src/freedreno/vulkan/tu_device.c

   1 /*
   2  * Copyright © 2016 Red Hat.
   3  * Copyright © 2016 Bas Nieuwenhuizen
   4  *
   5  * based in part on anv driver which is:
   6  * Copyright © 2015 Intel Corporation
   7  *
   8  * Permission is hereby granted, free of charge, to any person obtaining a
   9  * copy of this software and associated documentation files (the "Software"),
  10  * to deal in the Software without restriction, including without limitation
  11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  12  * and/or sell copies of the Software, and to permit persons to whom the
  13  * Software is furnished to do so, subject to the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the next
  16  * paragraph) shall be included in all copies or substantial portions of the
  17  * Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  22  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  23  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  24  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  25  * DEALINGS IN THE SOFTWARE.
  26  */
  27
  28 #include "tu_private.h"
  29
  30 #include <fcntl.h>
  31 #include <libsync.h>
  32 #include <stdbool.h>
  33 #include <string.h>
  34 #include <sys/mman.h>
  35 #include <sys/sysinfo.h>
  36 #include <unistd.h>
  37 #include <xf86drm.h>
  38
  39 #include "compiler/glsl_types.h"
  40 #include "util/debug.h"
  41 #include "util/disk_cache.h"
  42 #include "util/u_atomic.h"
  43 #include "vk_format.h"
  44 #include "vk_util.h"
  45
  46 #include "drm-uapi/msm_drm.h"
  47
  48 /* for fd_get_driver/device_uuid() */
  49 #include "freedreno/common/freedreno_uuid.h"
  50
  51 static void
  52 tu_semaphore_remove_temp(struct tu_device *device,
  53                          struct tu_semaphore *sem);
  54
  55 static int
  56 tu_device_get_cache_uuid(uint16_t family, void *uuid)
  57 {
  58    uint32_t mesa_timestamp;
  59    uint16_t f = family;
  60    memset(uuid, 0, VK_UUID_SIZE);
  61    if (!disk_cache_get_function_timestamp(tu_device_get_cache_uuid,
  62                                           &mesa_timestamp))
  63       return -1;
  64
  65    memcpy(uuid, &mesa_timestamp, 4);
  66    memcpy((char *) uuid + 4, &f, 2);
  67    snprintf((char *) uuid + 6, VK_UUID_SIZE - 10, "tu");
  68    return 0;
  69 }
  70
  71 static VkResult
  72 tu_bo_init(struct tu_device *dev,
  73            struct tu_bo *bo,
  74            uint32_t gem_handle,
  75            uint64_t size)
  76 {
  77    uint64_t iova = tu_gem_info_iova(dev, gem_handle);
  78    if (!iova)
  79       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
  80
  81    *bo = (struct tu_bo) {
  82       .gem_handle = gem_handle,
  83       .size = size,
  84       .iova = iova,
  85    };
  86
  87    return VK_SUCCESS;
  88 }
  89
  90 VkResult
  91 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size)
  92 {
  93    /* TODO: Choose better flags. As of 2018-11-12, freedreno/drm/msm_bo.c
  94     * always sets `flags = MSM_BO_WC`, and we copy that behavior here.
  95     */
  96    uint32_t gem_handle = tu_gem_new(dev, size, MSM_BO_WC);
  97    if (!gem_handle)
  98       return vk_error(dev->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
  99
 100    VkResult result = tu_bo_init(dev, bo, gem_handle, size);
 101    if (result != VK_SUCCESS) {
 102       tu_gem_close(dev, gem_handle);
 103       return vk_error(dev->instance, result);
 104    }
 105
 106    return VK_SUCCESS;
 107 }
 108
 109 VkResult
 110 tu_bo_init_dmabuf(struct tu_device *dev,
 111                   struct tu_bo *bo,
 112                   uint64_t size,
 113                   int fd)
 114 {
 115    uint32_t gem_handle = tu_gem_import_dmabuf(dev, fd, size);
 116    if (!gem_handle)
 117       return vk_error(dev->instance, VK_ERROR_INVALID_EXTERNAL_HANDLE);
 118
 119    VkResult result = tu_bo_init(dev, bo, gem_handle, size);
 120    if (result != VK_SUCCESS) {
 121       tu_gem_close(dev, gem_handle);
 122       return vk_error(dev->instance, result);
 123    }
 124
 125    return VK_SUCCESS;
 126 }
 127
 128 int
 129 tu_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo)
 130 {
 131    return tu_gem_export_dmabuf(dev, bo->gem_handle);
 132 }
 133
 134 VkResult
 135 tu_bo_map(struct tu_device *dev, struct tu_bo *bo)
 136 {
 137    if (bo->map)
 138       return VK_SUCCESS;
 139
 140    uint64_t offset = tu_gem_info_offset(dev, bo->gem_handle);
 141    if (!offset)
 142       return vk_error(dev->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
 143
 144    /* TODO: Should we use the wrapper os_mmap() like Freedreno does? */
 145    void *map = mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
 146                     dev->physical_device->local_fd, offset);
 147    if (map == MAP_FAILED)
 148       return vk_error(dev->instance, VK_ERROR_MEMORY_MAP_FAILED);
 149
 150    bo->map = map;
 151    return VK_SUCCESS;
 152 }
 153
 154 void
 155 tu_bo_finish(struct tu_device *dev, struct tu_bo *bo)
 156 {
 157    assert(bo->gem_handle);
 158
 159    if (bo->map)
 160       munmap(bo->map, bo->size);
 161
 162    tu_gem_close(dev, bo->gem_handle);
 163 }
 164
 165 static VkResult
 166 tu_physical_device_init(struct tu_physical_device *device,
 167                         struct tu_instance *instance,
 168                         drmDevicePtr drm_device)
 169 {
 170    const char *path = drm_device->nodes[DRM_NODE_RENDER];
 171    VkResult result = VK_SUCCESS;
 172    drmVersionPtr version;
 173    int fd;
 174    int master_fd = -1;
 175
 176    fd = open(path, O_RDWR | O_CLOEXEC);
 177    if (fd < 0) {
 178       return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 179                        "failed to open device %s", path);
 180    }
 181
 182    /* Version 1.3 added MSM_INFO_IOVA. */
 183    const int min_version_major = 1;
 184    const int min_version_minor = 3;
 185
 186    version = drmGetVersion(fd);
 187    if (!version) {
 188       close(fd);
 189       return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 190                        "failed to query kernel driver version for device %s",
 191                        path);
 192    }
 193
 194    if (strcmp(version->name, "msm")) {
 195       drmFreeVersion(version);
 196       close(fd);
 197       return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 198                        "device %s does not use the msm kernel driver", path);
 199    }
 200
 201    if (version->version_major != min_version_major ||
 202        version->version_minor < min_version_minor) {
 203       result = vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
 204                          "kernel driver for device %s has version %d.%d, "
 205                          "but Vulkan requires version >= %d.%d",
 206                          path, version->version_major, version->version_minor,
 207                          min_version_major, min_version_minor);
 208       drmFreeVersion(version);
 209       close(fd);
 210       return result;
 211    }
 212
 213    device->msm_major_version = version->version_major;
 214    device->msm_minor_version = version->version_minor;
 215
 216    drmFreeVersion(version);
 217
 218    if (instance->debug_flags & TU_DEBUG_STARTUP)
 219       tu_logi("Found compatible device '%s'.", path);
 220
 221    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 222    device->instance = instance;
 223    assert(strlen(path) < ARRAY_SIZE(device->path));
 224    strncpy(device->path, path, ARRAY_SIZE(device->path));
 225
 226    if (instance->enabled_extensions.KHR_display) {
 227       master_fd =
 228          open(drm_device->nodes[DRM_NODE_PRIMARY], O_RDWR | O_CLOEXEC);
 229       if (master_fd >= 0) {
 230          /* TODO: free master_fd is accel is not working? */
 231       }
 232    }
 233
 234    device->master_fd = master_fd;
 235    device->local_fd = fd;
 236
 237    if (tu_drm_get_gpu_id(device, &device->gpu_id)) {
 238       if (instance->debug_flags & TU_DEBUG_STARTUP)
 239          tu_logi("Could not query the GPU ID");
 240       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 241                          "could not get GPU ID");
 242       goto fail;
 243    }
 244
 245    if (tu_drm_get_gmem_size(device, &device->gmem_size)) {
 246       if (instance->debug_flags & TU_DEBUG_STARTUP)
 247          tu_logi("Could not query the GMEM size");
 248       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 249                          "could not get GMEM size");
 250       goto fail;
 251    }
 252
 253    if (tu_drm_get_gmem_base(device, &device->gmem_base)) {
 254       if (instance->debug_flags & TU_DEBUG_STARTUP)
 255          tu_logi("Could not query the GMEM size");
 256       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 257                          "could not get GMEM size");
 258       goto fail;
 259    }
 260
 261    memset(device->name, 0, sizeof(device->name));
 262    sprintf(device->name, "FD%d", device->gpu_id);
 263
 264    switch (device->gpu_id) {
 265    case 618:
 266       device->ccu_offset_gmem = 0x7c000; /* 0x7e000 in some cases? */
 267       device->ccu_offset_bypass = 0x10000;
 268       device->tile_align_w = 64;
 269       device->magic.PC_UNKNOWN_9805 = 0x0;
 270       device->magic.SP_UNKNOWN_A0F8 = 0x0;
 271       break;
 272    case 630:
 273    case 640:
 274       device->ccu_offset_gmem = 0xf8000;
 275       device->ccu_offset_bypass = 0x20000;
 276       device->tile_align_w = 64;
 277       device->magic.PC_UNKNOWN_9805 = 0x1;
 278       device->magic.SP_UNKNOWN_A0F8 = 0x1;
 279       break;
 280    case 650:
 281       device->ccu_offset_gmem = 0x114000;
 282       device->ccu_offset_bypass = 0x30000;
 283       device->tile_align_w = 96;
 284       device->magic.PC_UNKNOWN_9805 = 0x2;
 285       device->magic.SP_UNKNOWN_A0F8 = 0x2;
 286       break;
 287    default:
 288       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 289                          "device %s is unsupported", device->name);
 290       goto fail;
 291    }
 292    if (tu_device_get_cache_uuid(device->gpu_id, device->cache_uuid)) {
 293       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
 294                          "cannot generate UUID");
 295       goto fail;
 296    }
 297
 298    /* The gpu id is already embedded in the uuid so we just pass "tu"
 299     * when creating the cache.
