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/strtod.h"
  43 #include "vk_format.h"
  44 #include "vk_util.h"
  45
  46 #include "drm-uapi/msm_drm.h"
  47
  48 static int
  49 tu_device_get_cache_uuid(uint16_t family, void *uuid)
  50 {
  51    uint32_t mesa_timestamp;
  52    uint16_t f = family;
  53    memset(uuid, 0, VK_UUID_SIZE);
  54    if (!disk_cache_get_function_timestamp(tu_device_get_cache_uuid,
  55                                           &mesa_timestamp))
  56       return -1;
  57
  58    memcpy(uuid, &mesa_timestamp, 4);
  59    memcpy((char *) uuid + 4, &f, 2);
  60    snprintf((char *) uuid + 6, VK_UUID_SIZE - 10, "tu");
  61    return 0;
  62 }
  63
  64 static void
  65 tu_get_driver_uuid(void *uuid)
  66 {
  67    memset(uuid, 0, VK_UUID_SIZE);
  68    snprintf(uuid, VK_UUID_SIZE, "freedreno");
  69 }
  70
  71 static void
  72 tu_get_device_uuid(void *uuid)
  73 {
  74    memset(uuid, 0, VK_UUID_SIZE);
  75 }
  76
  77 static VkResult
  78 tu_bo_init(struct tu_device *dev,
  79            struct tu_bo *bo,
  80            uint32_t gem_handle,
  81            uint64_t size)
  82 {
  83    uint64_t iova = tu_gem_info_iova(dev, gem_handle);
  84    if (!iova)
  85       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
  86
  87    *bo = (struct tu_bo) {
  88       .gem_handle = gem_handle,
  89       .size = size,
  90       .iova = iova,
  91    };
  92
  93    return VK_SUCCESS;
  94 }
  95
  96 VkResult
  97 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size)
  98 {
  99    /* TODO: Choose better flags. As of 2018-11-12, freedreno/drm/msm_bo.c
 100     * always sets `flags = MSM_BO_WC`, and we copy that behavior here.
 101     */
 102    uint32_t gem_handle = tu_gem_new(dev, size, MSM_BO_WC);
 103    if (!gem_handle)
 104       return vk_error(dev->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
 105
 106    VkResult result = tu_bo_init(dev, bo, gem_handle, size);
 107    if (result != VK_SUCCESS) {
 108       tu_gem_close(dev, gem_handle);
 109       return vk_error(dev->instance, result);
 110    }
 111
 112    return VK_SUCCESS;
 113 }
 114
 115 VkResult
 116 tu_bo_init_dmabuf(struct tu_device *dev,
 117                   struct tu_bo *bo,
 118                   uint64_t size,
 119                   int fd)
 120 {
 121    uint32_t gem_handle = tu_gem_import_dmabuf(dev, fd, size);
 122    if (!gem_handle)
 123       return vk_error(dev->instance, VK_ERROR_INVALID_EXTERNAL_HANDLE);
 124
 125    VkResult result = tu_bo_init(dev, bo, gem_handle, size);
 126    if (result != VK_SUCCESS) {
 127       tu_gem_close(dev, gem_handle);
 128       return vk_error(dev->instance, result);
 129    }
 130
 131    return VK_SUCCESS;
 132 }
 133
 134 int
 135 tu_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo)
 136 {
 137    return tu_gem_export_dmabuf(dev, bo->gem_handle);
 138 }
 139
 140 VkResult
 141 tu_bo_map(struct tu_device *dev, struct tu_bo *bo)
 142 {
 143    if (bo->map)
 144       return VK_SUCCESS;
 145
 146    uint64_t offset = tu_gem_info_offset(dev, bo->gem_handle);
 147    if (!offset)
 148       return vk_error(dev->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
 149
 150    /* TODO: Should we use the wrapper os_mmap() like Freedreno does? */
 151    void *map = mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
 152                     dev->physical_device->local_fd, offset);
 153    if (map == MAP_FAILED)
 154       return vk_error(dev->instance, VK_ERROR_MEMORY_MAP_FAILED);
 155
 156    bo->map = map;
 157    return VK_SUCCESS;
 158 }
 159
 160 void
 161 tu_bo_finish(struct tu_device *dev, struct tu_bo *bo)
 162 {
 163    assert(bo->gem_handle);
 164
 165    if (bo->map)
 166       munmap(bo->map, bo->size);
 167
 168    tu_gem_close(dev, bo->gem_handle);
 169 }
 170
 171 static VkResult
 172 tu_physical_device_init(struct tu_physical_device *device,
 173                         struct tu_instance *instance,
 174                         drmDevicePtr drm_device)
 175 {
 176    const char *path = drm_device->nodes[DRM_NODE_RENDER];
 177    VkResult result = VK_SUCCESS;
 178    drmVersionPtr version;
 179    int fd;
 180    int master_fd = -1;
 181
 182    fd = open(path, O_RDWR | O_CLOEXEC);
 183    if (fd < 0) {
 184       return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 185                        "failed to open device %s", path);
 186    }
 187
 188    /* Version 1.3 added MSM_INFO_IOVA. */
 189    const int min_version_major = 1;
 190    const int min_version_minor = 3;
 191
 192    version = drmGetVersion(fd);
 193    if (!version) {
 194       close(fd);
 195       return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 196                        "failed to query kernel driver version for device %s",
 197                        path);
 198    }
 199
 200    if (strcmp(version->name, "msm")) {
 201       drmFreeVersion(version);
 202       close(fd);
 203       return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 204                        "device %s does not use the msm kernel driver", path);
 205    }
 206
 207    if (version->version_major != min_version_major ||
 208        version->version_minor < min_version_minor) {
 209       result = vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 210                          "kernel driver for device %s has version %d.%d, "
 211                          "but Vulkan requires version >= %d.%d",
 212                          path, version->version_major, version->version_minor,
 213                          min_version_major, min_version_minor);
 214       drmFreeVersion(version);
 215       close(fd);
 216       return result;
 217    }
 218
 219    drmFreeVersion(version);
 220
 221    if (instance->debug_flags & TU_DEBUG_STARTUP)
 222       tu_logi("Found compatible device '%s'.", path);
 223
 224    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 225    device->instance = instance;
 226    assert(strlen(path) < ARRAY_SIZE(device->path));
 227    strncpy(device->path, path, ARRAY_SIZE(device->path));
 228
 229    if (instance->enabled_extensions.KHR_display) {
 230       master_fd =
 231          open(drm_device->nodes[DRM_NODE_PRIMARY], O_RDWR | O_CLOEXEC);
 232       if (master_fd >= 0) {
 233          /* TODO: free master_fd is accel is not working? */
 234       }
 235    }
 236
 237    device->master_fd = master_fd;
 238    device->local_fd = fd;
 239
 240    if (tu_drm_get_gpu_id(device, &device->gpu_id)) {
 241       if (instance->debug_flags & TU_DEBUG_STARTUP)
 242          tu_logi("Could not query the GPU ID");
 243       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 244                          "could not get GPU ID");
 245       goto fail;
 246    }
 247
 248    if (tu_drm_get_gmem_size(device, &device->gmem_size)) {
 249       if (instance->debug_flags & TU_DEBUG_STARTUP)
 250          tu_logi("Could not query the GMEM size");
 251       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 252                          "could not get GMEM size");
 253       goto fail;
 254    }
 255
 256    memset(device->name, 0, sizeof(device->name));
 257    sprintf(device->name, "FD%d", device->gpu_id);
 258
 259    switch (device->gpu_id) {
 260    case 630:
 261       device->tile_align_w = 32;
 262       device->tile_align_h = 32;
 263       break;
 264    default:
 265       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 266                          "device %s is unsupported", device->name);
 267       goto fail;
 268    }
 269    if (tu_device_get_cache_uuid(device->gpu_id, device->cache_uuid)) {
 270       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 271                          "cannot generate UUID");
 272       goto fail;
 273    }
 274
 275    /* The gpu id is already embedded in the uuid so we just pass "tu"
 276     * when creating the cache.
