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 618:
 260       device->tile_align_w = 64;
 261       device->tile_align_h = 16;
 262       device->magic.RB_UNKNOWN_8E04_blit = 0x00100000;
 263       device->magic.RB_CCU_CNTL_gmem     = 0x3e400004;
 264       device->magic.PC_UNKNOWN_9805 = 0x0;
 265       device->magic.SP_UNKNOWN_A0F8 = 0x0;
 266       break;
 267    case 630:
 268    case 640:
 269       device->tile_align_w = 64;
 270       device->tile_align_h = 16;
 271       device->magic.RB_UNKNOWN_8E04_blit = 0x01000000;
 272       device->magic.RB_CCU_CNTL_gmem     = 0x7c400004;
 273       device->magic.PC_UNKNOWN_9805 = 0x1;
 274       device->magic.SP_UNKNOWN_A0F8 = 0x1;
 275       break;
 276    default:
 277       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 278                          "device %s is unsupported", device->name);
 279       goto fail;
 280    }
 281    if (tu_device_get_cache_uuid(device->gpu_id, device->cache_uuid)) {
 282       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 283                          "cannot generate UUID");
 284       goto fail;
 285    }
 286
 287    /* The gpu id is already embedded in the uuid so we just pass "tu"
 288     * when creating the cache.
 289     */
 290    char buf[VK_UUID_SIZE * 2 + 1];
 291    disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
 292    device->disk_cache = disk_cache_create(device->name, buf, 0);
 293
 294    fprintf(stderr, "WARNING: tu is not a conformant vulkan implementation, "
 295                    "testing use only.\n");
 296
 297    tu_get_driver_uuid(&device->device_uuid);
 298    tu_get_device_uuid(&device->device_uuid);
 299
 300    tu_fill_device_extension_table(device, &device->supported_extensions);
 301
 302    if (result != VK_SUCCESS) {
 303       vk_error(instance, result);
 304       goto fail;
 305    }
 306
 307    result = tu_wsi_init(device);
 308    if (result != VK_SUCCESS) {
 309       vk_error(instance, result);
 310       goto fail;
 311    }
 312
 313    return VK_SUCCESS;
 314
 315 fail:
 316    close(fd);
 317    if (master_fd != -1)
 318       close(master_fd);
 319    return result;
 320 }
 321
 322 static void
 323 tu_physical_device_finish(struct tu_physical_device *device)
 324 {
 325    tu_wsi_finish(device);
 326
 327    disk_cache_destroy(device->disk_cache);
 328    close(device->local_fd);
 329    if (device->master_fd != -1)
 330       close(device->master_fd);
 331 }
 332
 333 static void *
 334 default_alloc_func(void *pUserData,
 335                    size_t size,
 336                    size_t align,
 337                    VkSystemAllocationScope allocationScope)
 338 {
 339    return malloc(size);
 340 }
 341
 342 static void *
 343 default_realloc_func(void *pUserData,
 344                      void *pOriginal,
 345                      size_t size,
 346                      size_t align,
 347                      VkSystemAllocationScope allocationScope)
 348 {
 349    return realloc(pOriginal, size);
 350 }
 351
 352 static void
 353 default_free_func(void *pUserData, void *pMemory)
 354 {
 355    free(pMemory);
 356 }
 357
 358 static const VkAllocationCallbacks default_alloc = {
 359    .pUserData = NULL,
 360    .pfnAllocation = default_alloc_func,
 361    .pfnReallocation = default_realloc_func,
 362    .pfnFree = default_free_func,
 363 };
 364
 365 static const struct debug_control tu_debug_options[] = {
 366    { "startup", TU_DEBUG_STARTUP },
 367    { "nir", TU_DEBUG_NIR },
 368    { "ir3", TU_DEBUG_IR3 },
 369    { "nobin", TU_DEBUG_NOBIN },
 370    { NULL, 0 }
 371 };
 372
 373 const char *
 374 tu_get_debug_option_name(int id)
 375 {
 376    assert(id < ARRAY_SIZE(tu_debug_options) - 1);
 377    return tu_debug_options[id].string;
 378 }
 379
 380 static int
 381 tu_get_instance_extension_index(const char *name)
 382 {
 383    for (unsigned i = 0; i < TU_INSTANCE_EXTENSION_COUNT; ++i) {
 384       if (strcmp(name, tu_instance_extensions[i].extensionName) == 0)
 385          return i;
 386    }
 387    return -1;
 388 }
 389
 390 VkResult
 391 tu_CreateInstance(const VkInstanceCreateInfo *pCreateInfo,
 392                   const VkAllocationCallbacks *pAllocator,
 393                   VkInstance *pInstance)
 394 {
 395    struct tu_instance *instance;
 396    VkResult result;
 397
 398    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 399
 400    uint32_t client_version;
 401    if (pCreateInfo->pApplicationInfo &&
 402        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 403       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 404    } else {
 405       tu_EnumerateInstanceVersion(&client_version);
 406    }
 407
 408    instance = vk_zalloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 409                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 410    if (!instance)
 411       return vk_error(NULL, VK_ERROR_OUT_OF_HOST_MEMORY);
 412
 413    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 414
 415    if (pAllocator)
 416       instance->alloc = *pAllocator;
 417    else
 418       instance->alloc = default_alloc;
 419
 420    instance->api_version = client_version;
 421    instance->physical_device_count = -1;
 422
 423    instance->debug_flags =
 424       parse_debug_string(getenv("TU_DEBUG"), tu_debug_options);
 425
 426    if (instance->debug_flags & TU_DEBUG_STARTUP)
 427       tu_logi("Created an instance");
 428
 429    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 430       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
 431       int index = tu_get_instance_extension_index(ext_name);
 432
 433       if (index < 0 || !tu_supported_instance_extensions.extensions[index]) {
 434          vk_free2(&default_alloc, pAllocator, instance);
 435          return vk_error(instance, VK_ERROR_EXTENSION_NOT_PRESENT);
 436       }
 437
 438       instance->enabled_extensions.extensions[index] = true;
 439    }
 440
 441    result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
 442    if (result != VK_SUCCESS) {
 443       vk_free2(&default_alloc, pAllocator, instance);
 444       return vk_error(instance, result);
 445    }
 446
 447    glsl_type_singleton_init_or_ref();
 448
 449    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 450
 451    *pInstance = tu_instance_to_handle(instance);
 452
 453    return VK_SUCCESS;
 454 }
 455
 456 void
 457 tu_DestroyInstance(VkInstance _instance,
 458                    const VkAllocationCallbacks *pAllocator)
 459 {
 460    TU_FROM_HANDLE(tu_instance, instance, _instance);
 461
 462    if (!instance)
 463       return;
 464
 465    for (int i = 0; i < instance->physical_device_count; ++i) {
 466       tu_physical_device_finish(instance->physical_devices + i);
 467    }
 468
 469    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 470
 471    glsl_type_singleton_decref();
 472
 473    vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
 474
 475    vk_free(&instance->alloc, instance);
 476 }
 477
 478 static VkResult
 479 tu_enumerate_devices(struct tu_instance *instance)
 480 {
 481    /* TODO: Check for more devices ? */
 482    drmDevicePtr devices[8];
 483    VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
 484    int max_devices;
 485
 486    instance->physical_device_count = 0;
 487
 488    max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
 489
 490    if (instance->debug_flags & TU_DEBUG_STARTUP)
 491       tu_logi("Found %d drm nodes", max_devices);
 492
 493    if (max_devices < 1)
 494       return vk_error(instance, VK_ERROR_INCOMPATIBLE_DRIVER);
 495
 496    for (unsigned i = 0; i < (unsigned) max_devices; i++) {
 497       if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
 498           devices[i]->bustype == DRM_BUS_PLATFORM) {
 499
 500          result = tu_physical_device_init(
 501             instance->physical_devices + instance->physical_device_count,
 502             instance, devices[i]);
 503          if (result == VK_SUCCESS)
 504             ++instance->physical_device_count;
 505          else if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 506             break;
 507       }
 508    }
 509    drmFreeDevices(devices, max_devices);
 510
 511    return result;
 512 }
 513
 514 VkResult
 515 tu_EnumeratePhysicalDevices(VkInstance _instance,
 516                             uint32_t *pPhysicalDeviceCount,
 517                             VkPhysicalDevice *pPhysicalDevices)
 518 {
 519    TU_FROM_HANDLE(tu_instance, instance, _instance);
 520    VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
 521
 522    VkResult result;
 523
 524    if (instance->physical_device_count < 0) {
 525       result = tu_enumerate_devices(instance);
 526       if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
 527          return result;
 528    }
 529
 530    for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
 531       vk_outarray_append(&out, p)
 532       {
 533          *p = tu_physical_device_to_handle(instance->physical_devices + i);
 534       }
 535    }
 536
 537    return vk_outarray_status(&out);
 538 }
 539
 540 VkResult
 541 tu_EnumeratePhysicalDeviceGroups(
 542    VkInstance _instance,
 543    uint32_t *pPhysicalDeviceGroupCount,
 544    VkPhysicalDeviceGroupProperties *pPhysicalDeviceGroupProperties)
 545 {
 546    TU_FROM_HANDLE(tu_instance, instance, _instance);
