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    if (tu_drm_get_gmem_base(device, &device->gmem_base)) {
 256       if (instance->debug_flags & TU_DEBUG_STARTUP)
 257          tu_logi("Could not query the GMEM size");
 258       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 259                          "could not get GMEM size");
 260       goto fail;
 261    }
 262
 263    memset(device->name, 0, sizeof(device->name));
 264    sprintf(device->name, "FD%d", device->gpu_id);
 265
 266    switch (device->gpu_id) {
 267    case 618:
 268       device->tile_align_w = 64;
 269       device->tile_align_h = 16;
 270       device->magic.RB_UNKNOWN_8E04_blit = 0x00100000;
 271       device->magic.RB_CCU_CNTL_gmem     = 0x3e400004;
 272       device->magic.PC_UNKNOWN_9805 = 0x0;
 273       device->magic.SP_UNKNOWN_A0F8 = 0x0;
 274       break;
 275    case 630:
 276    case 640:
 277       device->tile_align_w = 64;
 278       device->tile_align_h = 16;
 279       device->magic.RB_UNKNOWN_8E04_blit = 0x01000000;
 280       device->magic.RB_CCU_CNTL_gmem     = 0x7c400004;
 281       device->magic.PC_UNKNOWN_9805 = 0x1;
 282       device->magic.SP_UNKNOWN_A0F8 = 0x1;
 283       break;
 284    default:
 285       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 286                          "device %s is unsupported", device->name);
 287       goto fail;
 288    }
 289    if (tu_device_get_cache_uuid(device->gpu_id, device->cache_uuid)) {
 290       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 291                          "cannot generate UUID");
 292       goto fail;
 293    }
 294
 295    /* The gpu id is already embedded in the uuid so we just pass "tu"
 296     * when creating the cache.
 297     */
 298    char buf[VK_UUID_SIZE * 2 + 1];
 299    disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
 300    device->disk_cache = disk_cache_create(device->name, buf, 0);
 301
 302    fprintf(stderr, "WARNING: tu is not a conformant vulkan implementation, "
 303                    "testing use only.\n");
 304
 305    tu_get_driver_uuid(&device->device_uuid);
 306    tu_get_device_uuid(&device->device_uuid);
 307
 308    tu_fill_device_extension_table(device, &device->supported_extensions);
 309
 310    if (result != VK_SUCCESS) {
 311       vk_error(instance, result);
 312       goto fail;
 313    }
 314
 315    result = tu_wsi_init(device);
 316    if (result != VK_SUCCESS) {
 317       vk_error(instance, result);
 318       goto fail;
 319    }
 320
 321    return VK_SUCCESS;
 322
 323 fail:
 324    close(fd);
 325    if (master_fd != -1)
 326       close(master_fd);
 327    return result;
 328 }
 329
 330 static void
 331 tu_physical_device_finish(struct tu_physical_device *device)
 332 {
 333    tu_wsi_finish(device);
 334
 335    disk_cache_destroy(device->disk_cache);
 336    close(device->local_fd);
 337    if (device->master_fd != -1)
 338       close(device->master_fd);
 339 }
 340
 341 static void *
 342 default_alloc_func(void *pUserData,
 343                    size_t size,
 344                    size_t align,
 345                    VkSystemAllocationScope allocationScope)
 346 {
 347    return malloc(size);
 348 }
 349
 350 static void *
 351 default_realloc_func(void *pUserData,
 352                      void *pOriginal,
 353                      size_t size,
 354                      size_t align,
 355                      VkSystemAllocationScope allocationScope)
 356 {
 357    return realloc(pOriginal, size);
 358 }
 359
 360 static void
 361 default_free_func(void *pUserData, void *pMemory)
 362 {
 363    free(pMemory);
 364 }
 365
 366 static const VkAllocationCallbacks default_alloc = {
 367    .pUserData = NULL,
 368    .pfnAllocation = default_alloc_func,
 369    .pfnReallocation = default_realloc_func,
 370    .pfnFree = default_free_func,
 371 };
 372
 373 static const struct debug_control tu_debug_options[] = {
 374    { "startup", TU_DEBUG_STARTUP },
 375    { "nir", TU_DEBUG_NIR },
 376    { "ir3", TU_DEBUG_IR3 },
 377    { "nobin", TU_DEBUG_NOBIN },
 378    { "sysmem", TU_DEBUG_SYSMEM },
 379    { "forcebin", TU_DEBUG_FORCEBIN },
 380    { NULL, 0 }
 381 };
 382
 383 const char *
 384 tu_get_debug_option_name(int id)
 385 {
 386    assert(id < ARRAY_SIZE(tu_debug_options) - 1);
 387    return tu_debug_options[id].string;
 388 }
 389
 390 static int
 391 tu_get_instance_extension_index(const char *name)
 392 {
 393    for (unsigned i = 0; i < TU_INSTANCE_EXTENSION_COUNT; ++i) {
 394       if (strcmp(name, tu_instance_extensions[i].extensionName) == 0)
 395          return i;
 396    }
 397    return -1;
 398 }
 399
 400 VkResult
 401 tu_CreateInstance(const VkInstanceCreateInfo *pCreateInfo,
 402                   const VkAllocationCallbacks *pAllocator,
 403                   VkInstance *pInstance)
 404 {
 405    struct tu_instance *instance;
 406    VkResult result;
 407
 408    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 409
 410    uint32_t client_version;
 411    if (pCreateInfo->pApplicationInfo &&
 412        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 413       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 414    } else {
 415       tu_EnumerateInstanceVersion(&client_version);
 416    }
 417
 418    instance = vk_zalloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 419                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 420    if (!instance)
 421       return vk_error(NULL, VK_ERROR_OUT_OF_HOST_MEMORY);
 422
 423    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 424
 425    if (pAllocator)
 426       instance->alloc = *pAllocator;
 427    else
 428       instance->alloc = default_alloc;
 429
 430    instance->api_version = client_version;
 431    instance->physical_device_count = -1;
 432
 433    instance->debug_flags =
 434       parse_debug_string(getenv("TU_DEBUG"), tu_debug_options);
 435
 436    if (instance->debug_flags & TU_DEBUG_STARTUP)
 437       tu_logi("Created an instance");
 438
 439    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 440       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
 441       int index = tu_get_instance_extension_index(ext_name);
 442
 443       if (index < 0 || !tu_supported_instance_extensions.extensions[index]) {
 444          vk_free2(&default_alloc, pAllocator, instance);
 445          return vk_error(instance, VK_ERROR_EXTENSION_NOT_PRESENT);
 446       }
 447
 448       instance->enabled_extensions.extensions[index] = true;
 449    }
 450
 451    result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
 452    if (result != VK_SUCCESS) {
 453       vk_free2(&default_alloc, pAllocator, instance);
 454       return vk_error(instance, result);
 455    }
 456
 457    glsl_type_singleton_init_or_ref();
 458
 459    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 460
 461    *pInstance = tu_instance_to_handle(instance);
 462
 463    return VK_SUCCESS;
 464 }
 465
 466 void
 467 tu_DestroyInstance(VkInstance _instance,
 468                    const VkAllocationCallbacks *pAllocator)
 469 {
 470    TU_FROM_HANDLE(tu_instance, instance, _instance);
 471
 472    if (!instance)
 473       return;
 474
 475    for (int i = 0; i < instance->physical_device_count; ++i) {
 476       tu_physical_device_finish(instance->physical_devices + i);
 477    }
 478
 479    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 480
 481    glsl_type_singleton_decref();
 482
 483    vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
 484
 485    vk_free(&instance->alloc, instance);
 486 }
 487
 488 static VkResult
 489 tu_enumerate_devices(struct tu_instance *instance)
 490 {
 491    /* TODO: Check for more devices ? */
 492    drmDevicePtr devices[8];
 493    VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
 494    int max_devices;
 495
 496    instance->physical_device_count = 0;
 497
 498    max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
 499
 500    if (instance->debug_flags & TU_DEBUG_STARTUP)
 501       tu_logi("Found %d drm nodes", max_devices);
 502
 503    if (max_devices < 1)
 504       return vk_error(instance, VK_ERROR_INCOMPATIBLE_DRIVER);
 505
 506    for (unsigned i = 0; i < (unsigned) max_devices; i++) {
 507       if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
 508           devices[i]->bustype == DRM_BUS_PLATFORM) {
 509
 510          result = tu_physical_device_init(
 511             instance->physical_devices + instance->physical_device_count,
 512             instance, devices[i]);
 513          if (result == VK_SUCCESS)
 514             ++instance->physical_device_count;
 515          else if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 516             break;
 517       }
 518    }
 519    drmFreeDevices(devices, max_devices);
 520
 521    return result;
 522 }
 523
 524 VkResult
 525 tu_EnumeratePhysicalDevices(VkInstance _instance,
 526                             uint32_t *pPhysicalDeviceCount,
 527                             VkPhysicalDevice *pPhysicalDevices)
 528 {
 529    TU_FROM_HANDLE(tu_instance, instance, _instance);
 530    VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
 531
 532    VkResult result;
 533
 534    if (instance->physical_device_count < 0) {
 535       result = tu_enumerate_devices(instance);
 536       if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
 537          return result;
 538    }
 539
 540    for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
 541       vk_outarray_append(&out, p)
 542       {
 543          *p = tu_physical_device_to_handle(instance->physical_devices + i);
