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       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
 716          VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
 717             (VkPhysicalDeviceTransformFeedbackFeaturesEXT *) ext;
 718          features->transformFeedback = true;
 719          features->geometryStreams = false;
 720          break;
 721       }
 722       default:
 723          break;
 724       }
 725    }
 726    return tu_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 727 }
 728
 729 void
 730 tu_GetPhysicalDeviceProperties(VkPhysicalDevice physicalDevice,
 731                                VkPhysicalDeviceProperties *pProperties)
 732 {
 733    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 734    VkSampleCountFlags sample_counts = VK_SAMPLE_COUNT_1_BIT |
 735       VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT | VK_SAMPLE_COUNT_8_BIT;
 736
 737    /* make sure that the entire descriptor set is addressable with a signed
 738     * 32-bit int. So the sum of all limits scaled by descriptor size has to
 739     * be at most 2 GiB. the combined image & samples object count as one of
 740     * both. This limit is for the pipeline layout, not for the set layout, but
 741     * there is no set limit, so we just set a pipeline limit. I don't think
 742     * any app is going to hit this soon. */
 743    size_t max_descriptor_set_size =
 744       ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS) /
 745       (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
 746        32 /* storage buffer, 32 due to potential space wasted on alignment */ +
 747        32 /* sampler, largest when combined with image */ +
 748        64 /* sampled image */ + 64 /* storage image */);
 749
 750    VkPhysicalDeviceLimits limits = {
 751       .maxImageDimension1D = (1 << 14),
 752       .maxImageDimension2D = (1 << 14),
 753       .maxImageDimension3D = (1 << 11),
 754       .maxImageDimensionCube = (1 << 14),
 755       .maxImageArrayLayers = (1 << 11),
 756       .maxTexelBufferElements = 128 * 1024 * 1024,
 757       .maxUniformBufferRange = UINT32_MAX,
 758       .maxStorageBufferRange = MAX_STORAGE_BUFFER_RANGE,
 759       .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
 760       .maxMemoryAllocationCount = UINT32_MAX,
 761       .maxSamplerAllocationCount = 64 * 1024,
 762       .bufferImageGranularity = 64,          /* A cache line */
 763       .sparseAddressSpaceSize = 0xffffffffu, /* buffer max size */
 764       .maxBoundDescriptorSets = MAX_SETS,
 765       .maxPerStageDescriptorSamplers = max_descriptor_set_size,
 766       .maxPerStageDescriptorUniformBuffers = max_descriptor_set_size,
 767       .maxPerStageDescriptorStorageBuffers = max_descriptor_set_size,
 768       .maxPerStageDescriptorSampledImages = max_descriptor_set_size,
 769       .maxPerStageDescriptorStorageImages = max_descriptor_set_size,
 770       .maxPerStageDescriptorInputAttachments = max_descriptor_set_size,
 771       .maxPerStageResources = max_descriptor_set_size,
 772       .maxDescriptorSetSamplers = max_descriptor_set_size,
 773       .maxDescriptorSetUniformBuffers = max_descriptor_set_size,
 774       .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS,
 775       .maxDescriptorSetStorageBuffers = max_descriptor_set_size,
 776       .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS,
 777       .maxDescriptorSetSampledImages = max_descriptor_set_size,
 778       .maxDescriptorSetStorageImages = max_descriptor_set_size,
 779       .maxDescriptorSetInputAttachments = max_descriptor_set_size,
 780       .maxVertexInputAttributes = 32,
 781       .maxVertexInputBindings = 32,
 782       .maxVertexInputAttributeOffset = 2047,
 783       .maxVertexInputBindingStride = 2048,
 784       .maxVertexOutputComponents = 128,
 785       .maxTessellationGenerationLevel = 64,
 786       .maxTessellationPatchSize = 32,
 787       .maxTessellationControlPerVertexInputComponents = 128,
 788       .maxTessellationControlPerVertexOutputComponents = 128,
 789       .maxTessellationControlPerPatchOutputComponents = 120,
 790       .maxTessellationControlTotalOutputComponents = 4096,
 791       .maxTessellationEvaluationInputComponents = 128,
 792       .maxTessellationEvaluationOutputComponents = 128,
 793       .maxGeometryShaderInvocations = 127,
 794       .maxGeometryInputComponents = 64,
 795       .maxGeometryOutputComponents = 128,
 796       .maxGeometryOutputVertices = 256,
 797       .maxGeometryTotalOutputComponents = 1024,
 798       .maxFragmentInputComponents = 128,
 799       .maxFragmentOutputAttachments = 8,
 800       .maxFragmentDualSrcAttachments = 1,
 801       .maxFragmentCombinedOutputResources = 8,
 802       .maxComputeSharedMemorySize = 32768,
 803       .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
 804       .maxComputeWorkGroupInvocations = 2048,
 805       .maxComputeWorkGroupSize = { 2048, 2048, 2048 },
 806       .subPixelPrecisionBits = 4 /* FIXME */,
 807       .subTexelPrecisionBits = 4 /* FIXME */,
 808       .mipmapPrecisionBits = 4 /* FIXME */,
 809       .maxDrawIndexedIndexValue = UINT32_MAX,
 810       .maxDrawIndirectCount = UINT32_MAX,
 811       .maxSamplerLodBias = 16,
 812       .maxSamplerAnisotropy = 16,
 813       .maxViewports = MAX_VIEWPORTS,
 814       .maxViewportDimensions = { (1 << 14), (1 << 14) },
 815       .viewportBoundsRange = { INT16_MIN, INT16_MAX },
 816       .viewportSubPixelBits = 8,
 817       .minMemoryMapAlignment = 4096, /* A page */
 818       .minTexelBufferOffsetAlignment = 64,
 819       .minUniformBufferOffsetAlignment = 4,
 820       .minStorageBufferOffsetAlignment = 4,
 821       .minTexelOffset = -32,
 822       .maxTexelOffset = 31,
 823       .minTexelGatherOffset = -32,
 824       .maxTexelGatherOffset = 31,
 825       .minInterpolationOffset = -2,
 826       .maxInterpolationOffset = 2,
 827       .subPixelInterpolationOffsetBits = 8,
 828       .maxFramebufferWidth = (1 << 14),
 829       .maxFramebufferHeight = (1 << 14),
 830       .maxFramebufferLayers = (1 << 10),
 831       .framebufferColorSampleCounts = sample_counts,
 832       .framebufferDepthSampleCounts = sample_counts,
 833       .framebufferStencilSampleCounts = sample_counts,
 834       .framebufferNoAttachmentsSampleCounts = sample_counts,
 835       .maxColorAttachments = MAX_RTS,
 836       .sampledImageColorSampleCounts = sample_counts,
