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