 544       }
 545    }
 546
 547    return vk_outarray_status(&out);
 548 }
 549
 550 VkResult
 551 tu_EnumeratePhysicalDeviceGroups(
 552    VkInstance _instance,
 553    uint32_t *pPhysicalDeviceGroupCount,
 554    VkPhysicalDeviceGroupProperties *pPhysicalDeviceGroupProperties)
 555 {
 556    TU_FROM_HANDLE(tu_instance, instance, _instance);
 557    VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
 558                     pPhysicalDeviceGroupCount);
 559    VkResult result;
 560
 561    if (instance->physical_device_count < 0) {
 562       result = tu_enumerate_devices(instance);
 563       if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
 564          return result;
 565    }
 566
 567    for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
 568       vk_outarray_append(&out, p)
 569       {
 570          p->physicalDeviceCount = 1;
 571          p->physicalDevices[0] =
 572             tu_physical_device_to_handle(instance->physical_devices + i);
 573          p->subsetAllocation = false;
 574       }
 575    }
 576
 577    return vk_outarray_status(&out);
 578 }
 579
 580 void
 581 tu_GetPhysicalDeviceFeatures(VkPhysicalDevice physicalDevice,
 582                              VkPhysicalDeviceFeatures *pFeatures)
 583 {
 584    memset(pFeatures, 0, sizeof(*pFeatures));
 585
 586    *pFeatures = (VkPhysicalDeviceFeatures) {
 587       .robustBufferAccess = false,
 588       .fullDrawIndexUint32 = true,
 589       .imageCubeArray = false,
 590       .independentBlend = true,
 591       .geometryShader = false,
 592       .tessellationShader = false,
 593       .sampleRateShading = true,
 594       .dualSrcBlend = true,
 595       .logicOp = true,
 596       .multiDrawIndirect = false,
 597       .drawIndirectFirstInstance = false,
 598       .depthClamp = false,
 599       .depthBiasClamp = false,
 600       .fillModeNonSolid = false,
 601       .depthBounds = false,
 602       .wideLines = false,
 603       .largePoints = false,
 604       .alphaToOne = false,
 605       .multiViewport = false,
 606       .samplerAnisotropy = true,
 607       .textureCompressionETC2 = true,
 608       .textureCompressionASTC_LDR = true,
 609       .textureCompressionBC = true,
 610       .occlusionQueryPrecise = true,
 611       .pipelineStatisticsQuery = false,
 612       .vertexPipelineStoresAndAtomics = false,
 613       .fragmentStoresAndAtomics = false,
 614       .shaderTessellationAndGeometryPointSize = false,
 615       .shaderImageGatherExtended = false,
 616       .shaderStorageImageExtendedFormats = false,
 617       .shaderStorageImageMultisample = false,
 618       .shaderUniformBufferArrayDynamicIndexing = false,
 619       .shaderSampledImageArrayDynamicIndexing = false,
 620       .shaderStorageBufferArrayDynamicIndexing = false,
 621       .shaderStorageImageArrayDynamicIndexing = false,
 622       .shaderStorageImageReadWithoutFormat = false,
 623       .shaderStorageImageWriteWithoutFormat = false,
 624       .shaderClipDistance = false,
 625       .shaderCullDistance = false,
 626       .shaderFloat64 = false,
 627       .shaderInt64 = false,
 628       .shaderInt16 = false,
 629       .sparseBinding = false,
 630       .variableMultisampleRate = false,
 631       .inheritedQueries = false,
 632    };
 633 }
 634
 635 void
 636 tu_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,
 637                               VkPhysicalDeviceFeatures2 *pFeatures)
 638 {
 639    vk_foreach_struct(ext, pFeatures->pNext)
 640    {
 641       switch (ext->sType) {
 642       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
 643          VkPhysicalDeviceVariablePointersFeatures *features = (void *) ext;
 644          features->variablePointersStorageBuffer = false;
 645          features->variablePointers = false;
 646          break;
 647       }
 648       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
 649          VkPhysicalDeviceMultiviewFeatures *features =
 650             (VkPhysicalDeviceMultiviewFeatures *) ext;
 651          features->multiview = false;
 652          features->multiviewGeometryShader = false;
 653          features->multiviewTessellationShader = false;
 654          break;
 655       }
 656       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
 657          VkPhysicalDeviceShaderDrawParametersFeatures *features =
 658             (VkPhysicalDeviceShaderDrawParametersFeatures *) ext;
 659          features->shaderDrawParameters = false;
 660          break;
 661       }
 662       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
 663          VkPhysicalDeviceProtectedMemoryFeatures *features =
 664             (VkPhysicalDeviceProtectedMemoryFeatures *) ext;
 665          features->protectedMemory = false;
 666          break;
 667       }
 668       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
 669          VkPhysicalDevice16BitStorageFeatures *features =
 670             (VkPhysicalDevice16BitStorageFeatures *) ext;
 671          features->storageBuffer16BitAccess = false;
 672          features->uniformAndStorageBuffer16BitAccess = false;
 673          features->storagePushConstant16 = false;
 674          features->storageInputOutput16 = false;
 675          break;
 676       }
 677       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
 678          VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
 679             (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
 680          features->samplerYcbcrConversion = false;
 681          break;
 682       }
 683       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
 684          VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
 685             (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *) ext;
 686          features->shaderInputAttachmentArrayDynamicIndexing = false;
 687          features->shaderUniformTexelBufferArrayDynamicIndexing = false;
 688          features->shaderStorageTexelBufferArrayDynamicIndexing = false;
 689          features->shaderUniformBufferArrayNonUniformIndexing = false;
 690          features->shaderSampledImageArrayNonUniformIndexing = false;
 691          features->shaderStorageBufferArrayNonUniformIndexing = false;
 692          features->shaderStorageImageArrayNonUniformIndexing = false;
 693          features->shaderInputAttachmentArrayNonUniformIndexing = false;
 694          features->shaderUniformTexelBufferArrayNonUniformIndexing = false;
 695          features->shaderStorageTexelBufferArrayNonUniformIndexing = false;
 696          features->descriptorBindingUniformBufferUpdateAfterBind = false;
 697          features->descriptorBindingSampledImageUpdateAfterBind = false;
 698          features->descriptorBindingStorageImageUpdateAfterBind = false;
 699          features->descriptorBindingStorageBufferUpdateAfterBind = false;
 700          features->descriptorBindingUniformTexelBufferUpdateAfterBind = false;
 701          features->descriptorBindingStorageTexelBufferUpdateAfterBind = false;
 702          features->descriptorBindingUpdateUnusedWhilePending = false;
 703          features->descriptorBindingPartiallyBound = false;
 704          features->descriptorBindingVariableDescriptorCount = false;
 705          features->runtimeDescriptorArray = false;
 706          break;
 707       }
 708       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
 709          VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
 710             (VkPhysicalDeviceConditionalRenderingFeaturesEXT *) ext;
 711          features->conditionalRendering = false;
 712          features->inheritedConditionalRendering = false;
 713          break;
 714       }
 715       default:
 716          break;
 717       }
 718    }
 719    return tu_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 720 }
 721
 722 void
 723 tu_GetPhysicalDeviceProperties(VkPhysicalDevice physicalDevice,
 724                                VkPhysicalDeviceProperties *pProperties)
 725 {
 726    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 727    VkSampleCountFlags sample_counts = VK_SAMPLE_COUNT_1_BIT |
 728       VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT | VK_SAMPLE_COUNT_8_BIT;
 729
 730    /* make sure that the entire descriptor set is addressable with a signed
 731     * 32-bit int. So the sum of all limits scaled by descriptor size has to
 732     * be at most 2 GiB. the combined image & samples object count as one of
 733     * both. This limit is for the pipeline layout, not for the set layout, but
 734     * there is no set limit, so we just set a pipeline limit. I don't think
 735     * any app is going to hit this soon. */
 736    size_t max_descriptor_set_size =
 737       ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS) /
 738       (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
 739        32 /* storage buffer, 32 due to potential space wasted on alignment */ +
 740        32 /* sampler, largest when combined with image */ +
 741        64 /* sampled image */ + 64 /* storage image */);
 742
 743    VkPhysicalDeviceLimits limits = {
 744       .maxImageDimension1D = (1 << 14),
 745       .maxImageDimension2D = (1 << 14),
 746       .maxImageDimension3D = (1 << 11),
 747       .maxImageDimensionCube = (1 << 14),
 748       .maxImageArrayLayers = (1 << 11),
 749       .maxTexelBufferElements = 128 * 1024 * 1024,
 750       .maxUniformBufferRange = UINT32_MAX,
 751       .maxStorageBufferRange = MAX_STORAGE_BUFFER_RANGE,
 752       .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