 837       .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 838       .sampledImageDepthSampleCounts = sample_counts,
 839       .sampledImageStencilSampleCounts = sample_counts,
 840       .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 841       .maxSampleMaskWords = 1,
 842       .timestampComputeAndGraphics = false, /* FINISHME */
 843       .timestampPeriod = 1,
 844       .maxClipDistances = 8,
 845       .maxCullDistances = 8,
 846       .maxCombinedClipAndCullDistances = 8,
 847       .discreteQueuePriorities = 1,
 848       .pointSizeRange = { 0.125, 255.875 },
 849       .lineWidthRange = { 0.0, 7.9921875 },
 850       .pointSizeGranularity = (1.0 / 8.0),
 851       .lineWidthGranularity = (1.0 / 128.0),
 852       .strictLines = false, /* FINISHME */
 853       .standardSampleLocations = true,
 854       .optimalBufferCopyOffsetAlignment = 128,
 855       .optimalBufferCopyRowPitchAlignment = 128,
 856       .nonCoherentAtomSize = 64,
 857    };
 858
 859    *pProperties = (VkPhysicalDeviceProperties) {
 860       .apiVersion = tu_physical_device_api_version(pdevice),
 861       .driverVersion = vk_get_driver_version(),
 862       .vendorID = 0, /* TODO */
 863       .deviceID = 0,
 864       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 865       .limits = limits,
 866       .sparseProperties = { 0 },
 867    };
 868
 869    strcpy(pProperties->deviceName, pdevice->name);
 870    memcpy(pProperties->pipelineCacheUUID, pdevice->cache_uuid, VK_UUID_SIZE);
 871 }
 872
 873 void
 874 tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
 875                                 VkPhysicalDeviceProperties2 *pProperties)
 876 {
 877    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 878    tu_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 879
 880    vk_foreach_struct(ext, pProperties->pNext)
 881    {
 882       switch (ext->sType) {
 883       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 884          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 885             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 886          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 887          break;
 888       }
 889       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
 890          VkPhysicalDeviceIDProperties *properties =
 891             (VkPhysicalDeviceIDProperties *) ext;
 892          memcpy(properties->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
 893          memcpy(properties->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
 894          properties->deviceLUIDValid = false;
 895          break;
 896       }
 897       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
 898          VkPhysicalDeviceMultiviewProperties *properties =
 899             (VkPhysicalDeviceMultiviewProperties *) ext;
 900          properties->maxMultiviewViewCount = MAX_VIEWS;
 901          properties->maxMultiviewInstanceIndex = INT_MAX;
 902          break;
 903       }
 904       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
 905          VkPhysicalDevicePointClippingProperties *properties =
 906             (VkPhysicalDevicePointClippingProperties *) ext;
 907          properties->pointClippingBehavior =
 908             VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
 909          break;
 910       }
 911       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
 912          VkPhysicalDeviceMaintenance3Properties *properties =
 913             (VkPhysicalDeviceMaintenance3Properties *) ext;
 914          /* Make sure everything is addressable by a signed 32-bit int, and
 915           * our largest descriptors are 96 bytes. */
 916          properties->maxPerSetDescriptors = (1ull << 31) / 96;
 917          /* Our buffer size fields allow only this much */
 918          properties->maxMemoryAllocationSize = 0xFFFFFFFFull;
 919          break;
 920       }
 921       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
 922          VkPhysicalDeviceTransformFeedbackPropertiesEXT *properties =
 923             (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
 924
 925          properties->maxTransformFeedbackStreams = IR3_MAX_SO_STREAMS;
 926          properties->maxTransformFeedbackBuffers = IR3_MAX_SO_BUFFERS;
 927          properties->maxTransformFeedbackBufferSize = UINT32_MAX;
 928          properties->maxTransformFeedbackStreamDataSize = 512;
 929          properties->maxTransformFeedbackBufferDataSize = 512;
 930          properties->maxTransformFeedbackBufferDataStride = 512;
 931          /* TODO: enable xfb query */
 932          properties->transformFeedbackQueries = false;
 933          properties->transformFeedbackStreamsLinesTriangles = false;
 934          properties->transformFeedbackRasterizationStreamSelect = false;
 935          properties->transformFeedbackDraw = true;
 936          break;
 937       }
 938       default:
 939          break;
 940       }
 941    }
 942 }
 943
 944 static const VkQueueFamilyProperties tu_queue_family_properties = {
 945    .queueFlags =
 946       VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT,
 947    .queueCount = 1,
 948    .timestampValidBits = 0, /* FINISHME */
 949    .minImageTransferGranularity = { 1, 1, 1 },
 950 };
 951
 952 void
 953 tu_GetPhysicalDeviceQueueFamilyProperties(
 954    VkPhysicalDevice physicalDevice,
 955    uint32_t *pQueueFamilyPropertyCount,
 956    VkQueueFamilyProperties *pQueueFamilyProperties)
 957 {
 958    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 959
 960    vk_outarray_append(&out, p) { *p = tu_queue_family_properties; }
 961 }
 962
 963 void
 964 tu_GetPhysicalDeviceQueueFamilyProperties2(
 965    VkPhysicalDevice physicalDevice,
 966    uint32_t *pQueueFamilyPropertyCount,
 967    VkQueueFamilyProperties2 *pQueueFamilyProperties)
 968 {
 969    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 970
 971    vk_outarray_append(&out, p)
 972    {
 973       p->queueFamilyProperties = tu_queue_family_properties;
 974    }
 975 }
 976
 977 static uint64_t
 978 tu_get_system_heap_size()
 979 {
 980    struct sysinfo info;
 981    sysinfo(&info);
 982
 983    uint64_t total_ram = (uint64_t) info.totalram * (uint64_t) info.mem_unit;
 984
 985    /* We don't want to burn too much ram with the GPU.  If the user has 4GiB
 986     * or less, we use at most half.  If they have more than 4GiB, we use 3/4.