 277     */
 278    char buf[VK_UUID_SIZE * 2 + 1];
 279    disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
 280    device->disk_cache = disk_cache_create(device->name, buf, 0);
 281
 282    fprintf(stderr, "WARNING: tu is not a conformant vulkan implementation, "
 283                    "testing use only.\n");
 284
 285    tu_get_driver_uuid(&device->device_uuid);
 286    tu_get_device_uuid(&device->device_uuid);
 287
 288    tu_fill_device_extension_table(device, &device->supported_extensions);
 289
 290    if (result != VK_SUCCESS) {
 291       vk_error(instance, result);
 292       goto fail;
 293    }
 294
 295    result = tu_wsi_init(device);
 296    if (result != VK_SUCCESS) {
 297       vk_error(instance, result);
 298       goto fail;
 299    }
 300
 301    return VK_SUCCESS;
 302
 303 fail:
 304    close(fd);
 305    if (master_fd != -1)
 306       close(master_fd);
 307    return result;
 308 }
 309
 310 static void
 311 tu_physical_device_finish(struct tu_physical_device *device)
 312 {
 313    tu_wsi_finish(device);
 314
 315    disk_cache_destroy(device->disk_cache);
 316    close(device->local_fd);
 317    if (device->master_fd != -1)
 318       close(device->master_fd);
 319 }
 320
 321 static void *
 322 default_alloc_func(void *pUserData,
 323                    size_t size,
 324                    size_t align,
 325                    VkSystemAllocationScope allocationScope)
 326 {
 327    return malloc(size);
 328 }
 329
 330 static void *
 331 default_realloc_func(void *pUserData,
 332                      void *pOriginal,
 333                      size_t size,
 334                      size_t align,
 335                      VkSystemAllocationScope allocationScope)
 336 {
 337    return realloc(pOriginal, size);
 338 }
 339
 340 static void
 341 default_free_func(void *pUserData, void *pMemory)
 342 {
 343    free(pMemory);
 344 }
 345
 346 static const VkAllocationCallbacks default_alloc = {
 347    .pUserData = NULL,
 348    .pfnAllocation = default_alloc_func,
 349    .pfnReallocation = default_realloc_func,
 350    .pfnFree = default_free_func,
 351 };
 352
 353 static const struct debug_control tu_debug_options[] = {
 354    { "startup", TU_DEBUG_STARTUP },
 355    { "nir", TU_DEBUG_NIR },
 356    { "ir3", TU_DEBUG_IR3 },
 357    { NULL, 0 }
 358 };
 359
 360 const char *
 361 tu_get_debug_option_name(int id)
 362 {
 363    assert(id < ARRAY_SIZE(tu_debug_options) - 1);
 364    return tu_debug_options[id].string;
 365 }
 366
 367 static int
 368 tu_get_instance_extension_index(const char *name)
 369 {
 370    for (unsigned i = 0; i < TU_INSTANCE_EXTENSION_COUNT; ++i) {
 371       if (strcmp(name, tu_instance_extensions[i].extensionName) == 0)
 372          return i;
 373    }
 374    return -1;
 375 }
 376
 377 VkResult
 378 tu_CreateInstance(const VkInstanceCreateInfo *pCreateInfo,
 379                   const VkAllocationCallbacks *pAllocator,
 380                   VkInstance *pInstance)
 381 {
 382    struct tu_instance *instance;
 383    VkResult result;
 384
 385    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 386
 387    uint32_t client_version;
 388    if (pCreateInfo->pApplicationInfo &&
 389        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 390       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 391    } else {
 392       tu_EnumerateInstanceVersion(&client_version);
 393    }
 394
 395    instance = vk_zalloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 396                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 397    if (!instance)
 398       return vk_error(NULL, VK_ERROR_OUT_OF_HOST_MEMORY);
 399
 400    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 401
 402    if (pAllocator)
 403       instance->alloc = *pAllocator;
 404    else
 405       instance->alloc = default_alloc;
 406
 407    instance->api_version = client_version;
 408    instance->physical_device_count = -1;
 409
 410    instance->debug_flags =
 411       parse_debug_string(getenv("TU_DEBUG"), tu_debug_options);
 412
 413    if (instance->debug_flags & TU_DEBUG_STARTUP)
 414       tu_logi("Created an instance");
 415
 416    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 417       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
 418       int index = tu_get_instance_extension_index(ext_name);
 419
 420       if (index < 0 || !tu_supported_instance_extensions.extensions[index]) {
 421          vk_free2(&default_alloc, pAllocator, instance);
 422          return vk_error(instance, VK_ERROR_EXTENSION_NOT_PRESENT);
 423       }
 424
 425       instance->enabled_extensions.extensions[index] = true;
 426    }
 427
 428    result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
 429    if (result != VK_SUCCESS) {
 430       vk_free2(&default_alloc, pAllocator, instance);
 431       return vk_error(instance, result);
 432    }
 433
 434    _mesa_locale_init();
 435    glsl_type_singleton_init_or_ref();
 436
 437    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 438
 439    *pInstance = tu_instance_to_handle(instance);
 440
 441    return VK_SUCCESS;
 442 }
 443
 444 void
 445 tu_DestroyInstance(VkInstance _instance,
 446                    const VkAllocationCallbacks *pAllocator)
 447 {
 448    TU_FROM_HANDLE(tu_instance, instance, _instance);
 449
 450    if (!instance)
 451       return;
 452
 453    for (int i = 0; i < instance->physical_device_count; ++i) {
 454       tu_physical_device_finish(instance->physical_devices + i);
 455    }
 456
 457    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 458
 459    glsl_type_singleton_decref();
 460    _mesa_locale_fini();
 461
 462    vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
 463
 464    vk_free(&instance->alloc, instance);
 465 }
 466
 467 static VkResult
 468 tu_enumerate_devices(struct tu_instance *instance)
 469 {
 470    /* TODO: Check for more devices ? */
 471    drmDevicePtr devices[8];
 472    VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
 473    int max_devices;
 474
 475    instance->physical_device_count = 0;
 476
 477    max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
 478
 479    if (instance->debug_flags & TU_DEBUG_STARTUP)
 480       tu_logi("Found %d drm nodes", max_devices);
 481
 482    if (max_devices < 1)
 483       return vk_error(instance, VK_ERROR_INCOMPATIBLE_DRIVER);
 484
 485    for (unsigned i = 0; i < (unsigned) max_devices; i++) {
 486       if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
 487           devices[i]->bustype == DRM_BUS_PLATFORM) {
 488
 489          result = tu_physical_device_init(
 490             instance->physical_devices + instance->physical_device_count,
 491             instance, devices[i]);
 492          if (result == VK_SUCCESS)
 493             ++instance->physical_device_count;
 494          else if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 495             break;
 496       }
 497    }
 498    drmFreeDevices(devices, max_devices);
 499
 500    return result;
 501 }
 502
 503 VkResult
 504 tu_EnumeratePhysicalDevices(VkInstance _instance,
 505                             uint32_t *pPhysicalDeviceCount,
 506                             VkPhysicalDevice *pPhysicalDevices)
 507 {
 508    TU_FROM_HANDLE(tu_instance, instance, _instance);
 509    VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
 510
 511    VkResult result;
 512
 513    if (instance->physical_device_count < 0) {
 514       result = tu_enumerate_devices(instance);
 515       if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
 516          return result;
 517    }
 518
 519    for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
 520       vk_outarray_append(&out, p)
 521       {
 522          *p = tu_physical_device_to_handle(instance->physical_devices + i);
 523       }
 524    }
 525
 526    return vk_outarray_status(&out);
 527 }
 528
 529 VkResult
 530 tu_EnumeratePhysicalDeviceGroups(
 531    VkInstance _instance,
 532    uint32_t *pPhysicalDeviceGroupCount,
 533    VkPhysicalDeviceGroupProperties *pPhysicalDeviceGroupProperties)
 534 {
 535    TU_FROM_HANDLE(tu_instance, instance, _instance);
 536    VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
 537                     pPhysicalDeviceGroupCount);
 538    VkResult result;
 539
 540    if (instance->physical_device_count < 0) {
 541       result = tu_enumerate_devices(instance);
 542       if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
 543          return result;
 544    }
 545
 546    for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
 547       vk_outarray_append(&out, p)
 548       {
 549          p->physicalDeviceCount = 1;
 550          p->physicalDevices[0] =
 551             tu_physical_device_to_handle(instance->physical_devices + i);
 552          p->subsetAllocation = false;
 553       }
 554    }
 555
 556    return vk_outarray_status(&out);
 557 }
 558
 559 void
 560 tu_GetPhysicalDeviceFeatures(VkPhysicalDevice physicalDevice,
 561                              VkPhysicalDeviceFeatures *pFeatures)
 562 {
 563    memset(pFeatures, 0, sizeof(*pFeatures));
 564
 565    *pFeatures = (VkPhysicalDeviceFeatures) {
 566       .robustBufferAccess = false,
 567       .fullDrawIndexUint32 = false,
 568       .imageCubeArray = false,
 569       .independentBlend = false,
 570       .geometryShader = false,
 571       .tessellationShader = false,
 572       .sampleRateShading = false,
 573       .dualSrcBlend = false,
 574       .logicOp = false,
 575       .multiDrawIndirect = false,
 576       .drawIndirectFirstInstance = false,
 577       .depthClamp = false,
 578       .depthBiasClamp = false,
 579       .fillModeNonSolid = false,
 580       .depthBounds = false,
 581       .wideLines = false,
 582       .largePoints = false,
 583       .alphaToOne = false,
 584       .multiViewport = false,
 585       .samplerAnisotropy = true,
 586       .textureCompressionETC2 = true,
 587       .textureCompressionASTC_LDR = true,
 588       .textureCompressionBC = true,
 589       .occlusionQueryPrecise = false,
 590       .pipelineStatisticsQuery = false,
 591       .vertexPipelineStoresAndAtomics = false,
 592       .fragmentStoresAndAtomics = false,
 593       .shaderTessellationAndGeometryPointSize = false,
 594       .shaderImageGatherExtended = false,
 595       .shaderStorageImageExtendedFormats = false,
 596       .shaderStorageImageMultisample = false,
 597       .shaderUniformBufferArrayDynamicIndexing = false,
 598       .shaderSampledImageArrayDynamicIndexing = false,