 753       .maxMemoryAllocationCount = UINT32_MAX,
 754       .maxSamplerAllocationCount = 64 * 1024,
 755       .bufferImageGranularity = 64,          /* A cache line */
 756       .sparseAddressSpaceSize = 0xffffffffu, /* buffer max size */
 757       .maxBoundDescriptorSets = MAX_SETS,
 758       .maxPerStageDescriptorSamplers = max_descriptor_set_size,
 759       .maxPerStageDescriptorUniformBuffers = max_descriptor_set_size,
 760       .maxPerStageDescriptorStorageBuffers = max_descriptor_set_size,
 761       .maxPerStageDescriptorSampledImages = max_descriptor_set_size,
 762       .maxPerStageDescriptorStorageImages = max_descriptor_set_size,
 763       .maxPerStageDescriptorInputAttachments = max_descriptor_set_size,
 764       .maxPerStageResources = max_descriptor_set_size,
 765       .maxDescriptorSetSamplers = max_descriptor_set_size,
 766       .maxDescriptorSetUniformBuffers = max_descriptor_set_size,
 767       .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS,
 768       .maxDescriptorSetStorageBuffers = max_descriptor_set_size,
 769       .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS,
 770       .maxDescriptorSetSampledImages = max_descriptor_set_size,
 771       .maxDescriptorSetStorageImages = max_descriptor_set_size,
 772       .maxDescriptorSetInputAttachments = max_descriptor_set_size,
 773       .maxVertexInputAttributes = 32,
 774       .maxVertexInputBindings = 32,
 775       .maxVertexInputAttributeOffset = 2047,
 776       .maxVertexInputBindingStride = 2048,
 777       .maxVertexOutputComponents = 128,
 778       .maxTessellationGenerationLevel = 64,
 779       .maxTessellationPatchSize = 32,
 780       .maxTessellationControlPerVertexInputComponents = 128,
 781       .maxTessellationControlPerVertexOutputComponents = 128,
 782       .maxTessellationControlPerPatchOutputComponents = 120,
 783       .maxTessellationControlTotalOutputComponents = 4096,
 784       .maxTessellationEvaluationInputComponents = 128,
 785       .maxTessellationEvaluationOutputComponents = 128,
 786       .maxGeometryShaderInvocations = 127,
 787       .maxGeometryInputComponents = 64,
 788       .maxGeometryOutputComponents = 128,
 789       .maxGeometryOutputVertices = 256,
 790       .maxGeometryTotalOutputComponents = 1024,
 791       .maxFragmentInputComponents = 128,
 792       .maxFragmentOutputAttachments = 8,
 793       .maxFragmentDualSrcAttachments = 1,
 794       .maxFragmentCombinedOutputResources = 8,
 795       .maxComputeSharedMemorySize = 32768,
 796       .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
 797       .maxComputeWorkGroupInvocations = 2048,
 798       .maxComputeWorkGroupSize = { 2048, 2048, 2048 },
 799       .subPixelPrecisionBits = 4 /* FIXME */,
 800       .subTexelPrecisionBits = 4 /* FIXME */,
 801       .mipmapPrecisionBits = 4 /* FIXME */,
 802       .maxDrawIndexedIndexValue = UINT32_MAX,
 803       .maxDrawIndirectCount = UINT32_MAX,
 804       .maxSamplerLodBias = 16,
 805       .maxSamplerAnisotropy = 16,
 806       .maxViewports = MAX_VIEWPORTS,
 807       .maxViewportDimensions = { (1 << 14), (1 << 14) },
 808       .viewportBoundsRange = { INT16_MIN, INT16_MAX },
 809       .viewportSubPixelBits = 8,
 810       .minMemoryMapAlignment = 4096, /* A page */
 811       .minTexelBufferOffsetAlignment = 64,
 812       .minUniformBufferOffsetAlignment = 4,
 813       .minStorageBufferOffsetAlignment = 4,
 814       .minTexelOffset = -32,
 815       .maxTexelOffset = 31,
 816       .minTexelGatherOffset = -32,
 817       .maxTexelGatherOffset = 31,
 818       .minInterpolationOffset = -2,
 819       .maxInterpolationOffset = 2,
 820       .subPixelInterpolationOffsetBits = 8,
 821       .maxFramebufferWidth = (1 << 14),
 822       .maxFramebufferHeight = (1 << 14),
 823       .maxFramebufferLayers = (1 << 10),
 824       .framebufferColorSampleCounts = sample_counts,
 825       .framebufferDepthSampleCounts = sample_counts,
 826       .framebufferStencilSampleCounts = sample_counts,
 827       .framebufferNoAttachmentsSampleCounts = sample_counts,
 828       .maxColorAttachments = MAX_RTS,
 829       .sampledImageColorSampleCounts = sample_counts,
 830       .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 831       .sampledImageDepthSampleCounts = sample_counts,
 832       .sampledImageStencilSampleCounts = sample_counts,
 833       .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 834       .maxSampleMaskWords = 1,
 835       .timestampComputeAndGraphics = false, /* FINISHME */
 836       .timestampPeriod = 1,
 837       .maxClipDistances = 8,
 838       .maxCullDistances = 8,
 839       .maxCombinedClipAndCullDistances = 8,
 840       .discreteQueuePriorities = 1,
 841       .pointSizeRange = { 0.125, 255.875 },
 842       .lineWidthRange = { 0.0, 7.9921875 },
 843       .pointSizeGranularity = (1.0 / 8.0),
 844       .lineWidthGranularity = (1.0 / 128.0),
 845       .strictLines = false, /* FINISHME */
 846       .standardSampleLocations = true,
 847       .optimalBufferCopyOffsetAlignment = 128,
 848       .optimalBufferCopyRowPitchAlignment = 128,
 849       .nonCoherentAtomSize = 64,
 850    };
 851
 852    *pProperties = (VkPhysicalDeviceProperties) {
 853       .apiVersion = tu_physical_device_api_version(pdevice),
 854       .driverVersion = vk_get_driver_version(),
 855       .vendorID = 0, /* TODO */
 856       .deviceID = 0,
 857       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 858       .limits = limits,
 859       .sparseProperties = { 0 },
 860    };
 861
 862    strcpy(pProperties->deviceName, pdevice->name);
 863    memcpy(pProperties->pipelineCacheUUID, pdevice->cache_uuid, VK_UUID_SIZE);
 864 }
 865
 866 void
 867 tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
 868                                 VkPhysicalDeviceProperties2 *pProperties)
 869 {
 870    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 871    tu_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 872
 873    vk_foreach_struct(ext, pProperties->pNext)
 874    {
 875       switch (ext->sType) {
 876       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 877          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 878             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 879          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 880          break;
 881       }
 882       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
 883          VkPhysicalDeviceIDProperties *properties =
 884             (VkPhysicalDeviceIDProperties *) ext;
 885          memcpy(properties->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
 886          memcpy(properties->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
 887          properties->deviceLUIDValid = false;
 888          break;
 889       }
 890       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
 891          VkPhysicalDeviceMultiviewProperties *properties =
 892             (VkPhysicalDeviceMultiviewProperties *) ext;
 893          properties->maxMultiviewViewCount = MAX_VIEWS;
 894          properties->maxMultiviewInstanceIndex = INT_MAX;
 895          break;
 896       }
 897       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
 898          VkPhysicalDevicePointClippingProperties *properties =
 899             (VkPhysicalDevicePointClippingProperties *) ext;
 900          properties->pointClippingBehavior =
 901             VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
 902          break;
 903       }
 904       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
 905          VkPhysicalDeviceMaintenance3Properties *properties =
 906             (VkPhysicalDeviceMaintenance3Properties *) ext;
 907          /* Make sure everything is addressable by a signed 32-bit int, and
 908           * our largest descriptors are 96 bytes. */
 909          properties->maxPerSetDescriptors = (1ull << 31) / 96;
 910          /* Our buffer size fields allow only this much */
 911          properties->maxMemoryAllocationSize = 0xFFFFFFFFull;
 912          break;
 913       }
 914       default:
 915          break;
 916       }
 917    }
 918 }
 919
 920 static const VkQueueFamilyProperties tu_queue_family_properties = {
 921    .queueFlags =
 922       VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT,
 923    .queueCount = 1,
 924    .timestampValidBits = 0, /* FINISHME */
 925    .minImageTransferGranularity = { 1, 1, 1 },
 926 };
 927
 928 void
 929 tu_GetPhysicalDeviceQueueFamilyProperties(
 930    VkPhysicalDevice physicalDevice,
 931    uint32_t *pQueueFamilyPropertyCount,
 932    VkQueueFamilyProperties *pQueueFamilyProperties)
 933 {
 934    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 935
 936    vk_outarray_append(&out, p) { *p = tu_queue_family_properties; }
 937 }
 938
 939 void
 940 tu_GetPhysicalDeviceQueueFamilyProperties2(
 941    VkPhysicalDevice physicalDevice,
 942    uint32_t *pQueueFamilyPropertyCount,
 943    VkQueueFamilyProperties2 *pQueueFamilyProperties)
 944 {
 945    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 946
 947    vk_outarray_append(&out, p)
 948    {
 949       p->queueFamilyProperties = tu_queue_family_properties;
 950    }
 951 }
 952
 953 static uint64_t
 954 tu_get_system_heap_size()
 955 {
 956    struct sysinfo info;
 957    sysinfo(&info);
 958
 959    uint64_t total_ram = (uint64_t) info.totalram * (uint64_t) info.mem_unit;
 960
 961    /* We don't want to burn too much ram with the GPU.  If the user has 4GiB
 962     * or less, we use at most half.  If they have more than 4GiB, we use 3/4.