 300     */
 301    char buf[VK_UUID_SIZE * 2 + 1];
 302    disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
 303    device->disk_cache = disk_cache_create(device->name, buf, 0);
 304
 305    fprintf(stderr, "WARNING: tu is not a conformant vulkan implementation, "
 306                    "testing use only.\n");
 307
 308    fd_get_driver_uuid(device->driver_uuid);
 309    fd_get_device_uuid(device->device_uuid, device->gpu_id);
 310
 311    tu_physical_device_get_supported_extensions(device, &device->supported_extensions);
 312
 313    if (result != VK_SUCCESS) {
 314       vk_error(instance, result);
 315       goto fail;
 316    }
 317
 318    result = tu_wsi_init(device);
 319    if (result != VK_SUCCESS) {
 320       vk_error(instance, result);
 321       goto fail;
 322    }
 323
 324    return VK_SUCCESS;
 325
 326 fail:
 327    close(fd);
 328    if (master_fd != -1)
 329       close(master_fd);
 330    return result;
 331 }
 332
 333 static void
 334 tu_physical_device_finish(struct tu_physical_device *device)
 335 {
 336    tu_wsi_finish(device);
 337
 338    disk_cache_destroy(device->disk_cache);
 339    close(device->local_fd);
 340    if (device->master_fd != -1)
 341       close(device->master_fd);
 342 }
 343
 344 static VKAPI_ATTR void *
 345 default_alloc_func(void *pUserData,
 346                    size_t size,
 347                    size_t align,
 348                    VkSystemAllocationScope allocationScope)
 349 {
 350    return malloc(size);
 351 }
 352
 353 static VKAPI_ATTR void *
 354 default_realloc_func(void *pUserData,
 355                      void *pOriginal,
 356                      size_t size,
 357                      size_t align,
 358                      VkSystemAllocationScope allocationScope)
 359 {
 360    return realloc(pOriginal, size);
 361 }
 362
 363 static VKAPI_ATTR void
 364 default_free_func(void *pUserData, void *pMemory)
 365 {
 366    free(pMemory);
 367 }
 368
 369 static const VkAllocationCallbacks default_alloc = {
 370    .pUserData = NULL,
 371    .pfnAllocation = default_alloc_func,
 372    .pfnReallocation = default_realloc_func,
 373    .pfnFree = default_free_func,
 374 };
 375
 376 static const struct debug_control tu_debug_options[] = {
 377    { "startup", TU_DEBUG_STARTUP },
 378    { "nir", TU_DEBUG_NIR },
 379    { "ir3", TU_DEBUG_IR3 },
 380    { "nobin", TU_DEBUG_NOBIN },
 381    { "sysmem", TU_DEBUG_SYSMEM },
 382    { "forcebin", TU_DEBUG_FORCEBIN },
 383    { "noubwc", TU_DEBUG_NOUBWC },
 384    { NULL, 0 }
 385 };
 386
 387 const char *
 388 tu_get_debug_option_name(int id)
 389 {
 390    assert(id < ARRAY_SIZE(tu_debug_options) - 1);
 391    return tu_debug_options[id].string;
 392 }
 393
 394 static int
 395 tu_get_instance_extension_index(const char *name)
 396 {
 397    for (unsigned i = 0; i < TU_INSTANCE_EXTENSION_COUNT; ++i) {
 398       if (strcmp(name, tu_instance_extensions[i].extensionName) == 0)
 399          return i;
 400    }
 401    return -1;
 402 }
 403
 404 VkResult
 405 tu_CreateInstance(const VkInstanceCreateInfo *pCreateInfo,
 406                   const VkAllocationCallbacks *pAllocator,
 407                   VkInstance *pInstance)
 408 {
 409    struct tu_instance *instance;
 410    VkResult result;
 411
 412    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 413
 414    uint32_t client_version;
 415    if (pCreateInfo->pApplicationInfo &&
 416        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 417       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 418    } else {
 419       tu_EnumerateInstanceVersion(&client_version);
 420    }
 421
 422    instance = vk_zalloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 423                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 424    if (!instance)
 425       return vk_error(NULL, VK_ERROR_OUT_OF_HOST_MEMORY);
 426
 427    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 428
 429    if (pAllocator)
 430       instance->alloc = *pAllocator;
 431    else
 432       instance->alloc = default_alloc;
 433
 434    instance->api_version = client_version;
 435    instance->physical_device_count = -1;
 436
 437    instance->debug_flags =
 438       parse_debug_string(getenv("TU_DEBUG"), tu_debug_options);
 439
 440    if (instance->debug_flags & TU_DEBUG_STARTUP)
 441       tu_logi("Created an instance");
 442
 443    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 444       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
 445       int index = tu_get_instance_extension_index(ext_name);
 446
 447       if (index < 0 || !tu_instance_extensions_supported.extensions[index]) {
 448          vk_free2(&default_alloc, pAllocator, instance);
 449          return vk_error(instance, VK_ERROR_EXTENSION_NOT_PRESENT);
 450       }
 451
 452       instance->enabled_extensions.extensions[index] = true;
 453    }
 454
 455    result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
 456    if (result != VK_SUCCESS) {
 457       vk_free2(&default_alloc, pAllocator, instance);
 458       return vk_error(instance, result);
 459    }
 460
 461    glsl_type_singleton_init_or_ref();
 462
 463    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 464
 465    *pInstance = tu_instance_to_handle(instance);
 466
 467    return VK_SUCCESS;
 468 }
 469
 470 void
 471 tu_DestroyInstance(VkInstance _instance,
 472                    const VkAllocationCallbacks *pAllocator)
 473 {
 474    TU_FROM_HANDLE(tu_instance, instance, _instance);
 475
 476    if (!instance)
 477       return;
 478
 479    for (int i = 0; i < instance->physical_device_count; ++i) {
 480       tu_physical_device_finish(instance->physical_devices + i);
 481    }
 482
 483    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 484
 485    glsl_type_singleton_decref();
 486
 487    vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
 488
 489    vk_free(&instance->alloc, instance);
 490 }
 491
 492 static VkResult
 493 tu_enumerate_devices(struct tu_instance *instance)
 494 {
 495    /* TODO: Check for more devices ? */
 496    drmDevicePtr devices[8];
 497    VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
 498    int max_devices;
 499
 500    instance->physical_device_count = 0;
 501
 502    max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
 503
 504    if (instance->debug_flags & TU_DEBUG_STARTUP) {
 505       if (max_devices < 0)
 506          tu_logi("drmGetDevices2 returned error: %s\n", strerror(max_devices));
 507       else
 508          tu_logi("Found %d drm nodes", max_devices);
 509    }
 510
 511    if (max_devices < 1)
 512       return vk_error(instance, VK_ERROR_INCOMPATIBLE_DRIVER);
 513
 514    for (unsigned i = 0; i < (unsigned) max_devices; i++) {
 515       if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
 516           devices[i]->bustype == DRM_BUS_PLATFORM) {
 517
 518          result = tu_physical_device_init(
 519             instance->physical_devices + instance->physical_device_count,
 520             instance, devices[i]);
 521          if (result == VK_SUCCESS)
 522             ++instance->physical_device_count;
 523          else if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 524             break;
 525       }
 526    }
 527    drmFreeDevices(devices, max_devices);
 528
 529    return result;
 530 }
 531
 532 VkResult
 533 tu_EnumeratePhysicalDevices(VkInstance _instance,
 534                             uint32_t *pPhysicalDeviceCount,
 535                             VkPhysicalDevice *pPhysicalDevices)
 536 {
 537    TU_FROM_HANDLE(tu_instance, instance, _instance);
 538    VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
 539
 540    VkResult result;
 541
 542    if (instance->physical_device_count < 0) {
 543       result = tu_enumerate_devices(instance);
 544       if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
 545          return result;
 546    }
 547
 548    for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
 549       vk_outarray_append(&out, p)
 550       {
 551          *p = tu_physical_device_to_handle(instance->physical_devices + i);
 552       }
 553    }
 554
 555    return vk_outarray_status(&out);
 556 }
 557
 558 VkResult
 559 tu_EnumeratePhysicalDeviceGroups(
 560    VkInstance _instance,
 561    uint32_t *pPhysicalDeviceGroupCount,
 562    VkPhysicalDeviceGroupProperties *pPhysicalDeviceGroupProperties)
 563 {
 564    TU_FROM_HANDLE(tu_instance, instance, _instance);
 565    VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
 566                     pPhysicalDeviceGroupCount);
 567    VkResult result;
 568
 569    if (instance->physical_device_count < 0) {
 570       result = tu_enumerate_devices(instance);
 571       if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
 572          return result;
 573    }
 574
 575    for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
 576       vk_outarray_append(&out, p)
 577       {
 578          p->physicalDeviceCount = 1;
 579          p->physicalDevices[0] =
 580             tu_physical_device_to_handle(instance->physical_devices + i);
 581          p->subsetAllocation = false;
 582       }
 583    }
 584
 585    return vk_outarray_status(&out);
 586 }
 587
 588 void
 589 tu_GetPhysicalDeviceFeatures(VkPhysicalDevice physicalDevice,