 599       .shaderStorageBufferArrayDynamicIndexing = false,
 600       .shaderStorageImageArrayDynamicIndexing = false,
 601       .shaderStorageImageReadWithoutFormat = false,
 602       .shaderStorageImageWriteWithoutFormat = false,
 603       .shaderClipDistance = false,
 604       .shaderCullDistance = false,
 605       .shaderFloat64 = false,
 606       .shaderInt64 = false,
 607       .shaderInt16 = false,
 608       .sparseBinding = false,
 609       .variableMultisampleRate = false,
 610       .inheritedQueries = false,
 611    };
 612 }
 613
 614 void
 615 tu_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,
 616                               VkPhysicalDeviceFeatures2 *pFeatures)
 617 {
 618    vk_foreach_struct(ext, pFeatures->pNext)
 619    {
 620       switch (ext->sType) {
 621       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
 622          VkPhysicalDeviceVariablePointersFeatures *features = (void *) ext;
 623          features->variablePointersStorageBuffer = false;
 624          features->variablePointers = false;
 625          break;
 626       }
 627       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
 628          VkPhysicalDeviceMultiviewFeatures *features =
 629             (VkPhysicalDeviceMultiviewFeatures *) ext;
 630          features->multiview = false;
 631          features->multiviewGeometryShader = false;
 632          features->multiviewTessellationShader = false;
 633          break;
 634       }
 635       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
 636          VkPhysicalDeviceShaderDrawParametersFeatures *features =
 637             (VkPhysicalDeviceShaderDrawParametersFeatures *) ext;
 638          features->shaderDrawParameters = false;
 639          break;
 640       }
 641       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
 642          VkPhysicalDeviceProtectedMemoryFeatures *features =
 643             (VkPhysicalDeviceProtectedMemoryFeatures *) ext;
 644          features->protectedMemory = false;
 645          break;
 646       }
 647       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
 648          VkPhysicalDevice16BitStorageFeatures *features =
 649             (VkPhysicalDevice16BitStorageFeatures *) ext;
 650          features->storageBuffer16BitAccess = false;
 651          features->uniformAndStorageBuffer16BitAccess = false;
 652          features->storagePushConstant16 = false;
 653          features->storageInputOutput16 = false;
 654          break;
 655       }
 656       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
 657          VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
 658             (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
 659          features->samplerYcbcrConversion = false;
 660          break;
 661       }
 662       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
 663          VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
 664             (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *) ext;
 665          features->shaderInputAttachmentArrayDynamicIndexing = false;
 666          features->shaderUniformTexelBufferArrayDynamicIndexing = false;
 667          features->shaderStorageTexelBufferArrayDynamicIndexing = false;
 668          features->shaderUniformBufferArrayNonUniformIndexing = false;
 669          features->shaderSampledImageArrayNonUniformIndexing = false;
 670          features->shaderStorageBufferArrayNonUniformIndexing = false;
 671          features->shaderStorageImageArrayNonUniformIndexing = false;
 672          features->shaderInputAttachmentArrayNonUniformIndexing = false;
 673          features->shaderUniformTexelBufferArrayNonUniformIndexing = false;
 674          features->shaderStorageTexelBufferArrayNonUniformIndexing = false;
 675          features->descriptorBindingUniformBufferUpdateAfterBind = false;
 676          features->descriptorBindingSampledImageUpdateAfterBind = false;
 677          features->descriptorBindingStorageImageUpdateAfterBind = false;
 678          features->descriptorBindingStorageBufferUpdateAfterBind = false;
 679          features->descriptorBindingUniformTexelBufferUpdateAfterBind = false;
 680          features->descriptorBindingStorageTexelBufferUpdateAfterBind = false;
 681          features->descriptorBindingUpdateUnusedWhilePending = false;
 682          features->descriptorBindingPartiallyBound = false;
 683          features->descriptorBindingVariableDescriptorCount = false;
 684          features->runtimeDescriptorArray = false;
 685          break;
 686       }
 687       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
 688          VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
 689             (VkPhysicalDeviceConditionalRenderingFeaturesEXT *) ext;
 690          features->conditionalRendering = false;
 691          features->inheritedConditionalRendering = false;
 692          break;
 693       }
 694       default:
 695          break;
 696       }
 697    }
 698    return tu_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 699 }
 700
 701 void
 702 tu_GetPhysicalDeviceProperties(VkPhysicalDevice physicalDevice,
 703                                VkPhysicalDeviceProperties *pProperties)
 704 {
 705    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 706    VkSampleCountFlags sample_counts = VK_SAMPLE_COUNT_1_BIT |
 707       VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT | VK_SAMPLE_COUNT_8_BIT;
 708
 709    /* make sure that the entire descriptor set is addressable with a signed
 710     * 32-bit int. So the sum of all limits scaled by descriptor size has to
 711     * be at most 2 GiB. the combined image & samples object count as one of
 712     * both. This limit is for the pipeline layout, not for the set layout, but
 713     * there is no set limit, so we just set a pipeline limit. I don't think
 714     * any app is going to hit this soon. */
 715    size_t max_descriptor_set_size =
 716       ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS) /
 717       (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
 718        32 /* storage buffer, 32 due to potential space wasted on alignment */ +
 719        32 /* sampler, largest when combined with image */ +
 720        64 /* sampled image */ + 64 /* storage image */);
 721
 722    VkPhysicalDeviceLimits limits = {
 723       .maxImageDimension1D = (1 << 14),
 724       .maxImageDimension2D = (1 << 14),
 725       .maxImageDimension3D = (1 << 11),
 726       .maxImageDimensionCube = (1 << 14),
 727       .maxImageArrayLayers = (1 << 11),
 728       .maxTexelBufferElements = 128 * 1024 * 1024,
 729       .maxUniformBufferRange = UINT32_MAX,
 730       .maxStorageBufferRange = UINT32_MAX,
 731       .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
 732       .maxMemoryAllocationCount = UINT32_MAX,
 733       .maxSamplerAllocationCount = 64 * 1024,
 734       .bufferImageGranularity = 64,          /* A cache line */
 735       .sparseAddressSpaceSize = 0xffffffffu, /* buffer max size */
 736       .maxBoundDescriptorSets = MAX_SETS,
 737       .maxPerStageDescriptorSamplers = max_descriptor_set_size,
 738       .maxPerStageDescriptorUniformBuffers = max_descriptor_set_size,
 739       .maxPerStageDescriptorStorageBuffers = max_descriptor_set_size,
 740       .maxPerStageDescriptorSampledImages = max_descriptor_set_size,
 741       .maxPerStageDescriptorStorageImages = max_descriptor_set_size,
 742       .maxPerStageDescriptorInputAttachments = max_descriptor_set_size,
 743       .maxPerStageResources = max_descriptor_set_size,
 744       .maxDescriptorSetSamplers = max_descriptor_set_size,
 745       .maxDescriptorSetUniformBuffers = max_descriptor_set_size,
 746       .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS,
 747       .maxDescriptorSetStorageBuffers = max_descriptor_set_size,
 748       .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS,
 749       .maxDescriptorSetSampledImages = max_descriptor_set_size,
 750       .maxDescriptorSetStorageImages = max_descriptor_set_size,
 751       .maxDescriptorSetInputAttachments = max_descriptor_set_size,
 752       .maxVertexInputAttributes = 32,
 753       .maxVertexInputBindings = 32,
 754       .maxVertexInputAttributeOffset = 2047,
 755       .maxVertexInputBindingStride = 2048,
 756       .maxVertexOutputComponents = 128,
 757       .maxTessellationGenerationLevel = 64,
 758       .maxTessellationPatchSize = 32,
 759       .maxTessellationControlPerVertexInputComponents = 128,
 760       .maxTessellationControlPerVertexOutputComponents = 128,
 761       .maxTessellationControlPerPatchOutputComponents = 120,
 762       .maxTessellationControlTotalOutputComponents = 4096,
 763       .maxTessellationEvaluationInputComponents = 128,
 764       .maxTessellationEvaluationOutputComponents = 128,
 765       .maxGeometryShaderInvocations = 127,
 766       .maxGeometryInputComponents = 64,
 767       .maxGeometryOutputComponents = 128,
 768       .maxGeometryOutputVertices = 256,
 769       .maxGeometryTotalOutputComponents = 1024,
 770       .maxFragmentInputComponents = 128,
 771       .maxFragmentOutputAttachments = 8,
 772       .maxFragmentDualSrcAttachments = 1,
 773       .maxFragmentCombinedOutputResources = 8,
 774       .maxComputeSharedMemorySize = 32768,
 775       .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
 776       .maxComputeWorkGroupInvocations = 2048,
 777       .maxComputeWorkGroupSize = { 2048, 2048, 2048 },
 778       .subPixelPrecisionBits = 4 /* FIXME */,
 779       .subTexelPrecisionBits = 4 /* FIXME */,
 780       .mipmapPrecisionBits = 4 /* FIXME */,
 781       .maxDrawIndexedIndexValue = UINT32_MAX,
 782       .maxDrawIndirectCount = UINT32_MAX,
 783       .maxSamplerLodBias = 16,
 784       .maxSamplerAnisotropy = 16,
 785       .maxViewports = MAX_VIEWPORTS,
 786       .maxViewportDimensions = { (1 << 14), (1 << 14) },
 787       .viewportBoundsRange = { INT16_MIN, INT16_MAX },
 788       .viewportSubPixelBits = 8,
 789       .minMemoryMapAlignment = 4096, /* A page */
 790       .minTexelBufferOffsetAlignment = 1,
 791       .minUniformBufferOffsetAlignment = 4,
 792       .minStorageBufferOffsetAlignment = 4,
 793       .minTexelOffset = -32,
 794       .maxTexelOffset = 31,
 795       .minTexelGatherOffset = -32,
 796       .maxTexelGatherOffset = 31,
 797       .minInterpolationOffset = -2,
 798       .maxInterpolationOffset = 2,
 799       .subPixelInterpolationOffsetBits = 8,
 800       .maxFramebufferWidth = (1 << 14),