 963     */
 964    uint64_t available_ram;
 965    if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
 966       available_ram = total_ram / 2;
 967    else
 968       available_ram = total_ram * 3 / 4;
 969
 970    return available_ram;
 971 }
 972
 973 void
 974 tu_GetPhysicalDeviceMemoryProperties(
 975    VkPhysicalDevice physicalDevice,
 976    VkPhysicalDeviceMemoryProperties *pMemoryProperties)
 977 {
 978    pMemoryProperties->memoryHeapCount = 1;
 979    pMemoryProperties->memoryHeaps[0].size = tu_get_system_heap_size();
 980    pMemoryProperties->memoryHeaps[0].flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT;
 981
 982    pMemoryProperties->memoryTypeCount = 1;
 983    pMemoryProperties->memoryTypes[0].propertyFlags =
 984       VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 985       VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 986       VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
 987    pMemoryProperties->memoryTypes[0].heapIndex = 0;
 988 }
 989
 990 void
 991 tu_GetPhysicalDeviceMemoryProperties2(
 992    VkPhysicalDevice physicalDevice,
 993    VkPhysicalDeviceMemoryProperties2 *pMemoryProperties)
 994 {
 995    return tu_GetPhysicalDeviceMemoryProperties(
 996       physicalDevice, &pMemoryProperties->memoryProperties);
 997 }
 998
 999 static VkResult
1000 tu_queue_init(struct tu_device *device,
1001               struct tu_queue *queue,
1002               uint32_t queue_family_index,
1003               int idx,
1004               VkDeviceQueueCreateFlags flags)
1005 {
1006    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1007    queue->device = device;
1008    queue->queue_family_index = queue_family_index;
1009    queue->queue_idx = idx;
1010    queue->flags = flags;
1011
1012    int ret = tu_drm_submitqueue_new(device, 0, &queue->msm_queue_id);
1013    if (ret)
1014       return VK_ERROR_INITIALIZATION_FAILED;
1015
1016    tu_fence_init(&queue->submit_fence, false);
1017
1018    return VK_SUCCESS;
1019 }
1020
1021 static void
1022 tu_queue_finish(struct tu_queue *queue)
1023 {
1024    tu_fence_finish(&queue->submit_fence);
1025    tu_drm_submitqueue_close(queue->device, queue->msm_queue_id);
1026 }
1027
1028 static int
1029 tu_get_device_extension_index(const char *name)
1030 {
1031    for (unsigned i = 0; i < TU_DEVICE_EXTENSION_COUNT; ++i) {
1032       if (strcmp(name, tu_device_extensions[i].extensionName) == 0)
1033          return i;
1034    }
1035    return -1;
1036 }
1037
1038 VkResult
1039 tu_CreateDevice(VkPhysicalDevice physicalDevice,
1040                 const VkDeviceCreateInfo *pCreateInfo,
1041                 const VkAllocationCallbacks *pAllocator,
1042                 VkDevice *pDevice)
1043 {
1044    TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
1045    VkResult result;
1046    struct tu_device *device;
1047
1048    /* Check enabled features */
1049    if (pCreateInfo->pEnabledFeatures) {
1050       VkPhysicalDeviceFeatures supported_features;
1051       tu_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1052       VkBool32 *supported_feature = (VkBool32 *) &supported_features;
1053       VkBool32 *enabled_feature = (VkBool32 *) pCreateInfo->pEnabledFeatures;
1054       unsigned num_features =
1055          sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1056       for (uint32_t i = 0; i < num_features; i++) {
1057          if (enabled_feature[i] && !supported_feature[i])
1058             return vk_error(physical_device->instance,
1059                             VK_ERROR_FEATURE_NOT_PRESENT);
1060       }
1061    }
1062
1063    device = vk_zalloc2(&physical_device->instance->alloc, pAllocator,
1064                        sizeof(*device), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1065    if (!device)
1066       return vk_error(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1067
1068    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1069    device->instance = physical_device->instance;
1070    device->physical_device = physical_device;
1071
1072    if (pAllocator)
1073       device->alloc = *pAllocator;
1074    else
1075       device->alloc = physical_device->instance->alloc;
1076
1077    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1078       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
1079       int index = tu_get_device_extension_index(ext_name);
1080       if (index < 0 ||
1081           !physical_device->supported_extensions.extensions[index]) {
1082          vk_free(&device->alloc, device);
1083          return vk_error(physical_device->instance,
1084                          VK_ERROR_EXTENSION_NOT_PRESENT);
1085       }
1086
1087       device->enabled_extensions.extensions[index] = true;
1088    }
1089
1090    for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1091       const VkDeviceQueueCreateInfo *queue_create =
1092          &pCreateInfo->pQueueCreateInfos[i];
1093       uint32_t qfi = queue_create->queueFamilyIndex;
1094       device->queues[qfi] = vk_alloc(
1095          &device->alloc, queue_create->queueCount * sizeof(struct tu_queue),
1096          8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1097       if (!device->queues[qfi]) {
1098          result = VK_ERROR_OUT_OF_HOST_MEMORY;
1099          goto fail_queues;
1100       }
1101
1102       memset(device->queues[qfi], 0,
1103              queue_create->queueCount * sizeof(struct tu_queue));
1104
1105       device->queue_count[qfi] = queue_create->queueCount;
1106
1107       for (unsigned q = 0; q < queue_create->queueCount; q++) {
1108          result = tu_queue_init(device, &device->queues[qfi][q], qfi, q,
1109                                 queue_create->flags);
1110          if (result != VK_SUCCESS)
1111             goto fail_queues;
1112       }
1113    }
1114
1115    device->compiler = ir3_compiler_create(NULL, physical_device->gpu_id);
1116    if (!device->compiler)
1117       goto fail_queues;
1118
1119 #define VSC_DATA_SIZE(pitch)  ((pitch) * 32 + 0x100)  /* extra size to store VSC_SIZE */
1120 #define VSC_DATA2_SIZE(pitch) ((pitch) * 32)
1121
1122    device->vsc_data_pitch = 0x440 * 4;
1123    device->vsc_data2_pitch = 0x1040 * 4;
1124
1125    result = tu_bo_init_new(device, &device->vsc_data, VSC_DATA_SIZE(device->vsc_data_pitch));
1126    if (result != VK_SUCCESS)
1127       goto fail_vsc_data;
1128
1129    result = tu_bo_init_new(device, &device->vsc_data2, VSC_DATA2_SIZE(device->vsc_data2_pitch));
1130    if (result != VK_SUCCESS)
1131       goto fail_vsc_data2;
1132
1133    VkPipelineCacheCreateInfo ci;
1134    ci.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
1135    ci.pNext = NULL;
1136    ci.flags = 0;
1137    ci.pInitialData = NULL;
1138    ci.initialDataSize = 0;
1139    VkPipelineCache pc;
1140    result =
1141       tu_CreatePipelineCache(tu_device_to_handle(device), &ci, NULL, &pc);
1142    if (result != VK_SUCCESS)
1143       goto fail_pipeline_cache;
1144
1145    device->mem_cache = tu_pipeline_cache_from_handle(pc);
1146
1147    *pDevice = tu_device_to_handle(device);
1148    return VK_SUCCESS;
1149
1150 fail_pipeline_cache:
1151    tu_bo_finish(device, &device->vsc_data2);
1152
1153 fail_vsc_data2:
1154    tu_bo_finish(device, &device->vsc_data);
1155
1156 fail_vsc_data:
1157    ralloc_free(device->compiler);
1158
1159 fail_queues:
1160    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1161       for (unsigned q = 0; q < device->queue_count[i]; q++)
1162          tu_queue_finish(&device->queues[i][q]);
1163       if (device->queue_count[i])
1164          vk_free(&device->alloc, device->queues[i]);
1165    }
1166
1167    vk_free(&device->alloc, device);
1168    return result;
1169 }
1170
1171 void
1172 tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
1173 {
1174    TU_FROM_HANDLE(tu_device, device, _device);
1175
1176    if (!device)
1177       return;
1178
1179    tu_bo_finish(device, &device->vsc_data);
1180    tu_bo_finish(device, &device->vsc_data2);
1181
1182    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1183       for (unsigned q = 0; q < device->queue_count[i]; q++)
1184          tu_queue_finish(&device->queues[i][q]);
1185       if (device->queue_count[i])
1186          vk_free(&device->alloc, device->queues[i]);
1187    }
1188
1189    /* the compiler does not use pAllocator */
1190    ralloc_free(device->compiler);
1191
1192    VkPipelineCache pc = tu_pipeline_cache_to_handle(device->mem_cache);
1193    tu_DestroyPipelineCache(tu_device_to_handle(device), pc, NULL);
1194
1195    vk_free(&device->alloc, device);
1196 }
1197
1198 VkResult
1199 tu_EnumerateInstanceLayerProperties(uint32_t *pPropertyCount,
1200                                     VkLayerProperties *pProperties)
1201 {
1202    *pPropertyCount = 0;
1203    return VK_SUCCESS;
1204 }
1205
1206 VkResult
1207 tu_EnumerateDeviceLayerProperties(VkPhysicalDevice physicalDevice,
1208                                   uint32_t *pPropertyCount,
1209                                   VkLayerProperties *pProperties)
1210 {
1211    *pPropertyCount = 0;
1212    return VK_SUCCESS;
1213 }
1214
1215 void
1216 tu_GetDeviceQueue2(VkDevice _device,
1217                    const VkDeviceQueueInfo2 *pQueueInfo,
1218                    VkQueue *pQueue)
1219 {
1220    TU_FROM_HANDLE(tu_device, device, _device);
1221    struct tu_queue *queue;
1222
1223    queue =
1224       &device->queues[pQueueInfo->queueFamilyIndex][pQueueInfo->queueIndex];
1225    if (pQueueInfo->flags != queue->flags) {
1226       /* From the Vulkan 1.1.70 spec:
1227        *
1228        * "The queue returned by vkGetDeviceQueue2 must have the same
1229        * flags value from this structure as that used at device
1230        * creation time in a VkDeviceQueueCreateInfo instance. If no
1231        * matching flags were specified at device creation time then
1232        * pQueue will return VK_NULL_HANDLE."