 547    VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
 548                     pPhysicalDeviceGroupCount);
 549    VkResult result;
 550
 551    if (instance->physical_device_count < 0) {
 552       result = tu_enumerate_devices(instance);
 553       if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
 554          return result;
 555    }
 556
 557    for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
 558       vk_outarray_append(&out, p)
 559       {
 560          p->physicalDeviceCount = 1;
 561          p->physicalDevices[0] =
 562             tu_physical_device_to_handle(instance->physical_devices + i);
 563          p->subsetAllocation = false;
 564       }
 565    }
 566
 567    return vk_outarray_status(&out);
 568 }
 569
 570 void
 571 tu_GetPhysicalDeviceFeatures(VkPhysicalDevice physicalDevice,
 572                              VkPhysicalDeviceFeatures *pFeatures)
 573 {
 574    memset(pFeatures, 0, sizeof(*pFeatures));
 575
 576    *pFeatures = (VkPhysicalDeviceFeatures) {
 577       .robustBufferAccess = false,
 578       .fullDrawIndexUint32 = false,
 579       .imageCubeArray = false,
 580       .independentBlend = false,
 581       .geometryShader = false,
 582       .tessellationShader = false,
 583       .sampleRateShading = false,
 584       .dualSrcBlend = false,
 585       .logicOp = false,
 586       .multiDrawIndirect = false,
 587       .drawIndirectFirstInstance = false,
 588       .depthClamp = false,
 589       .depthBiasClamp = false,
 590       .fillModeNonSolid = false,
 591       .depthBounds = false,
 592       .wideLines = false,
 593       .largePoints = false,
 594       .alphaToOne = false,
 595       .multiViewport = false,
 596       .samplerAnisotropy = true,
 597       .textureCompressionETC2 = true,
 598       .textureCompressionASTC_LDR = true,
 599       .textureCompressionBC = true,
 600       .occlusionQueryPrecise = true,
 601       .pipelineStatisticsQuery = false,
 602       .vertexPipelineStoresAndAtomics = false,
 603       .fragmentStoresAndAtomics = false,
 604       .shaderTessellationAndGeometryPointSize = false,
 605       .shaderImageGatherExtended = false,
 606       .shaderStorageImageExtendedFormats = false,
 607       .shaderStorageImageMultisample = false,
 608       .shaderUniformBufferArrayDynamicIndexing = false,
 609       .shaderSampledImageArrayDynamicIndexing = false,
 610       .shaderStorageBufferArrayDynamicIndexing = false,
 611       .shaderStorageImageArrayDynamicIndexing = false,
 612       .shaderStorageImageReadWithoutFormat = false,
 613       .shaderStorageImageWriteWithoutFormat = false,
 614       .shaderClipDistance = false,
 615       .shaderCullDistance = false,
 616       .shaderFloat64 = false,
 617       .shaderInt64 = false,
 618       .shaderInt16 = false,
 619       .sparseBinding = false,
 620       .variableMultisampleRate = false,
 621       .inheritedQueries = false,
 622    };
 623 }
 624
 625 void
 626 tu_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,
 627                               VkPhysicalDeviceFeatures2 *pFeatures)
 628 {
 629    vk_foreach_struct(ext, pFeatures->pNext)
 630    {
 631       switch (ext->sType) {
 632       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
 633          VkPhysicalDeviceVariablePointersFeatures *features = (void *) ext;
 634          features->variablePointersStorageBuffer = false;
 635          features->variablePointers = false;
 636          break;
 637       }
 638       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
 639          VkPhysicalDeviceMultiviewFeatures *features =
 640             (VkPhysicalDeviceMultiviewFeatures *) ext;
 641          features->multiview = false;
 642          features->multiviewGeometryShader = false;
 643          features->multiviewTessellationShader = false;
 644          break;
 645       }
 646       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
 647          VkPhysicalDeviceShaderDrawParametersFeatures *features =
 648             (VkPhysicalDeviceShaderDrawParametersFeatures *) ext;
 649          features->shaderDrawParameters = false;
 650          break;
 651       }
 652       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
 653          VkPhysicalDeviceProtectedMemoryFeatures *features =
 654             (VkPhysicalDeviceProtectedMemoryFeatures *) ext;
 655          features->protectedMemory = false;
 656          break;
 657       }
 658       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
 659          VkPhysicalDevice16BitStorageFeatures *features =
 660             (VkPhysicalDevice16BitStorageFeatures *) ext;
 661          features->storageBuffer16BitAccess = false;
 662          features->uniformAndStorageBuffer16BitAccess = false;
 663          features->storagePushConstant16 = false;
 664          features->storageInputOutput16 = false;
 665          break;
 666       }
 667       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
 668          VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
 669             (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
 670          features->samplerYcbcrConversion = false;
 671          break;
 672       }
 673       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
 674          VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
 675             (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *) ext;
 676          features->shaderInputAttachmentArrayDynamicIndexing = false;
 677          features->shaderUniformTexelBufferArrayDynamicIndexing = false;
 678          features->shaderStorageTexelBufferArrayDynamicIndexing = false;
 679          features->shaderUniformBufferArrayNonUniformIndexing = false;
 680          features->shaderSampledImageArrayNonUniformIndexing = false;
 681          features->shaderStorageBufferArrayNonUniformIndexing = false;
 682          features->shaderStorageImageArrayNonUniformIndexing = false;
 683          features->shaderInputAttachmentArrayNonUniformIndexing = false;
 684          features->shaderUniformTexelBufferArrayNonUniformIndexing = false;
 685          features->shaderStorageTexelBufferArrayNonUniformIndexing = false;
 686          features->descriptorBindingUniformBufferUpdateAfterBind = false;
 687          features->descriptorBindingSampledImageUpdateAfterBind = false;
 688          features->descriptorBindingStorageImageUpdateAfterBind = false;
 689          features->descriptorBindingStorageBufferUpdateAfterBind = false;
 690          features->descriptorBindingUniformTexelBufferUpdateAfterBind = false;
 691          features->descriptorBindingStorageTexelBufferUpdateAfterBind = false;
 692          features->descriptorBindingUpdateUnusedWhilePending = false;
 693          features->descriptorBindingPartiallyBound = false;
 694          features->descriptorBindingVariableDescriptorCount = false;
 695          features->runtimeDescriptorArray = false;
 696          break;
 697       }
 698       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
 699          VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
 700             (VkPhysicalDeviceConditionalRenderingFeaturesEXT *) ext;
 701          features->conditionalRendering = false;
 702          features->inheritedConditionalRendering = false;
 703          break;
 704       }
 705       default:
 706          break;
 707       }
 708    }
 709    return tu_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 710 }
 711
 712 void
 713 tu_GetPhysicalDeviceProperties(VkPhysicalDevice physicalDevice,
 714                                VkPhysicalDeviceProperties *pProperties)
 715 {
 716    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 717    VkSampleCountFlags sample_counts = VK_SAMPLE_COUNT_1_BIT |
 718       VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT | VK_SAMPLE_COUNT_8_BIT;
 719
 720    /* make sure that the entire descriptor set is addressable with a signed
 721     * 32-bit int. So the sum of all limits scaled by descriptor size has to
 722     * be at most 2 GiB. the combined image & samples object count as one of
 723     * both. This limit is for the pipeline layout, not for the set layout, but
 724     * there is no set limit, so we just set a pipeline limit. I don't think
 725     * any app is going to hit this soon. */
 726    size_t max_descriptor_set_size =
 727       ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS) /
 728       (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
 729        32 /* storage buffer, 32 due to potential space wasted on alignment */ +
 730        32 /* sampler, largest when combined with image */ +
 731        64 /* sampled image */ + 64 /* storage image */);
 732
 733    VkPhysicalDeviceLimits limits = {
 734       .maxImageDimension1D = (1 << 14),
 735       .maxImageDimension2D = (1 << 14),
 736       .maxImageDimension3D = (1 << 11),
 737       .maxImageDimensionCube = (1 << 14),
 738       .maxImageArrayLayers = (1 << 11),
 739       .maxTexelBufferElements = 128 * 1024 * 1024,
 740       .maxUniformBufferRange = UINT32_MAX,
 741       .maxStorageBufferRange = MAX_STORAGE_BUFFER_RANGE,
 742       .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
 743       .maxMemoryAllocationCount = UINT32_MAX,
 744       .maxSamplerAllocationCount = 64 * 1024,
 745       .bufferImageGranularity = 64,          /* A cache line */
