HANDLE) {
1530       TU_FROM_HANDLE(tu_fence, fence, _fence);
1531       tu_fence_copy(fence, &queue->submit_fence);
1532    }
1533
1534    return VK_SUCCESS;
1535 }
1536
1537 VkResult
1538 tu_QueueWaitIdle(VkQueue _queue)
1539 {
1540    TU_FROM_HANDLE(tu_queue, queue, _queue);
1541
1542    tu_fence_wait_idle(&queue->submit_fence);
1543
1544    return VK_SUCCESS;
1545 }
1546
1547 VkResult
1548 tu_DeviceWaitIdle(VkDevice _device)
1549 {
1550    TU_FROM_HANDLE(tu_device, device, _device);
1551
1552    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1553       for (unsigned q = 0; q < device->queue_count[i]; q++) {
1554          tu_QueueWaitIdle(tu_queue_to_handle(&device->queues[i][q]));
1555       }
1556    }
1557    return VK_SUCCESS;
1558 }
1559
1560 VkResult
1561 tu_ImportSemaphoreFdKHR(VkDevice _device,
1562                         const VkImportSemaphoreFdInfoKHR *pImportSemaphoreFdInfo)
1563 {
1564    tu_stub();
1565
1566    return VK_SUCCESS;
1567 }
1568
1569 VkResult
1570 tu_GetSemaphoreFdKHR(VkDevice _device,
1571                      const VkSemaphoreGetFdInfoKHR *pGetFdInfo,
1572                      int *pFd)
1573 {
1574    tu_stub();
1575
1576    return VK_SUCCESS;
1577 }
1578
1579 VkResult
1580 tu_ImportFenceFdKHR(VkDevice _device,
1581                     const VkImportFenceFdInfoKHR *pImportFenceFdInfo)
1582 {
1583    tu_stub();
1584
1585    return VK_SUCCESS;
1586 }
1587
1588 VkResult
1589 tu_GetFenceFdKHR(VkDevice _device,
1590                  const VkFenceGetFdInfoKHR *pGetFdInfo,
1591                  int *pFd)
1592 {
1593    tu_stub();
1594
1595    return VK_SUCCESS;
1596 }
1597
1598 VkResult
1599 tu_EnumerateInstanceExtensionProperties(const char *pLayerName,
1600                                         uint32_t *pPropertyCount,
1601                                         VkExtensionProperties *pProperties)
1602 {
1603    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1604
1605    /* We spport no lyaers */
1606    if (pLayerName)
1607       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1608
1609    for (int i = 0; i < TU_INSTANCE_EXTENSION_COUNT; i++) {
1610       if (tu_instance_extensions_supported.extensions[i]) {
1611          vk_outarray_append(&out, prop) { *prop = tu_instance_extensions[i]; }
1612       }
1613    }
1614
1615    return vk_outarray_status(&out);
1616 }
1617
1618 VkResult
1619 tu_EnumerateDeviceExtensionProperties(VkPhysicalDevice physicalDevice,
1620                                       const char *pLayerName,
1621                                       uint32_t *pPropertyCount,
1622                                       VkExtensionProperties *pProperties)
1623 {
1624    /* We spport no lyaers */
1625    TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
1626    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1627
1628    /* We spport no lyaers */
1629    if (pLayerName)
1630       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1631
1632    for (int i = 0; i < TU_DEVICE_EXTENSION_COUNT; i++) {
1633       if (device->supported_extensions.extensions[i]) {
1634          vk_outarray_append(&out, prop) { *prop = tu_device_extensions[i]; }
1635       }
1636    }
1637
1638    return vk_outarray_status(&out);
1639 }
1640
1641 PFN_vkVoidFunction
1642 tu_GetInstanceProcAddr(VkInstance _instance, const char *pName)
1643 {
1644    TU_FROM_HANDLE(tu_instance, instance, _instance);
1645
1646    return tu_lookup_entrypoint_checked(
1647       pName, instance ? instance->api_version : 0,
1648       instance ? &instance->enabled_extensions : NULL, NULL);
1649 }
1650
1651 /* The loader wants us to expose a second GetInstanceProcAddr function
1652  * to work around certain LD_PRELOAD issues seen in apps.