 590                              VkPhysicalDeviceFeatures *pFeatures)
 591 {
 592    memset(pFeatures, 0, sizeof(*pFeatures));
 593
 594    *pFeatures = (VkPhysicalDeviceFeatures) {
 595       .robustBufferAccess = true,
 596       .fullDrawIndexUint32 = true,
 597       .imageCubeArray = true,
 598       .independentBlend = true,
 599       .geometryShader = true,
 600       .tessellationShader = true,
 601       .sampleRateShading = true,
 602       .dualSrcBlend = true,
 603       .logicOp = true,
 604       .multiDrawIndirect = true,
 605       .drawIndirectFirstInstance = true,
 606       .depthClamp = true,
 607       .depthBiasClamp = false,
 608       .fillModeNonSolid = true,
 609       .depthBounds = true,
 610       .wideLines = false,
 611       .largePoints = true,
 612       .alphaToOne = true,
 613       .multiViewport = false,
 614       .samplerAnisotropy = true,
 615       .textureCompressionETC2 = true,
 616       .textureCompressionASTC_LDR = true,
 617       .textureCompressionBC = true,
 618       .occlusionQueryPrecise = true,
 619       .pipelineStatisticsQuery = false,
 620       .vertexPipelineStoresAndAtomics = false,
 621       .fragmentStoresAndAtomics = false,
 622       .shaderTessellationAndGeometryPointSize = false,
 623       .shaderImageGatherExtended = false,
 624       .shaderStorageImageExtendedFormats = false,
 625       .shaderStorageImageMultisample = false,
 626       .shaderUniformBufferArrayDynamicIndexing = false,
 627       .shaderSampledImageArrayDynamicIndexing = false,
 628       .shaderStorageBufferArrayDynamicIndexing = false,
 629       .shaderStorageImageArrayDynamicIndexing = false,
 630       .shaderStorageImageReadWithoutFormat = false,
 631       .shaderStorageImageWriteWithoutFormat = false,
 632       .shaderClipDistance = false,
 633       .shaderCullDistance = false,
 634       .shaderFloat64 = false,
 635       .shaderInt64 = false,
 636       .shaderInt16 = false,
 637       .sparseBinding = false,
 638       .variableMultisampleRate = false,
 639       .inheritedQueries = false,
 640    };
 641 }
 642
 643 void
 644 tu_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,
 645                               VkPhysicalDeviceFeatures2 *pFeatures)
 646 {
 647    vk_foreach_struct(ext, pFeatures->pNext)
 648    {
 649       switch (ext->sType) {
 650       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES: {
 651          VkPhysicalDeviceVulkan11Features *features = (void *) ext;
 652          features->storageBuffer16BitAccess            = false;
 653          features->uniformAndStorageBuffer16BitAccess  = false;
 654          features->storagePushConstant16               = false;
 655          features->storageInputOutput16                = false;
 656          features->multiview                           = false;
 657          features->multiviewGeometryShader             = false;
 658          features->multiviewTessellationShader         = false;
 659          features->variablePointersStorageBuffer       = false;
 660          features->variablePointers                    = false;
 661          features->protectedMemory                     = false;
 662          features->samplerYcbcrConversion              = true;
 663          features->shaderDrawParameters                = true;
 664          break;
 665       }
 666       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
 667          VkPhysicalDeviceVariablePointersFeatures *features = (void *) ext;
 668          features->variablePointersStorageBuffer = false;
 669          features->variablePointers = false;
 670          break;
 671       }
 672       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
 673          VkPhysicalDeviceMultiviewFeatures *features =
 674             (VkPhysicalDeviceMultiviewFeatures *) ext;
 675          features->multiview = false;
 676          features->multiviewGeometryShader = false;
 677          features->multiviewTessellationShader = false;
 678          break;
 679       }
 680       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
 681          VkPhysicalDeviceShaderDrawParametersFeatures *features =
 682             (VkPhysicalDeviceShaderDrawParametersFeatures *) ext;
 683          features->shaderDrawParameters = true;
 684          break;
 685       }
 686       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
 687          VkPhysicalDeviceProtectedMemoryFeatures *features =
 688             (VkPhysicalDeviceProtectedMemoryFeatures *) ext;
 689          features->protectedMemory = false;
 690          break;
 691       }
 692       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
 693          VkPhysicalDevice16BitStorageFeatures *features =
 694             (VkPhysicalDevice16BitStorageFeatures *) ext;
 695          features->storageBuffer16BitAccess = false;
 696          features->uniformAndStorageBuffer16BitAccess = false;
 697          features->storagePushConstant16 = false;
 698          features->storageInputOutput16 = false;
 699          break;
 700       }
 701       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
 702          VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
 703             (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
 704          features->samplerYcbcrConversion = true;
 705          break;
 706       }
 707       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
 708          VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
 709             (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *) ext;
 710          features->shaderInputAttachmentArrayDynamicIndexing = false;
 711          features->shaderUniformTexelBufferArrayDynamicIndexing = false;
 712          features->shaderStorageTexelBufferArrayDynamicIndexing = false;
 713          features->shaderUniformBufferArrayNonUniformIndexing = false;
 714          features->shaderSampledImageArrayNonUniformIndexing = false;
 715          features->shaderStorageBufferArrayNonUniformIndexing = false;
 716          features->shaderStorageImageArrayNonUniformIndexing = false;
 717          features->shaderInputAttachmentArrayNonUniformIndexing = false;
 718          features->shaderUniformTexelBufferArrayNonUniformIndexing = false;
 719          features->shaderStorageTexelBufferArrayNonUniformIndexing = false;
 720          features->descriptorBindingUniformBufferUpdateAfterBind = false;
 721          features->descriptorBindingSampledImageUpdateAfterBind = false;
 722          features->descriptorBindingStorageImageUpdateAfterBind = false;
 723          features->descriptorBindingStorageBufferUpdateAfterBind = false;
 724          features->descriptorBindingUniformTexelBufferUpdateAfterBind = false;
 725          features->descriptorBindingStorageTexelBufferUpdateAfterBind = false;
 726          features->descriptorBindingUpdateUnusedWhilePending = false;
 727          features->descriptorBindingPartiallyBound = false;
 728          features->descriptorBindingVariableDescriptorCount = false;
 729          features->runtimeDescriptorArray = false;
 730          break;
 731       }
 732       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
 733          VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
 734             (VkPhysicalDeviceConditionalRenderingFeaturesEXT *) ext;
 735          features->conditionalRendering = false;
 736          features->inheritedConditionalRendering = false;
 737          break;
 738       }
 739       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
 740          VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
 741             (VkPhysicalDeviceTransformFeedbackFeaturesEXT *) ext;
 742          features->transformFeedback = true;
 743          features->geometryStreams = false;
 744          break;
 745       }
 746       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
 747          VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
 748             (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
 749          features->indexTypeUint8 = true;
 750          break;
 751       }
 752       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
 753          VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
 754             (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
 755          features->vertexAttributeInstanceRateDivisor = true;
 756          features->vertexAttributeInstanceRateZeroDivisor = true;
 757          break;
 758       }
 759       default:
 760          break;
 761       }
 762    }
 763    return tu_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 764 }
 765
 766 void
 767 tu_GetPhysicalDeviceProperties(VkPhysicalDevice physicalDevice,
 768                                VkPhysicalDeviceProperties *pProperties)
 769 {
 770    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 771    VkSampleCountFlags sample_counts =
 772       VK_SAMPLE_COUNT_1_BIT | VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT;
 773
 774    /* I have no idea what the maximum size is, but the hardware supports very
 775     * large numbers of descriptors (at least 2^16). This limit is based on
 776     * CP_LOAD_STATE6, which has a 28-bit field for the DWORD offset, so that
 777     * we don't have to think about what to do if that overflows, but really
 778     * nothing is likely to get close to this.