 746       .sparseAddressSpaceSize = 0xffffffffu, /* buffer max size */
 747       .maxBoundDescriptorSets = MAX_SETS,
 748       .maxPerStageDescriptorSamplers = max_descriptor_set_size,
 749       .maxPerStageDescriptorUniformBuffers = max_descriptor_set_size,
 750       .maxPerStageDescriptorStorageBuffers = max_descriptor_set_size,
 751       .maxPerStageDescriptorSampledImages = max_descriptor_set_size,
 752       .maxPerStageDescriptorStorageImages = max_descriptor_set_size,
 753       .maxPerStageDescriptorInputAttachments = max_descriptor_set_size,
 754       .maxPerStageResources = max_descriptor_set_size,
 755       .maxDescriptorSetSamplers = max_descriptor_set_size,
 756       .maxDescriptorSetUniformBuffers = max_descriptor_set_size,
 757       .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS,
 758       .maxDescriptorSetStorageBuffers = max_descriptor_set_size,
 759       .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS,
 760       .maxDescriptorSetSampledImages = max_descriptor_set_size,
 761       .maxDescriptorSetStorageImages = max_descriptor_set_size,
 762       .maxDescriptorSetInputAttachments = max_descriptor_set_size,
 763       .maxVertexInputAttributes = 32,
 764       .maxVertexInputBindings = 32,
 765       .maxVertexInputAttributeOffset = 2047,
 766       .maxVertexInputBindingStride = 2048,
 767       .maxVertexOutputComponents = 128,
 768       .maxTessellationGenerationLevel = 64,
 769       .maxTessellationPatchSize = 32,
 770       .maxTessellationControlPerVertexInputComponents = 128,
 771       .maxTessellationControlPerVertexOutputComponents = 128,
 772       .maxTessellationControlPerPatchOutputComponents = 120,
 773       .maxTessellationControlTotalOutputComponents = 4096,
 774       .maxTessellationEvaluationInputComponents = 128,
 775       .maxTessellationEvaluationOutputComponents = 128,
 776       .maxGeometryShaderInvocations = 127,
 777       .maxGeometryInputComponents = 64,
 778       .maxGeometryOutputComponents = 128,
 779       .maxGeometryOutputVertices = 256,
 780       .maxGeometryTotalOutputComponents = 1024,
 781       .maxFragmentInputComponents = 128,
 782       .maxFragmentOutputAttachments = 8,
 783       .maxFragmentDualSrcAttachments = 1,
 784       .maxFragmentCombinedOutputResources = 8,
 785       .maxComputeSharedMemorySize = 32768,
 786       .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
 787       .maxComputeWorkGroupInvocations = 2048,
 788       .maxComputeWorkGroupSize = { 2048, 2048, 2048 },
 789       .subPixelPrecisionBits = 4 /* FIXME */,
 790       .subTexelPrecisionBits = 4 /* FIXME */,
 791       .mipmapPrecisionBits = 4 /* FIXME */,
 792       .maxDrawIndexedIndexValue = UINT32_MAX,
 793       .maxDrawIndirectCount = UINT32_MAX,
 794       .maxSamplerLodBias = 16,
 795       .maxSamplerAnisotropy = 16,
 796       .maxViewports = MAX_VIEWPORTS,
 797       .maxViewportDimensions = { (1 << 14), (1 << 14) },
 798       .viewportBoundsRange = { INT16_MIN, INT16_MAX },
 799       .viewportSubPixelBits = 8,
 800       .minMemoryMapAlignment = 4096, /* A page */
 801       .minTexelBufferOffsetAlignment = 64,
 802       .minUniformBufferOffsetAlignment = 4,
 803       .minStorageBufferOffsetAlignment = 4,
 804       .minTexelOffset = -32,
 805       .maxTexelOffset = 31,
 806       .minTexelGatherOffset = -32,
 807       .maxTexelGatherOffset = 31,
 808       .minInterpolationOffset = -2,
 809       .maxInterpolationOffset = 2,
 810       .subPixelInterpolationOffsetBits = 8,
 811       .maxFramebufferWidth = (1 << 14),
 812       .maxFramebufferHeight = (1 << 14),
 813       .maxFramebufferLayers = (1 << 10),
 814       .framebufferColorSampleCounts = sample_counts,
 815       .framebufferDepthSampleCounts = sample_counts,
 816       .framebufferStencilSampleCounts = sample_counts,
 817       .framebufferNoAttachmentsSampleCounts = sample_counts,
 818       .maxColorAttachments = MAX_RTS,
 819       .sampledImageColorSampleCounts = sample_counts,
 820       .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 821       .sampledImageDepthSampleCounts = sample_counts,
 822       .sampledImageStencilSampleCounts = sample_counts,
 823       .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 824       .maxSampleMaskWords = 1,
 825       .timestampComputeAndGraphics = false, /* FINISHME */
 826       .timestampPeriod = 1,
 827       .maxClipDistances = 8,
 828       .maxCullDistances = 8,
 829       .maxCombinedClipAndCullDistances = 8,
 830       .discreteQueuePriorities = 1,
 831       .pointSizeRange = { 0.125, 255.875 },
 832       .lineWidthRange = { 0.0, 7.9921875 },
 833       .pointSizeGranularity = (1.0 / 8.0),
 834       .lineWidthGranularity = (1.0 / 128.0),
 835       .strictLines = false, /* FINISHME */
 836       .standardSampleLocations = true,
 837       .optimalBufferCopyOffsetAlignment = 128,
 838       .optimalBufferCopyRowPitchAlignment = 128,
 839       .nonCoherentAtomSize = 64,
 840    };
 841
 842    *pProperties = (VkPhysicalDeviceProperties) {
 843       .apiVersion = tu_physical_device_api_version(pdevice),
 844       .driverVersion = vk_get_driver_version(),
 845       .vendorID = 0, /* TODO */
 846       .deviceID = 0,
 847       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 848       .limits = limits,
 849       .sparseProperties = { 0 },
 850    };
 851
 852    strcpy(pProperties->deviceName, pdevice->name);
 853    memcpy(pProperties->pipelineCacheUUID, pdevice->cache_uuid, VK_UUID_SIZE);
 854 }
 855
 856 void
 857 tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
 858                                 VkPhysicalDeviceProperties2 *pProperties)
 859 {
 860    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 861    tu_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 862
 863    vk_foreach_struct(ext, pProperties->pNext)
 864    {
 865       switch (ext->sType) {
 866       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 867          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 868             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 869          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 870          break;
 871       }
 872       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
 873          VkPhysicalDeviceIDProperties *properties =
 874             (VkPhysicalDeviceIDProperties *) ext;
 875          memcpy(properties->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
 876          memcpy(properties->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
 877          properties->deviceLUIDValid = false;
 878          break;
 879       }
 880       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
 881          VkPhysicalDeviceMultiviewProperties *properties =
 882             (VkPhysicalDeviceMultiviewProperties *) ext;
 883          properties->maxMultiviewViewCount = MAX_VIEWS;
 884          properties->maxMultiviewInstanceIndex = INT_MAX;
 885          break;
 886       }
 887       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
 888          VkPhysicalDevicePointClippingProperties *properties =
 889             (VkPhysicalDevicePointClippingProperties *) ext;
 890          properties->pointClippingBehavior =
 891             VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
 892          break;
 893       }
 894       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
 895          VkPhysicalDeviceMaintenance3Properties *properties =
 896             (VkPhysicalDeviceMaintenance3Properties *) ext;
 897          /* Make sure everything is addressable by a signed 32-bit int, and
 898           * our largest descriptors are 96 bytes. */
 899          properties->maxPerSetDescriptors = (1ull << 31) / 96;
 900          /* Our buffer size fields allow only this much */
 901          properties->maxMemoryAllocationSize = 0xFFFFFFFFull;
 902          break;
 903       }
 904       default:
 905          break;
 906       }
 907    }
 908 }
 909
 910 static const VkQueueFamilyProperties tu_queue_family_properties = {
 911    .queueFlags =
 912       VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT,
 913    .queueCount = 1,
 914    .timestampValidBits = 0, /* FINISHME */
 915    .minImageTransferGranularity = { 1, 1, 1 },
 916 };
 917
 918 void
 919 tu_GetPhysicalDeviceQueueFamilyProperties(
 920    VkPhysicalDevice physicalDevice,
 921    uint32_t *pQueueFamilyPropertyCount,
 922    VkQueueFamilyProperties *pQueueFamilyProperties)
 923 {
 924    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 925
 926    vk_outarray_append(&out, p) { *p = tu_queue_family_properties; }
 927 }
 928
 929 void
 930 tu_GetPhysicalDeviceQueueFamilyProperties2(
 931    VkPhysicalDevice physicalDevice,
 932    uint32_t *pQueueFamilyPropertyCount,
 933    VkQueueFamilyProperties2 *pQueueFamilyProperties)
 934 {
 935    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 936
 937    vk_outarray_append(&out, p)
 938    {
 939       p->queueFamilyProperties = tu_queue_family_properties;
 940    }
 941 }
 942
 943 static uint64_t
 944 tu_get_system_heap_size()
 945 {
 946    struct sysinfo info;
 947    sysinfo(&info);
 948
 949    uint64_t total_ram = (uint64_t) info.totalram * (uint64_t) info.mem_unit;
 950
 951    /* We don't want to burn too much ram with the GPU.  If the user has 4GiB
 952     * or less, we use at most half.  If they have more than 4GiB, we use 3/4.