1653  */
1654 PUBLIC
1655 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1656 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName);
1657
1658 PUBLIC
1659 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1660 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName)
1661 {
1662    return tu_GetInstanceProcAddr(instance, pName);
1663 }
1664
1665 PFN_vkVoidFunction
1666 tu_GetDeviceProcAddr(VkDevice _device, const char *pName)
1667 {
1668    TU_FROM_HANDLE(tu_device, device, _device);
1669
1670    return tu_lookup_entrypoint_checked(pName, device->instance->api_version,
1671                                        &device->instance->enabled_extensions,
1672                                        &device->enabled_extensions);
1673 }
1674
1675 static VkResult
1676 tu_alloc_memory(struct tu_device *device,
1677                 const VkMemoryAllocateInfo *pAllocateInfo,
1678                 const VkAllocationCallbacks *pAllocator,
1679                 VkDeviceMemory *pMem)
1680 {
1681    struct tu_device_memory *mem;
1682    VkResult result;
1683
1684    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1685
1686    if (pAllocateInfo->allocationSize == 0) {
1687       /* Apparently, this is allowed */
1688       *pMem = VK_NULL_HANDLE;
1689       return VK_SUCCESS;
1690    }
1691
1692    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1693                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1694    if (mem == NULL)
1695       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1696
1697    const VkImportMemoryFdInfoKHR *fd_info =
1698       vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
1699    if (fd_info && !fd_info->handleType)
1700       fd_info = NULL;
1701
1702    if (fd_info) {
1703       assert(fd_info->handleType ==
1704                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
1705              fd_info->handleType ==
1706                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1707
1708       /*
1709        * TODO Importing the same fd twice gives us the same handle without
1710        * reference counting.  We need to maintain a per-instance handle-to-bo
1711        * table and add reference count to tu_bo.
1712        */
1713       result = tu_bo_init_dmabuf(device, &mem->bo,
1714                                  pAllocateInfo->allocationSize, fd_info->fd);
1715       if (result == VK_SUCCESS) {
1716          /* take ownership and close the fd */
1717          close(fd_info->fd);
1718       }
1719    } else {
1720       result =
1721          tu_bo_init_new(device, &mem->bo, pAllocateInfo->allocationSize);
1722    }
1723
1724    if (result != VK_SUCCESS) {
1725       vk_free2(&device->alloc, pAllocator, mem);
1726       return result;
1727    }
1728
1729    mem->size = pAllocateInfo->allocationSize;
1730    mem->type_index = pAllocateInfo->memoryTypeIndex;
1731
1732    mem->map = NULL;
1733    mem->user_ptr = NULL;
1734
1735    *pMem = tu_device_memory_to_handle(mem);
1736
1737    return VK_SUCCESS;
1738 }
1739
1740 VkResult
1741 tu_AllocateMemory(VkDevice _device,
1742                   const VkMemoryAllocateInfo *pAllocateInfo,
1743                   const VkAllocationCallbacks *pAllocator,
1744                   VkDeviceMemory *pMem)
1745 {
1746    TU_FROM_HANDLE(tu_device, device, _device);
1747    return tu_alloc_memory(device, pAllocateInfo, pAllocator, pMem);
1748 }
1749
1750 void
1751 tu_FreeMemory(VkDevice _device,
1752               VkDeviceMemory _mem,
1753               const VkAllocationCallbacks *pAllocator)
1754 {
1755    TU_FROM_HANDLE(tu_device, device, _device);
1756    TU_FROM_HANDLE(tu_device_memory, mem, _mem);
1757
1758    if (mem == NULL)
1759       return;
1760
1761    tu_bo_finish(device, &mem->bo);
1762    vk_free2(&device->alloc, pAllocator, mem);
1763 }
1764
1765 VkResult