 779     */
 780    const size_t max_descriptor_set_size = (1 << 28) / A6XX_TEX_CONST_DWORDS;
 781
 782    VkPhysicalDeviceLimits limits = {
 783       .maxImageDimension1D = (1 << 14),
 784       .maxImageDimension2D = (1 << 14),
 785       .maxImageDimension3D = (1 << 11),
 786       .maxImageDimensionCube = (1 << 14),
 787       .maxImageArrayLayers = (1 << 11),
 788       .maxTexelBufferElements = 128 * 1024 * 1024,
 789       .maxUniformBufferRange = MAX_UNIFORM_BUFFER_RANGE,
 790       .maxStorageBufferRange = MAX_STORAGE_BUFFER_RANGE,
 791       .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
 792       .maxMemoryAllocationCount = UINT32_MAX,
 793       .maxSamplerAllocationCount = 64 * 1024,
 794       .bufferImageGranularity = 64,          /* A cache line */
 795       .sparseAddressSpaceSize = 0xffffffffu, /* buffer max size */
 796       .maxBoundDescriptorSets = MAX_SETS,
 797       .maxPerStageDescriptorSamplers = max_descriptor_set_size,
 798       .maxPerStageDescriptorUniformBuffers = max_descriptor_set_size,
 799       .maxPerStageDescriptorStorageBuffers = max_descriptor_set_size,
 800       .maxPerStageDescriptorSampledImages = max_descriptor_set_size,
 801       .maxPerStageDescriptorStorageImages = max_descriptor_set_size,
 802       .maxPerStageDescriptorInputAttachments = MAX_RTS,
 803       .maxPerStageResources = max_descriptor_set_size,
 804       .maxDescriptorSetSamplers = max_descriptor_set_size,
 805       .maxDescriptorSetUniformBuffers = max_descriptor_set_size,
 806       .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS,
 807       .maxDescriptorSetStorageBuffers = max_descriptor_set_size,
 808       .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS,
 809       .maxDescriptorSetSampledImages = max_descriptor_set_size,
 810       .maxDescriptorSetStorageImages = max_descriptor_set_size,
 811       .maxDescriptorSetInputAttachments = MAX_RTS,
 812       .maxVertexInputAttributes = 32,
 813       .maxVertexInputBindings = 32,
 814       .maxVertexInputAttributeOffset = 4095,
 815       .maxVertexInputBindingStride = 2048,
 816       .maxVertexOutputComponents = 128,
 817       .maxTessellationGenerationLevel = 64,
 818       .maxTessellationPatchSize = 32,
 819       .maxTessellationControlPerVertexInputComponents = 128,
 820       .maxTessellationControlPerVertexOutputComponents = 128,
 821       .maxTessellationControlPerPatchOutputComponents = 120,
 822       .maxTessellationControlTotalOutputComponents = 4096,
 823       .maxTessellationEvaluationInputComponents = 128,
 824       .maxTessellationEvaluationOutputComponents = 128,
 825       .maxGeometryShaderInvocations = 32,
 826       .maxGeometryInputComponents = 64,
 827       .maxGeometryOutputComponents = 128,
 828       .maxGeometryOutputVertices = 256,
 829       .maxGeometryTotalOutputComponents = 1024,
 830       .maxFragmentInputComponents = 124,
 831       .maxFragmentOutputAttachments = 8,
 832       .maxFragmentDualSrcAttachments = 1,
 833       .maxFragmentCombinedOutputResources = 8,
 834       .maxComputeSharedMemorySize = 32768,
 835       .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
 836       .maxComputeWorkGroupInvocations = 2048,
 837       .maxComputeWorkGroupSize = { 2048, 2048, 2048 },
 838       .subPixelPrecisionBits = 8,
 839       .subTexelPrecisionBits = 8,
 840       .mipmapPrecisionBits = 8,
 841       .maxDrawIndexedIndexValue = UINT32_MAX,
 842       .maxDrawIndirectCount = UINT32_MAX,
 843       .maxSamplerLodBias = 4095.0 / 256.0, /* [-16, 15.99609375] */
 844       .maxSamplerAnisotropy = 16,
 845       .maxViewports = MAX_VIEWPORTS,
 846       .maxViewportDimensions = { (1 << 14), (1 << 14) },
 847       .viewportBoundsRange = { INT16_MIN, INT16_MAX },
 848       .viewportSubPixelBits = 8,
 849       .minMemoryMapAlignment = 4096, /* A page */
 850       .minTexelBufferOffsetAlignment = 64,
 851       .minUniformBufferOffsetAlignment = 64,
 852       .minStorageBufferOffsetAlignment = 64,
 853       .minTexelOffset = -16,
 854       .maxTexelOffset = 15,
 855       .minTexelGatherOffset = -32,
 856       .maxTexelGatherOffset = 31,
 857       .minInterpolationOffset = -0.5,
 858       .maxInterpolationOffset = 0.4375,
 859       .subPixelInterpolationOffsetBits = 4,
 860       .maxFramebufferWidth = (1 << 14),
 861       .maxFramebufferHeight = (1 << 14),
 862       .maxFramebufferLayers = (1 << 10),
 863       .framebufferColorSampleCounts = sample_counts,
 864       .framebufferDepthSampleCounts = sample_counts,
 865       .framebufferStencilSampleCounts = sample_counts,
 866       .framebufferNoAttachmentsSampleCounts = sample_counts,
 867       .maxColorAttachments = MAX_RTS,
 868       .sampledImageColorSampleCounts = sample_counts,
 869       .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 870       .sampledImageDepthSampleCounts = sample_counts,
 871       .sampledImageStencilSampleCounts = sample_counts,
 872       .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
 873       .maxSampleMaskWords = 1,
 874       .timestampComputeAndGraphics = true,
 875       .timestampPeriod = 1000000000.0 / 19200000.0, /* CP_ALWAYS_ON_COUNTER is fixed 19.2MHz */
 876       .maxClipDistances = 8,
 877       .maxCullDistances = 8,
 878       .maxCombinedClipAndCullDistances = 8,
 879       .discreteQueuePriorities = 1,
 880       .pointSizeRange = { 1, 4092 },
 881       .lineWidthRange = { 0.0, 7.9921875 },
 882       .pointSizeGranularity =   0.0625,
 883       .lineWidthGranularity = (1.0 / 128.0),
 884       .strictLines = false, /* FINISHME */
 885       .standardSampleLocations = true,
 886       .optimalBufferCopyOffsetAlignment = 128,
 887       .optimalBufferCopyRowPitchAlignment = 128,
 888       .nonCoherentAtomSize = 64,
 889    };
 890
 891    *pProperties = (VkPhysicalDeviceProperties) {
 892       .apiVersion = tu_physical_device_api_version(pdevice),
 893       .driverVersion = vk_get_driver_version(),
 894       .vendorID = 0, /* TODO */
 895       .deviceID = 0,
 896       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 897       .limits = limits,
 898       .sparseProperties = { 0 },
 899    };
 900
 901    strcpy(pProperties->deviceName, pdevice->name);
 902    memcpy(pProperties->pipelineCacheUUID, pdevice->cache_uuid, VK_UUID_SIZE);
 903 }
 904
 905 void
 906 tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
 907                                 VkPhysicalDeviceProperties2 *pProperties)
 908 {
 909    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
 910    tu_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 911
 912    vk_foreach_struct(ext, pProperties->pNext)