 953     */
 954    uint64_t available_ram;
 955    if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
 956       available_ram = total_ram / 2;
 957    else
 958       available_ram = total_ram * 3 / 4;
 959
 960    return available_ram;
 961 }
 962
 963 void
 964 tu_GetPhysicalDeviceMemoryProperties(
 965    VkPhysicalDevice physicalDevice,
 966    VkPhysicalDeviceMemoryProperties *pMemoryProperties)
 967 {
 968    pMemoryProperties->memoryHeapCount = 1;
 969    pMemoryProperties->memoryHeaps[0].size = tu_get_system_heap_size();
 970    pMemoryProperties->memoryHeaps[0].flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT;
 971
 972    pMemoryProperties->memoryTypeCount = 1;
 973    pMemoryProperties->memoryTypes[0].propertyFlags =
 974       VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 975       VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 976       VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
 977    pMemoryProperties->memoryTypes[0].heapIndex = 0;
 978 }
 979
 980 void
 981 tu_GetPhysicalDeviceMemoryProperties2(
 982    VkPhysicalDevice physicalDevice,
 983    VkPhysicalDeviceMemoryProperties2 *pMemoryProperties)
 984 {
 985    return tu_GetPhysicalDeviceMemoryProperties(
 986       physicalDevice, &pMemoryProperties->memoryProperties);
 987 }
 988
 989 static VkResult
 990 tu_queue_init(struct tu_device *device,
 991               struct tu_queue *queue,
 992               uint32_t queue_family_index,
 993               int idx,
 994               VkDeviceQueueCreateFlags flags)
 995 {
 996    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 997    queue->device = device;
 998    queue->queue_family_index = queue_family_index;
 999    queue->queue_idx = idx;
1000    queue->flags = flags;
1001
1002    int ret = tu_drm_submitqueue_new(device, 0, &queue->msm_queue_id);
1003    if (ret)
1004       return VK_ERROR_INITIALIZATION_FAILED;
1005
1006    tu_fence_init(&queue->submit_fence, false);
1007
1008    return VK_SUCCESS;
1009 }
1010
1011 static void
1012 tu_queue_finish(struct tu_queue *queue)
1013 {
1014    tu_fence_finish(&queue->submit_fence);
1015    tu_drm_submitqueue_close(queue->device, queue->msm_queue_id);
1016 }
1017
1018 static int
1019 tu_get_device_extension_index(const char *name)
1020 {
1021    for (unsigned i = 0; i < TU_DEVICE_EXTENSION_COUNT; ++i) {
1022       if (strcmp(name, tu_device_extensions[i].extensionName) == 0)
1023          return i;
1024    }
1025    return -1;
1026 }
1027
1028 VkResult
1029 tu_CreateDevice(VkPhysicalDevice physicalDevice,
1030                 const VkDeviceCreateInfo *pCreateInfo,
1031                 const VkAllocationCallbacks *pAllocator,
1032                 VkDevice *pDevice)
1033 {
1034    TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
1035    VkResult result;
1036    struct tu_device *device;
1037
1038    /* Check enabled features */
1039    if (pCreateInfo->pEnabledFeatures) {
1040       VkPhysicalDeviceFeatures supported_features;
1041       tu_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1042       VkBool32 *supported_feature = (VkBool32 *) &supported_features;
1043       VkBool32 *enabled_feature = (VkBool32 *) pCreateInfo->pEnabledFeatures;
1044       unsigned num_features =
1045          sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1046       for (uint32_t i = 0; i < num_features; i++) {
1047          if (enabled_feature[i] && !supported_feature[i])
1048             return vk_error(physical_device->instance,
1049                             VK_ERROR_FEATURE_NOT_PRESENT);
1050       }
1051    }
1052
1053    device = vk_zalloc2(&physical_device->instance->alloc, pAllocator,
1054                        sizeof(*device), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1055    if (!device)
1056       return vk_error(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1057
1058    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1059    device->instance = physical_device->instance;
1060    device->physical_device = physical_device;
1061
1062    if (pAllocator)
1063       device->alloc = *pAllocator;
1064    else
1065       device->alloc = physical_device->instance->alloc;
1066
1067    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1068       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
1069       int index = tu_get_device_extension_index(ext_name);
1070       if (index < 0 ||
1071           !physical_device->supported_extensions.extensions[index]) {
1072          vk_free(&device->alloc, device);
1073          return vk_error(physical_device->instance,
1074                          VK_ERROR_EXTENSION_NOT_PRESENT);
1075       }
1076
1077       device->enabled_extensions.extensions[index] = true;
1078    }
1079
1080    for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1081       const VkDeviceQueueCreateInfo *queue_create =
1082          &pCreateInfo->pQueueCreateInfos[i];
1083       uint32_t qfi = queue_create->queueFamilyIndex;
1084       device->queues[qfi] = vk_alloc(
1085          &device->alloc, queue_create->queueCount * sizeof(struct tu_queue),
1086          8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1087       if (!device->queues[qfi]) {
1088          result = VK_ERROR_OUT_OF_HOST_MEMORY;
1089          goto fail_queues;
1090       }
1091
1092       memset(device->queues[qfi], 0,
1093              queue_create->queueCount * sizeof(struct tu_queue));
1094
1095       device->queue_count[qfi] = queue_create->queueCount;
1096
1097       for (unsigned q = 0; q < queue_create->queueCount; q++) {
1098          result = tu_queue_init(device, &device->queues[qfi][q], qfi, q,
1099                                 queue_create->flags);
1100          if (result != VK_SUCCESS)
1101             goto fail_queues;
1102       }
1103    }
1104
1105    device->compiler = ir3_compiler_create(NULL, physical_device->gpu_id);
1106    if (!device->compiler)
1107       goto fail_queues;
1108
1109 #define VSC_DATA_SIZE(pitch)  ((pitch) * 32 + 0x100)  /* extra size to store VSC_SIZE */
1110 #define VSC_DATA2_SIZE(pitch) ((pitch) * 32)
1111
1112    device->vsc_data_pitch = 0x440 * 4;
1113    device->vsc_data2_pitch = 0x1040 * 4;
1114
1115    result = tu_bo_init_new(device, &device->vsc_data, VSC_DATA_SIZE(device->vsc_data_pitch));
1116    if (result != VK_SUCCESS)
1117       goto fail_vsc_data;
1118
1119    result = tu_bo_init_new(device, &device->vsc_data2, VSC_DATA2_SIZE(device->vsc_data2_pitch));
1120    if (result != VK_SUCCESS)
1121       goto fail_vsc_data2;
1122
1123    VkPipelineCacheCreateInfo ci;
1124    ci.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
1125    ci.pNext = NULL;
1126    ci.flags = 0;
1127    ci.pInitialData = NULL;
1128    ci.initialDataSize = 0;
1129    VkPipelineCache pc;
1130    result =
1131       tu_CreatePipelineCache(tu_device_to_handle(device), &ci, NULL, &pc);
1132    if (result != VK_SUCCESS)
1133       goto fail_pipeline_cache;
1134
1135    device->mem_cache = tu_pipeline_cache_from_handle(pc);
1136
1137    *pDevice = tu_device_to_handle(device);
1138    return VK_SUCCESS;
1139
1140 fail_pipeline_cache:
1141    tu_bo_finish(device, &device->vsc_data2);
1142
1143 fail_vsc_data2:
1144    tu_bo_finish(device, &device->vsc_data);
1145
1146 fail_vsc_data:
1147    ralloc_free(device->compiler);
1148
1149 fail_queues:
1150    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1151       for (unsigned q = 0; q < device->queue_count[i]; q++)
1152          tu_queue_finish(&device->queues[i][q]);
1153       if (device->queue_count[i])
1154          vk_free(&device->alloc, device->queues[i]);
1155    }
1156
1157    vk_free(&device->alloc, device);
1158    return result;
1159 }
1160
1161 void
1162 tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
1163 {
1164    TU_FROM_HANDLE(tu_device, device, _device);
1165
1166    if (!device)
1167       return;
1168
1169    tu_bo_finish(device, &device->vsc_data);
1170    tu_bo_finish(device, &device->vsc_data2);
1171
1172    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1173       for (unsigned q = 0; q < device->queue_count[i]; q++)
1174          tu_queue_finish(&device->queues[i][q]);
1175       if (device->queue_count[i])
1176          vk_free(&device->alloc, device->queues[i]);
1177    }
1178
1179    /* the compiler does not use pAllocator */
1180    ralloc_free(device->compiler);
1181
1182    VkPipelineCache pc = tu_pipeline_cache_to_handle(device->mem_cache);
1183    tu_DestroyPipelineCache(tu_device_to_handle(device), pc, NULL);
1184
1185    vk_free(&device->alloc, device);
1186 }
1187
1188 VkResult
1189 tu_EnumerateInstanceLayerProperties(uint32_t *pPropertyCount,
1190                                     VkLayerProperties *pProperties)
1191 {
1192    *pPropertyCount = 0;
1193    return VK_SUCCESS;
1194 }
1195
1196 VkResult
1197 tu_EnumerateDeviceLayerProperties(VkPhysicalDevice physicalDevice,
1198                                   uint32_t *pPropertyCount,
1199                                   VkLayerProperties *pProperties)
1200 {
1201    *pPropertyCount = 0;
1202    return VK_SUCCESS;
1203 }
1204
1205 void
1206 tu_GetDeviceQueue2(VkDevice _device,
1207                    const VkDeviceQueueInfo2 *pQueueInfo,
1208                    VkQueue *pQueue)
1209 {
1210    TU_FROM_HANDLE(tu_device, device, _device);
1211    struct tu_queue *queue;
1212
1213    queue =
1214       &device->queues[pQueueInfo->queueFamilyIndex][pQueueInfo->queueIndex];
1215    if (pQueueInfo->flags != queue->flags) {
1216       /* From the Vulkan 1.1.70 spec:
1217        *
1218        * "The queue returned by vkGetDeviceQueue2 must have the same
1219        * flags value from this structure as that used at device
1220        * creation time in a VkDeviceQueueCreateInfo instance. If no
1221        * matching flags were specified at device creation time then
1222        * pQueue will return VK_NULL_HANDLE."