1766 tu_MapMemory(VkDevice _device,
1767              VkDeviceMemory _memory,
1768              VkDeviceSize offset,
1769              VkDeviceSize size,
1770              VkMemoryMapFlags flags,
1771              void **ppData)
1772 {
1773    TU_FROM_HANDLE(tu_device, device, _device);
1774    TU_FROM_HANDLE(tu_device_memory, mem, _memory);
1775    VkResult result;
1776
1777    if (mem == NULL) {
1778       *ppData = NULL;
1779       return VK_SUCCESS;
1780    }
1781
1782    if (mem->user_ptr) {
1783       *ppData = mem->user_ptr;
1784    } else if (!mem->map) {
1785       result = tu_bo_map(device, &mem->bo);
1786       if (result != VK_SUCCESS)
1787          return result;
1788       *ppData = mem->map = mem->bo.map;
1789    } else
1790       *ppData = mem->map;
1791
1792    if (*ppData) {
1793       *ppData += offset;
1794       return VK_SUCCESS;
1795    }
1796
1797    return vk_error(device->instance, VK_ERROR_MEMORY_MAP_FAILED);
1798 }
1799
1800 void
1801 tu_UnmapMemory(VkDevice _device, VkDeviceMemory _memory)
1802 {
1803    /* I do not see any unmapping done by the freedreno Gallium driver. */
1804 }
1805
1806 VkResult
1807 tu_FlushMappedMemoryRanges(VkDevice _device,
1808                            uint32_t memoryRangeCount,
1809                            const VkMappedMemoryRange *pMemoryRanges)
1810 {
1811    return VK_SUCCESS;
1812 }
1813
1814 VkResult
1815 tu_InvalidateMappedMemoryRanges(VkDevice _device,
1816                                 uint32_t memoryRangeCount,
1817                                 const VkMappedMemoryRange *pMemoryRanges)
1818 {
1819    return VK_SUCCESS;
1820 }
1821
1822 void
1823 tu_GetBufferMemoryRequirements(VkDevice _device,
1824                                VkBuffer _buffer,
1825                                VkMemoryRequirements *pMemoryRequirements)
1826 {
1827    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
1828
1829    pMemoryRequirements->memoryTypeBits = 1;
1830    pMemoryRequirements->alignment = 64;
1831    pMemoryRequirements->size =
1832       align64(buffer->size, pMemoryRequirements->alignment);
1833 }
1834
1835 void
1836 tu_GetBufferMemoryRequirements2(
1837    VkDevice device,
1838    const VkBufferMemoryRequirementsInfo2 *pInfo,
1839    VkMemoryRequirements2 *pMemoryRequirements)
1840 {
1841    tu_GetBufferMemoryRequirements(device, pInfo->buffer,
1842                                   &pMemoryRequirements->memoryRequirements);
1843 }
1844
1845 void
1846 tu_GetImageMemoryRequirements(VkDevice _device,
1847                               VkImage _image,
1848                               VkMemoryRequirements *pMemoryRequirements)
1849 {
1850    TU_FROM_HANDLE(tu_image, image, _image);
1851
1852    pMemoryRequirements->memoryTypeBits = 1;
1853    pMemoryRequirements->size = image->layout.size;
1854    pMemoryRequirements->alignment = image->layout.base_align;
1855 }
1856
1857 void
1858 tu_GetImageMemoryRequirements2(VkDevice device,
1859                                const VkImageMemoryRequirementsInfo2 *pInfo,
1860                                VkMemoryRequirements2 *pMemoryRequirements)
1861 {
1862    tu_GetImageMemoryRequirements(device, pInfo->image,
1863                                  &pMemoryRequirements->memoryRequirements);
1864 }
1865
1866 void
1867 tu_GetImageSparseMemoryRequirements(
1868    VkDevice device,
1869    VkImage image,
1870    uint32_t *pSparseMemoryRequirementCount,
1871    VkSparseImageMemoryRequirements *pSparseMemoryRequirements)
1872 {
1873    tu_stub();
1874 }
1875
1876 void
1877 tu_GetImageSparseMemoryRequirements2(
1878    VkDevice device,
1879    const VkImageSparseMemoryRequirementsInfo2 *pInfo,
1880    uint32_t *pSparseMemoryRequirementCount,
1881    VkSparseImageMemoryRequirements2 *pSparseMemoryRequirements)
1882 {
1883    tu_stub();
1884 }
1885
1886 void