1233        */
1234       *pQueue = VK_NULL_HANDLE;
1235       return;
1236    }
1237
1238    *pQueue = tu_queue_to_handle(queue);
1239 }
1240
1241 void
1242 tu_GetDeviceQueue(VkDevice _device,
1243                   uint32_t queueFamilyIndex,
1244                   uint32_t queueIndex,
1245                   VkQueue *pQueue)
1246 {
1247    const VkDeviceQueueInfo2 info =
1248       (VkDeviceQueueInfo2) { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
1249                              .queueFamilyIndex = queueFamilyIndex,
1250                              .queueIndex = queueIndex };
1251
1252    tu_GetDeviceQueue2(_device, &info, pQueue);
1253 }
1254
1255 VkResult
1256 tu_QueueSubmit(VkQueue _queue,
1257                uint32_t submitCount,
1258                const VkSubmitInfo *pSubmits,
1259                VkFence _fence)
1260 {
1261    TU_FROM_HANDLE(tu_queue, queue, _queue);
1262
1263    for (uint32_t i = 0; i < submitCount; ++i) {
1264       const VkSubmitInfo *submit = pSubmits + i;
1265       const bool last_submit = (i == submitCount - 1);
1266       struct tu_bo_list bo_list;
1267       tu_bo_list_init(&bo_list);
1268
1269       uint32_t entry_count = 0;
1270       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1271          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1272          entry_count += cmdbuf->cs.entry_count;
1273       }
1274
1275       struct drm_msm_gem_submit_cmd cmds[entry_count];
1276       uint32_t entry_idx = 0;
1277       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1278          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1279          struct tu_cs *cs = &cmdbuf->cs;
1280          for (unsigned i = 0; i < cs->entry_count; ++i, ++entry_idx) {
1281             cmds[entry_idx].type = MSM_SUBMIT_CMD_BUF;
1282             cmds[entry_idx].submit_idx =
1283                tu_bo_list_add(&bo_list, cs->entries[i].bo,
1284                               MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
1285             cmds[entry_idx].submit_offset = cs->entries[i].offset;
1286             cmds[entry_idx].size = cs->entries[i].size;
1287             cmds[entry_idx].pad = 0;
1288             cmds[entry_idx].nr_relocs = 0;
1289             cmds[entry_idx].relocs = 0;
1290          }
1291
1292          tu_bo_list_merge(&bo_list, &cmdbuf->bo_list);
1293       }
1294
1295       uint32_t flags = MSM_PIPE_3D0;
1296       if (last_submit) {
1297          flags |= MSM_SUBMIT_FENCE_FD_OUT;
1298       }
1299
1300       struct drm_msm_gem_submit req = {
1301          .flags = flags,
1302          .queueid = queue->msm_queue_id,
1303          .bos = (uint64_t)(uintptr_t) bo_list.bo_infos,
1304          .nr_bos = bo_list.count,
1305          .cmds = (uint64_t)(uintptr_t)cmds,
1306          .nr_cmds = entry_count,
1307       };
1308
1309       int ret = drmCommandWriteRead(queue->device->physical_device->local_fd,
1310                                     DRM_MSM_GEM_SUBMIT,
1311                                     &req, sizeof(req));
1312       if (ret) {
1313          fprintf(stderr, "submit failed: %s\n", strerror(errno));
1314          abort();
1315       }
1316
1317       tu_bo_list_destroy(&bo_list);
1318
1319       if (last_submit) {
1320          /* no need to merge fences as queue execution is serialized */
1321          tu_fence_update_fd(&queue->submit_fence, req.fence_fd);
1322       }
1323    }
1324
1325    if (_fence != VK_NULL_HANDLE) {
1326       TU_FROM_HANDLE(tu_fence, fence, _fence);
1327       tu_fence_copy(fence, &queue->submit_fence);
1328    }
1329
1330    return VK_SUCCESS;
1331 }
1332
1333 VkResult
1334 tu_QueueWaitIdle(VkQueue _queue)
1335 {
1336    TU_FROM_HANDLE(tu_queue, queue, _queue);
1337
1338    tu_fence_wait_idle(&queue->submit_fence);
1339
1340    return VK_SUCCESS;
1341 }
1342
1343 VkResult
1344 tu_DeviceWaitIdle(VkDevice _device)
1345 {
1346    TU_FROM_HANDLE(tu_device, device, _device);
1347
1348    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1349       for (unsigned q = 0; q < device->queue_count[i]; q++) {
1350          tu_QueueWaitIdle(tu_queue_to_handle(&device->queues[i][q]));
1351       }
1352    }
1353    return VK_SUCCESS;
1354 }
1355
1356 VkResult
1357 tu_EnumerateInstanceExtensionProperties(const char *pLayerName,
1358                                         uint32_t *pPropertyCount,
1359                                         VkExtensionProperties *pProperties)
1360 {
1361    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1362
1363    /* We spport no lyaers */
1364    if (pLayerName)
1365       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1366
1367    for (int i = 0; i < TU_INSTANCE_EXTENSION_COUNT; i++) {
1368       if (tu_supported_instance_extensions.extensions[i]) {
1369          vk_outarray_append(&out, prop) { *prop = tu_instance_extensions[i]; }
1370       }
1371    }
1372
1373    return vk_outarray_status(&out);
1374 }
1375
1376 VkResult
1377 tu_EnumerateDeviceExtensionProperties(VkPhysicalDevice physicalDevice,
1378                                       const char *pLayerName,
1379                                       uint32_t *pPropertyCount,
1380                                       VkExtensionProperties *pProperties)
1381 {
1382    /* We spport no lyaers */
1383    TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
1384    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1385
1386    /* We spport no lyaers */
1387    if (pLayerName)
1388       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1389
1390    for (int i = 0; i < TU_DEVICE_EXTENSION_COUNT; i++) {
1391       if (device->supported_extensions.extensions[i]) {
1392          vk_outarray_append(&out, prop) { *prop = tu_device_extensions[i]; }
1393       }
1394    }
1395
1396    return vk_outarray_status(&out);
1397 }
1398
1399 PFN_vkVoidFunction
1400 tu_GetInstanceProcAddr(VkInstance _instance, const char *pName)
1401 {
1402    TU_FROM_HANDLE(tu_instance, instance, _instance);
1403
1404    return tu_lookup_entrypoint_checked(
1405       pName, instance ? instance->api_version : 0,
1406       instance ? &instance->enabled_extensions : NULL, NULL);
1407 }
1408
1409 /* The loader wants us to expose a second GetInstanceProcAddr function
1410  * to work around certain LD_PRELOAD issues seen in apps.
1411  */
1412 PUBLIC
1413 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1414 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName);
1415
1416 PUBLIC
1417 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1418 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName)
1419 {
1420    return tu_GetInstanceProcAddr(instance, pName);
1421 }
1422
1423 PFN_vkVoidFunction
1424 tu_GetDeviceProcAddr(VkDevice _device, const char *pName)
1425 {
1426    TU_FROM_HANDLE(tu_device, device, _device);
1427
1428    return tu_lookup_entrypoint_checked(pName, device->instance->api_version,
1429                                        &device->instance->enabled_extensions,
1430                                        &device->enabled_extensions);
1431 }
1432
1433 static VkResult
1434 tu_alloc_memory(struct tu_device *device,
1435                 const VkMemoryAllocateInfo *pAllocateInfo,
1436                 const VkAllocationCallbacks *pAllocator,
1437                 VkDeviceMemory *pMem)
1438 {
1439    struct tu_device_memory *mem;
1440    VkResult result;
1441
1442    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1443
1444    if (pAllocateInfo->allocationSize == 0) {
1445       /* Apparently, this is allowed */
1446       *pMem = VK_NULL_HANDLE;
1447       return VK_SUCCESS;
1448    }
1449
1450    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1451                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1452    if (mem == NULL)
1453       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1454
1455    const VkImportMemoryFdInfoKHR *fd_info =
1456       vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
1457    if (fd_info && !fd_info->handleType)
1458       fd_info = NULL;
1459
1460    if (fd_info) {
1461       assert(fd_info->handleType ==
1462                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
1463              fd_info->handleType ==
1464                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1465
1466       /*
1467        * TODO Importing the same fd twice gives us the same handle without
1468        * reference counting.  We need to maintain a per-instance handle-to-bo
1469        * table and add reference count to tu_bo.