1223        */
1224       *pQueue = VK_NULL_HANDLE;
1225       return;
1226    }
1227
1228    *pQueue = tu_queue_to_handle(queue);
1229 }
1230
1231 void
1232 tu_GetDeviceQueue(VkDevice _device,
1233                   uint32_t queueFamilyIndex,
1234                   uint32_t queueIndex,
1235                   VkQueue *pQueue)
1236 {
1237    const VkDeviceQueueInfo2 info =
1238       (VkDeviceQueueInfo2) { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
1239                              .queueFamilyIndex = queueFamilyIndex,
1240                              .queueIndex = queueIndex };
1241
1242    tu_GetDeviceQueue2(_device, &info, pQueue);
1243 }
1244
1245 VkResult
1246 tu_QueueSubmit(VkQueue _queue,
1247                uint32_t submitCount,
1248                const VkSubmitInfo *pSubmits,
1249                VkFence _fence)
1250 {
1251    TU_FROM_HANDLE(tu_queue, queue, _queue);
1252
1253    for (uint32_t i = 0; i < submitCount; ++i) {
1254       const VkSubmitInfo *submit = pSubmits + i;
1255       const bool last_submit = (i == submitCount - 1);
1256       struct tu_bo_list bo_list;
1257       tu_bo_list_init(&bo_list);
1258
1259       uint32_t entry_count = 0;
1260       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1261          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1262          entry_count += cmdbuf->cs.entry_count;
1263       }
1264
1265       struct drm_msm_gem_submit_cmd cmds[entry_count];
1266       uint32_t entry_idx = 0;
1267       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1268          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1269          struct tu_cs *cs = &cmdbuf->cs;
1270          for (unsigned i = 0; i < cs->entry_count; ++i, ++entry_idx) {
1271             cmds[entry_idx].type = MSM_SUBMIT_CMD_BUF;
1272             cmds[entry_idx].submit_idx =
1273                tu_bo_list_add(&bo_list, cs->entries[i].bo,
1274                               MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
1275             cmds[entry_idx].submit_offset = cs->entries[i].offset;
1276             cmds[entry_idx].size = cs->entries[i].size;
1277             cmds[entry_idx].pad = 0;
1278             cmds[entry_idx].nr_relocs = 0;
1279             cmds[entry_idx].relocs = 0;
1280          }
1281
1282          tu_bo_list_merge(&bo_list, &cmdbuf->bo_list);
1283       }
1284
1285       uint32_t flags = MSM_PIPE_3D0;
1286       if (last_submit) {
1287          flags |= MSM_SUBMIT_FENCE_FD_OUT;
1288       }
1289
1290       struct drm_msm_gem_submit req = {
1291          .flags = flags,
1292          .queueid = queue->msm_queue_id,
1293          .bos = (uint64_t)(uintptr_t) bo_list.bo_infos,
1294          .nr_bos = bo_list.count,
1295          .cmds = (uint64_t)(uintptr_t)cmds,
1296          .nr_cmds = entry_count,
1297       };
1298
1299       int ret = drmCommandWriteRead(queue->device->physical_device->local_fd,
1300                                     DRM_MSM_GEM_SUBMIT,
1301                                     &req, sizeof(req));
1302       if (ret) {
1303          fprintf(stderr, "submit failed: %s\n", strerror(errno));
1304          abort();
1305       }
1306
1307       tu_bo_list_destroy(&bo_list);
1308
1309       if (last_submit) {
1310          /* no need to merge fences as queue execution is serialized */
1311          tu_fence_update_fd(&queue->submit_fence, req.fence_fd);
1312       }
1313    }
1314
1315    if (_fence != VK_NULL_HANDLE) {
1316       TU_FROM_HANDLE(tu_fence, fence, _fence);
1317       tu_fence_copy(fence, &queue->submit_fence);
1318    }
1319
1320    return VK_SUCCESS;
1321 }
1322
1323 VkResult
1324 tu_QueueWaitIdle(VkQueue _queue)
1325 {
1326    TU_FROM_HANDLE(tu_queue, queue, _queue);
1327
1328    tu_fence_wait_idle(&queue->submit_fence);
1329
1330    return VK_SUCCESS;
1331 }
1332
1333 VkResult
1334 tu_DeviceWaitIdle(VkDevice _device)
1335 {
1336    TU_FROM_HANDLE(tu_device, device, _device);
1337
1338    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1339       for (unsigned q = 0; q < device->queue_count[i]; q++) {
1340          tu_QueueWaitIdle(tu_queue_to_handle(&device->queues[i][q]));
1341       }
1342    }
1343    return VK_SUCCESS;
1344 }
1345
1346 VkResult
1347 tu_EnumerateInstanceExtensionProperties(const char *pLayerName,
1348                                         uint32_t *pPropertyCount,
1349                                         VkExtensionProperties *pProperties)
1350 {
1351    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1352
1353    /* We spport no lyaers */
1354    if (pLayerName)
1355       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1356
1357    for (int i = 0; i < TU_INSTANCE_EXTENSION_COUNT; i++) {
1358       if (tu_supported_instance_extensions.extensions[i]) {
1359          vk_outarray_append(&out, prop) { *prop = tu_instance_extensions[i]; }
1360       }
1361    }
1362
1363    return vk_outarray_status(&out);
1364 }
1365
1366 VkResult
1367 tu_EnumerateDeviceExtensionProperties(VkPhysicalDevice physicalDevice,
1368                                       const char *pLayerName,
1369                                       uint32_t *pPropertyCount,
1370                                       VkExtensionProperties *pProperties)
1371 {
1372    /* We spport no lyaers */
1373    TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
1374    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1375
1376    /* We spport no lyaers */
1377    if (pLayerName)
1378       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1379
1380    for (int i = 0; i < TU_DEVICE_EXTENSION_COUNT; i++) {
1381       if (device->supported_extensions.extensions[i]) {
1382          vk_outarray_append(&out, prop) { *prop = tu_device_extensions[i]; }
1383       }
1384    }
1385
1386    return vk_outarray_status(&out);
1387 }
1388
1389 PFN_vkVoidFunction
1390 tu_GetInstanceProcAddr(VkInstance _instance, const char *pName)
1391 {
1392    TU_FROM_HANDLE(tu_instance, instance, _instance);
1393
1394    return tu_lookup_entrypoint_checked(
1395       pName, instance ? instance->api_version : 0,
1396       instance ? &instance->enabled_extensions : NULL, NULL);
1397 }
1398
1399 /* The loader wants us to expose a second GetInstanceProcAddr function
1400  * to work around certain LD_PRELOAD issues seen in apps.
1401  */
1402 PUBLIC
1403 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1404 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName);
1405
1406 PUBLIC
1407 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1408 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName)
1409 {
1410    return tu_GetInstanceProcAddr(instance, pName);
1411 }
1412
1413 PFN_vkVoidFunction
1414 tu_GetDeviceProcAddr(VkDevice _device, const char *pName)
1415 {
1416    TU_FROM_HANDLE(tu_device, device, _device);
1417
1418    return tu_lookup_entrypoint_checked(pName, device->instance->api_version,
1419                                        &device->instance->enabled_extensions,
1420                                        &device->enabled_extensions);
1421 }
1422
1423 static VkResult
1424 tu_alloc_memory(struct tu_device *device,
1425                 const VkMemoryAllocateInfo *pAllocateInfo,
1426                 const VkAllocationCallbacks *pAllocator,
1427                 VkDeviceMemory *pMem)
1428 {
1429    struct tu_device_memory *mem;
1430    VkResult result;
1431
1432    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1433
1434    if (pAllocateInfo->allocationSize == 0) {
1435       /* Apparently, this is allowed */
1436       *pMem = VK_NULL_HANDLE;
1437       return VK_SUCCESS;
1438    }
1439
1440    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1441                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1442    if (mem == NULL)
1443       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1444
1445    const VkImportMemoryFdInfoKHR *fd_info =
1446       vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
1447    if (fd_info && !fd_info->handleType)
1448       fd_info = NULL;
1449
1450    if (fd_info) {
1451       assert(fd_info->handleType ==
1452                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
1453              fd_info->handleType ==
1454                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1455
1456       /*
1457        * TODO Importing the same fd twice gives us the same handle without
1458        * reference counting.  We need to maintain a per-instance handle-to-bo
1459        * table and add reference count to tu_bo.