1887 tu_GetDeviceMemoryCommitment(VkDevice device,
1888                              VkDeviceMemory memory,
1889                              VkDeviceSize *pCommittedMemoryInBytes)
1890 {
1891    *pCommittedMemoryInBytes = 0;
1892 }
1893
1894 VkResult
1895 tu_BindBufferMemory2(VkDevice device,
1896                      uint32_t bindInfoCount,
1897                      const VkBindBufferMemoryInfo *pBindInfos)
1898 {
1899    for (uint32_t i = 0; i < bindInfoCount; ++i) {
1900       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
1901       TU_FROM_HANDLE(tu_buffer, buffer, pBindInfos[i].buffer);
1902
1903       if (mem) {
1904          buffer->bo = &mem->bo;
1905          buffer->bo_offset = pBindInfos[i].memoryOffset;
1906       } else {
1907          buffer->bo = NULL;
1908       }
1909    }
1910    return VK_SUCCESS;
1911 }
1912
1913 VkResult
1914 tu_BindBufferMemory(VkDevice device,
1915                     VkBuffer buffer,
1916                     VkDeviceMemory memory,
1917                     VkDeviceSize memoryOffset)
1918 {
1919    const VkBindBufferMemoryInfo info = {
1920       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
1921       .buffer = buffer,
1922       .memory = memory,
1923       .memoryOffset = memoryOffset
1924    };
1925
1926    return tu_BindBufferMemory2(device, 1, &info);
1927 }
1928
1929 VkResult
1930 tu_BindImageMemory2(VkDevice device,
1931                     uint32_t bindInfoCount,
1932                     const VkBindImageMemoryInfo *pBindInfos)
1933 {
1934    for (uint32_t i = 0; i < bindInfoCount; ++i) {
1935       TU_FROM_HANDLE(tu_image, image, pBindInfos[i].image);
1936       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
1937
1938       if (mem) {
1939          image->bo = &mem->bo;
1940          image->bo_offset = pBindInfos[i].memoryOffset;
1941       } else {
1942          image->bo = NULL;
1943          image->bo_offset = 0;
1944       }
1945    }
1946
1947    return VK_SUCCESS;
1948 }
1949
1950 VkResult
1951 tu_BindImageMemory(VkDevice device,
1952                    VkImage image,
1953                    VkDeviceMemory memory,
1954                    VkDeviceSize memoryOffset)
1955 {
1956    const VkBindImageMemoryInfo info = {
1957       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
1958       .image = image,
1959       .memory = memory,
1960       .memoryOffset = memoryOffset
1961    };
1962
1963    return tu_BindImageMemory2(device, 1, &info);
1964 }
1965
1966 VkResult
1967 tu_QueueBindSparse(VkQueue _queue,
1968                    uint32_t bindInfoCount,
1969                    const VkBindSparseInfo *pBindInfo,
1970                    VkFence _fence)
1971 {
1972    return VK_SUCCESS;
1973 }
1974
1975 // Queue semaphore functions
1976
1977 VkResult
1978 tu_CreateSemaphore(VkDevice _device,
1979                    const VkSemaphoreCreateInfo *pCreateInfo,
1980                    const VkAllocationCallbacks *pAllocator,
1981                    VkSemaphore *pSemaphore)
1982 {
1983    TU_FROM_HANDLE(tu_device, device, _device);
1984
1985    struct tu_semaphore *sem =
1986       vk_alloc2(&device->alloc, pAllocator, sizeof(*sem), 8,
1987                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1988    if (!sem)
1989       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1990
1991    *pSemaphore = tu_semaphore_to_handle(sem);
1992    return VK_SUCCESS;
1993 }
1994
1995 void
1996 tu_DestroySemaphore(VkDevice _device,
1997                     VkSemaphore _semaphore,
1998                     const VkAllocationCallbacks *pAllocator)
1999 {
2000    TU_FROM_HANDLE(tu_device, device, _device);
2001    TU_FROM_HANDLE(tu_semaphore, sem, _semaphore);
2002    if (!_semaphore)
2003       return;
2004
2005    vk_free2(&device->alloc, pAllocator, sem);
2006 }
2007
2008 VkResult
2009 tu_CreateEvent(VkDevice _device,