 801       .maxFramebufferHeight = (1 << 14),
 802       .maxFramebufferLayers = (1 << 10),
 803       .framebufferColorSampleCounts = sample_counts,
 804       .framebufferDepthSampleCounts = sample_counts,
 805       .framebufferStencilSampleCounts = sample_counts,
 806       .framebufferNoAttachmentsSampleCounts = sample_counts,
 807       .maxColorAttachments = MAX_RTS,
 808       .sampledImageColorSampleCounts = sample_counts,
 809       .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 810       .sampledImageDepthSampleCounts = sample_counts,
 811       .sampledImageStencilSampleCounts = sample_counts,
 812       .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 813       .maxSampleMaskWords = 1,
 814       .timestampComputeAndGraphics = true,
 815       .timestampPeriod = 1,
 816       .maxClipDistances = 8,
 817       .maxCullDistances = 8,
 818       .maxCombinedClipAndCullDistances = 8,
 819       .discreteQueuePriorities = 1,
 820       .pointSizeRange = { 0.125, 255.875 },
 821       .lineWidthRange = { 0.0, 7.9921875 },
 822       .pointSizeGranularity = (1.0 / 8.0),
 823       .lineWidthGranularity = (1.0 / 128.0),
 824       .strictLines = false, /* FINISHME */
 825       .standardSampleLocations = true,
 826       .optimalBufferCopyOffsetAlignment = 128,
 827       .optimalBufferCopyRowPitchAlignment = 128,
 828       .nonCoherentAtomSize = 64,
 829    };
 830
 831    *pProperties = (VkPhysicalDeviceProperties) {
 832       .apiVersion = tu_physical_device_api_version(pdevice),
 833       .driverVersion = vk_get_driver_version(),
 834       .vendorID = 0, /* TODO */
 835       .deviceID = 0,
 836       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 837       .limits = limits,
 838       .sparseProperties = { 0 },
 839    };
 840
 841    strcpy(pProperties->deviceName, pdevice->name);
 842    memcpy(pProperties->pipelineCacheUUID, pdevice->cache_uuid, VK_UUID_SIZE);
 843 }
 844
 845 void
 846 tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
 847                                 VkPhysicalDeviceProperties2 *pProperties)
 848 {
 849    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 850    tu_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 851
 852    vk_foreach_struct(ext, pProperties->pNext)
 853    {
 854       switch (ext->sType) {
 855       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 856          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 857             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 858          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 859          break;
 860       }
 861       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
 862          VkPhysicalDeviceIDProperties *properties =
 863             (VkPhysicalDeviceIDProperties *) ext;
 864          memcpy(properties->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
 865          memcpy(properties->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
 866          properties->deviceLUIDValid = false;
 867          break;
 868       }
 869       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
 870          VkPhysicalDeviceMultiviewProperties *properties =
 871             (VkPhysicalDeviceMultiviewProperties *) ext;
 872          properties->maxMultiviewViewCount = MAX_VIEWS;
 873          properties->maxMultiviewInstanceIndex = INT_MAX;
 874          break;
 875       }
 876       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
 877          VkPhysicalDevicePointClippingProperties *properties =
 878             (VkPhysicalDevicePointClippingProperties *) ext;
 879          properties->pointClippingBehavior =
 880             VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
 881          break;
 882       }
 883       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
 884          VkPhysicalDeviceMaintenance3Properties *properties =
 885             (VkPhysicalDeviceMaintenance3Properties *) ext;
 886          /* Make sure everything is addressable by a signed 32-bit int, and
 887           * our largest descriptors are 96 bytes. */
 888          properties->maxPerSetDescriptors = (1ull << 31) / 96;
 889          /* Our buffer size fields allow only this much */
 890          properties->maxMemoryAllocationSize = 0xFFFFFFFFull;
 891          break;
 892       }
 893       default:
 894          break;
 895       }
 896    }
 897 }
 898
 899 static const VkQueueFamilyProperties tu_queue_family_properties = {
 900    .queueFlags =
 901       VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT,
 902    .queueCount = 1,
 903    .timestampValidBits = 64,
 904    .minImageTransferGranularity = { 1, 1, 1 },
 905 };
 906
 907 void
 908 tu_GetPhysicalDeviceQueueFamilyProperties(
 909    VkPhysicalDevice physicalDevice,
 910    uint32_t *pQueueFamilyPropertyCount,
 911    VkQueueFamilyProperties *pQueueFamilyProperties)
 912 {
 913    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 914
 915    vk_outarray_append(&out, p) { *p = tu_queue_family_properties; }
 916 }
 917
 918 void
 919 tu_GetPhysicalDeviceQueueFamilyProperties2(
 920    VkPhysicalDevice physicalDevice,
 921    uint32_t *pQueueFamilyPropertyCount,
 922    VkQueueFamilyProperties2 *pQueueFamilyProperties)
 923 {
 924    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 925
 926    vk_outarray_append(&out, p)
 927    {
 928       p->queueFamilyProperties = tu_queue_family_properties;
 929    }
 930 }
 931
 932 static uint64_t
 933 tu_get_system_heap_size()
 934 {
 935    struct sysinfo info;
 936    sysinfo(&info);
 937
 938    uint64_t total_ram = (uint64_t) info.totalram * (uint64_t) info.mem_unit;
 939
 940    /* We don't want to burn too much ram with the GPU.  If the user has 4GiB
 941     * or less, we use at most half.  If they have more than 4GiB, we use 3/4.
 942     */
 943    uint64_t available_ram;
 944    if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
 945       available_ram = total_ram / 2;
 946    else
 947       available_ram = total_ram * 3 / 4;
 948
 949    return available_ram;
 950 }
 951
 952 void
 953 tu_GetPhysicalDeviceMemoryProperties(
 954    VkPhysicalDevice physicalDevice,
 955    VkPhysicalDeviceMemoryProperties *pMemoryProperties)
 956 {
 957    pMemoryProperties->memoryHeapCount = 1;
 958    pMemoryProperties->memoryHeaps[0].size = tu_get_system_heap_size();
 959    pMemoryProperties->memoryHeaps[0].flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT;
 960
 961    pMemoryProperties->memoryTypeCount = 1;
 962    pMemoryProperties->memoryTypes[0].propertyFlags =
 963       VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 964       VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 965       VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
 966    pMemoryProperties->memoryTypes[0].heapIndex = 0;
 967 }
 968
 969 void
 970 tu_GetPhysicalDeviceMemoryProperties2(
 971    VkPhysicalDevice physicalDevice,
 972    VkPhysicalDeviceMemoryProperties2 *pMemoryProperties)
 973 {
 974    return tu_GetPhysicalDeviceMemoryProperties(
 975       physicalDevice, &pMemoryProperties->memoryProperties);
 976 }
 977
 978 static VkResult
 979 tu_queue_init(struct tu_device *device,
 980               struct tu_queue *queue,
 981               uint32_t queue_family_index,
 982               int idx,
 983               VkDeviceQueueCreateFlags flags)
 984 {
 985    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 986    queue->device = device;
 987    queue->queue_family_index = queue_family_index;
 988    queue->queue_idx = idx;
 989    queue->flags = flags;
 990
 991    int ret = tu_drm_submitqueue_new(device, 0, &queue->msm_queue_id);
 992    if (ret)
 993       return VK_ERROR_INITIALIZATION_FAILED;
 994
 995    tu_fence_init(&queue->submit_fence, false);
 996
 997    return VK_SUCCESS;
 998 }
 999
1000 static void
1001 tu_queue_finish(struct tu_queue *queue)
1002 {
1003    tu_fence_finish(&queue->submit_fence);
1004    tu_drm_submitqueue_close(queue->device, queue->msm_queue_id);
1005 }
1006
1007 static int
1008 tu_get_device_extension_index(const char *name)
1009 {
1010    for (unsigned i = 0; i < TU_DEVICE_EXTENSION_COUNT; ++i) {
1011       if (strcmp(name, tu_device_extensions[i].extensionName) == 0)
1012          return i;
1013    }
1014    return -1;
1015 }
1016
1017 VkResult
1018 tu_CreateDevice(VkPhysicalDevice physicalDevice,
1019                 const VkDeviceCreateInfo *pCreateInfo,
1020                 const VkAllocationCallbacks *pAllocator,
1021                 VkDevice *pDevice)
1022 {
1023    TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
1024    VkResult result;
1025    struct tu_device *device;
1026
1027    /* Check enabled features */
1028    if (pCreateInfo->pEnabledFeatures) {
1029       VkPhysicalDeviceFeatures supported_features;
1030       tu_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1031       VkBool32 *supported_feature = (VkBool32 *) &supported_features;
1032       VkBool32 *enabled_feature = (VkBool32 *) pCreateInfo->pEnabledFeatures;
1033       unsigned num_features =
1034          sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1035       for (uint32_t i = 0; i < num_features; i++) {
1036          if (enabled_feature[i] && !supported_feature[i])
1037             return vk_error(physical_device->instance,
1038                             VK_ERROR_FEATURE_NOT_PRESENT);
1039       }
1040    }
1041
1042    device = vk_zalloc2(&physical_device->instance->alloc, pAllocator,
1043                        sizeof(*device), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1044    if (!device)
1045       return vk_error(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1046
1047    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1048    device->instance = physical_device->instance;
1049    device->physical_device = physical_device;
1050
1051    if (pAllocator)
1052       device->alloc = *pAllocator;
1053    else
1054       device->alloc = physical_device->instance->alloc;
1055
1056    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1057       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
1058       int index = tu_get_device_extension_index(ext_name);
1059       if (index < 0 ||
1060           !physical_device->supported_extensions.extensions[index]) {
1061          vk_free(&device->alloc, device);
1062          return vk_error(physical_device->instance,
1063                          VK_ERROR_EXTENSION_NOT_PRESENT);
1064       }
1065
1066       device->enabled_extensions.extensions[index] = true;
1067    }
1068