1460        */
1461       result = tu_bo_init_dmabuf(device, &mem->bo,
1462                                  pAllocateInfo->allocationSize, fd_info->fd);
1463       if (result == VK_SUCCESS) {
1464          /* take ownership and close the fd */
1465          close(fd_info->fd);
1466       }
1467    } else {
1468       result =
1469          tu_bo_init_new(device, &mem->bo, pAllocateInfo->allocationSize);
1470    }
1471
1472    if (result != VK_SUCCESS) {
1473       vk_free2(&device->alloc, pAllocator, mem);
1474       return result;
1475    }
1476
1477    mem->size = pAllocateInfo->allocationSize;
1478    mem->type_index = pAllocateInfo->memoryTypeIndex;
1479
1480    mem->map = NULL;
1481    mem->user_ptr = NULL;
1482
1483    *pMem = tu_device_memory_to_handle(mem);
1484
1485    return VK_SUCCESS;
1486 }
1487
1488 VkResult
1489 tu_AllocateMemory(VkDevice _device,
1490                   const VkMemoryAllocateInfo *pAllocateInfo,
1491                   const VkAllocationCallbacks *pAllocator,
1492                   VkDeviceMemory *pMem)
1493 {
1494    TU_FROM_HANDLE(tu_device, device, _device);
1495    return tu_alloc_memory(device, pAllocateInfo, pAllocator, pMem);
1496 }
1497
1498 void
1499 tu_FreeMemory(VkDevice _device,
1500               VkDeviceMemory _mem,
1501               const VkAllocationCallbacks *pAllocator)
1502 {
1503    TU_FROM_HANDLE(tu_device, device, _device);
1504    TU_FROM_HANDLE(tu_device_memory, mem, _mem);
1505
1506    if (mem == NULL)
1507       return;
1508
1509    tu_bo_finish(device, &mem->bo);
1510    vk_free2(&device->alloc, pAllocator, mem);
1511 }
1512
1513 VkResult
1514 tu_MapMemory(VkDevice _device,
1515              VkDeviceMemory _memory,
1516              VkDeviceSize offset,
1517              VkDeviceSize size,
1518              VkMemoryMapFlags flags,
1519              void **ppData)
1520 {
1521    TU_FROM_HANDLE(tu_device, device, _device);
1522    TU_FROM_HANDLE(tu_device_memory, mem, _memory);
1523    VkResult result;
1524
1525    if (mem == NULL) {
1526       *ppData = NULL;
1527       return VK_SUCCESS;
1528    }
1529
1530    if (mem->user_ptr) {
1531       *ppData = mem->user_ptr;
1532    } else if (!mem->map) {
1533       result = tu_bo_map(device, &mem->bo);
1534       if (result != VK_SUCCESS)
1535          return result;
1536       *ppData = mem->map = mem->bo.map;
1537    } else
1538       *ppData = mem->map;
1539
1540    if (*ppData) {
1541       *ppData += offset;
1542       return VK_SUCCESS;
1543    }
1544
1545    return vk_error(device->instance, VK_ERROR_MEMORY_MAP_FAILED);
1546 }
1547
1548 void
1549 tu_UnmapMemory(VkDevice _device, VkDeviceMemory _memory)
1550 {
1551    /* I do not see any unmapping done by the freedreno Gallium driver. */
1552 }
1553
1554 VkResult
1555 tu_FlushMappedMemoryRanges(VkDevice _device,
1556                            uint32_t memoryRangeCount,
1557                            const VkMappedMemoryRange *pMemoryRanges)
1558 {
1559    return VK_SUCCESS;
1560 }
1561
1562 VkResult
1563 tu_InvalidateMappedMemoryRanges(VkDevice _device,
1564                                 uint32_t memoryRangeCount,
1565                                 const VkMappedMemoryRange *pMemoryRanges)
1566 {
1567    return VK_SUCCESS;
1568 }
1569
1570 void
1571 tu_GetBufferMemoryRequirements(VkDevice _device,
1572                                VkBuffer _buffer,
1573                                VkMemoryRequirements *pMemoryRequirements)
1574 {
1575    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
1576
1577    pMemoryRequirements->memoryTypeBits = 1;
1578    pMemoryRequirements->alignment = 16;
1579    pMemoryRequirements->size =
1580       align64(buffer->size, pMemoryRequirements->alignment);
1581 }
1582
1583 void
1584 tu_GetBufferMemoryRequirements2(
1585    VkDevice device,
1586    const VkBufferMemoryRequirementsInfo2 *pInfo,
1587    VkMemoryRequirements2 *pMemoryRequirements)
1588 {
1589    tu_GetBufferMemoryRequirements(device, pInfo->buffer,
1590                                   &pMemoryRequirements->memoryRequirements);
1591 }
1592
1593 void
1594 tu_GetImageMemoryRequirements(VkDevice _device,
1595                               VkImage _image,
1596                               VkMemoryRequirements *pMemoryRequirements)
1597 {
1598    TU_FROM_HANDLE(tu_image, image, _image);
1599
1600    pMemoryRequirements->memoryTypeBits = 1;
1601    pMemoryRequirements->size = image->layout.size;
1602    pMemoryRequirements->alignment = image->alignment;
1603 }
1604
1605 void
1606 tu_GetImageMemoryRequirements2(VkDevice device,
1607                                const VkImageMemoryRequirementsInfo2 *pInfo,
1608                                VkMemoryRequirements2 *pMemoryRequirements)
1609 {
1610    tu_GetImageMemoryRequirements(device, pInfo->image,
1611                                  &pMemoryRequirements->memoryRequirements);
1612 }
1613
1614 void
1615 tu_GetImageSparseMemoryRequirements(
1616    VkDevice device,
1617    VkImage image,
1618    uint32_t *pSparseMemoryRequirementCount,
1619    VkSparseImageMemoryRequirements *pSparseMemoryRequirements)
1620 {
1621    tu_stub();
1622 }
1623
1624 void
1625 tu_GetImageSparseMemoryRequirements2(
1626    VkDevice device,
1627    const VkImageSparseMemoryRequirementsInfo2 *pInfo,
1628    uint32_t *pSparseMemoryRequirementCount,
1629    VkSparseImageMemoryRequirements2 *pSparseMemoryRequirements)
1630 {
1631    tu_stub();
1632 }
1633
1634 void
1635 tu_GetDeviceMemoryCommitment(VkDevice device,
1636                              VkDeviceMemory memory,
1637                              VkDeviceSize *pCommittedMemoryInBytes)
1638 {
1639    *pCommittedMemoryInBytes = 0;
1640 }
1641
1642 VkResult
1643 tu_BindBufferMemory2(VkDevice device,
1644                      uint32_t bindInfoCount,
1645                      const VkBindBufferMemoryInfo *pBindInfos)
1646 {
1647    for (uint32_t i = 0; i < bindInfoCount; ++i) {
1648       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
1649       TU_FROM_HANDLE(tu_buffer, buffer, pBindInfos[i].buffer);
1650
1651       if (mem) {
1652          buffer->bo = &mem->bo;
1653          buffer->bo_offset = pBindInfos[i].memoryOffset;
1654       } else {
1655          buffer->bo = NULL;
1656       }
1657    }
1658    return VK_SUCCESS;
1659 }
1660
1661 VkResult
1662 tu_BindBufferMemory(VkDevice device,
1663                     VkBuffer buffer,
1664                     VkDeviceMemory memory,
1665                     VkDeviceSize memoryOffset)
1666 {
1667    const VkBindBufferMemoryInfo info = {
1668       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
1669       .buffer = buffer,
1670       .memory = memory,
1671       .memoryOffset = memoryOffset
1672    };
1673
1674    return tu_BindBufferMemory2(device, 1, &info);
1675 }
1676
1677 VkResult
1678 tu_BindImageMemory2(VkDevice device,
1679                     uint32_t bindInfoCount,
1680                     const VkBindImageMemoryInfo *pBindInfos)
1681 {
1682    for (uint32_t i = 0; i < bindInfoCount; ++i) {
1683       TU_FROM_HANDLE(tu_image, image, pBindInfos[i].image);
1684       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
1685
1686       if (mem) {
1687          image->bo = &mem->bo;
1688          image->bo_offset = pBindInfos[i].memoryOffset;
1689       } else {
1690          image->bo = NULL;
1691          image->bo_offset = 0;
1692       }
1693    }
1694
1695    return VK_SUCCESS;
1696 }
1697
1698 VkResult
1699 tu_BindImageMemory(VkDevice device,
1700                    VkImage image,
1701                    VkDeviceMemory memory,
1702                    VkDeviceSize memoryOffset)
1703 {
1704    const VkBindImageMemoryInfo info = {
1705       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
1706       .image = image,
1707       .memory = memory,
1708       .memoryOffset = memoryOffset
1709    };
1710
1711    return tu_BindImageMemory2(device, 1, &info);
1712 }
1713
1714 VkResult
1715 tu_QueueBindSparse(VkQueue _queue,
1716                    uint32_t bindInfoCount,
1717                    const VkBindSparseInfo *pBindInfo,
1718                    VkFence _fence)
1719 {
1720    return VK_SUCCESS;
1721 }
1722
1723 // Queue semaphore functions
1724
1725 VkResult
1726 tu_CreateSemaphore(VkDevice _device,
1727                    const VkSemaphoreCreateInfo *pCreateInfo,