2010                const VkEventCreateInfo *pCreateInfo,
2011                const VkAllocationCallbacks *pAllocator,
2012                VkEvent *pEvent)
2013 {
2014    TU_FROM_HANDLE(tu_device, device, _device);
2015    struct tu_event *event =
2016       vk_alloc2(&device->alloc, pAllocator, sizeof(*event), 8,
2017                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2018
2019    if (!event)
2020       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2021
2022    VkResult result = tu_bo_init_new(device, &event->bo, 0x1000);
2023    if (result != VK_SUCCESS)
2024       goto fail_alloc;
2025
2026    result = tu_bo_map(device, &event->bo);
2027    if (result != VK_SUCCESS)
2028       goto fail_map;
2029
2030    *pEvent = tu_event_to_handle(event);
2031
2032    return VK_SUCCESS;
2033
2034 fail_map:
2035    tu_bo_finish(device, &event->bo);
2036 fail_alloc:
2037    vk_free2(&device->alloc, pAllocator, event);
2038    return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2039 }
2040
2041 void
2042 tu_DestroyEvent(VkDevice _device,
2043                 VkEvent _event,
2044                 const VkAllocationCallbacks *pAllocator)
2045 {
2046    TU_FROM_HANDLE(tu_device, device, _device);
2047    TU_FROM_HANDLE(tu_event, event, _event);
2048
2049    if (!event)
2050       return;
2051
2052    tu_bo_finish(device, &event->bo);
2053    vk_free2(&device->alloc, pAllocator, event);
2054 }
2055
2056 VkResult
2057 tu_GetEventStatus(VkDevice _device, VkEvent _event)
2058 {
2059    TU_FROM_HANDLE(tu_event, event, _event);
2060
2061    if (*(uint64_t*) event->bo.map == 1)
2062       return VK_EVENT_SET;
2063    return VK_EVENT_RESET;
2064 }
2065
2066 VkResult
2067 tu_SetEvent(VkDevice _device, VkEvent _event)
2068 {
2069    TU_FROM_HANDLE(tu_event, event, _event);
2070    *(uint64_t*) event->bo.map = 1;
2071
2072    return VK_SUCCESS;
2073 }
2074
2075 VkResult
2076 tu_ResetEvent(VkDevice _device, VkEvent _event)
2077 {
2078    TU_FROM_HANDLE(tu_event, event, _event);
2079    *(uint64_t*) event->bo.map = 0;
2080
2081    return VK_SUCCESS;
2082 }
2083
2084 VkResult
2085 tu_CreateBuffer(VkDevice _device,
2086                 const VkBufferCreateInfo *pCreateInfo,
2087                 const VkAllocationCallbacks *pAllocator,
2088                 VkBuffer *pBuffer)
2089 {
2090    TU_FROM_HANDLE(tu_device, device, _device);
2091    struct tu_buffer *buffer;
2092
2093    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2094
2095    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2096                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2097    if (buffer == NULL)
2098       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2099
2100    buffer->size = pCreateInfo->size;
2101    buffer->usage = pCreateInfo->usage;
2102    buffer->flags = pCreateInfo->flags;
2103
2104    *pBuffer = tu_buffer_to_handle(buffer);
2105
2106    return VK_SUCCESS;
2107 }
2108
2109 void
2110 tu_DestroyBuffer(VkDevice _device,
2111                  VkBuffer _buffer,
2112                  const VkAllocationCallbacks *pAllocator)
2113 {
2114    TU_FROM_HANDLE(tu_device, device, _device);
2115    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
2116
2117    if (!buffer)
2118       return;
2119
2120    vk_free2(&device->alloc, pAllocator, buffer);
2121 }
2122
2123 VkResult
2124 tu_CreateFramebuffer(VkDevice _device,
2125                      const VkFramebufferCreateInfo *pCreateInfo,
2126                      const VkAllocationCallbacks *pAllocator,
2127                      VkFramebuffer *pFramebuffer)
2128 {
2129    TU_FROM_HANDLE(tu_device, device, _device);
2130    struct tu_framebuffer *framebuffer;
2131
2132    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2133
2134    size_t size = sizeof(*framebuffer) + sizeof(struct tu_attachment_info) *