 987     */
 988    uint64_t available_ram;
 989    if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
 990       available_ram = total_ram / 2;
 991    else
 992       available_ram = total_ram * 3 / 4;
 993
 994    return available_ram;
 995 }
 996
 997 void
 998 tu_GetPhysicalDeviceMemoryProperties(
 999    VkPhysicalDevice physicalDevice,
1000    VkPhysicalDeviceMemoryProperties *pMemoryProperties)
1001 {
1002    pMemoryProperties->memoryHeapCount = 1;
1003    pMemoryProperties->memoryHeaps[0].size = tu_get_system_heap_size();
1004    pMemoryProperties->memoryHeaps[0].flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT;
1005
1006    pMemoryProperties->memoryTypeCount = 1;
1007    pMemoryProperties->memoryTypes[0].propertyFlags =
1008       VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
1009       VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
1010       VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
1011    pMemoryProperties->memoryTypes[0].heapIndex = 0;
1012 }
1013
1014 void
1015 tu_GetPhysicalDeviceMemoryProperties2(
1016    VkPhysicalDevice physicalDevice,
1017    VkPhysicalDeviceMemoryProperties2 *pMemoryProperties)
1018 {
1019    return tu_GetPhysicalDeviceMemoryProperties(
1020       physicalDevice, &pMemoryProperties->memoryProperties);
1021 }
1022
1023 static VkResult
1024 tu_queue_init(struct tu_device *device,
1025               struct tu_queue *queue,
1026               uint32_t queue_family_index,
1027               int idx,
1028               VkDeviceQueueCreateFlags flags)
1029 {
1030    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1031    queue->device = device;
1032    queue->queue_family_index = queue_family_index;
1033    queue->queue_idx = idx;
1034    queue->flags = flags;
1035
1036    int ret = tu_drm_submitqueue_new(device, 0, &queue->msm_queue_id);
1037    if (ret)
1038       return VK_ERROR_INITIALIZATION_FAILED;
1039
1040    tu_fence_init(&queue->submit_fence, false);
1041
1042    return VK_SUCCESS;
1043 }
1044
1045 static void
1046 tu_queue_finish(struct tu_queue *queue)
1047 {
1048    tu_fence_finish(&queue->submit_fence);
1049    tu_drm_submitqueue_close(queue->device, queue->msm_queue_id);
1050 }
1051
1052 static int
1053 tu_get_device_extension_index(const char *name)
1054 {
1055    for (unsigned i = 0; i < TU_DEVICE_EXTENSION_COUNT; ++i) {
1056       if (strcmp(name, tu_device_extensions[i].extensionName) == 0)
1057          return i;
1058    }
1059    return -1;
1060 }
1061
1062 VkResult
1063 tu_CreateDevice(VkPhysicalDevice physicalDevice,
1064                 const VkDeviceCreateInfo *pCreateInfo,
1065                 const VkAllocationCallbacks *pAllocator,
1066                 VkDevice *pDevice)
1067 {
1068    TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
1069    VkResult result;
1070    struct tu_device *device;
1071
1072    /* Check enabled features */
1073    if (pCreateInfo->pEnabledFeatures) {
1074       VkPhysicalDeviceFeatures supported_features;
1075       tu_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1076       VkBool32 *supported_feature = (VkBool32 *) &supported_features;
1077       VkBool32 *enabled_feature = (VkBool32 *) pCreateInfo->pEnabledFeatures;
1078       unsigned num_features =
1079          sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1080       for (uint32_t i = 0; i < num_features; i++) {
1081          if (enabled_feature[i] && !supported_feature[i])
1082             return vk_error(physical_device->instance,
1083                             VK_ERROR_FEATURE_NOT_PRESENT);
1084       }
1085    }
1086
1087    device = vk_zalloc2(&physical_device->instance->alloc, pAllocator,
1088                        sizeof(*device), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1089    if (!device)
1090       return vk_error(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1091
1092    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1093    device->instance = physical_device->instance;
1094    device->physical_device = physical_device;
1095
1096    if (pAllocator)
1097       device->alloc = *pAllocator;
1098    else
1099       device->alloc = physical_device->instance->alloc;
1100
1101    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1102       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
1103       int index = tu_get_device_extension_index(ext_name);
1104       if (index < 0 ||
1105           !physical_device->supported_extensions.extensions[index]) {
1106          vk_free(&device->alloc, device);
1107          return vk_error(physical_device->instance,
1108                          VK_ERROR_EXTENSION_NOT_PRESENT);
1109       }
1110
1111       device->enabled_extensions.extensions[index] = true;
1112    }
1113
1114    for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1115       const VkDeviceQueueCreateInfo *queue_create =
1116          &pCreateInfo->pQueueCreateInfos[i];
1117       uint32_t qfi = queue_create->queueFamilyIndex;
1118       device->queues[qfi] = vk_alloc(
1119          &device->alloc, queue_create->queueCount * sizeof(struct tu_queue),
1120          8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1121       if (!device->queues[qfi]) {
1122          result = VK_ERROR_OUT_OF_HOST_MEMORY;
1123          goto fail_queues;
1124       }
1125
1126       memset(device->queues[qfi], 0,
1127              queue_create->queueCount * sizeof(struct tu_queue));
1128
1129       device->queue_count[qfi] = queue_create->queueCount;
1130
1131       for (unsigned q = 0; q < queue_create->queueCount; q++) {
1132          result = tu_queue_init(device, &device->queues[qfi][q], qfi, q,
1133                                 queue_create->flags);
1134          if (result != VK_SUCCESS)
1135             goto fail_queues;
1136       }
1137    }
1138
1139    device->compiler = ir3_compiler_create(NULL, physical_device->gpu_id);
1140    if (!device->compiler)
1141       goto fail_queues;
1142
1143 #define VSC_DATA_SIZE(pitch)  ((pitch) * 32 + 0x100)  /* extra size to store VSC_SIZE */
1144 #define VSC_DATA2_SIZE(pitch) ((pitch) * 32)
1145
1146    device->vsc_data_pitch = 0x440 * 4;
1147    device->vsc_data2_pitch = 0x1040 * 4;
1148
1149    result = tu_bo_init_new(device, &device->vsc_data, VSC_DATA_SIZE(device->vsc_data_pitch));
1150    if (result != VK_SUCCESS)
1151       goto fail_vsc_data;
1152
1153    result = tu_bo_init_new(device, &device->vsc_data2, VSC_DATA2_SIZE(device->vsc_data2_pitch));
1154    if (result != VK_SUCCESS)
1155       goto fail_vsc_data2;
1156
1157    VkPipelineCacheCreateInfo ci;
1158    ci.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
1159    ci.pNext = NULL;
1160    ci.flags = 0;
1161    ci.pInitialData = NULL;
1162    ci.initialDataSize = 0;
1163    VkPipelineCache pc;
1164    result =
1165       tu_CreatePipelineCache(tu_device_to_handle(device), &ci, NULL, &pc);
1166    if (result != VK_SUCCESS)
1167       goto fail_pipeline_cache;
1168
1169    device->mem_cache = tu_pipeline_cache_from_handle(pc);
1170
1171    *pDevice = tu_device_to_handle(device);
1172    return VK_SUCCESS;
1173
1174 fail_pipeline_cache:
1175    tu_bo_finish(device, &device->vsc_data2);
1176
1177 fail_vsc_data2:
1178    tu_bo_finish(device, &device->vsc_data);
1179
1180 fail_vsc_data:
1181    ralloc_free(device->compiler);
1182
1183 fail_queues:
1184    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1185       for (unsigned q = 0; q < device->queue_count[i]; q++)
1186          tu_queue_finish(&device->queues[i][q]);
1187       if (device->queue_count[i])
1188          vk_free(&device->alloc, device->queues[i]);
1189    }
1190
1191    vk_free(&device->alloc, device);
1192    return result;
1193 }
1194
1195 void
1196 tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
1197 {
1198    TU_FROM_HANDLE(tu_device, device, _device);
1199
1200    if (!device)
1201       return;
1202
1203    tu_bo_finish(device, &device->vsc_data);
1204    tu_bo_finish(device, &device->vsc_data2);
1205
1206    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1207       for (unsigned q = 0; q < device->queue_count[i]; q++)
1208          tu_queue_finish(&device->queues[i][q]);
1209       if (device->queue_count[i])
1210          vk_free(&device->alloc, device->queues[i]);
1211    }
1212
1213    /* the compiler does not use pAllocator */
1214    ralloc_free(device->compiler);
1215
1216    VkPipelineCache pc = tu_pipeline_cache_to_handle(device->mem_cache);
1217    tu_DestroyPipelineCache(tu_device_to_handle(device), pc, NULL);
1218
1219    vk_free(&device->alloc, device);
1220 }
1221
1222 VkResult
1223 tu_EnumerateInstanceLayerProperties(uint32_t *pPropertyCount,
1224                                     VkLayerProperties *pProperties)
1225 {
1226    *pPropertyCount = 0;
1227    return VK_SUCCESS;
1228 }
1229
1230 VkResult
1231 tu_EnumerateDeviceLayerProperties(VkPhysicalDevice physicalDevice,
1232                                   uint32_t *pPropertyCount,
1233                                   VkLayerProperties *pProperties)
1234 {
1235    *pPropertyCount = 0;
1236    return VK_SUCCESS;
1237 }
1238
1239 void
1240 tu_GetDeviceQueue2(VkDevice _device,
1241                    const VkDeviceQueueInfo2 *pQueueInfo,
1242                    VkQueue *pQueue)
1243 {
1244    TU_FROM_HANDLE(tu_device, device, _device);
1245    struct tu_queue *queue;
1246
1247    queue =
1248       &device->queues[pQueueInfo->queueFamilyIndex][pQueueInfo->queueIndex];
1249    if (pQueueInfo->flags != queue->flags) {
1250       /* From the Vulkan 1.1.70 spec:
1251        *
1252        * "The queue returned by vkGetDeviceQueue2 must have the same
1253        * flags value from this structure as that used at device
1254        * creation time in a VkDeviceQueueCreateInfo instance. If no
1255        * matching flags were specified at device creation time then
1256        * pQueue will return VK_NULL_HANDLE."