1728                    const VkAllocationCallbacks *pAllocator,
1729                    VkSemaphore *pSemaphore)
1730 {
1731    TU_FROM_HANDLE(tu_device, device, _device);
1732
1733    struct tu_semaphore *sem =
1734       vk_alloc2(&device->alloc, pAllocator, sizeof(*sem), 8,
1735                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1736    if (!sem)
1737       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1738
1739    *pSemaphore = tu_semaphore_to_handle(sem);
1740    return VK_SUCCESS;
1741 }
1742
1743 void
1744 tu_DestroySemaphore(VkDevice _device,
1745                     VkSemaphore _semaphore,
1746                     const VkAllocationCallbacks *pAllocator)
1747 {
1748    TU_FROM_HANDLE(tu_device, device, _device);
1749    TU_FROM_HANDLE(tu_semaphore, sem, _semaphore);
1750    if (!_semaphore)
1751       return;
1752
1753    vk_free2(&device->alloc, pAllocator, sem);
1754 }
1755
1756 VkResult
1757 tu_CreateEvent(VkDevice _device,
1758                const VkEventCreateInfo *pCreateInfo,
1759                const VkAllocationCallbacks *pAllocator,
1760                VkEvent *pEvent)
1761 {
1762    TU_FROM_HANDLE(tu_device, device, _device);
1763    struct tu_event *event =
1764       vk_alloc2(&device->alloc, pAllocator, sizeof(*event), 8,
1765                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1766
1767    if (!event)
1768       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1769
1770    VkResult result = tu_bo_init_new(device, &event->bo, 0x1000);
1771    if (result != VK_SUCCESS)
1772       goto fail_alloc;
1773
1774    result = tu_bo_map(device, &event->bo);
1775    if (result != VK_SUCCESS)
1776       goto fail_map;
1777
1778    *pEvent = tu_event_to_handle(event);
1779
1780    return VK_SUCCESS;
1781
1782 fail_map:
1783    tu_bo_finish(device, &event->bo);
1784 fail_alloc:
1785    vk_free2(&device->alloc, pAllocator, event);
1786    return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1787 }
1788
1789 void
1790 tu_DestroyEvent(VkDevice _device,
1791                 VkEvent _event,
1792                 const VkAllocationCallbacks *pAllocator)
1793 {
1794    TU_FROM_HANDLE(tu_device, device, _device);
1795    TU_FROM_HANDLE(tu_event, event, _event);
1796
1797    if (!event)
1798       return;
1799
1800    tu_bo_finish(device, &event->bo);
1801    vk_free2(&device->alloc, pAllocator, event);
1802 }
1803
1804 VkResult
1805 tu_GetEventStatus(VkDevice _device, VkEvent _event)
1806 {
1807    TU_FROM_HANDLE(tu_event, event, _event);
1808
1809    if (*(uint64_t*) event->bo.map == 1)
1810       return VK_EVENT_SET;
1811    return VK_EVENT_RESET;
1812 }
1813
1814 VkResult
1815 tu_SetEvent(VkDevice _device, VkEvent _event)
1816 {
1817    TU_FROM_HANDLE(tu_event, event, _event);
1818    *(uint64_t*) event->bo.map = 1;
1819
1820    return VK_SUCCESS;
1821 }
1822
1823 VkResult
1824 tu_ResetEvent(VkDevice _device, VkEvent _event)
1825 {
1826    TU_FROM_HANDLE(tu_event, event, _event);
1827    *(uint64_t*) event->bo.map = 0;
1828
1829    return VK_SUCCESS;
1830 }
1831
1832 VkResult
1833 tu_CreateBuffer(VkDevice _device,
1834                 const VkBufferCreateInfo *pCreateInfo,
1835                 const VkAllocationCallbacks *pAllocator,
1836                 VkBuffer *pBuffer)
1837 {
1838    TU_FROM_HANDLE(tu_device, device, _device);
1839    struct tu_buffer *buffer;
1840
1841    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1842
1843    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1844                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1845    if (buffer == NULL)
1846       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1847
1848    buffer->size = pCreateInfo->size;
1849    buffer->usage = pCreateInfo->usage;
1850    buffer->flags = pCreateInfo->flags;
1851
1852    *pBuffer = tu_buffer_to_handle(buffer);
1853
1854    return VK_SUCCESS;
1855 }
1856
1857 void
1858 tu_DestroyBuffer(VkDevice _device,
1859                  VkBuffer _buffer,
1860                  const VkAllocationCallbacks *pAllocator)
1861 {
1862    TU_FROM_HANDLE(tu_device, device, _device);
1863    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
1864
1865    if (!buffer)
1866       return;
1867
1868    vk_free2(&device->alloc, pAllocator, buffer);
1869 }
1870
1871 static uint32_t
1872 tu_surface_max_layer_count(struct tu_image_view *iview)
1873 {
1874    return iview->type == VK_IMAGE_VIEW_TYPE_3D
1875              ? iview->extent.depth
1876              : (iview->base_layer + iview->layer_count);
1877 }
1878
1879 VkResult
1880 tu_CreateFramebuffer(VkDevice _device,
1881                      const VkFramebufferCreateInfo *pCreateInfo,
1882                      const VkAllocationCallbacks *pAllocator,
1883                      VkFramebuffer *pFramebuffer)
1884 {
1885    TU_FROM_HANDLE(tu_device, device, _device);
1886    struct tu_framebuffer *framebuffer;
1887
1888    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
1889
1890    size_t size = sizeof(*framebuffer) + sizeof(struct tu_attachment_info) *
1891                                            pCreateInfo->attachmentCount;
1892    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
1893                            VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1894    if (framebuffer == NULL)
1895       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1896
1897    framebuffer->attachment_count = pCreateInfo->attachmentCount;
1898    framebuffer->width = pCreateInfo->width;
1899    framebuffer->height = pCreateInfo->height;
1900    framebuffer->layers = pCreateInfo->layers;
1901    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
1902       VkImageView _iview = pCreateInfo->pAttachments[i];
1903       struct tu_image_view *iview = tu_image_view_from_handle(_iview);
1904       framebuffer->attachments[i].attachment = iview;
1905
1906       framebuffer->width = MIN2(framebuffer->width, iview->extent.width);
1907       framebuffer->height = MIN2(framebuffer->height, iview->extent.height);
1908       framebuffer->layers =
1909          MIN2(framebuffer->layers, tu_surface_max_layer_count(iview));
1910    }
1911
1912    *pFramebuffer = tu_framebuffer_to_handle(framebuffer);
1913    return VK_SUCCESS;
1914 }
1915
1916 void
1917 tu_DestroyFramebuffer(VkDevice _device,
1918                       VkFramebuffer _fb,
1919                       const VkAllocationCallbacks *pAllocator)
1920 {
1921    TU_FROM_HANDLE(tu_device, device, _device);
1922    TU_FROM_HANDLE(tu_framebuffer, fb, _fb);
1923
1924    if (!fb)
1925       return;
1926    vk_free2(&device->alloc, pAllocator, fb);
1927 }
1928
1929 static enum a6xx_tex_clamp
1930 tu6_tex_wrap(VkSamplerAddressMode address_mode, bool *needs_border)
1931 {
1932    switch (address_mode) {
1933    case VK_SAMPLER_ADDRESS_MODE_REPEAT:
1934       return A6XX_TEX_REPEAT;
1935    case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT:
1936       return A6XX_TEX_MIRROR_REPEAT;
1937    case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE:
1938       return A6XX_TEX_CLAMP_TO_EDGE;
1939    case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER:
1940       *needs_border = true;
1941       return A6XX_TEX_CLAMP_TO_BORDER;
1942    case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE:
1943       /* only works for PoT.. need to emulate otherwise! */
1944       return A6XX_TEX_MIRROR_CLAMP;
1945    default:
1946       unreachable("illegal tex wrap mode");
1947       break;
1948    }
1949 }
1950
1951 static enum a6xx_tex_filter
1952 tu6_tex_filter(VkFilter filter, unsigned aniso)
1953 {
1954    switch (filter) {
1955    case VK_FILTER_NEAREST:
1956       return A6XX_TEX_NEAREST;
1957    case VK_FILTER_LINEAR:
1958       return aniso ? A6XX_TEX_ANISO : A6XX_TEX_LINEAR;
1959    case VK_FILTER_CUBIC_IMG:
1960    default:
1961       unreachable("illegal texture filter");
1962       break;
1963    }
1964 }
1965
1966 static inline enum adreno_compare_func
1967 tu6_compare_func(VkCompareOp op)
1968 {
1969    return (enum adreno_compare_func) op;
1970 }
1971
1972 static void
1973 tu_init_sampler(struct tu_device *device,
1974                 struct tu_sampler *sampler,
1975                 const VkSamplerCreateInfo *pCreateInfo)
1976 {
1977    unsigned aniso = pCreateInfo->anisotropyEnable ?