2135                                            pCreateInfo->attachmentCount;
2136    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2137                            VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2138    if (framebuffer == NULL)
2139       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2140
2141    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2142    framebuffer->width = pCreateInfo->width;
2143    framebuffer->height = pCreateInfo->height;
2144    framebuffer->layers = pCreateInfo->layers;
2145    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2146       VkImageView _iview = pCreateInfo->pAttachments[i];
2147       struct tu_image_view *iview = tu_image_view_from_handle(_iview);
2148       framebuffer->attachments[i].attachment = iview;
2149    }
2150
2151    *pFramebuffer = tu_framebuffer_to_handle(framebuffer);
2152    return VK_SUCCESS;
2153 }
2154
2155 void
2156 tu_DestroyFramebuffer(VkDevice _device,
2157                       VkFramebuffer _fb,
2158                       const VkAllocationCallbacks *pAllocator)
2159 {
2160    TU_FROM_HANDLE(tu_device, device, _device);
2161    TU_FROM_HANDLE(tu_framebuffer, fb, _fb);
2162
2163    if (!fb)
2164       return;
2165    vk_free2(&device->alloc, pAllocator, fb);
2166 }
2167
2168 static void
2169 tu_init_sampler(struct tu_device *device,
2170                 struct tu_sampler *sampler,
2171                 const VkSamplerCreateInfo *pCreateInfo)
2172 {
2173    const struct VkSamplerReductionModeCreateInfo *reduction =
2174       vk_find_struct_const(pCreateInfo->pNext, SAMPLER_REDUCTION_MODE_CREATE_INFO);
2175    const struct VkSamplerYcbcrConversionInfo *ycbcr_conversion =
2176       vk_find_struct_const(pCreateInfo->pNext,  SAMPLER_YCBCR_CONVERSION_INFO);
2177
2178    unsigned aniso = pCreateInfo->anisotropyEnable ?
2179       util_last_bit(MIN2((uint32_t)pCreateInfo->maxAnisotropy >> 1, 8)) : 0;
2180    bool miplinear = (pCreateInfo->mipmapMode == VK_SAMPLER_MIPMAP_MODE_LINEAR);
2181    float min_lod = CLAMP(pCreateInfo->minLod, 0.0f, 4095.0f / 256.0f);
2182    float max_lod = CLAMP(pCreateInfo->maxLod, 0.0f, 4095.0f / 256.0f);
2183
2184    sampler->descriptor[0] =
2185       COND(miplinear, A6XX_TEX_SAMP_0_MIPFILTER_LINEAR_NEAR) |
2186       A6XX_TEX_SAMP_0_XY_MAG(tu6_tex_filter(pCreateInfo->magFilter, aniso)) |
2187       A6XX_TEX_SAMP_0_XY_MIN(tu6_tex_filter(pCreateInfo->minFilter, aniso)) |
2188       A6XX_TEX_SAMP_0_ANISO(aniso) |
2189       A6XX_TEX_SAMP_0_WRAP_S(tu6_tex_wrap(pCreateInfo->addressModeU)) |
2190       A6XX_TEX_SAMP_0_WRAP_T(tu6_tex_wrap(pCreateInfo->addressModeV)) |
2191       A6XX_TEX_SAMP_0_WRAP_R(tu6_tex_wrap(pCreateInfo->addressModeW)) |
2192       A6XX_TEX_SAMP_0_LOD_BIAS(pCreateInfo->mipLodBias);
2193    sampler->descriptor[1] =
2194       /* COND(!cso->seamless_cube_map, A6XX_TEX_SAMP_1_CUBEMAPSEAMLESSFILTOFF) | */
2195       COND(pCreateInfo->unnormalizedCoordinates, A6XX_TEX_SAMP_1_UNNORM_COORDS) |
2196       A6XX_TEX_SAMP_1_MIN_LOD(min_lod) |
2197       A6XX_TEX_SAMP_1_MAX_LOD(max_lod) |
2198       COND(pCreateInfo->compareEnable,
2199            A6XX_TEX_SAMP_1_COMPARE_FUNC(tu6_compare_func(pCreateInfo->compareOp)));
2200    /* This is an offset into the border_color BO, which we fill with all the
2201     * possible Vulkan border colors in the correct order, so we can just use
2202     * the Vulkan enum with no translation necessary.
2203     */
2204    sampler->descriptor[2] =
2205       A6XX_TEX_SAMP_2_BCOLOR_OFFSET((unsigned) pCreateInfo->borderColor *
2206                                     sizeof(struct bcolor_entry));
2207    sampler->descriptor[3] = 0;
2208
2209    if (reduction) {
2210       sampler->descriptor[2] |= A6XX_TEX_SAMP_2_REDUCTION_MODE(
2211          tu6_reduction_mode(reduction->reductionMode));
2212    }
2213
2214    sampler->ycbcr_sampler = ycbcr_conversion ?