1257        */
1258       *pQueue = VK_NULL_HANDLE;
1259       return;
1260    }
1261
1262    *pQueue = tu_queue_to_handle(queue);
1263 }
1264
1265 void
1266 tu_GetDeviceQueue(VkDevice _device,
1267                   uint32_t queueFamilyIndex,
1268                   uint32_t queueIndex,
1269                   VkQueue *pQueue)
1270 {
1271    const VkDeviceQueueInfo2 info =
1272       (VkDeviceQueueInfo2) { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
1273                              .queueFamilyIndex = queueFamilyIndex,
1274                              .queueIndex = queueIndex };
1275
1276    tu_GetDeviceQueue2(_device, &info, pQueue);
1277 }
1278
1279 VkResult
1280 tu_QueueSubmit(VkQueue _queue,
1281                uint32_t submitCount,
1282                const VkSubmitInfo *pSubmits,
1283                VkFence _fence)
1284 {
1285    TU_FROM_HANDLE(tu_queue, queue, _queue);
1286
1287    for (uint32_t i = 0; i < submitCount; ++i) {
1288       const VkSubmitInfo *submit = pSubmits + i;
1289       const bool last_submit = (i == submitCount - 1);
1290       struct tu_bo_list bo_list;
1291       tu_bo_list_init(&bo_list);
1292
1293       uint32_t entry_count = 0;
1294       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1295          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1296          entry_count += cmdbuf->cs.entry_count;
1297       }
1298
1299       struct drm_msm_gem_submit_cmd cmds[entry_count];
1300       uint32_t entry_idx = 0;
1301       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1302          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1303          struct tu_cs *cs = &cmdbuf->cs;
1304          for (unsigned i = 0; i < cs->entry_count; ++i, ++entry_idx) {
1305             cmds[entry_idx].type = MSM_SUBMIT_CMD_BUF;
1306             cmds[entry_idx].submit_idx =
1307                tu_bo_list_add(&bo_list, cs->entries[i].bo,
1308                               MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
1309             cmds[entry_idx].submit_offset = cs->entries[i].offset;
1310             cmds[entry_idx].size = cs->entries[i].size;
1311             cmds[entry_idx].pad = 0;
1312             cmds[entry_idx].nr_relocs = 0;
1313             cmds[entry_idx].relocs = 0;
1314          }
1315
1316          tu_bo_list_merge(&bo_list, &cmdbuf->bo_list);
1317       }
1318
1319       uint32_t flags = MSM_PIPE_3D0;
1320       if (last_submit) {
1321          flags |= MSM_SUBMIT_FENCE_FD_OUT;
1322       }
1323
1324       struct drm_msm_gem_submit req = {
1325          .flags = flags,
1326          .queueid = queue->msm_queue_id,
1327          .bos = (uint64_t)(uintptr_t) bo_list.bo_infos,
1328          .nr_bos = bo_list.count,
1329          .cmds = (uint64_t)(uintptr_t)cmds,
1330          .nr_cmds = entry_count,
1331       };
1332
1333       int ret = drmCommandWriteRead(queue->device->physical_device->local_fd,
1334                                     DRM_MSM_GEM_SUBMIT,
1335                                     &req, sizeof(req));
1336       if (ret) {
1337          fprintf(stderr, "submit failed: %s\n", strerror(errno));
1338          abort();
1339       }
1340
1341       tu_bo_list_destroy(&bo_list);
1342
1343       if (last_submit) {
1344          /* no need to merge fences as queue execution is serialized */
1345          tu_fence_update_fd(&queue->submit_fence, req.fence_fd);
1346       }
1347    }
1348
1349    if (_fence != VK_NULL_HANDLE) {
1350       TU_FROM_HANDLE(tu_fence, fence, _fence);
1351       tu_fence_copy(fence, &queue->submit_fence);
1352    }
1353
1354    return VK_SUCCESS;
1355 }
1356
1357 VkResult
1358 tu_QueueWaitIdle(VkQueue _queue)
1359 {
1360    TU_FROM_HANDLE(tu_queue, queue, _queue);
1361
1362    tu_fence_wait_idle(&queue->submit_fence);
1363
1364    return VK_SUCCESS;
1365 }
1366
1367 VkResult
1368 tu_DeviceWaitIdle(VkDevice _device)
1369 {
1370    TU_FROM_HANDLE(tu_device, device, _device);
1371
1372    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1373       for (unsigned q = 0; q < device->queue_count[i]; q++) {
1374          tu_QueueWaitIdle(tu_queue_to_handle(&device->queues[i][q]));
1375       }
1376    }
1377    return VK_SUCCESS;
1378 }
1379
1380 VkResult
1381 tu_EnumerateInstanceExtensionProperties(const char *pLayerName,
1382                                         uint32_t *pPropertyCount,
1383                                         VkExtensionProperties *pProperties)
1384 {
1385    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1386
1387    /* We spport no lyaers */
1388    if (pLayerName)
1389       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1390
1391    for (int i = 0; i < TU_INSTANCE_EXTENSION_COUNT; i++) {
1392       if (tu_supported_instance_extensions.extensions[i]) {
1393          vk_outarray_append(&out, prop) { *prop = tu_instance_extensions[i]; }
1394       }
1395    }
1396
1397    return vk_outarray_status(&out);
1398 }
1399
1400 VkResult
1401 tu_EnumerateDeviceExtensionProperties(VkPhysicalDevice physicalDevice,
1402                                       const char *pLayerName,
1403                                       uint32_t *pPropertyCount,
1404                                       VkExtensionProperties *pProperties)
1405 {
1406    /* We spport no lyaers */
1407    TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
1408    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1409
1410    /* We spport no lyaers */
1411    if (pLayerName)
1412       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1413
1414    for (int i = 0; i < TU_DEVICE_EXTENSION_COUNT; i++) {
1415       if (device->supported_extensions.extensions[i]) {
1416          vk_outarray_append(&out, prop) { *prop = tu_device_extensions[i]; }
1417       }
1418    }
1419
1420    return vk_outarray_status(&out);
1421 }
1422
1423 PFN_vkVoidFunction
1424 tu_GetInstanceProcAddr(VkInstance _instance, const char *pName)
1425 {
1426    TU_FROM_HANDLE(tu_instance, instance, _instance);
1427
1428    return tu_lookup_entrypoint_checked(
1429       pName, instance ? instance->api_version : 0,
1430       instance ? &instance->enabled_extensions : NULL, NULL);
1431 }
1432
1433 /* The loader wants us to expose a second GetInstanceProcAddr function
1434  * to work around certain LD_PRELOAD issues seen in apps.