1470        */
1471       result = tu_bo_init_dmabuf(device, &mem->bo,
1472                                  pAllocateInfo->allocationSize, fd_info->fd);
1473       if (result == VK_SUCCESS) {
1474          /* take ownership and close the fd */
1475          close(fd_info->fd);
1476       }
1477    } else {
1478       result =
1479          tu_bo_init_new(device, &mem->bo, pAllocateInfo->allocationSize);
1480    }
1481
1482    if (result != VK_SUCCESS) {
1483       vk_free2(&device->alloc, pAllocator, mem);
1484       return result;
1485    }
1486
1487    mem->size = pAllocateInfo->allocationSize;
1488    mem->type_index = pAllocateInfo->memoryTypeIndex;
1489
1490    mem->map = NULL;
1491    mem->user_ptr = NULL;
1492
1493    *pMem = tu_device_memory_to_handle(mem);
1494
1495    return VK_SUCCESS;
1496 }
1497
1498 VkResult
1499 tu_AllocateMemory(VkDevice _device,
1500                   const VkMemoryAllocateInfo *pAllocateInfo,
1501                   const VkAllocationCallbacks *pAllocator,
1502                   VkDeviceMemory *pMem)
1503 {
1504    TU_FROM_HANDLE(tu_device, device, _device);
1505    return tu_alloc_memory(device, pAllocateInfo, pAllocator, pMem);
1506 }
1507
1508 void
1509 tu_FreeMemory(VkDevice _device,
1510               VkDeviceMemory _mem,
1511               const VkAllocationCallbacks *pAllocator)
1512 {
1513    TU_FROM_HANDLE(tu_device, device, _device);
1514    TU_FROM_HANDLE(tu_device_memory, mem, _mem);
1515
1516    if (mem == NULL)
1517       return;
1518
1519    tu_bo_finish(device, &mem->bo);
1520    vk_free2(&device->alloc, pAllocator, mem);
1521 }
1522
1523 VkResult
1524 tu_MapMemory(VkDevice _device,
1525              VkDeviceMemory _memory,
1526              VkDeviceSize offset,
1527              VkDeviceSize size,
1528              VkMemoryMapFlags flags,
1529              void **ppData)
1530 {
1531    TU_FROM_HANDLE(tu_device, device, _device);
1532    TU_FROM_HANDLE(tu_device_memory, mem, _memory);
1533    VkResult result;
1534
1535    if (mem == NULL) {
1536       *ppData = NULL;
1537       return VK_SUCCESS;
1538    }
1539
1540    if (mem->user_ptr) {
1541       *ppData = mem->user_ptr;
1542    } else if (!mem->map) {
1543       result = tu_bo_map(device, &mem->bo);
1544       if (result != VK_SUCCESS)
1545          return result;
1546       *ppData = mem->map = mem->bo.map;
1547    } else
1548       *ppData = mem->map;
1549
1550    if (*ppData) {
1551       *ppData += offset;
1552       return VK_SUCCESS;
1553    }
1554
1555    return vk_error(device->instance, VK_ERROR_MEMORY_MAP_FAILED);
1556 }
1557
1558 void
1559 tu_UnmapMemory(VkDevice _device, VkDeviceMemory _memory)
1560 {
1561    /* I do not see any unmapping done by the freedreno Gallium driver. */
1562 }
1563
1564 VkResult
1565 tu_FlushMappedMemoryRanges(VkDevice _device,
1566                            uint32_t memoryRangeCount,
1567                            const VkMappedMemoryRange *pMemoryRanges)
1568 {
1569    return VK_SUCCESS;
1570 }
1571
1572 VkResult
1573 tu_InvalidateMappedMemoryRanges(VkDevice _device,
1574                                 uint32_t memoryRangeCount,
1575                                 const VkMappedMemoryRange *pMemoryRanges)
1576 {
1577    return VK_SUCCESS;
1578 }
1579
1580 void
1581 tu_GetBufferMemoryRequirements(VkDevice _device,
1582                                VkBuffer _buffer,
1583                                VkMemoryRequirements *pMemoryRequirements)
1584 {
1585    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
1586
1587    pMemoryRequirements->memoryTypeBits = 1;
1588    pMemoryRequirements->alignment = 16;
1589    pMemoryRequirements->size =
1590       align64(buffer->size, pMemoryRequirements->alignment);
1591 }
1592
1593 void
1594 tu_GetBufferMemoryRequirements2(
1595    VkDevice device,
1596    const VkBufferMemoryRequirementsInfo2 *pInfo,
1597    VkMemoryRequirements2 *pMemoryRequirements)
1598 {
1599    tu_GetBufferMemoryRequirements(device, pInfo->buffer,
1600                                   &pMemoryRequirements->memoryRequirements);
1601 }
1602
1603 void
1604 tu_GetImageMemoryRequirements(VkDevice _device,
1605                               VkImage _image,
1606                               VkMemoryRequirements *pMemoryRequirements)
1607 {
1608    TU_FROM_HANDLE(tu_image, image, _image);
1609
1610    pMemoryRequirements->memoryTypeBits = 1;
1611    pMemoryRequirements->size = image->layout.size;
1612    pMemoryRequirements->alignment = image->alignment;
1613 }
1614
1615 void
1616 tu_GetImageMemoryRequirements2(VkDevice device,
1617                                const VkImageMemoryRequirementsInfo2 *pInfo,
1618                                VkMemoryRequirements2 *pMemoryRequirements)
1619 {
1620    tu_GetImageMemoryRequirements(device, pInfo->image,
1621                                  &pMemoryRequirements->memoryRequirements);
1622 }
1623
1624 void
1625 tu_GetImageSparseMemoryRequirements(
1626    VkDevice device,
1627    VkImage image,
1628    uint32_t *pSparseMemoryRequirementCount,
1629    VkSparseImageMemoryRequirements *pSparseMemoryRequirements)
1630 {
1631    tu_stub();
1632 }
1633
1634 void
1635 tu_GetImageSparseMemoryRequirements2(
1636    VkDevice device,
1637    const VkImageSparseMemoryRequirementsInfo2 *pInfo,
1638    uint32_t *pSparseMemoryRequirementCount,
1639    VkSparseImageMemoryRequirements2 *pSparseMemoryRequirements)
1640 {
1641    tu_stub();
1642 }
1643
1644 void
1645 tu_GetDeviceMemoryCommitment(VkDevice device,
1646                              VkDeviceMemory memory,
1647                              VkDeviceSize *pCommittedMemoryInBytes)
1648 {
1649    *pCommittedMemoryInBytes = 0;
1650 }
1651
1652 VkResult
1653 tu_BindBufferMemory2(VkDevice device,
1654                      uint32_t bindInfoCount,
1655                      const VkBindBufferMemoryInfo *pBindInfos)
1656 {
1657    for (uint32_t i = 0; i < bindInfoCount; ++i) {
1658       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
1659       TU_FROM_HANDLE(tu_buffer, buffer, pBindInfos[i].buffer);
1660
1661       if (mem) {
1662          buffer->bo = &mem->bo;
1663          buffer->bo_offset = pBindInfos[i].memoryOffset;
1664       } else {
1665          buffer->bo = NULL;
1666       }
1667    }
1668    return VK_SUCCESS;
1669 }
1670
1671 VkResult
1672 tu_BindBufferMemory(VkDevice device,
1673                     VkBuffer buffer,
1674                     VkDeviceMemory memory,
1675                     VkDeviceSize memoryOffset)
1676 {
1677    const VkBindBufferMemoryInfo info = {
1678       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
1679       .buffer = buffer,
1680       .memory = memory,
1681       .memoryOffset = memoryOffset
1682    };
1683
1684    return tu_BindBufferMemory2(device, 1, &info);
1685 }
1686
1687 VkResult
1688 tu_BindImageMemory2(VkDevice device,
1689                     uint32_t bindInfoCount,
1690                     const VkBindImageMemoryInfo *pBindInfos)
1691 {
1692    for (uint32_t i = 0; i < bindInfoCount; ++i) {
1693       TU_FROM_HANDLE(tu_image, image, pBindInfos[i].image);
1694       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
1695
1696       if (mem) {
1697          image->bo = &mem->bo;
1698          image->bo_offset = pBindInfos[i].memoryOffset;
1699       } else {
1700          image->bo = NULL;
1701          image->bo_offset = 0;
1702       }
1703    }
1704
1705    return VK_SUCCESS;
1706 }
1707
1708 VkResult
1709 tu_BindImageMemory(VkDevice device,
1710                    VkImage image,
1711                    VkDeviceMemory memory,
1712                    VkDeviceSize memoryOffset)
1713 {
1714    const VkBindImageMemoryInfo info = {
1715       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
1716       .image = image,
1717       .memory = memory,
1718       .memoryOffset = memoryOffset
1719    };
1720
1721    return tu_BindImageMemory2(device, 1, &info);
1722 }
1723
1724 VkResult
1725 tu_QueueBindSparse(VkQueue _queue,
1726                    uint32_t bindInfoCount,
1727                    const VkBindSparseInfo *pBindInfo,
1728                    VkFence _fence)
1729 {
1730    return VK_SUCCESS;
1731 }
1732
1733 // Queue semaphore functions
1734
1735 VkResult
1736 tu_CreateSemaphore(VkDevice _device,
1737                    const VkSemaphoreCreateInfo *pCreateInfo,