2215       tu_sampler_ycbcr_conversion_from_handle(ycbcr_conversion->conversion) : NULL;
2216
2217    if (sampler->ycbcr_sampler &&
2218        sampler->ycbcr_sampler->chroma_filter == VK_FILTER_LINEAR) {
2219       sampler->descriptor[2] |= A6XX_TEX_SAMP_2_CHROMA_LINEAR;
2220    }
2221
2222    /* TODO:
2223     * A6XX_TEX_SAMP_1_MIPFILTER_LINEAR_FAR disables mipmapping, but vk has no NONE mipfilter?
2224     */
2225 }
2226
2227 VkResult
2228 tu_CreateSampler(VkDevice _device,
2229                  const VkSamplerCreateInfo *pCreateInfo,
2230                  const VkAllocationCallbacks *pAllocator,
2231                  VkSampler *pSampler)
2232 {
2233    TU_FROM_HANDLE(tu_device, device, _device);
2234    struct tu_sampler *sampler;
2235
2236    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO);
2237
2238    sampler = vk_alloc2(&device->alloc, pAllocator, sizeof(*sampler), 8,
2239                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2240    if (!sampler)
2241       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2242
2243    tu_init_sampler(device, sampler, pCreateInfo);
2244    *pSampler = tu_sampler_to_handle(sampler);
2245
2246    return VK_SUCCESS;
2247 }
2248
2249 void
2250 tu_DestroySampler(VkDevice _device,
2251                   VkSampler _sampler,
2252                   const VkAllocationCallbacks *pAllocator)
2253 {
2254    TU_FROM_HANDLE(tu_device, device, _device);
2255    TU_FROM_HANDLE(tu_sampler, sampler, _sampler);
2256
2257    if (!sampler)
2258       return;
2259    vk_free2(&device->alloc, pAllocator, sampler);
2260 }
2261
2262 /* vk_icd.h does not declare this function, so we declare it here to
2263  * suppress Wmissing-prototypes.
2264  */
2265 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2266 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion);
2267
2268 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2269 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion)
2270 {
2271    /* For the full details on loader interface versioning, see
2272     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2273     * What follows is a condensed summary, to help you navigate the large and
2274     * confusing official doc.
2275     *
2276     *   - Loader interface v0 is incompatible with later versions. We don't
2277     *     support it.
2278     *
2279     *   - In loader interface v1:
2280     *       - The first ICD entrypoint called by the loader is
2281     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2282     *         entrypoint.
2283     *       - The ICD must statically expose no other Vulkan symbol unless it
2284     * is linked with -Bsymbolic.
2285     *       - Each dispatchable Vulkan handle created by the ICD must be
2286     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2287     *         ICD must initialize VK_LOADER_DATA.loadMagic to
2288     * ICD_LOADER_MAGIC.
2289     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2290     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2291     *         such loader-managed surfaces.
2292     *
2293     *    - Loader interface v2 differs from v1 in:
2294     *       - The first ICD entrypoint called by the loader is
2295     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2296     *         statically expose this entrypoint.
2297     *
2298     *    - Loader interface v3 differs from v2 in:
2299     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2300     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2301     *          because the loader no longer does so.