1435  */
1436 PUBLIC
1437 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1438 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName);
1439
1440 PUBLIC
1441 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1442 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName)
1443 {
1444    return tu_GetInstanceProcAddr(instance, pName);
1445 }
1446
1447 PFN_vkVoidFunction
1448 tu_GetDeviceProcAddr(VkDevice _device, const char *pName)
1449 {
1450    TU_FROM_HANDLE(tu_device, device, _device);
1451
1452    return tu_lookup_entrypoint_checked(pName, device->instance->api_version,
1453                                        &device->instance->enabled_extensions,
1454                                        &device->enabled_extensions);
1455 }
1456
1457 static VkResult
1458 tu_alloc_memory(struct tu_device *device,
1459                 const VkMemoryAllocateInfo *pAllocateInfo,
1460                 const VkAllocationCallbacks *pAllocator,
1461                 VkDeviceMemory *pMem)
1462 {
1463    struct tu_device_memory *mem;
1464    VkResult result;
1465
1466    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1467
1468    if (pAllocateInfo->allocationSize == 0) {
1469       /* Apparently, this is allowed */
1470       *pMem = VK_NULL_HANDLE;
1471       return VK_SUCCESS;
1472    }
1473
1474    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1475                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1476    if (mem == NULL)
1477       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1478
1479    const VkImportMemoryFdInfoKHR *fd_info =
1480       vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
1481    if (fd_info && !fd_info->handleType)
1482       fd_info = NULL;
1483
1484    if (fd_info) {
1485       assert(fd_info->handleType ==
1486                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
1487              fd_info->handleType ==
1488                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1489
1490       /*
1491        * TODO Importing the same fd twice gives us the same handle without
1492        * reference counting.  We need to maintain a per-instance handle-to-bo
1493        * table and add reference count to tu_bo.
1494        */
1495       result = tu_bo_init_dmabuf(device, &mem->bo,
1496                                  pAllocateInfo->allocationSize, fd_info->fd);
1497       if (result == VK_SUCCESS) {
1498          /* take ownership and close the fd */
1499          close(fd_info->fd);
1500       }
1501    } else {
1502       result =
1503          tu_bo_init_new(device, &mem->bo, pAllocateInfo->allocationSize);
1504    }
1505
1506    if (result != VK_SUCCESS) {
1507       vk_free2(&device->alloc, pAllocator, mem);
1508       return result;
1509    }
1510
1511    mem->size = pAllocateInfo->allocationSize;
1512    mem->type_index = pAllocateInfo->memoryTypeIndex;
1513
1514    mem->map = NULL;
1515    mem->user_ptr = NULL;
1516
1517    *pMem = tu_device_memory_to_handle(mem);
1518
1519    return VK_SUCCESS;
1520 }
1521
1522 VkResult
1523 tu_AllocateMemory(VkDevice _device,
1524                   const VkMemoryAllocateInfo *pAllocateInfo,
1525                   const VkAllocationCallbacks *pAllocator,
1526                   VkDeviceMemory *pMem)
1527 {
1528    TU_FROM_HANDLE(tu_device, device, _device);
1529    return tu_alloc_memory(device, pAllocateInfo, pAllocator, pMem);
1530 }
1531
1532 void
1533 tu_FreeMemory(VkDevice _device,
1534               VkDeviceMemory _mem,
1535               const VkAllocationCallbacks *pAllocator)
1536 {
1537    TU_FROM_HANDLE(tu_device, device, _device);
1538    TU_FROM_HANDLE(tu_device_memory, mem, _mem);
1539
1540    if (mem == NULL)
1541       return;
1542
1543    tu_bo_finish(device, &mem->bo);
1544    vk_free2(&device->alloc, pAllocator, mem);
1545 }
1546
1547 VkResult
1548 tu_MapMemory(VkDevice _device,
1549              VkDeviceMemory _memory,
1550              VkDeviceSize offset,
1551              VkDeviceSize size,
1552              VkMemoryMapFlags flags,
1553              void **ppData)
1554 {
1555    TU_FROM_HANDLE(tu_device, device, _device);
1556    TU_FROM_HANDLE(tu_device_memory, mem, _memory);
1557    VkResult result;
1558
1559    if (mem == NULL) {
1560       *ppData = NULL;
1561       return VK_SUCCESS;
1562    }
1563
1564    if (mem->user_ptr) {
1565       *ppData = mem->user_ptr;
1566    } else if (!mem->map) {
1567       result = tu_bo_map(device, &mem->bo);
1568       if (result != VK_SUCCESS)
1569          return result;
1570       *ppData = mem->map = mem->bo.map;
1571    } else
1572       *ppData = mem->map;
1573
1574    if (*ppData) {
1575       *ppData += offset;
1576       return VK_SUCCESS;
1577    }
1578
1579    return vk_error(device->instance, VK_ERROR_MEMORY_MAP_FAILED);
1580 }
1581
1582 void
1583 tu_UnmapMemory(VkDevice _device, VkDeviceMemory _memory)
1584 {
1585    /* I do not see any unmapping done by the freedreno Gallium driver. */
1586 }
1587
1588 VkResult
1589 tu_FlushMappedMemoryRanges(VkDevice _device,
1590                            uint32_t memoryRangeCount,
1591                            const VkMappedMemoryRange *pMemoryRanges)
1592 {
1593    return VK_SUCCESS;
1594 }
1595
1596 VkResult
1597 tu_InvalidateMappedMemoryRanges(VkDevice _device,
1598                                 uint32_t memoryRangeCount,
1599                                 const VkMappedMemoryRange *pMemoryRanges)
1600 {
1601    return VK_SUCCESS;
1602 }
1603
1604 void
1605 tu_GetBufferMemoryRequirements(VkDevice _device,
1606                                VkBuffer _buffer,
1607                                VkMemoryRequirements *pMemoryRequirements)
1608 {
1609    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
1610
1611    pMemoryRequirements->memoryTypeBits = 1;
1612    pMemoryRequirements->alignment = 16;
1613    pMemoryRequirements->size =
1614       align64(buffer->size, pMemoryRequirements->alignment);
1615 }
1616
1617 void
1618 tu_GetBufferMemoryRequirements2(
1619    VkDevice device,
1620    const VkBufferMemoryRequirementsInfo2 *pInfo,
1621    VkMemoryRequirements2 *pMemoryRequirements)
1622 {
1623    tu_GetBufferMemoryRequirements(device, pInfo->buffer,
1624                                   &pMemoryRequirements->memoryRequirements);
1625 }
1626
1627 void
1628 tu_GetImageMemoryRequirements(VkDevice _device,
1629                               VkImage _image,
1630                               VkMemoryRequirements *pMemoryRequirements)
1631 {
1632    TU_FROM_HANDLE(tu_image, image, _image);
1633
1634    pMemoryRequirements->memoryTypeBits = 1;
1635    pMemoryRequirements->size = image->layout.size;
1636    pMemoryRequirements->alignment = image->alignment;
1637 }
1638
1639 void
1640 tu_GetImageMemoryRequirements2(VkDevice device,