1069    for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1070       const VkDeviceQueueCreateInfo *queue_create =
1071          &pCreateInfo->pQueueCreateInfos[i];
1072       uint32_t qfi = queue_create->queueFamilyIndex;
1073       device->queues[qfi] = vk_alloc(
1074          &device->alloc, queue_create->queueCount * sizeof(struct tu_queue),
1075          8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1076       if (!device->queues[qfi]) {
1077          result = VK_ERROR_OUT_OF_HOST_MEMORY;
1078          goto fail;
1079       }
1080
1081       memset(device->queues[qfi], 0,
1082              queue_create->queueCount * sizeof(struct tu_queue));
1083
1084       device->queue_count[qfi] = queue_create->queueCount;
1085
1086       for (unsigned q = 0; q < queue_create->queueCount; q++) {
1087          result = tu_queue_init(device, &device->queues[qfi][q], qfi, q,
1088                                 queue_create->flags);
1089          if (result != VK_SUCCESS)
1090             goto fail;
1091       }
1092    }
1093
1094    device->compiler = ir3_compiler_create(NULL, physical_device->gpu_id);
1095    if (!device->compiler)
1096       goto fail;
1097
1098    VkPipelineCacheCreateInfo ci;
1099    ci.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
1100    ci.pNext = NULL;
1101    ci.flags = 0;
1102    ci.pInitialData = NULL;
1103    ci.initialDataSize = 0;
1104    VkPipelineCache pc;
1105    result =
1106       tu_CreatePipelineCache(tu_device_to_handle(device), &ci, NULL, &pc);
1107    if (result != VK_SUCCESS)
1108       goto fail;
1109
1110    device->mem_cache = tu_pipeline_cache_from_handle(pc);
1111
1112    *pDevice = tu_device_to_handle(device);
1113    return VK_SUCCESS;
1114
1115 fail:
1116    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1117       for (unsigned q = 0; q < device->queue_count[i]; q++)
1118          tu_queue_finish(&device->queues[i][q]);
1119       if (device->queue_count[i])
1120          vk_free(&device->alloc, device->queues[i]);
1121    }
1122
1123    if (device->compiler)
1124       ralloc_free(device->compiler);
1125
1126    vk_free(&device->alloc, device);
1127    return result;
1128 }
1129
1130 void
1131 tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
1132 {
1133    TU_FROM_HANDLE(tu_device, device, _device);
1134
1135    if (!device)
1136       return;
1137
1138    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1139       for (unsigned q = 0; q < device->queue_count[i]; q++)
1140          tu_queue_finish(&device->queues[i][q]);
1141       if (device->queue_count[i])
1142          vk_free(&device->alloc, device->queues[i]);
1143    }
1144
1145    /* the compiler does not use pAllocator */
1146    ralloc_free(device->compiler);
1147
1148    VkPipelineCache pc = tu_pipeline_cache_to_handle(device->mem_cache);
1149    tu_DestroyPipelineCache(tu_device_to_handle(device), pc, NULL);
1150
1151    vk_free(&device->alloc, device);
1152 }
1153
1154 VkResult
1155 tu_EnumerateInstanceLayerProperties(uint32_t *pPropertyCount,
1156                                     VkLayerProperties *pProperties)
1157 {
1158    *pPropertyCount = 0;
1159    return VK_SUCCESS;
1160 }
1161
1162 VkResult
1163 tu_EnumerateDeviceLayerProperties(VkPhysicalDevice physicalDevice,
1164                                   uint32_t *pPropertyCount,
1165                                   VkLayerProperties *pProperties)
1166 {
1167    *pPropertyCount = 0;
1168    return VK_SUCCESS;
1169 }
1170
1171 void
1172 tu_GetDeviceQueue2(VkDevice _device,
1173                    const VkDeviceQueueInfo2 *pQueueInfo,
1174                    VkQueue *pQueue)
1175 {
1176    TU_FROM_HANDLE(tu_device, device, _device);
1177    struct tu_queue *queue;
1178
1179    queue =
1180       &device->queues[pQueueInfo->queueFamilyIndex][pQueueInfo->queueIndex];
1181    if (pQueueInfo->flags != queue->flags) {
1182       /* From the Vulkan 1.1.70 spec:
1183        *
1184        * "The queue returned by vkGetDeviceQueue2 must have the same
1185        * flags value from this structure as that used at device
1186        * creation time in a VkDeviceQueueCreateInfo instance. If no
1187        * matching flags were specified at device creation time then
1188        * pQueue will return VK_NULL_HANDLE."
1189        */
1190       *pQueue = VK_NULL_HANDLE;
1191       return;
1192    }
1193
1194    *pQueue = tu_queue_to_handle(queue);
1195 }
1196
1197 void
1198 tu_GetDeviceQueue(VkDevice _device,
1199                   uint32_t queueFamilyIndex,
1200                   uint32_t queueIndex,
1201                   VkQueue *pQueue)
1202 {
1203    const VkDeviceQueueInfo2 info =
1204       (VkDeviceQueueInfo2) { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
1205                              .queueFamilyIndex = queueFamilyIndex,
1206                              .queueIndex = queueIndex };
1207
1208    tu_GetDeviceQueue2(_device, &info, pQueue);
1209 }
1210
1211 VkResult
1212 tu_QueueSubmit(VkQueue _queue,
1213                uint32_t submitCount,
1214                const VkSubmitInfo *pSubmits,
1215                VkFence _fence)
1216 {
1217    TU_FROM_HANDLE(tu_queue, queue, _queue);
1218
1219    for (uint32_t i = 0; i < submitCount; ++i) {
1220       const VkSubmitInfo *submit = pSubmits + i;
1221       const bool last_submit = (i == submitCount - 1);
1222       struct tu_bo_list bo_list;
1223       tu_bo_list_init(&bo_list);
1224
1225       uint32_t entry_count = 0;
1226       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1227          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1228          entry_count += cmdbuf->cs.entry_count;
1229       }
1230
1231       struct drm_msm_gem_submit_cmd cmds[entry_count];
1232       uint32_t entry_idx = 0;
1233       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1234          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1235          struct tu_cs *cs = &cmdbuf->cs;
1236          for (unsigned i = 0; i < cs->entry_count; ++i, ++entry_idx) {
1237             cmds[entry_idx].type = MSM_SUBMIT_CMD_BUF;
1238             cmds[entry_idx].submit_idx =
1239                tu_bo_list_add(&bo_list, cs->entries[i].bo,
1240                               MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
1241             cmds[entry_idx].submit_offset = cs->entries[i].offset;
1242             cmds[entry_idx].size = cs->entries[i].size;
1243             cmds[entry_idx].pad = 0;
1244             cmds[entry_idx].nr_relocs = 0;
1245             cmds[entry_idx].relocs = 0;
1246          }
1247
1248          tu_bo_list_merge(&bo_list, &cmdbuf->bo_list);
1249       }
1250
1251       uint32_t flags = MSM_PIPE_3D0;
1252       if (last_submit) {
1253          flags |= MSM_SUBMIT_FENCE_FD_OUT;
1254       }
1255
1256       struct drm_msm_gem_submit req = {
1257          .flags = flags,
1258          .queueid = queue->msm_queue_id,
1259          .bos = (uint64_t)(uintptr_t) bo_list.bo_infos,
1260          .nr_bos = bo_list.count,
1261          .cmds = (uint64_t)(uintptr_t)cmds,
1262          .nr_cmds = entry_count,
1263       };
1264
1265       int ret = drmCommandWriteRead(queue->device->physical_device->local_fd,
1266                                     DRM_MSM_GEM_SUBMIT,
1267                                     &req, sizeof(req));
1268       if (ret) {
1269          fprintf(stderr, "submit failed: %s\n", strerror(errno));
1270          abort();
1271       }
1272
1273       tu_bo_list_destroy(&bo_list);
1274
1275       if (last_submit) {
1276          /* no need to merge fences as queue execution is serialized */
1277          tu_fence_update_fd(&queue->submit_fence, req.fence_fd);
1278       }
1279    }
1280
1281    if (_fence != VK_NULL_HANDLE) {
1282       TU_FROM_HANDLE(tu_fence, fence, _fence);
1283       tu_fence_copy(fence, &queue->submit_fence);
1284    }
1285
1286    return VK_SUCCESS;
1287 }
1288
1289 VkResult
1290 tu_QueueWaitIdle(VkQueue _queue)
1291 {
1292    TU_FROM_HANDLE(tu_queue, queue, _queue);
1293
1294    tu_fence_wait_idle(&queue->submit_fence);
1295
1296    return VK_SUCCESS;
1297 }
1298
1299 VkResult
1300 tu_DeviceWaitIdle(VkDevice _device)
1301 {
1302    TU_FROM_HANDLE(tu_device, device, _device);
1303
1304    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1305       for (unsigned q = 0; q < device->queue_count[i]; q++) {
1306          tu_QueueWaitIdle(tu_queue_to_handle(&device->queues[i][q]));
1307       }
1308    }
1309    return VK_SUCCESS;
1310 }
1311
1312 VkResult
1313 tu_EnumerateInstanceExtensionProperties(const char *pLayerName,
1314                                         uint32_t *pPropertyCount,
1315                                         VkExtensionProperties *pProperties)
1316 {
1317    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1318
1319    /* We spport no lyaers */
1320    if (pLayerName)
1321       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1322
1323    for (int i = 0; i < TU_INSTANCE_EXTENSION_COUNT; i++) {
1324       if (tu_supported_instance_extensions.extensions[i]) {
1325          vk_outarray_append(&out, prop) { *prop = tu_instance_extensions[i]; }
1326       }
1327    }
1328
1329    return vk_outarray_status(&out);
1330 }
1331
1332 VkResult
1333 tu_EnumerateDeviceExtensionProperties(VkPhysicalDevice physicalDevice,
1334                                       const char *pLayerName,
1335                                       uint32_t *pPropertyCount,
1336                                       VkExtensionProperties *pProperties)
1337 {
1338    /* We spport no lyaers */
1339    TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
1340    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1341
1342    /* We spport no lyaers */
1343    if (pLayerName)
1344       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1345
1346    for (int i = 0; i < TU_DEVICE_EXTENSION_COUNT; i++) {
1347       if (device->supported_extensions.extensions[i]) {
1348          vk_outarray_append(&out, prop) { *prop = tu_device_extensions[i]; }
1349       }
1350    }
1351
1352    return vk_outarray_status(&out);
1353 }
1354
1355 PFN_vkVoidFunction
1356 tu_GetInstanceProcAddr(VkInstance _instance, const char *pName)
1357 {
1358    TU_FROM_HANDLE(tu_instance, instance, _instance);
1359
1360    return tu_lookup_entrypoint_checked(
1361       pName, instance ? instance->api_version : 0,
1362       instance ? &instance->enabled_extensions : NULL, NULL);
1363 }
1364
1365 /* The loader wants us to expose a second GetInstanceProcAddr function
1366  * to work around certain LD_PRELOAD issues seen in apps.