2302     */
2303    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2304    return VK_SUCCESS;
2305 }
2306
2307 VkResult
2308 tu_GetMemoryFdKHR(VkDevice _device,
2309                   const VkMemoryGetFdInfoKHR *pGetFdInfo,
2310                   int *pFd)
2311 {
2312    TU_FROM_HANDLE(tu_device, device, _device);
2313    TU_FROM_HANDLE(tu_device_memory, memory, pGetFdInfo->memory);
2314
2315    assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2316
2317    /* At the moment, we support only the below handle types. */
2318    assert(pGetFdInfo->handleType ==
2319              VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2320           pGetFdInfo->handleType ==
2321              VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2322
2323    int prime_fd = tu_bo_export_dmabuf(device, &memory->bo);
2324    if (prime_fd < 0)
2325       return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
2326
2327    *pFd = prime_fd;
2328    return VK_SUCCESS;
2329 }
2330
2331 VkResult
2332 tu_GetMemoryFdPropertiesKHR(VkDevice _device,
2333                             VkExternalMemoryHandleTypeFlagBits handleType,
2334                             int fd,
2335                             VkMemoryFdPropertiesKHR *pMemoryFdProperties)
2336 {
2337    assert(handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2338    pMemoryFdProperties->memoryTypeBits = 1;
2339    return VK_SUCCESS;
2340 }
2341
2342 void
2343 tu_GetPhysicalDeviceExternalSemaphoreProperties(
2344    VkPhysicalDevice physicalDevice,
2345    const VkPhysicalDeviceExternalSemaphoreInfo *pExternalSemaphoreInfo,
2346    VkExternalSemaphoreProperties *pExternalSemaphoreProperties)
2347 {
2348    pExternalSemaphoreProperties->exportFromImportedHandleTypes = 0;
2349    pExternalSemaphoreProperties->compatibleHandleTypes = 0;
2350    pExternalSemaphoreProperties->externalSemaphoreFeatures = 0;
2351 }
2352
2353 void
2354 tu_GetPhysicalDeviceExternalFenceProperties(
2355    VkPhysicalDevice physicalDevice,
2356    const VkPhysicalDeviceExternalFenceInfo *pExternalFenceInfo,
2357    VkExternalFenceProperties *pExternalFenceProperties)
2358 {
2359    pExternalFenceProperties->exportFromImportedHandleTypes = 0;
2360    pExternalFenceProperties->compatibleHandleTypes = 0;
2361    pExternalFenceProperties->externalFenceFeatures = 0;
2362 }
2363
2364 VkResult
2365 tu_CreateDebugReportCallbackEXT(
2366    VkInstance _instance,
2367    const VkDebugReportCallbackCreateInfoEXT *pCreateInfo,
2368    const VkAllocationCallbacks *pAllocator,
2369    VkDebugReportCallbackEXT *pCallback)
2370 {
2371    TU_FROM_HANDLE(tu_instance, instance, _instance);
2372    return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2373                                           pCreateInfo, pAllocator,
2374                                           &instance->alloc, pCallback);
2375 }
2376
2377 void
2378 tu_DestroyDebugReportCallbackEXT(VkInstance _instance,
2379                                  VkDebugReportCallbackEXT _callback,
2380                                  const VkAllocationCallbacks *pAllocator)
2381 {
2382    TU_FROM_HANDLE(tu_instance, instance, _instance);
2383    vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2384                                     _callback, pAllocator, &instance->alloc);
2385 }
2386
2387 void
2388 tu_DebugReportMessageEXT(VkInstance _instance,
2389                          VkDebugReportFlagsEXT flags,
2390                          VkDebugReportObjectTypeEXT objectType,
2391                          uint64_t object,
2392                          size_t location,
2393                          int32_t messageCode,
2394                          const char *pLayerPrefix,
2395                          const char *pMessage)
2396 {
2397    TU_FROM_HANDLE(tu_instance, instance, _instance);
2398    vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2399                    object, location, messageCode, pLayerPrefix, pMessage);
2400 }
2401
2402 void
2403 tu_GetDeviceGroupPeerMemoryFeatures(
2404    VkDevice device,
2405    uint32_t heapIndex,
2406    uint32_t localDeviceIndex,
2407    uint32_t remoteDeviceIndex,
2408    VkPeerMemoryFeatureFlags *pPeerMemoryFeatures)
2409 {
2410    assert(localDeviceIndex == remoteDeviceIndex);
2411
2412    *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2413                           VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2414                           VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2415                           VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2416 }
2417
2418 void tu_GetPhysicalDeviceMultisamplePropertiesEXT(
2419    VkPhysicalDevice                            physicalDevice,
2420    VkSampleCountFlagBits                       samples,
2421    VkMultisamplePropertiesEXT*                 pMultisampleProperties)
2422 {
2423    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
2424
2425    if (samples <= VK_SAMPLE_COUNT_4_BIT && pdevice->supported_extensions.EXT_sample_locations)
2426       pMultisampleProperties->maxSampleLocationGridSize = (VkExtent2D){ 1, 1 };
2427    else
2428       pMultisampleProperties->maxSampleLocationGridSize = (VkExtent2D){ 0, 0 };
2429 }