1641                                const VkImageMemoryRequirementsInfo2 *pInfo,
1642                                VkMemoryRequirements2 *pMemoryRequirements)
1643 {
1644    tu_GetImageMemoryRequirements(device, pInfo->image,
1645                                  &pMemoryRequirements->memoryRequirements);
1646 }
1647
1648 void
1649 tu_GetImageSparseMemoryRequirements(
1650    VkDevice device,
1651    VkImage image,
1652    uint32_t *pSparseMemoryRequirementCount,
1653    VkSparseImageMemoryRequirements *pSparseMemoryRequirements)
1654 {
1655    tu_stub();
1656 }
1657
1658 void
1659 tu_GetImageSparseMemoryRequirements2(
1660    VkDevice device,
1661    const VkImageSparseMemoryRequirementsInfo2 *pInfo,
1662    uint32_t *pSparseMemoryRequirementCount,
1663    VkSparseImageMemoryRequirements2 *pSparseMemoryRequirements)
1664 {
1665    tu_stub();
1666 }
1667
1668 void
1669 tu_GetDeviceMemoryCommitment(VkDevice device,
1670                              VkDeviceMemory memory,
1671                              VkDeviceSize *pCommittedMemoryInBytes)
1672 {
1673    *pCommittedMemoryInBytes = 0;
1674 }
1675
1676 VkResult
1677 tu_BindBufferMemory2(VkDevice device,
1678                      uint32_t bindInfoCount,
1679                      const VkBindBufferMemoryInfo *pBindInfos)
1680 {
1681    for (uint32_t i = 0; i < bindInfoCount; ++i) {
1682       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
1683       TU_FROM_HANDLE(tu_buffer, buffer, pBindInfos[i].buffer);
1684
1685       if (mem) {
1686          buffer->bo = &mem->bo;
1687          buffer->bo_offset = pBindInfos[i].memoryOffset;
1688       } else {
1689          buffer->bo = NULL;
1690       }
1691    }
1692    return VK_SUCCESS;
1693 }
1694
1695 VkResult
1696 tu_BindBufferMemory(VkDevice device,
1697                     VkBuffer buffer,
1698                     VkDeviceMemory memory,
1699                     VkDeviceSize memoryOffset)
1700 {
1701    const VkBindBufferMemoryInfo info = {
1702       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
1703       .buffer = buffer,
1704       .memory = memory,
1705       .memoryOffset = memoryOffset
1706    };
1707
1708    return tu_BindBufferMemory2(device, 1, &info);
1709 }
1710
1711 VkResult
1712 tu_BindImageMemory2(VkDevice device,
1713                     uint32_t bindInfoCount,
1714                     const VkBindImageMemoryInfo *pBindInfos)
1715 {
1716    for (uint32_t i = 0; i < bindInfoCount; ++i) {
1717       TU_FROM_HANDLE(tu_image, image, pBindInfos[i].image);
1718       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
1719
1720       if (mem) {
1721          image->bo = &mem->bo;
1722          image->bo_offset = pBindInfos[i].memoryOffset;
1723       } else {
1724          image->bo = NULL;
1725          image->bo_offset = 0;
1726       }
1727    }
1728
1729    return VK_SUCCESS;
1730 }
1731
1732 VkResult
1733 tu_BindImageMemory(VkDevice device,
1734                    VkImage image,
1735                    VkDeviceMemory memory,
1736                    VkDeviceSize memoryOffset)
1737 {
1738    const VkBindImageMemoryInfo info = {
1739       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
1740       .image = image,
1741       .memory = memory,
1742       .memoryOffset = memoryOffset
1743    };
1744
1745    return tu_BindImageMemory2(device, 1, &info);
1746 }
1747
1748 VkResult
1749 tu_QueueBindSparse(VkQueue _queue,
1750                    uint32_t bindInfoCount,
1751                    const VkBindSparseInfo *pBindInfo,
1752                    VkFence _fence)
1753 {
1754    return VK_SUCCESS;
1755 }
1756
1757 // Queue semaphore functions
1758
1759 VkResult
1760 tu_CreateSemaphore(VkDevice _device,
1761                    const VkSemaphoreCreateInfo *pCreateInfo,
1762                    const VkAllocationCallbacks *pAllocator,
1763                    VkSemaphore *pSemaphore)
1764 {
1765    TU_FROM_HANDLE(tu_device, device, _device);
1766
1767    struct tu_semaphore *sem =
1768       vk_alloc2(&device->alloc, pAllocator, sizeof(*sem), 8,
1769                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1770    if (!sem)
1771       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1772
1773    *pSemaphore = tu_semaphore_to_handle(sem);
1774    return VK_SUCCESS;
1775 }
1776
1777 void
1778 tu_DestroySemaphore(VkDevice _device,
1779                     VkSemaphore _semaphore,
1780                     const VkAllocationCallbacks *pAllocator)
1781 {
1782    TU_FROM_HANDLE(tu_device, device, _device);
1783    TU_FROM_HANDLE(tu_semaphore, sem, _semaphore);
1784    if (!_semaphore)
1785       return;
1786
1787    vk_free2(&device->alloc, pAllocator, sem);
1788 }
1789
1790 VkResult
1791 tu_CreateEvent(VkDevice _device,
1792                const VkEventCreateInfo *pCreateInfo,
1793                const VkAllocationCallbacks *pAllocator,
1794                VkEvent *pEvent)
1795 {
1796    TU_FROM_HANDLE(tu_device, device, _device);
1797    struct tu_event *event =
1798       vk_alloc2(&device->alloc, pAllocator, sizeof(*event), 8,
1799                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1800
1801    if (!event)
1802       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1803
1804    VkResult result = tu_bo_init_new(device, &event->bo, 0x1000);
1805    if (result != VK_SUCCESS)
1806       goto fail_alloc;
1807
1808    result = tu_bo_map(device, &event->bo);
1809    if (result != VK_SUCCESS)
1810       goto fail_map;
1811
1812    *pEvent = tu_event_to_handle(event);
1813
1814    return VK_SUCCESS;
1815
1816 fail_map:
1817    tu_bo_finish(device, &event->bo);
1818 fail_alloc:
1819    vk_free2(&device->alloc, pAllocator, event);
1820    return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1821 }
1822
1823 void
1824 tu_DestroyEvent(VkDevice _device,
1825                 VkEvent _event,
1826                 const VkAllocationCallbacks *pAllocator)
1827 {
1828    TU_FROM_HANDLE(tu_device, device, _device);
1829    TU_FROM_HANDLE(tu_event, event, _event);
1830
1831    if (!event)
1832       return;
1833
1834    tu_bo_finish(device, &event->bo);
1835    vk_free2(&device->alloc, pAllocator, event);
1836 }
1837
1838 VkResult
1839 tu_GetEventStatus(VkDevice _device, VkEvent _event)
1840 {
1841    TU_FROM_HANDLE(tu_event, event, _event);
1842
1843    if (*(uint64_t*) event->bo.map == 1)
1844       return VK_EVENT_SET;
1845    return VK_EVENT_RESET;
1846 }
1847
1848 VkResult
1849 tu_SetEvent(VkDevice _device, VkEvent _event)
1850 {
1851    TU_FROM_HANDLE(tu_event, event, _event);
1852    *(uint64_t*) event->bo.map = 1;
1853
1854    return VK_SUCCESS;
1855 }
1856
1857 VkResult
1858 tu_ResetEvent(VkDevice _device, VkEvent _event)
1859 {
1860    TU_FROM_HANDLE(tu_event, event, _event);
1861    *(uint64_t*) event->bo.map = 0;