1367  */
1368 PUBLIC
1369 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1370 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName);
1371
1372 PUBLIC
1373 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1374 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName)
1375 {
1376    return tu_GetInstanceProcAddr(instance, pName);
1377 }
1378
1379 PFN_vkVoidFunction
1380 tu_GetDeviceProcAddr(VkDevice _device, const char *pName)
1381 {
1382    TU_FROM_HANDLE(tu_device, device, _device);
1383
1384    return tu_lookup_entrypoint_checked(pName, device->instance->api_version,
1385                                        &device->instance->enabled_extensions,
1386                                        &device->enabled_extensions);
1387 }
1388
1389 static VkResult
1390 tu_alloc_memory(struct tu_device *device,
1391                 const VkMemoryAllocateInfo *pAllocateInfo,
1392                 const VkAllocationCallbacks *pAllocator,
1393                 VkDeviceMemory *pMem)
1394 {
1395    struct tu_device_memory *mem;
1396    VkResult result;
1397
1398    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1399
1400    if (pAllocateInfo->allocationSize == 0) {
1401       /* Apparently, this is allowed */
1402       *pMem = VK_NULL_HANDLE;
1403       return VK_SUCCESS;
1404    }
1405
1406    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1407                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1408    if (mem == NULL)
1409       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1410
1411    const VkImportMemoryFdInfoKHR *fd_info =
1412       vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
1413    if (fd_info && !fd_info->handleType)
1414       fd_info = NULL;
1415
1416    if (fd_info) {
1417       assert(fd_info->handleType ==
1418                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
1419              fd_info->handleType ==
1420                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1421
1422       /*
1423        * TODO Importing the same fd twice gives us the same handle without
1424        * reference counting.  We need to maintain a per-instance handle-to-bo
1425        * table and add reference count to tu_bo.
1426        */
1427       result = tu_bo_init_dmabuf(device, &mem->bo,
1428                                  pAllocateInfo->allocationSize, fd_info->fd);
1429       if (result == VK_SUCCESS) {
1430          /* take ownership and close the fd */
1431          close(fd_info->fd);
1432       }
1433    } else {
1434       result =
1435          tu_bo_init_new(device, &mem->bo, pAllocateInfo->allocationSize);
1436    }
1437
1438    if (result != VK_SUCCESS) {
1439       vk_free2(&device->alloc, pAllocator, mem);
1440       return result;
1441    }
1442
1443    mem->size = pAllocateInfo->allocationSize;
1444    mem->type_index = pAllocateInfo->memoryTypeIndex;
1445
1446    mem->map = NULL;
1447    mem->user_ptr = NULL;
1448
1449    *pMem = tu_device_memory_to_handle(mem);
1450
1451    return VK_SUCCESS;
1452 }
1453
1454 VkResult
1455 tu_AllocateMemory(VkDevice _device,
1456                   const VkMemoryAllocateInfo *pAllocateInfo,
1457                   const VkAllocationCallbacks *pAllocator,
1458                   VkDeviceMemory *pMem)
1459 {
1460    TU_FROM_HANDLE(tu_device, device, _device);
1461    return tu_alloc_memory(device, pAllocateInfo, pAllocator, pMem);
1462 }
1463
1464 void
1465 tu_FreeMemory(VkDevice _device,
1466               VkDeviceMemory _mem,
1467               const VkAllocationCallbacks *pAllocator)
1468 {
1469    TU_FROM_HANDLE(tu_device, device, _device);
1470    TU_FROM_HANDLE(tu_device_memory, mem, _mem);
1471
1472    if (mem == NULL)
1473       return;
1474
1475    tu_bo_finish(device, &mem->bo);
1476    vk_free2(&device->alloc, pAllocator, mem);
1477 }
1478
1479 VkResult
1480 tu_MapMemory(VkDevice _device,
1481              VkDeviceMemory _memory,
1482              VkDeviceSize offset,
1483              VkDeviceSize size,
1484              VkMemoryMapFlags flags,
1485              void **ppData)
1486 {
1487    TU_FROM_HANDLE(tu_device, device, _device);
1488    TU_FROM_HANDLE(tu_device_memory, mem, _memory);
1489    VkResult result;
1490
1491    if (mem == NULL) {
1492       *ppData = NULL;
1493       return VK_SUCCESS;
1494    }
1495
1496    if (mem->user_ptr) {
1497       *ppData = mem->user_ptr;
1498    } else if (!mem->map) {
1499       result = tu_bo_map(device, &mem->bo);
1500       if (result != VK_SUCCESS)
1501          return result;
1502       *ppData = mem->map = mem->bo.map;
1503    } else
1504       *ppData = mem->map;
1505
1506    if (*ppData) {
1507       *ppData += offset;
1508       return VK_SUCCESS;
1509    }
1510
1511    return vk_error(device->instance, VK_ERROR_MEMORY_MAP_FAILED);
1512 }
1513
1514 void
1515 tu_UnmapMemory(VkDevice _device, VkDeviceMemory _memory)
1516 {
1517    /* I do not see any unmapping done by the freedreno Gallium driver. */
1518 }
1519
1520 VkResult
1521 tu_FlushMappedMemoryRanges(VkDevice _device,
1522                            uint32_t memoryRangeCount,
1523                            const VkMappedMemoryRange *pMemoryRanges)
1524 {
1525    return VK_SUCCESS;
1526 }
1527
1528 VkResult
1529 tu_InvalidateMappedMemoryRanges(VkDevice _device,
1530                                 uint32_t memoryRangeCount,
1531                                 const VkMappedMemoryRange *pMemoryRanges)
1532 {
1533    return VK_SUCCESS;
1534 }
1535
1536 void
1537 tu_GetBufferMemoryRequirements(VkDevice _device,
1538                                VkBuffer _buffer,
1539                                VkMemoryRequirements *pMemoryRequirements)
1540 {
1541    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
1542
1543    pMemoryRequirements->memoryTypeBits = 1;
1544    pMemoryRequirements->alignment = 16;
1545    pMemoryRequirements->size =
1546       align64(buffer->size, pMemoryRequirements->alignment);
1547 }
1548
1549 void
1550 tu_GetBufferMemoryRequirements2(
1551    VkDevice device,
1552    const VkBufferMemoryRequirementsInfo2 *pInfo,
1553    VkMemoryRequirements2 *pMemoryRequirements)
1554 {
1555    tu_GetBufferMemoryRequirements(device, pInfo->buffer,
1556                                   &pMemoryRequirements->memoryRequirements);
1557 }
1558
1559 void
1560 tu_GetImageMemoryRequirements(VkDevice _device,
1561                               VkImage _image,
1562                               VkMemoryRequirements *pMemoryRequirements)
1563 {
1564    TU_FROM_HANDLE(tu_image, image, _image);
1565
1566    pMemoryRequirements->memoryTypeBits = 1;
1567    pMemoryRequirements->size = image->size;
1568    pMemoryRequirements->alignment = image->alignment;
1569 }
1570
1571 void
1572 tu_GetImageMemoryRequirements2(VkDevice device,
1573                                const VkImageMemoryRequirementsInfo2 *pInfo,
1574                                VkMemoryRequirements2 *pMemoryRequirements)
1575 {
1576    tu_GetImageMemoryRequirements(device, pInfo->image,
1577                                  &pMemoryRequirements->memoryRequirements);
1578 }
1579
1580 void
1581 tu_GetImageSparseMemoryRequirements(
1582    VkDevice device,
1583    VkImage image,
1584    uint32_t *pSparseMemoryRequirementCount,
1585    VkSparseImageMemoryRequirements *pSparseMemoryRequirements)
1586 {
1587    tu_stub();
1588 }
1589
1590 void
1591 tu_GetImageSparseMemoryRequirements2(
1592    VkDevice device,
1593    const VkImageSparseMemoryRequirementsInfo2 *pInfo,
1594    uint32_t *pSparseMemoryRequirementCount,
1595    VkSparseImageMemoryRequirements2 *pSparseMemoryRequirements)
1596 {
1597    tu_stub();
1598 }
1599
1600 void
1601 tu_GetDeviceMemoryCommitment(VkDevice device,
1602                              VkDeviceMemory memory,
1603                              VkDeviceSize *pCommittedMemoryInBytes)
1604 {
1605    *pCommittedMemoryInBytes = 0;
1606 }
1607
1608 VkResult
1609 tu_BindBufferMemory2(VkDevice device,
1610                      uint32_t bindInfoCount,
1611                      const VkBindBufferMemoryInfo *pBindInfos)
1612 {
1613    for (uint32_t i = 0; i < bindInfoCount; ++i) {
1614       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
1615       TU_FROM_HANDLE(tu_buffer, buffer, pBindInfos[i].buffer);
1616
1617       if (mem) {
1618          buffer->bo = &mem->bo;
1619          buffer->bo_offset = pBindInfos[i].memoryOffset;
1620       } else {
1621          buffer->bo = NULL;
1622       }
1623    }
1624    return VK_SUCCESS;
1625 }
1626
1627 VkResult
1628 tu_BindBufferMemory(VkDevice device,
1629                     VkBuffer buffer,
1630                     VkDeviceMemory memory,
1631                     VkDeviceSize memoryOffset)
1632 {
1633    const VkBindBufferMemoryInfo info = {
1634       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
1635       .buffer = buffer,
1636       .memory = memory,
1637       .memoryOffset = memoryOffset
1638    };
1639
1640    return tu_BindBufferMemory2(device, 1, &info);
1641 }
1642
1643 VkResult
1644 tu_BindImageMemory2(VkDevice device,
1645                     uint32_t bindInfoCount,
1646                     const VkBindImageMemoryInfo *pBindInfos)
1647 {
1648    for (uint32_t i = 0; i < bindInfoCount; ++i) {
1649       TU_FROM_HANDLE(tu_image, image, pBindInfos[i].image);