 913    {
 914       switch (ext->sType) {
 915       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 916          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 917             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 918          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 919          break;
 920       }
 921       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
 922          VkPhysicalDeviceIDProperties *properties =
 923             (VkPhysicalDeviceIDProperties *) ext;
 924          memcpy(properties->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
 925          memcpy(properties->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
 926          properties->deviceLUIDValid = false;
 927          break;
 928       }
 929       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
 930          VkPhysicalDeviceMultiviewProperties *properties =
 931             (VkPhysicalDeviceMultiviewProperties *) ext;
 932          properties->maxMultiviewViewCount = MAX_VIEWS;
 933          properties->maxMultiviewInstanceIndex = INT_MAX;
 934          break;
 935       }
 936       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
 937          VkPhysicalDevicePointClippingProperties *properties =
 938             (VkPhysicalDevicePointClippingProperties *) ext;
 939          properties->pointClippingBehavior =
 940             VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
 941          break;
 942       }
 943       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
 944          VkPhysicalDeviceMaintenance3Properties *properties =
 945             (VkPhysicalDeviceMaintenance3Properties *) ext;
 946          /* Make sure everything is addressable by a signed 32-bit int, and
 947           * our largest descriptors are 96 bytes. */
 948          properties->maxPerSetDescriptors = (1ull << 31) / 96;
 949          /* Our buffer size fields allow only this much */
 950          properties->maxMemoryAllocationSize = 0xFFFFFFFFull;
 951          break;
 952       }
 953       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
 954          VkPhysicalDeviceTransformFeedbackPropertiesEXT *properties =
 955             (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
 956
 957          properties->maxTransformFeedbackStreams = IR3_MAX_SO_STREAMS;
 958          properties->maxTransformFeedbackBuffers = IR3_MAX_SO_BUFFERS;
 959          properties->maxTransformFeedbackBufferSize = UINT32_MAX;
 960          properties->maxTransformFeedbackStreamDataSize = 512;
 961          properties->maxTransformFeedbackBufferDataSize = 512;
 962          properties->maxTransformFeedbackBufferDataStride = 512;
 963          properties->transformFeedbackQueries = true;
 964          properties->transformFeedbackStreamsLinesTriangles = false;
 965          properties->transformFeedbackRasterizationStreamSelect = false;
 966          properties->transformFeedbackDraw = true;
 967          break;
 968       }
 969       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLE_LOCATIONS_PROPERTIES_EXT: {
 970          VkPhysicalDeviceSampleLocationsPropertiesEXT *properties =
 971             (VkPhysicalDeviceSampleLocationsPropertiesEXT *)ext;
 972          properties->sampleLocationSampleCounts = 0;
 973          if (pdevice->supported_extensions.EXT_sample_locations) {
 974             properties->sampleLocationSampleCounts =
 975                VK_SAMPLE_COUNT_1_BIT | VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT;
 976          }
 977          properties->maxSampleLocationGridSize = (VkExtent2D) { 1 , 1 };
 978          properties->sampleLocationCoordinateRange[0] = 0.0f;
 979          properties->sampleLocationCoordinateRange[1] = 0.9375f;
 980          properties->sampleLocationSubPixelBits = 4;
 981          properties->variableSampleLocations = true;
 982          break;
 983       }
 984       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES: {
 985          VkPhysicalDeviceSamplerFilterMinmaxProperties *properties =
 986             (VkPhysicalDeviceSamplerFilterMinmaxProperties *)ext;
 987          properties->filterMinmaxImageComponentMapping = true;
 988          properties->filterMinmaxSingleComponentFormats = true;
 989          break;
 990       }
 991       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
 992          VkPhysicalDeviceSubgroupProperties *properties =
 993             (VkPhysicalDeviceSubgroupProperties *)ext;
 994          properties->subgroupSize = 64;
 995          properties->supportedStages = VK_SHADER_STAGE_COMPUTE_BIT;
 996          properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
 997                                            VK_SUBGROUP_FEATURE_VOTE_BIT;
 998          properties->quadOperationsInAllStages = false;
 999          break;
1000       }
1001       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1002          VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1003             (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1004          props->maxVertexAttribDivisor = UINT32_MAX;
1005          break;
1006       }
1007       default:
1008          break;
1009       }
1010    }
1011 }
1012
1013 static const VkQueueFamilyProperties tu_queue_family_properties = {
1014    .queueFlags =
1015       VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT,
1016    .queueCount = 1,
1017    .timestampValidBits = 48,
1018    .minImageTransferGranularity = { 1, 1, 1 },
1019 };
1020
1021 void
1022 tu_GetPhysicalDeviceQueueFamilyProperties(
1023    VkPhysicalDevice physicalDevice,
1024    uint32_t *pQueueFamilyPropertyCount,
1025    VkQueueFamilyProperties *pQueueFamilyProperties)
1026 {
1027    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1028
1029    vk_outarray_append(&out, p) { *p = tu_queue_family_properties; }
1030 }
1031
1032 void
1033 tu_GetPhysicalDeviceQueueFamilyProperties2(
1034    VkPhysicalDevice physicalDevice,
1035    uint32_t *pQueueFamilyPropertyCount,
1036    VkQueueFamilyProperties2 *pQueueFamilyProperties)
1037 {
1038    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1039
1040    vk_outarray_append(&out, p)
1041    {
1042       p->queueFamilyProperties = tu_queue_family_properties;
1043    }
1044 }
1045
1046 static uint64_t
1047 tu_get_system_heap_size()
1048 {
1049    struct sysinfo info;
1050    sysinfo(&info);
1051
1052    uint64_t total_ram = (uint64_t) info.totalram * (uint64_t) info.mem_unit;
1053
1054    /* We don't want to burn too much ram with the GPU.  If the user has 4GiB
1055     * or less, we use at most half.  If they have more than 4GiB, we use 3/4.