1862
1863    return VK_SUCCESS;
1864 }
1865
1866 VkResult
1867 tu_CreateBuffer(VkDevice _device,
1868                 const VkBufferCreateInfo *pCreateInfo,
1869                 const VkAllocationCallbacks *pAllocator,
1870                 VkBuffer *pBuffer)
1871 {
1872    TU_FROM_HANDLE(tu_device, device, _device);
1873    struct tu_buffer *buffer;
1874
1875    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1876
1877    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1878                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1879    if (buffer == NULL)
1880       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1881
1882    buffer->size = pCreateInfo->size;
1883    buffer->usage = pCreateInfo->usage;
1884    buffer->flags = pCreateInfo->flags;
1885
1886    *pBuffer = tu_buffer_to_handle(buffer);
1887
1888    return VK_SUCCESS;
1889 }
1890
1891 void
1892 tu_DestroyBuffer(VkDevice _device,
1893                  VkBuffer _buffer,
1894                  const VkAllocationCallbacks *pAllocator)
1895 {
1896    TU_FROM_HANDLE(tu_device, device, _device);
1897    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
1898
1899    if (!buffer)
1900       return;
1901
1902    vk_free2(&device->alloc, pAllocator, buffer);
1903 }
1904
1905 static uint32_t
1906 tu_surface_max_layer_count(struct tu_image_view *iview)
1907 {
1908    return iview->type == VK_IMAGE_VIEW_TYPE_3D
1909              ? iview->extent.depth
1910              : (iview->base_layer + iview->layer_count);
1911 }
1912
1913 VkResult
1914 tu_CreateFramebuffer(VkDevice _device,
1915                      const VkFramebufferCreateInfo *pCreateInfo,
1916                      const VkAllocationCallbacks *pAllocator,
1917                      VkFramebuffer *pFramebuffer)
1918 {
1919    TU_FROM_HANDLE(tu_device, device, _device);
1920    struct tu_framebuffer *framebuffer;
1921
1922    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
1923
1924    size_t size = sizeof(*framebuffer) + sizeof(struct tu_attachment_info) *
1925                                            pCreateInfo->attachmentCount;
1926    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
1927                            VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1928    if (framebuffer == NULL)
1929       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1930
1931    framebuffer->attachment_count = pCreateInfo->attachmentCount;
1932    framebuffer->width = pCreateInfo->width;
1933    framebuffer->height = pCreateInfo->height;
1934    framebuffer->layers = pCreateInfo->layers;
1935    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
1936       VkImageView _iview = pCreateInfo->pAttachments[i];
1937       struct tu_image_view *iview = tu_image_view_from_handle(_iview);
1938       framebuffer->attachments[i].attachment = iview;
1939
1940       framebuffer->width = MIN2(framebuffer->width, iview->extent.width);
1941       framebuffer->height = MIN2(framebuffer->height, iview->extent.height);
1942       framebuffer->layers =
1943          MIN2(framebuffer->layers, tu_surface_max_layer_count(iview));
1944    }
1945
1946    *pFramebuffer = tu_framebuffer_to_handle(framebuffer);
1947    return VK_SUCCESS;
1948 }
1949
1950 void
1951 tu_DestroyFramebuffer(VkDevice _device,
1952                       VkFramebuffer _fb,
1953                       const VkAllocationCallbacks *pAllocator)
1954 {
1955    TU_FROM_HANDLE(tu_device, device, _device);
1956    TU_FROM_HANDLE(tu_framebuffer, fb, _fb);
1957
1958    if (!fb)
1959       return;
1960    vk_free2(&device->alloc, pAllocator, fb);
1961 }
1962
1963 static enum a6xx_tex_clamp
1964 tu6_tex_wrap(VkSamplerAddressMode address_mode, bool *needs_border)
1965 {
1966    switch (address_mode) {
1967    case VK_SAMPLER_ADDRESS_MODE_REPEAT:
1968       return A6XX_TEX_REPEAT;
1969    case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT:
1970       return A6XX_TEX_MIRROR_REPEAT;
1971    case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE:
1972       return A6XX_TEX_CLAMP_TO_EDGE;
1973    case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER:
1974       *needs_border = true;
1975       return A6XX_TEX_CLAMP_TO_BORDER;
1976    case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE:
1977       /* only works for PoT.. need to emulate otherwise! */
1978       return A6XX_TEX_MIRROR_CLAMP;
1979    default:
1980       unreachable("illegal tex wrap mode");
1981       break;
1982    }
1983 }
1984
1985 static enum a6xx_tex_filter
1986 tu6_tex_filter(VkFilter filter, unsigned aniso)
1987 {
1988    switch (filter) {
1989    case VK_FILTER_NEAREST:
1990       return A6XX_TEX_NEAREST;
1991    case VK_FILTER_LINEAR:
1992       return aniso ? A6XX_TEX_ANISO : A6XX_TEX_LINEAR;
1993    case VK_FILTER_CUBIC_IMG:
1994    default:
1995       unreachable("illegal texture filter");
1996       break;
1997    }
1998 }
1999
2000 static inline enum adreno_compare_func
2001 tu6_compare_func(VkCompareOp op)
2002 {
2003    return (enum adreno_compare_func) op;
2004 }
2005
2006 static void
2007 tu_init_sampler(struct tu_device *device,
2008                 struct tu_sampler *sampler,
2009                 const VkSamplerCreateInfo *pCreateInfo)
2010 {
2011    unsigned aniso = pCreateInfo->anisotropyEnable ?
2012       util_last_bit(MIN2((uint32_t)pCreateInfo->maxAnisotropy >> 1, 8)) : 0;
2013    bool miplinear = (pCreateInfo->mipmapMode == VK_SAMPLER_MIPMAP_MODE_LINEAR);
2014    bool needs_border = false;
2015
2016    sampler->state[0] =
2017       COND(miplinear, A6XX_TEX_SAMP_0_MIPFILTER_LINEAR_NEAR) |
2018       A6XX_TEX_SAMP_0_XY_MAG(tu6_tex_filter(pCreateInfo->magFilter, aniso)) |
2019       A6XX_TEX_SAMP_0_XY_MIN(tu6_tex_filter(pCreateInfo->minFilter, aniso)) |
2020       A6XX_TEX_SAMP_0_ANISO(aniso) |
2021       A6XX_TEX_SAMP_0_WRAP_S(tu6_tex_wrap(pCreateInfo->addressModeU, &needs_border)) |
2022       A6XX_TEX_SAMP_0_WRAP_T(tu6_tex_wrap(pCreateInfo->addressModeV, &needs_border)) |
2023       A6XX_TEX_SAMP_0_WRAP_R(tu6_tex_wrap(pCreateInfo->addressModeW, &needs_border)) |
2024       A6XX_TEX_SAMP_0_LOD_BIAS(pCreateInfo->mipLodBias);
2025    sampler->state[1] =
2026       /* COND(!cso->seamless_cube_map, A6XX_TEX_SAMP_1_CUBEMAPSEAMLESSFILTOFF) | */
2027       COND(pCreateInfo->unnormalizedCoordinates, A6XX_TEX_SAMP_1_UNNORM_COORDS) |
2028       A6XX_TEX_SAMP_1_MIN_LOD(pCreateInfo->minLod) |
2029       A6XX_TEX_SAMP_1_MAX_LOD(pCreateInfo->maxLod) |
2030       COND(pCreateInfo->compareEnable,
2031            A6XX_TEX_SAMP_1_COMPARE_FUNC(tu6_compare_func(pCreateInfo->compareOp)));
2032    sampler->state[2] = 0;
2033    sampler->state[3] = 0;
2034
2035    /* TODO:
2036     * A6XX_TEX_SAMP_1_MIPFILTER_LINEAR_FAR disables mipmapping, but vk has no NONE mipfilter?