1650       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
1651
1652       if (mem) {
1653          image->bo = &mem->bo;
1654          image->bo_offset = pBindInfos[i].memoryOffset;
1655       } else {
1656          image->bo = NULL;
1657          image->bo_offset = 0;
1658       }
1659    }
1660
1661    return VK_SUCCESS;
1662 }
1663
1664 VkResult
1665 tu_BindImageMemory(VkDevice device,
1666                    VkImage image,
1667                    VkDeviceMemory memory,
1668                    VkDeviceSize memoryOffset)
1669 {
1670    const VkBindImageMemoryInfo info = {
1671       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
1672       .image = image,
1673       .memory = memory,
1674       .memoryOffset = memoryOffset
1675    };
1676
1677    return tu_BindImageMemory2(device, 1, &info);
1678 }
1679
1680 VkResult
1681 tu_QueueBindSparse(VkQueue _queue,
1682                    uint32_t bindInfoCount,
1683                    const VkBindSparseInfo *pBindInfo,
1684                    VkFence _fence)
1685 {
1686    return VK_SUCCESS;
1687 }
1688
1689 // Queue semaphore functions
1690
1691 VkResult
1692 tu_CreateSemaphore(VkDevice _device,
1693                    const VkSemaphoreCreateInfo *pCreateInfo,
1694                    const VkAllocationCallbacks *pAllocator,
1695                    VkSemaphore *pSemaphore)
1696 {
1697    TU_FROM_HANDLE(tu_device, device, _device);
1698
1699    struct tu_semaphore *sem =
1700       vk_alloc2(&device->alloc, pAllocator, sizeof(*sem), 8,
1701                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1702    if (!sem)
1703       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1704
1705    *pSemaphore = tu_semaphore_to_handle(sem);
1706    return VK_SUCCESS;
1707 }
1708
1709 void
1710 tu_DestroySemaphore(VkDevice _device,
1711                     VkSemaphore _semaphore,
1712                     const VkAllocationCallbacks *pAllocator)
1713 {
1714    TU_FROM_HANDLE(tu_device, device, _device);
1715    TU_FROM_HANDLE(tu_semaphore, sem, _semaphore);
1716    if (!_semaphore)
1717       return;
1718
1719    vk_free2(&device->alloc, pAllocator, sem);
1720 }
1721
1722 VkResult
1723 tu_CreateEvent(VkDevice _device,
1724                const VkEventCreateInfo *pCreateInfo,
1725                const VkAllocationCallbacks *pAllocator,
1726                VkEvent *pEvent)
1727 {
1728    TU_FROM_HANDLE(tu_device, device, _device);
1729    struct tu_event *event =
1730       vk_alloc2(&device->alloc, pAllocator, sizeof(*event), 8,
1731                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1732
1733    if (!event)
1734       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1735
1736    VkResult result = tu_bo_init_new(device, &event->bo, 0x1000);
1737    if (result != VK_SUCCESS)
1738       goto fail_alloc;
1739
1740    result = tu_bo_map(device, &event->bo);
1741    if (result != VK_SUCCESS)
1742       goto fail_map;
1743
1744    *pEvent = tu_event_to_handle(event);
1745
1746    return VK_SUCCESS;
1747
1748 fail_map:
1749    tu_bo_finish(device, &event->bo);
1750 fail_alloc:
1751    vk_free2(&device->alloc, pAllocator, event);
1752    return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1753 }
1754
1755 void
1756 tu_DestroyEvent(VkDevice _device,
1757                 VkEvent _event,
1758                 const VkAllocationCallbacks *pAllocator)
1759 {
1760    TU_FROM_HANDLE(tu_device, device, _device);
1761    TU_FROM_HANDLE(tu_event, event, _event);
1762
1763    if (!event)
1764       return;
1765    vk_free2(&device->alloc, pAllocator, event);
1766 }
1767
1768 VkResult
1769 tu_GetEventStatus(VkDevice _device, VkEvent _event)
1770 {
1771    TU_FROM_HANDLE(tu_event, event, _event);
1772
1773    if (*(uint64_t*) event->bo.map == 1)
1774       return VK_EVENT_SET;
1775    return VK_EVENT_RESET;
1776 }
1777
1778 VkResult
1779 tu_SetEvent(VkDevice _device, VkEvent _event)
1780 {
1781    TU_FROM_HANDLE(tu_event, event, _event);
1782    *(uint64_t*) event->bo.map = 1;
1783
1784    return VK_SUCCESS;
1785 }
1786
1787 VkResult
1788 tu_ResetEvent(VkDevice _device, VkEvent _event)
1789 {
1790    TU_FROM_HANDLE(tu_event, event, _event);
1791    *(uint64_t*) event->bo.map = 0;
1792
1793    return VK_SUCCESS;
1794 }
1795
1796 VkResult
1797 tu_CreateBuffer(VkDevice _device,
1798                 const VkBufferCreateInfo *pCreateInfo,
1799                 const VkAllocationCallbacks *pAllocator,
1800                 VkBuffer *pBuffer)
1801 {
1802    TU_FROM_HANDLE(tu_device, device, _device);
1803    struct tu_buffer *buffer;
1804
1805    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1806
1807    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1808                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1809    if (buffer == NULL)
1810       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1811
1812    buffer->size = pCreateInfo->size;
1813    buffer->usage = pCreateInfo->usage;
1814    buffer->flags = pCreateInfo->flags;
1815
1816    *pBuffer = tu_buffer_to_handle(buffer);
1817
1818    return VK_SUCCESS;
1819 }
1820
1821 void
1822 tu_DestroyBuffer(VkDevice _device,
1823                  VkBuffer _buffer,
1824                  const VkAllocationCallbacks *pAllocator)
1825 {
1826    TU_FROM_HANDLE(tu_device, device, _device);
1827    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
1828
1829    if (!buffer)
1830       return;
1831
1832    vk_free2(&device->alloc, pAllocator, buffer);
1833 }
1834
1835 static uint32_t
1836 tu_surface_max_layer_count(struct tu_image_view *iview)
1837 {
1838    return iview->type == VK_IMAGE_VIEW_TYPE_3D
1839              ? iview->extent.depth
1840              : (iview->base_layer + iview->layer_count);
1841 }
1842
1843 VkResult
1844 tu_CreateFramebuffer(VkDevice _device,
1845                      const VkFramebufferCreateInfo *pCreateInfo,
1846                      const VkAllocationCallbacks *pAllocator,
1847                      VkFramebuffer *pFramebuffer)
1848 {
1849    TU_FROM_HANDLE(tu_device, device, _device);
1850    struct tu_framebuffer *framebuffer;
1851
1852    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
1853
1854    size_t size = sizeof(*framebuffer) + sizeof(struct tu_attachment_info) *
1855                                            pCreateInfo->attachmentCount;
1856    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
1857                            VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1858    if (framebuffer == NULL)
1859       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1860
1861    framebuffer->attachment_count = pCreateInfo->attachmentCount;
1862    framebuffer->width = pCreateInfo->width;
1863    framebuffer->height = pCreateInfo->height;
1864    framebuffer->layers = pCreateInfo->layers;
1865    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
1866       VkImageView _iview = pCreateInfo->pAttachments[i];
1867       struct tu_image_view *iview = tu_image_view_from_handle(_iview);
1868       framebuffer->attachments[i].attachment = iview;
1869
1870       framebuffer->width = MIN2(framebuffer->width, iview->extent.width);
1871       framebuffer->height = MIN2(framebuffer->height, iview->extent.height);
1872       framebuffer->layers =
1873          MIN2(framebuffer->layers, tu_surface_max_layer_count(iview));
1874    }
1875
1876    *pFramebuffer = tu_framebuffer_to_handle(framebuffer);
1877    return VK_SUCCESS;
1878 }
1879
1880 void
1881 tu_DestroyFramebuffer(VkDevice _device,
1882                       VkFramebuffer _fb,
1883                       const VkAllocationCallbacks *pAllocator)
1884 {
1885    TU_FROM_HANDLE(tu_device, device, _device);
1886    TU_FROM_HANDLE(tu_framebuffer, fb, _fb);
1887
1888    if (!fb)
1889       return;
1890    vk_free2(&device->alloc, pAllocator, fb);
1891 }
1892
1893 static enum a6xx_tex_clamp
1894 tu6_tex_wrap(VkSamplerAddressMode address_mode, bool *needs_border)
1895 {
1896    switch (address_mode) {
1897    case VK_SAMPLER_ADDRESS_MODE_REPEAT:
1898       return A6XX_TEX_REPEAT;
1899    case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT:
1900       return A6XX_TEX_MIRROR_REPEAT;
1901    case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE:
1902       return A6XX_TEX_CLAMP_TO_EDGE;
1903    case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER:
1904       *needs_border = true;
1905       return A6XX_TEX_CLAMP_TO_BORDER;
1906    case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE:
1907       /* only works for PoT.. need to emulate otherwise! */
1908       return A6XX_TEX_MIRROR_CLAMP;
1909    default:
1910       unreachable("illegal tex wrap mode");
1911       break;
1912    }
1913 }
1914
1915 static enum a6xx_tex_filter
1916 tu6_tex_filter(VkFilter filter, unsigned aniso)
1917 {
1918    switch (filter) {
1919    case VK_FILTER_NEAREST:
1920       return A6XX_TEX_NEAREST;
1921    case VK_FILTER_LINEAR:
1922       return aniso ? A6XX_TEX_ANISO : A6XX_TEX_LINEAR;
1923    case VK_FILTER_CUBIC_IMG:
1924    default:
1925       unreachable("illegal texture filter");
1926       break;
1927    }
1928 }
1929
1930 static void
1931 tu_init_sampler(struct tu_device *device,
1932                 struct tu_sampler *sampler,
1933                 const VkSamplerCreateInfo *pCreateInfo)
1934 {
1935    unsigned aniso = pCreateInfo->anisotropyEnable ?