1738                    const VkAllocationCallbacks *pAllocator,
1739                    VkSemaphore *pSemaphore)
1740 {
1741    TU_FROM_HANDLE(tu_device, device, _device);
1742
1743    struct tu_semaphore *sem =
1744       vk_alloc2(&device->alloc, pAllocator, sizeof(*sem), 8,
1745                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1746    if (!sem)
1747       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1748
1749    *pSemaphore = tu_semaphore_to_handle(sem);
1750    return VK_SUCCESS;
1751 }
1752
1753 void
1754 tu_DestroySemaphore(VkDevice _device,
1755                     VkSemaphore _semaphore,
1756                     const VkAllocationCallbacks *pAllocator)
1757 {
1758    TU_FROM_HANDLE(tu_device, device, _device);
1759    TU_FROM_HANDLE(tu_semaphore, sem, _semaphore);
1760    if (!_semaphore)
1761       return;
1762
1763    vk_free2(&device->alloc, pAllocator, sem);
1764 }
1765
1766 VkResult
1767 tu_CreateEvent(VkDevice _device,
1768                const VkEventCreateInfo *pCreateInfo,
1769                const VkAllocationCallbacks *pAllocator,
1770                VkEvent *pEvent)
1771 {
1772    TU_FROM_HANDLE(tu_device, device, _device);
1773    struct tu_event *event =
1774       vk_alloc2(&device->alloc, pAllocator, sizeof(*event), 8,
1775                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1776
1777    if (!event)
1778       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1779
1780    VkResult result = tu_bo_init_new(device, &event->bo, 0x1000);
1781    if (result != VK_SUCCESS)
1782       goto fail_alloc;
1783
1784    result = tu_bo_map(device, &event->bo);
1785    if (result != VK_SUCCESS)
1786       goto fail_map;
1787
1788    *pEvent = tu_event_to_handle(event);
1789
1790    return VK_SUCCESS;
1791
1792 fail_map:
1793    tu_bo_finish(device, &event->bo);
1794 fail_alloc:
1795    vk_free2(&device->alloc, pAllocator, event);
1796    return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1797 }
1798
1799 void
1800 tu_DestroyEvent(VkDevice _device,
1801                 VkEvent _event,
1802                 const VkAllocationCallbacks *pAllocator)
1803 {
1804    TU_FROM_HANDLE(tu_device, device, _device);
1805    TU_FROM_HANDLE(tu_event, event, _event);
1806
1807    if (!event)
1808       return;
1809
1810    tu_bo_finish(device, &event->bo);
1811    vk_free2(&device->alloc, pAllocator, event);
1812 }
1813
1814 VkResult
1815 tu_GetEventStatus(VkDevice _device, VkEvent _event)
1816 {
1817    TU_FROM_HANDLE(tu_event, event, _event);
1818
1819    if (*(uint64_t*) event->bo.map == 1)
1820       return VK_EVENT_SET;
1821    return VK_EVENT_RESET;
1822 }
1823
1824 VkResult
1825 tu_SetEvent(VkDevice _device, VkEvent _event)
1826 {
1827    TU_FROM_HANDLE(tu_event, event, _event);
1828    *(uint64_t*) event->bo.map = 1;
1829
1830    return VK_SUCCESS;
1831 }
1832
1833 VkResult
1834 tu_ResetEvent(VkDevice _device, VkEvent _event)
1835 {
1836    TU_FROM_HANDLE(tu_event, event, _event);
1837    *(uint64_t*) event->bo.map = 0;
1838
1839    return VK_SUCCESS;
1840 }
1841
1842 VkResult
1843 tu_CreateBuffer(VkDevice _device,
1844                 const VkBufferCreateInfo *pCreateInfo,
1845                 const VkAllocationCallbacks *pAllocator,
1846                 VkBuffer *pBuffer)
1847 {
1848    TU_FROM_HANDLE(tu_device, device, _device);
1849    struct tu_buffer *buffer;
1850
1851    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1852
1853    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1854                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1855    if (buffer == NULL)
1856       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1857
1858    buffer->size = pCreateInfo->size;
1859    buffer->usage = pCreateInfo->usage;
1860    buffer->flags = pCreateInfo->flags;
1861
1862    *pBuffer = tu_buffer_to_handle(buffer);
1863
1864    return VK_SUCCESS;
1865 }
1866
1867 void
1868 tu_DestroyBuffer(VkDevice _device,
1869                  VkBuffer _buffer,
1870                  const VkAllocationCallbacks *pAllocator)
1871 {
1872    TU_FROM_HANDLE(tu_device, device, _device);
1873    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
1874
1875    if (!buffer)
1876       return;
1877
1878    vk_free2(&device->alloc, pAllocator, buffer);
1879 }
1880
1881 static uint32_t
1882 tu_surface_max_layer_count(struct tu_image_view *iview)
1883 {
1884    return iview->type == VK_IMAGE_VIEW_TYPE_3D
1885              ? iview->extent.depth
1886              : (iview->base_layer + iview->layer_count);
1887 }
1888
1889 VkResult
1890 tu_CreateFramebuffer(VkDevice _device,
1891                      const VkFramebufferCreateInfo *pCreateInfo,
1892                      const VkAllocationCallbacks *pAllocator,
1893                      VkFramebuffer *pFramebuffer)
1894 {
1895    TU_FROM_HANDLE(tu_device, device, _device);
1896    struct tu_framebuffer *framebuffer;
1897
1898    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
1899
1900    size_t size = sizeof(*framebuffer) + sizeof(struct tu_attachment_info) *
1901                                            pCreateInfo->attachmentCount;
1902    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
1903                            VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1904    if (framebuffer == NULL)
1905       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1906
1907    framebuffer->attachment_count = pCreateInfo->attachmentCount;
1908    framebuffer->width = pCreateInfo->width;
1909    framebuffer->height = pCreateInfo->height;
1910    framebuffer->layers = pCreateInfo->layers;
1911    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
1912       VkImageView _iview = pCreateInfo->pAttachments[i];
1913       struct tu_image_view *iview = tu_image_view_from_handle(_iview);
1914       framebuffer->attachments[i].attachment = iview;
1915
1916       framebuffer->width = MIN2(framebuffer->width, iview->extent.width);
1917       framebuffer->height = MIN2(framebuffer->height, iview->extent.height);
1918       framebuffer->layers =
1919          MIN2(framebuffer->layers, tu_surface_max_layer_count(iview));
1920    }
1921
1922    *pFramebuffer = tu_framebuffer_to_handle(framebuffer);
1923    return VK_SUCCESS;
1924 }
1925
1926 void
1927 tu_DestroyFramebuffer(VkDevice _device,
1928                       VkFramebuffer _fb,
1929                       const VkAllocationCallbacks *pAllocator)
1930 {
1931    TU_FROM_HANDLE(tu_device, device, _device);
1932    TU_FROM_HANDLE(tu_framebuffer, fb, _fb);
1933
1934    if (!fb)
1935       return;
1936    vk_free2(&device->alloc, pAllocator, fb);
1937 }
1938
1939 static enum a6xx_tex_clamp
1940 tu6_tex_wrap(VkSamplerAddressMode address_mode, bool *needs_border)
1941 {
1942    switch (address_mode) {
1943    case VK_SAMPLER_ADDRESS_MODE_REPEAT:
1944       return A6XX_TEX_REPEAT;
1945    case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT:
1946       return A6XX_TEX_MIRROR_REPEAT;
1947    case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE:
1948       return A6XX_TEX_CLAMP_TO_EDGE;
1949    case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER:
1950       *needs_border = true;
1951       return A6XX_TEX_CLAMP_TO_BORDER;
1952    case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE:
1953       /* only works for PoT.. need to emulate otherwise! */
1954       return A6XX_TEX_MIRROR_CLAMP;
1955    default:
1956       unreachable("illegal tex wrap mode");
1957       break;
1958    }
1959 }
1960
1961 static enum a6xx_tex_filter
1962 tu6_tex_filter(VkFilter filter, unsigned aniso)
1963 {
1964    switch (filter) {
1965    case VK_FILTER_NEAREST:
1966       return A6XX_TEX_NEAREST;
1967    case VK_FILTER_LINEAR:
1968       return aniso ? A6XX_TEX_ANISO : A6XX_TEX_LINEAR;
1969    case VK_FILTER_CUBIC_IMG:
1970    default:
1971       unreachable("illegal texture filter");
1972       break;
1973    }
1974 }
1975
1976 static inline enum adreno_compare_func
1977 tu6_compare_func(VkCompareOp op)
1978 {
1979    return (enum adreno_compare_func) op;
1980 }
1981
1982 static void
1983 tu_init_sampler(struct tu_device *device,
1984                 struct tu_sampler *sampler,
1985                 const VkSamplerCreateInfo *pCreateInfo)
1986 {
1987    unsigned aniso = pCreateInfo->anisotropyEnable ?