1056     */
1057    uint64_t available_ram;
1058    if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
1059       available_ram = total_ram / 2;
1060    else
1061       available_ram = total_ram * 3 / 4;
1062
1063    return available_ram;
1064 }
1065
1066 void
1067 tu_GetPhysicalDeviceMemoryProperties(
1068    VkPhysicalDevice physicalDevice,
1069    VkPhysicalDeviceMemoryProperties *pMemoryProperties)
1070 {
1071    pMemoryProperties->memoryHeapCount = 1;
1072    pMemoryProperties->memoryHeaps[0].size = tu_get_system_heap_size();
1073    pMemoryProperties->memoryHeaps[0].flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT;
1074
1075    pMemoryProperties->memoryTypeCount = 1;
1076    pMemoryProperties->memoryTypes[0].propertyFlags =
1077       VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
1078       VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
1079       VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
1080    pMemoryProperties->memoryTypes[0].heapIndex = 0;
1081 }
1082
1083 void
1084 tu_GetPhysicalDeviceMemoryProperties2(
1085    VkPhysicalDevice physicalDevice,
1086    VkPhysicalDeviceMemoryProperties2 *pMemoryProperties)
1087 {
1088    return tu_GetPhysicalDeviceMemoryProperties(
1089       physicalDevice, &pMemoryProperties->memoryProperties);
1090 }
1091
1092 static VkResult
1093 tu_queue_init(struct tu_device *device,
1094               struct tu_queue *queue,
1095               uint32_t queue_family_index,
1096               int idx,
1097               VkDeviceQueueCreateFlags flags)
1098 {
1099    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1100    queue->device = device;
1101    queue->queue_family_index = queue_family_index;
1102    queue->queue_idx = idx;
1103    queue->flags = flags;
1104
1105    int ret = tu_drm_submitqueue_new(device, 0, &queue->msm_queue_id);
1106    if (ret)
1107       return VK_ERROR_INITIALIZATION_FAILED;
1108
1109    tu_fence_init(&queue->submit_fence, false);
1110
1111    return VK_SUCCESS;
1112 }
1113
1114 static void
1115 tu_queue_finish(struct tu_queue *queue)
1116 {
1117    tu_fence_finish(&queue->submit_fence);
1118    tu_drm_submitqueue_close(queue->device, queue->msm_queue_id);
1119 }
1120
1121 static int
1122 tu_get_device_extension_index(const char *name)
1123 {
1124    for (unsigned i = 0; i < TU_DEVICE_EXTENSION_COUNT; ++i) {
1125       if (strcmp(name, tu_device_extensions[i].extensionName) == 0)
1126          return i;
1127    }
1128    return -1;
1129 }
1130
1131 struct PACKED bcolor_entry {
1132    uint32_t fp32[4];
1133    uint16_t ui16[4];
1134    int16_t  si16[4];
1135    uint16_t fp16[4];
1136    uint16_t rgb565;
1137    uint16_t rgb5a1;
1138    uint16_t rgba4;
1139    uint8_t __pad0[2];
1140    uint8_t  ui8[4];
1141    int8_t   si8[4];
1142    uint32_t rgb10a2;
1143    uint32_t z24; /* also s8? */
1144    uint16_t srgb[4];      /* appears to duplicate fp16[], but clamped, used for srgb */
1145    uint8_t  __pad1[56];
1146 } border_color[] = {
1147    [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = {},
1148    [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = {},
1149    [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = {
1150       .fp32[3] = 0x3f800000,
1151       .ui16[3] = 0xffff,
1152       .si16[3] = 0x7fff,
1153       .fp16[3] = 0x3c00,
1154       .rgb5a1 = 0x8000,
1155       .rgba4 = 0xf000,
1156       .ui8[3] = 0xff,
1157       .si8[3] = 0x7f,
1158       .rgb10a2 = 0xc0000000,
1159       .srgb[3] = 0x3c00,
1160    },
1161    [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = {
1162       .fp32[3] = 1,
1163       .fp16[3] = 1,
1164    },
1165    [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = {
1166       .fp32[0 ... 3] = 0x3f800000,
1167       .ui16[0 ... 3] = 0xffff,
1168       .si16[0 ... 3] = 0x7fff,
1169       .fp16[0 ... 3] = 0x3c00,
1170       .rgb565 = 0xffff,
1171       .rgb5a1 = 0xffff,
1172       .rgba4 = 0xffff,
1173       .ui8[0 ... 3] = 0xff,
1174       .si8[0 ... 3] = 0x7f,
1175       .rgb10a2 = 0xffffffff,
1176       .z24 = 0xffffff,
1177       .srgb[0 ... 3] = 0x3c00,
1178    },
1179    [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = {
1180       .fp32[0 ... 3] = 1,
1181       .fp16[0 ... 3] = 1,
1182    },
1183 };
1184
1185
1186 VkResult
1187 tu_CreateDevice(VkPhysicalDevice physicalDevice,
1188                 const VkDeviceCreateInfo *pCreateInfo,
1189                 const VkAllocationCallbacks *pAllocator,
1190                 VkDevice *pDevice)
1191 {
1192    TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
1193    VkResult result;
1194    struct tu_device *device;
1195
1196    /* Check enabled features */
1197    if (pCreateInfo->pEnabledFeatures) {
1198       VkPhysicalDeviceFeatures supported_features;
1199       tu_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1200       VkBool32 *supported_feature = (VkBool32 *) &supported_features;
1201       VkBool32 *enabled_feature = (VkBool32 *) pCreateInfo->pEnabledFeatures;
1202       unsigned num_features =
1203          sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1204       for (uint32_t i = 0; i < num_features; i++) {
1205          if (enabled_feature[i] && !supported_feature[i])
1206             return vk_error(physical_device->instance,
1207                             VK_ERROR_FEATURE_NOT_PRESENT);
1208       }
1209    }
1210
1211    device = vk_zalloc2(&physical_device->instance->alloc, pAllocator,
1212                        sizeof(*device), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1213    if (!device)
1214       return vk_error(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1215
1216    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1217    device->instance = physical_device->instance;
1218    device->physical_device = physical_device;
1219    device->_lost = false;
1220
1221    if (pAllocator)
1222       device->alloc = *pAllocator;
1223    else
1224       device->alloc = physical_device->instance->alloc;
1225
1226    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1227       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
1228       int index = tu_get_device_extension_index(ext_name);
1229       if (index < 0 ||
1230           !physical_device->supported_extensions.extensions[index]) {
1231          vk_free(&device->alloc, device);
1232          return vk_error(physical_device->instance,
1233                          VK_ERROR_EXTENSION_NOT_PRESENT);
1234       }
1235
1236       device->enabled_extensions.extensions[index] = true;
1237    }
1238
1239    for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1240       const VkDeviceQueueCreateInfo *queue_create =
1241          &pCreateInfo->pQueueCreateInfos[i];
1242       uint32_t qfi = queue_create->queueFamilyIndex;
1243       device->queues[qfi] = vk_alloc(
1244          &device->alloc, queue_create->queueCount * sizeof(struct tu_queue),
1245          8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1246       if (!device->queues[qfi]) {
1247          result = VK_ERROR_OUT_OF_HOST_MEMORY;
1248          goto fail_queues;
1249       }
1250
1251       memset(device->queues[qfi], 0,
1252              queue_create->queueCount * sizeof(struct tu_queue));
1253
1254       device->queue_count[qfi] = queue_create->queueCount;
1255
1256       for (unsigned q = 0; q < queue_create->queueCount; q++) {
1257          result = tu_queue_init(device, &device->queues[qfi][q], qfi, q,
1258                                 queue_create->flags);
1259          if (result != VK_SUCCESS)
1260             goto fail_queues;
1261       }
1262    }
1263
1264    device->compiler = ir3_compiler_create(NULL, physical_device->gpu_id);
1265    if (!device->compiler)
1266       goto fail_queues;
1267
1268 #define VSC_DRAW_STRM_SIZE(pitch)  ((pitch) * 32 + 0x100)  /* extra size to store VSC_SIZE */
1269 #define VSC_PRIM_STRM_SIZE(pitch) ((pitch) * 32)
1270
1271    device->vsc_draw_strm_pitch = 0x440 * 4;
1272    device->vsc_prim_strm_pitch = 0x1040 * 4;
1273
1274    result = tu_bo_init_new(device, &device->vsc_draw_strm, VSC_DRAW_STRM_SIZE(device->vsc_draw_strm_pitch));
1275    if (result != VK_SUCCESS)
1276       goto fail_vsc_data;
1277
1278    result = tu_bo_init_new(device, &device->vsc_prim_strm, VSC_PRIM_STRM_SIZE(device->vsc_prim_strm_pitch));
1279    if (result != VK_SUCCESS)
1280       goto fail_vsc_data2;
1281
1282    STATIC_ASSERT(sizeof(struct bcolor_entry) == 128);
1283    result = tu_bo_init_new(device, &device->border_color, sizeof(border_color));
1284    if (result != VK_SUCCESS)
1285       goto fail_border_color;
1286
1287    result = tu_bo_map(device, &device->border_color);
1288    if (result != VK_SUCCESS)
1289       goto fail_border_color_map;
1290
1291    memcpy(device->border_color.map, border_color, sizeof(border_color));
1292
1293    VkPipelineCacheCreateInfo ci;
1294    ci.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
1295    ci.pNext = NULL;
1296    ci.flags = 0;
1297    ci.pInitialData = NULL;
1298    ci.initialDataSize = 0;
1299    VkPipelineCache pc;
1300    result =
1301       tu_CreatePipelineCache(tu_device_to_handle(device), &ci, NULL, &pc);
1302    if (result != VK_SUCCESS)
1303       goto fail_pipeline_cache;
1304
1305    device->mem_cache = tu_pipeline_cache_from_handle(pc);
1306
1307    for (unsigned i = 0; i < ARRAY_SIZE(device->scratch_bos); i++)
1308       mtx_init(&device->scratch_bos[i].construct_mtx, mtx_plain);
1309
1310    *pDevice = tu_device_to_handle(device);
1311    return VK_SUCCESS;
1312
1313 fail_pipeline_cache:
1314 fail_border_color_map:
1315    tu_bo_finish(device, &device->border_color);
1316
1317 fail_border_color:
1318    tu_bo_finish(device, &device->vsc_prim_strm);
1319
1320 fail_vsc_data2:
1321    tu_bo_finish(device, &device->vsc_draw_strm);
1322
1323 fail_vsc_data:
1324    ralloc_free(device->compiler);
1325
1326 fail_queues:
1327    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1328       for (unsigned q = 0; q < device->queue_count[i]; q++)
1329          tu_queue_finish(&device->queues[i][q]);
1330       if (device->queue_count[i])
1331          vk_free(&device->alloc, device->queues[i]);
1332    }
1333
1334    vk_free(&device->alloc, device);
1335    return result;
1336 }
1337
1338 void
1339 tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
1340 {
1341    TU_FROM_HANDLE(tu_device, device, _device);
1342
1343    if (!device)
1344       return;
1345
1346    tu_bo_finish(device, &device->vsc_draw_strm);
1347    tu_bo_finish(device, &device->vsc_prim_strm);
1348
1349    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1350       for (unsigned q = 0; q < device->queue_count[i]; q++)
1351          tu_queue_finish(&device->queues[i][q]);
1352       if (device->queue_count[i])
1353          vk_free(&device->alloc, device->queues[i]);
1354    }
1355
1356    for (unsigned i = 0; i < ARRAY_SIZE(device->scratch_bos); i++) {
1357       if (device->scratch_bos[i].initialized)
1358          tu_bo_finish(device, &device->scratch_bos[i].bo);
1359    }
1360
1361    ir3_compiler_destroy(device->compiler);
1362
1363    VkPipelineCache pc = tu_pipeline_cache_to_handle(device->mem_cache);
1364    tu_DestroyPipelineCache(tu_device_to_handle(device), pc, NULL);
1365
1366    vk_free(&device->alloc, device);
1367 }
1368
1369 VkResult
1370 _tu_device_set_lost(struct tu_device *device,
1371                     const char *file, int line,
1372                     const char *msg, ...)