1936       util_last_bit(MIN2((uint32_t)pCreateInfo->maxAnisotropy >> 1, 8)) : 0;
1937    bool miplinear = (pCreateInfo->mipmapMode == VK_SAMPLER_MIPMAP_MODE_LINEAR);
1938    bool needs_border = false;
1939
1940    sampler->state[0] =
1941       COND(miplinear, A6XX_TEX_SAMP_0_MIPFILTER_LINEAR_NEAR) |
1942       A6XX_TEX_SAMP_0_XY_MAG(tu6_tex_filter(pCreateInfo->magFilter, aniso)) |
1943       A6XX_TEX_SAMP_0_XY_MIN(tu6_tex_filter(pCreateInfo->minFilter, aniso)) |
1944       A6XX_TEX_SAMP_0_ANISO(aniso) |
1945       A6XX_TEX_SAMP_0_WRAP_S(tu6_tex_wrap(pCreateInfo->addressModeU, &needs_border)) |
1946       A6XX_TEX_SAMP_0_WRAP_T(tu6_tex_wrap(pCreateInfo->addressModeV, &needs_border)) |
1947       A6XX_TEX_SAMP_0_WRAP_R(tu6_tex_wrap(pCreateInfo->addressModeW, &needs_border)) |
1948       A6XX_TEX_SAMP_0_LOD_BIAS(pCreateInfo->mipLodBias);
1949    sampler->state[1] =
1950       /* COND(!cso->seamless_cube_map, A6XX_TEX_SAMP_1_CUBEMAPSEAMLESSFILTOFF) | */
1951       COND(pCreateInfo->unnormalizedCoordinates, A6XX_TEX_SAMP_1_UNNORM_COORDS) |
1952       A6XX_TEX_SAMP_1_MIN_LOD(pCreateInfo->minLod) |
1953       A6XX_TEX_SAMP_1_MAX_LOD(pCreateInfo->maxLod) |
1954       COND(pCreateInfo->compareEnable, A6XX_TEX_SAMP_1_COMPARE_FUNC(pCreateInfo->compareOp));
1955    sampler->state[2] = 0;
1956    sampler->state[3] = 0;
1957
1958    /* TODO:
1959     * A6XX_TEX_SAMP_1_MIPFILTER_LINEAR_FAR disables mipmapping, but vk has no NONE mipfilter?
1960     * border color
1961     */
1962
1963    sampler->needs_border = needs_border;
1964 }
1965
1966 VkResult
1967 tu_CreateSampler(VkDevice _device,
1968                  const VkSamplerCreateInfo *pCreateInfo,
1969                  const VkAllocationCallbacks *pAllocator,
1970                  VkSampler *pSampler)
1971 {
1972    TU_FROM_HANDLE(tu_device, device, _device);
1973    struct tu_sampler *sampler;
1974
1975    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO);
1976
1977    sampler = vk_alloc2(&device->alloc, pAllocator, sizeof(*sampler), 8,
1978                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1979    if (!sampler)
1980       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1981
1982    tu_init_sampler(device, sampler, pCreateInfo);
1983    *pSampler = tu_sampler_to_handle(sampler);
1984
1985    return VK_SUCCESS;
1986 }
1987
1988 void
1989 tu_DestroySampler(VkDevice _device,
1990                   VkSampler _sampler,
1991                   const VkAllocationCallbacks *pAllocator)
1992 {
1993    TU_FROM_HANDLE(tu_device, device, _device);
1994    TU_FROM_HANDLE(tu_sampler, sampler, _sampler);
1995
1996    if (!sampler)
1997       return;
1998    vk_free2(&device->alloc, pAllocator, sampler);
1999 }
2000
2001 /* vk_icd.h does not declare this function, so we declare it here to
2002  * suppress Wmissing-prototypes.
2003  */
2004 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2005 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion);
2006
2007 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2008 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion)
2009 {
2010    /* For the full details on loader interface versioning, see
2011     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2012     * What follows is a condensed summary, to help you navigate the large and
2013     * confusing official doc.
2014     *
2015     *   - Loader interface v0 is incompatible with later versions. We don't
2016     *     support it.
2017     *
2018     *   - In loader interface v1:
2019     *       - The first ICD entrypoint called by the loader is
2020     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2021     *         entrypoint.
2022     *       - The ICD must statically expose no other Vulkan symbol unless it
2023     * is linked with -Bsymbolic.
2024     *       - Each dispatchable Vulkan handle created by the ICD must be
2025     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2026     *         ICD must initialize VK_LOADER_DATA.loadMagic to
2027     * ICD_LOADER_MAGIC.
2028     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2029     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2030     *         such loader-managed surfaces.
2031     *
2032     *    - Loader interface v2 differs from v1 in:
2033     *       - The first ICD entrypoint called by the loader is
2034     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2035     *         statically expose this entrypoint.
2036     *
2037     *    - Loader interface v3 differs from v2 in:
2038     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2039     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2040     *          because the loader no longer does so.
2041     */
2042    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2043    return VK_SUCCESS;
2044 }
2045
2046 VkResult
2047 tu_GetMemoryFdKHR(VkDevice _device,
2048                   const VkMemoryGetFdInfoKHR *pGetFdInfo,
2049                   int *pFd)
2050 {
2051    TU_FROM_HANDLE(tu_device, device, _device);
2052    TU_FROM_HANDLE(tu_device_memory, memory, pGetFdInfo->memory);
2053
2054    assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2055
2056    /* At the moment, we support only the below handle types. */
2057    assert(pGetFdInfo->handleType ==
2058              VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2059           pGetFdInfo->handleType ==
2060              VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2061
2062    int prime_fd = tu_bo_export_dmabuf(device, &memory->bo);
2063    if (prime_fd < 0)
2064       return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
2065
2066    *pFd = prime_fd;
2067    return VK_SUCCESS;
2068 }
2069
2070 VkResult
2071 tu_GetMemoryFdPropertiesKHR(VkDevice _device,
2072                             VkExternalMemoryHandleTypeFlagBits handleType,
2073                             int fd,
2074                             VkMemoryFdPropertiesKHR *pMemoryFdProperties)
2075 {
2076    assert(handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2077    pMemoryFdProperties->memoryTypeBits = 1;
2078    return VK_SUCCESS;
2079 }
2080
2081 void
2082 tu_GetPhysicalDeviceExternalSemaphoreProperties(
2083    VkPhysicalDevice physicalDevice,
2084    const VkPhysicalDeviceExternalSemaphoreInfo *pExternalSemaphoreInfo,
2085    VkExternalSemaphoreProperties *pExternalSemaphoreProperties)
2086 {
2087    pExternalSemaphoreProperties->exportFromImportedHandleTypes = 0;
2088    pExternalSemaphoreProperties->compatibleHandleTypes = 0;
2089    pExternalSemaphoreProperties->externalSemaphoreFeatures = 0;
2090 }
2091
2092 void
2093 tu_GetPhysicalDeviceExternalFenceProperties(
2094    VkPhysicalDevice physicalDevice,
2095    const VkPhysicalDeviceExternalFenceInfo *pExternalFenceInfo,
2096    VkExternalFenceProperties *pExternalFenceProperties)
2097 {
2098    pExternalFenceProperties->exportFromImportedHandleTypes = 0;
2099    pExternalFenceProperties->compatibleHandleTypes = 0;
2100    pExternalFenceProperties->externalFenceFeatures = 0;
2101 }
2102
2103 VkResult
2104 tu_CreateDebugReportCallbackEXT(
2105    VkInstance _instance,
2106    const VkDebugReportCallbackCreateInfoEXT *pCreateInfo,
2107    const VkAllocationCallbacks *pAllocator,
2108    VkDebugReportCallbackEXT *pCallback)
2109 {
2110    TU_FROM_HANDLE(tu_instance, instance, _instance);
2111    return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2112                                           pCreateInfo, pAllocator,
2113                                           &instance->alloc, pCallback);
2114 }
2115
2116 void
2117 tu_DestroyDebugReportCallbackEXT(VkInstance _instance,
2118                                  VkDebugReportCallbackEXT _callback,
2119                                  const VkAllocationCallbacks *pAllocator)
2120 {
2121    TU_FROM_HANDLE(tu_instance, instance, _instance);
2122    vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2123                                     _callback, pAllocator, &instance->alloc);
2124 }
2125
2126 void
2127 tu_DebugReportMessageEXT(VkInstance _instance,
2128                          VkDebugReportFlagsEXT flags,
2129                          VkDebugReportObjectTypeEXT objectType,
2130                          uint64_t object,
2131                          size_t location,
2132                          int32_t messageCode,
2133                          const char *pLayerPrefix,
2134                          const char *pMessage)
2135 {
2136    TU_FROM_HANDLE(tu_instance, instance, _instance);
2137    vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2138                    object, location, messageCode, pLayerPrefix, pMessage);
2139 }
2140
2141 void
2142 tu_GetDeviceGroupPeerMemoryFeatures(
2143    VkDevice device,
2144    uint32_t heapIndex,
2145    uint32_t localDeviceIndex,
2146    uint32_t remoteDeviceIndex,
2147    VkPeerMemoryFeatureFlags *pPeerMemoryFeatures)
2148 {
2149    assert(localDeviceIndex == remoteDeviceIndex);
2150
2151    *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2152                           VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2153                           VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2154                           VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2155 }