1988       util_last_bit(MIN2((uint32_t)pCreateInfo->maxAnisotropy >> 1, 8)) : 0;
1989    bool miplinear = (pCreateInfo->mipmapMode == VK_SAMPLER_MIPMAP_MODE_LINEAR);
1990    bool needs_border = false;
1991
1992    sampler->state[0] =
1993       COND(miplinear, A6XX_TEX_SAMP_0_MIPFILTER_LINEAR_NEAR) |
1994       A6XX_TEX_SAMP_0_XY_MAG(tu6_tex_filter(pCreateInfo->magFilter, aniso)) |
1995       A6XX_TEX_SAMP_0_XY_MIN(tu6_tex_filter(pCreateInfo->minFilter, aniso)) |
1996       A6XX_TEX_SAMP_0_ANISO(aniso) |
1997       A6XX_TEX_SAMP_0_WRAP_S(tu6_tex_wrap(pCreateInfo->addressModeU, &needs_border)) |
1998       A6XX_TEX_SAMP_0_WRAP_T(tu6_tex_wrap(pCreateInfo->addressModeV, &needs_border)) |
1999       A6XX_TEX_SAMP_0_WRAP_R(tu6_tex_wrap(pCreateInfo->addressModeW, &needs_border)) |
2000       A6XX_TEX_SAMP_0_LOD_BIAS(pCreateInfo->mipLodBias);
2001    sampler->state[1] =
2002       /* COND(!cso->seamless_cube_map, A6XX_TEX_SAMP_1_CUBEMAPSEAMLESSFILTOFF) | */
2003       COND(pCreateInfo->unnormalizedCoordinates, A6XX_TEX_SAMP_1_UNNORM_COORDS) |
2004       A6XX_TEX_SAMP_1_MIN_LOD(pCreateInfo->minLod) |
2005       A6XX_TEX_SAMP_1_MAX_LOD(pCreateInfo->maxLod) |
2006       COND(pCreateInfo->compareEnable,
2007            A6XX_TEX_SAMP_1_COMPARE_FUNC(tu6_compare_func(pCreateInfo->compareOp)));
2008    sampler->state[2] = 0;
2009    sampler->state[3] = 0;
2010
2011    /* TODO:
2012     * A6XX_TEX_SAMP_1_MIPFILTER_LINEAR_FAR disables mipmapping, but vk has no NONE mipfilter?
2013     * border color
2014     */
2015
2016    sampler->needs_border = needs_border;
2017    sampler->border = pCreateInfo->borderColor;
2018 }
2019
2020 VkResult
2021 tu_CreateSampler(VkDevice _device,
2022                  const VkSamplerCreateInfo *pCreateInfo,
2023                  const VkAllocationCallbacks *pAllocator,
2024                  VkSampler *pSampler)
2025 {
2026    TU_FROM_HANDLE(tu_device, device, _device);
2027    struct tu_sampler *sampler;
2028
2029    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO);
2030
2031    sampler = vk_alloc2(&device->alloc, pAllocator, sizeof(*sampler), 8,
2032                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2033    if (!sampler)
2034       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2035
2036    tu_init_sampler(device, sampler, pCreateInfo);
2037    *pSampler = tu_sampler_to_handle(sampler);
2038
2039    return VK_SUCCESS;
2040 }
2041
2042 void
2043 tu_DestroySampler(VkDevice _device,
2044                   VkSampler _sampler,
2045                   const VkAllocationCallbacks *pAllocator)
2046 {
2047    TU_FROM_HANDLE(tu_device, device, _device);
2048    TU_FROM_HANDLE(tu_sampler, sampler, _sampler);
2049
2050    if (!sampler)
2051       return;
2052    vk_free2(&device->alloc, pAllocator, sampler);
2053 }
2054
2055 /* vk_icd.h does not declare this function, so we declare it here to
2056  * suppress Wmissing-prototypes.
2057  */
2058 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2059 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion);
2060
2061 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2062 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion)
2063 {
2064    /* For the full details on loader interface versioning, see
2065     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2066     * What follows is a condensed summary, to help you navigate the large and
2067     * confusing official doc.
2068     *
2069     *   - Loader interface v0 is incompatible with later versions. We don't
2070     *     support it.
2071     *
2072     *   - In loader interface v1:
2073     *       - The first ICD entrypoint called by the loader is
2074     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2075     *         entrypoint.
2076     *       - The ICD must statically expose no other Vulkan symbol unless it
2077     * is linked with -Bsymbolic.
2078     *       - Each dispatchable Vulkan handle created by the ICD must be
2079     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2080     *         ICD must initialize VK_LOADER_DATA.loadMagic to
2081     * ICD_LOADER_MAGIC.
2082     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2083     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2084     *         such loader-managed surfaces.
2085     *
2086     *    - Loader interface v2 differs from v1 in:
2087     *       - The first ICD entrypoint called by the loader is
2088     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2089     *         statically expose this entrypoint.
2090     *
2091     *    - Loader interface v3 differs from v2 in:
2092     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2093     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2094     *          because the loader no longer does so.
2095     */
2096    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2097    return VK_SUCCESS;
2098 }
2099
2100 VkResult
2101 tu_GetMemoryFdKHR(VkDevice _device,
2102                   const VkMemoryGetFdInfoKHR *pGetFdInfo,
2103                   int *pFd)
2104 {
2105    TU_FROM_HANDLE(tu_device, device, _device);
2106    TU_FROM_HANDLE(tu_device_memory, memory, pGetFdInfo->memory);
2107
2108    assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2109
2110    /* At the moment, we support only the below handle types. */
2111    assert(pGetFdInfo->handleType ==
2112              VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2113           pGetFdInfo->handleType ==
2114              VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2115
2116    int prime_fd = tu_bo_export_dmabuf(device, &memory->bo);
2117    if (prime_fd < 0)
2118       return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
2119
2120    *pFd = prime_fd;
2121    return VK_SUCCESS;
2122 }
2123
2124 VkResult
2125 tu_GetMemoryFdPropertiesKHR(VkDevice _device,
2126                             VkExternalMemoryHandleTypeFlagBits handleType,
2127                             int fd,
2128                             VkMemoryFdPropertiesKHR *pMemoryFdProperties)
2129 {
2130    assert(handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2131    pMemoryFdProperties->memoryTypeBits = 1;
2132    return VK_SUCCESS;
2133 }
2134
2135 void
2136 tu_GetPhysicalDeviceExternalSemaphoreProperties(
2137    VkPhysicalDevice physicalDevice,
2138    const VkPhysicalDeviceExternalSemaphoreInfo *pExternalSemaphoreInfo,
2139    VkExternalSemaphoreProperties *pExternalSemaphoreProperties)
2140 {
2141    pExternalSemaphoreProperties->exportFromImportedHandleTypes = 0;
2142    pExternalSemaphoreProperties->compatibleHandleTypes = 0;
2143    pExternalSemaphoreProperties->externalSemaphoreFeatures = 0;
2144 }
2145
2146 void
2147 tu_GetPhysicalDeviceExternalFenceProperties(
2148    VkPhysicalDevice physicalDevice,
2149    const VkPhysicalDeviceExternalFenceInfo *pExternalFenceInfo,
2150    VkExternalFenceProperties *pExternalFenceProperties)
2151 {
2152    pExternalFenceProperties->exportFromImportedHandleTypes = 0;
2153    pExternalFenceProperties->compatibleHandleTypes = 0;
2154    pExternalFenceProperties->externalFenceFeatures = 0;
2155 }
2156
2157 VkResult
2158 tu_CreateDebugReportCallbackEXT(
2159    VkInstance _instance,
2160    const VkDebugReportCallbackCreateInfoEXT *pCreateInfo,
2161    const VkAllocationCallbacks *pAllocator,
2162    VkDebugReportCallbackEXT *pCallback)
2163 {
2164    TU_FROM_HANDLE(tu_instance, instance, _instance);
2165    return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2166                                           pCreateInfo, pAllocator,
2167                                           &instance->alloc, pCallback);
2168 }
2169
2170 void
2171 tu_DestroyDebugReportCallbackEXT(VkInstance _instance,
2172                                  VkDebugReportCallbackEXT _callback,
2173                                  const VkAllocationCallbacks *pAllocator)
2174 {
2175    TU_FROM_HANDLE(tu_instance, instance, _instance);
2176    vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2177                                     _callback, pAllocator, &instance->alloc);
2178 }
2179
2180 void
2181 tu_DebugReportMessageEXT(VkInstance _instance,
2182                          VkDebugReportFlagsEXT flags,
2183                          VkDebugReportObjectTypeEXT objectType,
2184                          uint64_t object,
2185                          size_t location,
2186                          int32_t messageCode,
2187                          const char *pLayerPrefix,
2188                          const char *pMessage)
2189 {
2190    TU_FROM_HANDLE(tu_instance, instance, _instance);
2191    vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2192                    object, location, messageCode, pLayerPrefix, pMessage);
2193 }
2194
2195 void
2196 tu_GetDeviceGroupPeerMemoryFeatures(
2197    VkDevice device,
2198    uint32_t heapIndex,
2199    uint32_t localDeviceIndex,
2200    uint32_t remoteDeviceIndex,
2201    VkPeerMemoryFeatureFlags *pPeerMemoryFeatures)
2202 {
2203    assert(localDeviceIndex == remoteDeviceIndex);
2204
2205    *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2206                           VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2207                           VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2208                           VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2209 }