1978       util_last_bit(MIN2((uint32_t)pCreateInfo->maxAnisotropy >> 1, 8)) : 0;
1979    bool miplinear = (pCreateInfo->mipmapMode == VK_SAMPLER_MIPMAP_MODE_LINEAR);
1980    bool needs_border = false;
1981
1982    sampler->state[0] =
1983       COND(miplinear, A6XX_TEX_SAMP_0_MIPFILTER_LINEAR_NEAR) |
1984       A6XX_TEX_SAMP_0_XY_MAG(tu6_tex_filter(pCreateInfo->magFilter, aniso)) |
1985       A6XX_TEX_SAMP_0_XY_MIN(tu6_tex_filter(pCreateInfo->minFilter, aniso)) |
1986       A6XX_TEX_SAMP_0_ANISO(aniso) |
1987       A6XX_TEX_SAMP_0_WRAP_S(tu6_tex_wrap(pCreateInfo->addressModeU, &needs_border)) |
1988       A6XX_TEX_SAMP_0_WRAP_T(tu6_tex_wrap(pCreateInfo->addressModeV, &needs_border)) |
1989       A6XX_TEX_SAMP_0_WRAP_R(tu6_tex_wrap(pCreateInfo->addressModeW, &needs_border)) |
1990       A6XX_TEX_SAMP_0_LOD_BIAS(pCreateInfo->mipLodBias);
1991    sampler->state[1] =
1992       /* COND(!cso->seamless_cube_map, A6XX_TEX_SAMP_1_CUBEMAPSEAMLESSFILTOFF) | */
1993       COND(pCreateInfo->unnormalizedCoordinates, A6XX_TEX_SAMP_1_UNNORM_COORDS) |
1994       A6XX_TEX_SAMP_1_MIN_LOD(pCreateInfo->minLod) |
1995       A6XX_TEX_SAMP_1_MAX_LOD(pCreateInfo->maxLod) |
1996       COND(pCreateInfo->compareEnable,
1997            A6XX_TEX_SAMP_1_COMPARE_FUNC(tu6_compare_func(pCreateInfo->compareOp)));
1998    sampler->state[2] = 0;
1999    sampler->state[3] = 0;
2000
2001    /* TODO:
2002     * A6XX_TEX_SAMP_1_MIPFILTER_LINEAR_FAR disables mipmapping, but vk has no NONE mipfilter?
2003     * border color
2004     */
2005
2006    sampler->needs_border = needs_border;
2007    sampler->border = pCreateInfo->borderColor;
2008 }
2009
2010 VkResult
2011 tu_CreateSampler(VkDevice _device,
2012                  const VkSamplerCreateInfo *pCreateInfo,
2013                  const VkAllocationCallbacks *pAllocator,
2014                  VkSampler *pSampler)
2015 {
2016    TU_FROM_HANDLE(tu_device, device, _device);
2017    struct tu_sampler *sampler;
2018
2019    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO);
2020
2021    sampler = vk_alloc2(&device->alloc, pAllocator, sizeof(*sampler), 8,
2022                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2023    if (!sampler)
2024       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2025
2026    tu_init_sampler(device, sampler, pCreateInfo);
2027    *pSampler = tu_sampler_to_handle(sampler);
2028
2029    return VK_SUCCESS;
2030 }
2031
2032 void
2033 tu_DestroySampler(VkDevice _device,
2034                   VkSampler _sampler,
2035                   const VkAllocationCallbacks *pAllocator)
2036 {
2037    TU_FROM_HANDLE(tu_device, device, _device);
2038    TU_FROM_HANDLE(tu_sampler, sampler, _sampler);
2039
2040    if (!sampler)
2041       return;
2042    vk_free2(&device->alloc, pAllocator, sampler);
2043 }
2044
2045 /* vk_icd.h does not declare this function, so we declare it here to
2046  * suppress Wmissing-prototypes.
2047  */
2048 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2049 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion);
2050
2051 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2052 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion)
2053 {
2054    /* For the full details on loader interface versioning, see
2055     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2056     * What follows is a condensed summary, to help you navigate the large and
2057     * confusing official doc.
2058     *
2059     *   - Loader interface v0 is incompatible with later versions. We don't
2060     *     support it.
2061     *
2062     *   - In loader interface v1:
2063     *       - The first ICD entrypoint called by the loader is
2064     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2065     *         entrypoint.
2066     *       - The ICD must statically expose no other Vulkan symbol unless it
2067     * is linked with -Bsymbolic.
2068     *       - Each dispatchable Vulkan handle created by the ICD must be
2069     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2070     *         ICD must initialize VK_LOADER_DATA.loadMagic to
2071     * ICD_LOADER_MAGIC.
2072     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2073     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2074     *         such loader-managed surfaces.
2075     *
2076     *    - Loader interface v2 differs from v1 in:
2077     *       - The first ICD entrypoint called by the loader is
2078     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2079     *         statically expose this entrypoint.
2080     *
2081     *    - Loader interface v3 differs from v2 in:
2082     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2083     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2084     *          because the loader no longer does so.
2085     */
2086    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2087    return VK_SUCCESS;
2088 }
2089
2090 VkResult
2091 tu_GetMemoryFdKHR(VkDevice _device,
2092                   const VkMemoryGetFdInfoKHR *pGetFdInfo,
2093                   int *pFd)
2094 {
2095    TU_FROM_HANDLE(tu_device, device, _device);
2096    TU_FROM_HANDLE(tu_device_memory, memory, pGetFdInfo->memory);
2097
2098    assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2099
2100    /* At the moment, we support only the below handle types. */
2101    assert(pGetFdInfo->handleType ==
2102              VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2103           pGetFdInfo->handleType ==
2104              VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2105
2106    int prime_fd = tu_bo_export_dmabuf(device, &memory->bo);
2107    if (prime_fd < 0)
2108       return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
2109
2110    *pFd = prime_fd;
2111    return VK_SUCCESS;
2112 }
2113
2114 VkResult
2115 tu_GetMemoryFdPropertiesKHR(VkDevice _device,
2116                             VkExternalMemoryHandleTypeFlagBits handleType,
2117                             int fd,
2118                             VkMemoryFdPropertiesKHR *pMemoryFdProperties)
2119 {
2120    assert(handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2121    pMemoryFdProperties->memoryTypeBits = 1;
2122    return VK_SUCCESS;
2123 }
2124
2125 void
2126 tu_GetPhysicalDeviceExternalSemaphoreProperties(
2127    VkPhysicalDevice physicalDevice,
2128    const VkPhysicalDeviceExternalSemaphoreInfo *pExternalSemaphoreInfo,
2129    VkExternalSemaphoreProperties *pExternalSemaphoreProperties)
2130 {
2131    pExternalSemaphoreProperties->exportFromImportedHandleTypes = 0;
2132    pExternalSemaphoreProperties->compatibleHandleTypes = 0;
2133    pExternalSemaphoreProperties->externalSemaphoreFeatures = 0;
2134 }
2135
2136 void
2137 tu_GetPhysicalDeviceExternalFenceProperties(
2138    VkPhysicalDevice physicalDevice,
2139    const VkPhysicalDeviceExternalFenceInfo *pExternalFenceInfo,
2140    VkExternalFenceProperties *pExternalFenceProperties)
2141 {
2142    pExternalFenceProperties->exportFromImportedHandleTypes = 0;
2143    pExternalFenceProperties->compatibleHandleTypes = 0;
2144    pExternalFenceProperties->externalFenceFeatures = 0;
2145 }
2146
2147 VkResult
2148 tu_CreateDebugReportCallbackEXT(
2149    VkInstance _instance,
2150    const VkDebugReportCallbackCreateInfoEXT *pCreateInfo,
2151    const VkAllocationCallbacks *pAllocator,
2152    VkDebugReportCallbackEXT *pCallback)
2153 {
2154    TU_FROM_HANDLE(tu_instance, instance, _instance);
2155    return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2156                                           pCreateInfo, pAllocator,
2157                                           &instance->alloc, pCallback);
2158 }
2159
2160 void
2161 tu_DestroyDebugReportCallbackEXT(VkInstance _instance,
2162                                  VkDebugReportCallbackEXT _callback,
2163                                  const VkAllocationCallbacks *pAllocator)
2164 {
2165    TU_FROM_HANDLE(tu_instance, instance, _instance);
2166    vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2167                                     _callback, pAllocator, &instance->alloc);
2168 }
2169
2170 void
2171 tu_DebugReportMessageEXT(VkInstance _instance,
2172                          VkDebugReportFlagsEXT flags,
2173                          VkDebugReportObjectTypeEXT objectType,
2174                          uint64_t object,
2175                          size_t location,
2176                          int32_t messageCode,
2177                          const char *pLayerPrefix,
2178                          const char *pMessage)
2179 {
2180    TU_FROM_HANDLE(tu_instance, instance, _instance);
2181    vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2182                    object, location, messageCode, pLayerPrefix, pMessage);
2183 }
2184
2185 void
2186 tu_GetDeviceGroupPeerMemoryFeatures(
2187    VkDevice device,
2188    uint32_t heapIndex,
2189    uint32_t localDeviceIndex,
2190    uint32_t remoteDeviceIndex,
2191    VkPeerMemoryFeatureFlags *pPeerMemoryFeatures)
2192 {
2193    assert(localDeviceIndex == remoteDeviceIndex);
2194
2195    *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2196                           VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2197                           VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2198                           VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2199 }