1373 {
1374    /* Set the flag indicating that waits should return in finite time even
1375     * after device loss.
1376     */
1377    p_atomic_inc(&device->_lost);
1378
1379    /* TODO: Report the log message through VkDebugReportCallbackEXT instead */
1380    fprintf(stderr, "%s:%d: ", file, line);
1381    va_list ap;
1382    va_start(ap, msg);
1383    vfprintf(stderr, msg, ap);
1384    va_end(ap);
1385
1386    if (env_var_as_boolean("TU_ABORT_ON_DEVICE_LOSS", false))
1387       abort();
1388
1389    return VK_ERROR_DEVICE_LOST;
1390 }
1391
1392 VkResult
1393 tu_get_scratch_bo(struct tu_device *dev, uint64_t size, struct tu_bo **bo)
1394 {
1395    unsigned size_log2 = MAX2(util_logbase2_ceil64(size), MIN_SCRATCH_BO_SIZE_LOG2);
1396    unsigned index = size_log2 - MIN_SCRATCH_BO_SIZE_LOG2;
1397    assert(index < ARRAY_SIZE(dev->scratch_bos));
1398
1399    for (unsigned i = index; i < ARRAY_SIZE(dev->scratch_bos); i++) {
1400       if (p_atomic_read(&dev->scratch_bos[i].initialized)) {
1401          /* Fast path: just return the already-allocated BO. */
1402          *bo = &dev->scratch_bos[i].bo;
1403          return VK_SUCCESS;
1404       }
1405    }
1406
1407    /* Slow path: actually allocate the BO. We take a lock because the process
1408     * of allocating it is slow, and we don't want to block the CPU while it
1409     * finishes.
1410    */
1411    mtx_lock(&dev->scratch_bos[index].construct_mtx);
1412
1413    /* Another thread may have allocated it already while we were waiting on
1414     * the lock. We need to check this in order to avoid double-allocating.
1415     */
1416    if (dev->scratch_bos[index].initialized) {
1417       mtx_unlock(&dev->scratch_bos[index].construct_mtx);
1418       *bo = &dev->scratch_bos[index].bo;
1419       return VK_SUCCESS;
1420    }
1421
1422    unsigned bo_size = 1ull << size_log2;
1423    VkResult result = tu_bo_init_new(dev, &dev->scratch_bos[index].bo, bo_size);
1424    if (result != VK_SUCCESS) {
1425       mtx_unlock(&dev->scratch_bos[index].construct_mtx);
1426       return result;
1427    }
1428
1429    p_atomic_set(&dev->scratch_bos[index].initialized, true);
1430
1431    mtx_unlock(&dev->scratch_bos[index].construct_mtx);
1432
1433    *bo = &dev->scratch_bos[index].bo;
1434    return VK_SUCCESS;
1435 }
1436
1437 VkResult
1438 tu_EnumerateInstanceLayerProperties(uint32_t *pPropertyCount,
1439                                     VkLayerProperties *pProperties)
1440 {
1441    *pPropertyCount = 0;
1442    return VK_SUCCESS;
1443 }
1444
1445 VkResult
1446 tu_EnumerateDeviceLayerProperties(VkPhysicalDevice physicalDevice,
1447                                   uint32_t *pPropertyCount,
1448                                   VkLayerProperties *pProperties)
1449 {
1450    *pPropertyCount = 0;
1451    return VK_SUCCESS;
1452 }
1453
1454 void
1455 tu_GetDeviceQueue2(VkDevice _device,
1456                    const VkDeviceQueueInfo2 *pQueueInfo,
1457                    VkQueue *pQueue)
1458 {
1459    TU_FROM_HANDLE(tu_device, device, _device);
1460    struct tu_queue *queue;
1461
1462    queue =
1463       &device->queues[pQueueInfo->queueFamilyIndex][pQueueInfo->queueIndex];
1464    if (pQueueInfo->flags != queue->flags) {
1465       /* From the Vulkan 1.1.70 spec:
1466        *
1467        * "The queue returned by vkGetDeviceQueue2 must have the same
1468        * flags value from this structure as that used at device
1469        * creation time in a VkDeviceQueueCreateInfo instance. If no
1470        * matching flags were specified at device creation time then
1471        * pQueue will return VK_NULL_HANDLE."
1472        */
1473       *pQueue = VK_NULL_HANDLE;
1474       return;
1475    }
1476
1477    *pQueue = tu_queue_to_handle(queue);
1478 }
1479
1480 void
1481 tu_GetDeviceQueue(VkDevice _device,
1482                   uint32_t queueFamilyIndex,
1483                   uint32_t queueIndex,
1484                   VkQueue *pQueue)
1485 {
1486    const VkDeviceQueueInfo2 info =
1487       (VkDeviceQueueInfo2) { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
1488                              .queueFamilyIndex = queueFamilyIndex,
1489                              .queueIndex = queueIndex };
1490
1491    tu_GetDeviceQueue2(_device, &info, pQueue);
1492 }
1493
1494 static VkResult
1495 tu_get_semaphore_syncobjs(const VkSemaphore *sems,
1496                           uint32_t sem_count,
1497                           bool wait,
1498                           struct drm_msm_gem_submit_syncobj **out,
1499                           uint32_t *out_count)
1500 {
1501    uint32_t syncobj_count = 0;
1502    struct drm_msm_gem_submit_syncobj *syncobjs;
1503
1504    for (uint32_t i = 0; i  < sem_count; ++i) {
1505       TU_FROM_HANDLE(tu_semaphore, sem, sems[i]);
1506
1507       struct tu_semaphore_part *part =
1508          sem->temporary.kind != TU_SEMAPHORE_NONE ?
1509             &sem->temporary : &sem->permanent;
1510
1511       if (part->kind == TU_SEMAPHORE_SYNCOBJ)
1512          ++syncobj_count;
1513    }
1514
1515    *out = NULL;
1516    *out_count = syncobj_count;
1517    if (!syncobj_count)
1518       return VK_SUCCESS;
1519
1520    *out = syncobjs = calloc(syncobj_count, sizeof (*syncobjs));
1521    if (!syncobjs)
1522       return VK_ERROR_OUT_OF_HOST_MEMORY;
1523
1524    for (uint32_t i = 0, j = 0; i  < sem_count; ++i) {
1525       TU_FROM_HANDLE(tu_semaphore, sem, sems[i]);
1526
1527       struct tu_semaphore_part *part =
1528          sem->temporary.kind != TU_SEMAPHORE_NONE ?