2037     * border color
2038     */
2039
2040    sampler->needs_border = needs_border;
2041    sampler->border = pCreateInfo->borderColor;
2042 }
2043
2044 VkResult
2045 tu_CreateSampler(VkDevice _device,
2046                  const VkSamplerCreateInfo *pCreateInfo,
2047                  const VkAllocationCallbacks *pAllocator,
2048                  VkSampler *pSampler)
2049 {
2050    TU_FROM_HANDLE(tu_device, device, _device);
2051    struct tu_sampler *sampler;
2052
2053    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO);
2054
2055    sampler = vk_alloc2(&device->alloc, pAllocator, sizeof(*sampler), 8,
2056                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2057    if (!sampler)
2058       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2059
2060    tu_init_sampler(device, sampler, pCreateInfo);
2061    *pSampler = tu_sampler_to_handle(sampler);
2062
2063    return VK_SUCCESS;
2064 }
2065
2066 void
2067 tu_DestroySampler(VkDevice _device,
2068                   VkSampler _sampler,
2069                   const VkAllocationCallbacks *pAllocator)
2070 {
2071    TU_FROM_HANDLE(tu_device, device, _device);
2072    TU_FROM_HANDLE(tu_sampler, sampler, _sampler);
2073
2074    if (!sampler)
2075       return;
2076    vk_free2(&device->alloc, pAllocator, sampler);
2077 }
2078
2079 /* vk_icd.h does not declare this function, so we declare it here to
2080  * suppress Wmissing-prototypes.
2081  */
2082 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2083 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion);
2084
2085 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2086 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion)
2087 {
2088    /* For the full details on loader interface versioning, see
2089     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2090     * What follows is a condensed summary, to help you navigate the large and
2091     * confusing official doc.
2092     *
2093     *   - Loader interface v0 is incompatible with later versions. We don't
2094     *     support it.
2095     *
2096     *   - In loader interface v1:
2097     *       - The first ICD entrypoint called by the loader is
2098     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2099     *         entrypoint.
2100     *       - The ICD must statically expose no other Vulkan symbol unless it
2101     * is linked with -Bsymbolic.
2102     *       - Each dispatchable Vulkan handle created by the ICD must be
2103     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2104     *         ICD must initialize VK_LOADER_DATA.loadMagic to
2105     * ICD_LOADER_MAGIC.
2106     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2107     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2108     *         such loader-managed surfaces.
2109     *
2110     *    - Loader interface v2 differs from v1 in:
2111     *       - The first ICD entrypoint called by the loader is
2112     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2113     *         statically expose this entrypoint.
2114     *
2115     *    - Loader interface v3 differs from v2 in:
2116     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2117     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2118     *          because the loader no longer does so.
2119     */
2120    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2121    return VK_SUCCESS;
2122 }
2123
2124 VkResult
2125 tu_GetMemoryFdKHR(VkDevice _device,
2126                   const VkMemoryGetFdInfoKHR *pGetFdInfo,
2127                   int *pFd)
2128 {
2129    TU_FROM_HANDLE(tu_device, device, _device);
2130    TU_FROM_HANDLE(tu_device_memory, memory, pGetFdInfo->memory);
2131
2132    assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2133
2134    /* At the moment, we support only the below handle types. */
2135    assert(pGetFdInfo->handleType ==
2136              VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2137           pGetFdInfo->handleType ==
2138              VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2139
2140    int prime_fd = tu_bo_export_dmabuf(device, &memory->bo);
2141    if (prime_fd < 0)
2142       return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
2143
2144    *pFd = prime_fd;
2145    return VK_SUCCESS;
2146 }
2147
2148 VkResult
2149 tu_GetMemoryFdPropertiesKHR(VkDevice _device,
2150                             VkExternalMemoryHandleTypeFlagBits handleType,
2151                             int fd,
2152                             VkMemoryFdPropertiesKHR *pMemoryFdProperties)
2153 {
2154    assert(handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2155    pMemoryFdProperties->memoryTypeBits = 1;
2156    return VK_SUCCESS;
2157 }
2158
2159 void
2160 tu_GetPhysicalDeviceExternalSemaphoreProperties(
2161    VkPhysicalDevice physicalDevice,
2162    const VkPhysicalDeviceExternalSemaphoreInfo *pExternalSemaphoreInfo,
2163    VkExternalSemaphoreProperties *pExternalSemaphoreProperties)
2164 {
2165    pExternalSemaphoreProperties->exportFromImportedHandleTypes = 0;
2166    pExternalSemaphoreProperties->compatibleHandleTypes = 0;
2167    pExternalSemaphoreProperties->externalSemaphoreFeatures = 0;
2168 }
2169
2170 void
2171 tu_GetPhysicalDeviceExternalFenceProperties(
2172    VkPhysicalDevice physicalDevice,
2173    const VkPhysicalDeviceExternalFenceInfo *pExternalFenceInfo,
2174    VkExternalFenceProperties *pExternalFenceProperties)
2175 {
2176    pExternalFenceProperties->exportFromImportedHandleTypes = 0;
2177    pExternalFenceProperties->compatibleHandleTypes = 0;
2178    pExternalFenceProperties->externalFenceFeatures = 0;
2179 }
2180
2181 VkResult
2182 tu_CreateDebugReportCallbackEXT(
2183    VkInstance _instance,
2184    const VkDebugReportCallbackCreateInfoEXT *pCreateInfo,
2185    const VkAllocationCallbacks *pAllocator,
2186    VkDebugReportCallbackEXT *pCallback)
2187 {
2188    TU_FROM_HANDLE(tu_instance, instance, _instance);
2189    return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2190                                           pCreateInfo, pAllocator,
2191                                           &instance->alloc, pCallback);
2192 }
2193
2194 void
2195 tu_DestroyDebugReportCallbackEXT(VkInstance _instance,
2196                                  VkDebugReportCallbackEXT _callback,
2197                                  const VkAllocationCallbacks *pAllocator)
2198 {
2199    TU_FROM_HANDLE(tu_instance, instance, _instance);
2200    vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2201                                     _callback, pAllocator, &instance->alloc);
2202 }
2203
2204 void
2205 tu_DebugReportMessageEXT(VkInstance _instance,
2206                          VkDebugReportFlagsEXT flags,
2207                          VkDebugReportObjectTypeEXT objectType,
2208                          uint64_t object,
2209                          size_t location,
2210                          int32_t messageCode,
2211                          const char *pLayerPrefix,
2212                          const char *pMessage)
2213 {
2214    TU_FROM_HANDLE(tu_instance, instance, _instance);
2215    vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2216                    object, location, messageCode, pLayerPrefix, pMessage);
2217 }
2218
2219 void
2220 tu_GetDeviceGroupPeerMemoryFeatures(
2221    VkDevice device,
2222    uint32_t heapIndex,
2223    uint32_t localDeviceIndex,
2224    uint32_t remoteDeviceIndex,
2225    VkPeerMemoryFeatureFlags *pPeerMemoryFeatures)
2226 {
2227    assert(localDeviceIndex == remoteDeviceIndex);
2228
2229    *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2230                           VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2231                           VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2232                           VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2233 }