1529             &sem->temporary : &sem->permanent;
1530
1531       if (part->kind == TU_SEMAPHORE_SYNCOBJ) {
1532          syncobjs[j].handle = part->syncobj;
1533          syncobjs[j].flags = wait ? MSM_SUBMIT_SYNCOBJ_RESET : 0;
1534          ++j;
1535       }
1536    }
1537
1538    return VK_SUCCESS;
1539 }
1540
1541
1542 static void
1543 tu_semaphores_remove_temp(struct tu_device *device,
1544                           const VkSemaphore *sems,
1545                           uint32_t sem_count)
1546 {
1547    for (uint32_t i = 0; i  < sem_count; ++i) {
1548       TU_FROM_HANDLE(tu_semaphore, sem, sems[i]);
1549       tu_semaphore_remove_temp(device, sem);
1550    }
1551 }
1552
1553 VkResult
1554 tu_QueueSubmit(VkQueue _queue,
1555                uint32_t submitCount,
1556                const VkSubmitInfo *pSubmits,
1557                VkFence _fence)
1558 {
1559    TU_FROM_HANDLE(tu_queue, queue, _queue);
1560    VkResult result;
1561
1562    for (uint32_t i = 0; i < submitCount; ++i) {
1563       const VkSubmitInfo *submit = pSubmits + i;
1564       const bool last_submit = (i == submitCount - 1);
1565       struct drm_msm_gem_submit_syncobj *in_syncobjs = NULL, *out_syncobjs = NULL;
1566       uint32_t nr_in_syncobjs, nr_out_syncobjs;
1567       struct tu_bo_list bo_list;
1568       tu_bo_list_init(&bo_list);
1569
1570       result = tu_get_semaphore_syncobjs(pSubmits[i].pWaitSemaphores,
1571                                          pSubmits[i].waitSemaphoreCount,
1572                                          false, &in_syncobjs, &nr_in_syncobjs);
1573       if (result != VK_SUCCESS) {
1574          return tu_device_set_lost(queue->device,
1575                                    "failed to allocate space for semaphore submission\n");
1576       }
1577
1578       result = tu_get_semaphore_syncobjs(pSubmits[i].pSignalSemaphores,
1579                                          pSubmits[i].signalSemaphoreCount,
1580                                          false, &out_syncobjs, &nr_out_syncobjs);
1581       if (result != VK_SUCCESS) {
1582          free(in_syncobjs);
1583          return tu_device_set_lost(queue->device,
1584                                    "failed to allocate space for semaphore submission\n");
1585       }
1586
1587       uint32_t entry_count = 0;
1588       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1589          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1590          entry_count += cmdbuf->cs.entry_count;
1591       }
1592
1593       struct drm_msm_gem_submit_cmd cmds[entry_count];
1594       uint32_t entry_idx = 0;
1595       for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
1596          TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
1597          struct tu_cs *cs = &cmdbuf->cs;
1598          for (unsigned i = 0; i < cs->entry_count; ++i, ++entry_idx) {
1599             cmds[entry_idx].type = MSM_SUBMIT_CMD_BUF;
1600             cmds[entry_idx].submit_idx =
1601                tu_bo_list_add(&bo_list, cs->entries[i].bo,
1602                               MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
1603             cmds[entry_idx].submit_offset = cs->entries[i].offset;
1604             cmds[entry_idx].size = cs->entries[i].size;
1605             cmds[entry_idx].pad = 0;
1606             cmds[entry_idx].nr_relocs = 0;
1607             cmds[entry_idx].relocs = 0;
1608          }
1609
1610          tu_bo_list_merge(&bo_list, &cmdbuf->bo_list);
1611       }
1612
1613       uint32_t flags = MSM_PIPE_3D0;
1614       if (nr_in_syncobjs) {
1615          flags |= MSM_SUBMIT_SYNCOBJ_IN;
1616       }
1617       if (nr_out_syncobjs) {
1618          flags |= MSM_SUBMIT_SYNCOBJ_OUT;
1619       }
1620
1621       if (last_submit) {
1622          flags |= MSM_SUBMIT_FENCE_FD_OUT;
1623       }
1624
1625       struct drm_msm_gem_submit req = {
1626          .flags = flags,
1627          .queueid = queue->msm_queue_id,
1628          .bos = (uint64_t)(uintptr_t) bo_list.bo_infos,
1629          .nr_bos = bo_list.count,
1630          .cmds = (uint64_t)(uintptr_t)cmds,
1631          .nr_cmds = entry_count,
1632          .in_syncobjs = (uint64_t)(uintptr_t)in_syncobjs,
1633          .out_syncobjs = (uint64_t)(uintptr_t)out_syncobjs,
1634          .nr_in_syncobjs = nr_in_syncobjs,
1635          .nr_out_syncobjs = nr_out_syncobjs,
1636          .syncobj_stride = sizeof(struct drm_msm_gem_submit_syncobj),
1637       };
1638
1639       int ret = drmCommandWriteRead(queue->device->physical_device->local_fd,
1640                                     DRM_MSM_GEM_SUBMIT,
1641                                     &req, sizeof(req));
1642       if (ret) {
1643          free(in_syncobjs);
1644          free(out_syncobjs);
1645          return tu_device_set_lost(queue->device, "submit failed: %s\n",
1646                                    strerror(errno));
1647       }
1648
1649       tu_bo_list_destroy(&bo_list);
1650       free(in_syncobjs);
1651       free(out_syncobjs);
1652
1653       tu_semaphores_remove_temp(queue->device, pSubmits[i].pWaitSemaphores,
1654                                 pSubmits[i].waitSemaphoreCount);
1655       if (last_submit) {
1656          /* no need to merge fences as queue execution is serialized */
1657          tu_fence_update_fd(&queue->submit_fence, req.fence_fd);
1658       } else if (last_submit) {
1659          close(req.fence_fd);
1660       }
1661    }
1662
1663    if (_fence != VK_NULL_HANDLE) {
1664       TU_FROM_HANDLE(tu_fence, fence, _fence);
1665       tu_fence_copy(fence, &queue->submit_fence);
1666    }
1667
1668    return VK_SUCCESS;
1669 }
1670
1671 VkResult
1672 tu_QueueWaitIdle(VkQueue _queue)
1673 {
1674    TU_FROM_HANDLE(tu_queue, queue, _queue);
1675
1676    if (tu_device_is_lost(queue->device))
1677       return VK_ERROR_DEVICE_LOST;
1678
1679    tu_fence_wait_idle(&queue->submit_fence);
1680
1681    return VK_SUCCESS;
1682 }
1683
1684 VkResult
1685 tu_DeviceWaitIdle(VkDevice _device)
1686 {
1687    TU_FROM_HANDLE(tu_device, device, _device);
1688
1689    if (tu_device_is_lost(device))
1690       return VK_ERROR_DEVICE_LOST;
1691
1692    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
1693       for (unsigned q = 0; q < device->queue_count[i]; q++) {
1694          tu_QueueWaitIdle(tu_queue_to_handle(&device->queues[i][q]));
1695       }
1696    }
1697    return VK_SUCCESS;
1698 }
1699
1700 VkResult
1701 tu_EnumerateInstanceExtensionProperties(const char *pLayerName,
1702                                         uint32_t *pPropertyCount,
1703                                         VkExtensionProperties *pProperties)
1704 {
1705    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1706
1707    /* We spport no lyaers */
1708    if (pLayerName)
1709       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1710
1711    for (int i = 0; i < TU_INSTANCE_EXTENSION_COUNT; i++) {
1712       if (tu_instance_extensions_supported.extensions[i]) {
1713          vk_outarray_append(&out, prop) { *prop = tu_instance_extensions[i]; }
1714       }
1715    }
1716
1717    return vk_outarray_status(&out);
1718 }
1719
1720 VkResult
1721 tu_EnumerateDeviceExtensionProperties(VkPhysicalDevice physicalDevice,
1722                                       const char *pLayerName,
1723                                       uint32_t *pPropertyCount,
1724                                       VkExtensionProperties *pProperties)
1725 {
1726    /* We spport no lyaers */
1727    TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
1728    VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1729
1730    /* We spport no lyaers */
1731    if (pLayerName)
1732       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1733
1734    for (int i = 0; i < TU_DEVICE_EXTENSION_COUNT; i++) {
1735       if (device->supported_extensions.extensions[i]) {
1736          vk_outarray_append(&out, prop) { *prop = tu_device_extensions[i]; }
1737       }
1738    }
1739
1740    return vk_outarray_status(&out);
1741 }
1742
1743 PFN_vkVoidFunction
1744 tu_GetInstanceProcAddr(VkInstance _instance, const char *pName)
1745 {
1746    TU_FROM_HANDLE(tu_instance, instance, _instance);
1747
1748    return tu_lookup_entrypoint_checked(
1749       pName, instance ? instance->api_version : 0,
1750       instance ? &instance->enabled_extensions : NULL, NULL);
1751 }
1752
1753 /* The loader wants us to expose a second GetInstanceProcAddr function
1754  * to work around certain LD_PRELOAD issues seen in apps.
1755  */
1756 PUBLIC
1757 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1758 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName);
1759
1760 PUBLIC
1761 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
1762 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName)
1763 {
1764    return tu_GetInstanceProcAddr(instance, pName);
1765 }
1766
1767 PFN_vkVoidFunction
1768 tu_GetDeviceProcAddr(VkDevice _device, const char *pName)
1769 {
1770    TU_FROM_HANDLE(tu_device, device, _device);
1771
1772    return tu_lookup_entrypoint_checked(pName, device->instance->api_version,
1773                                        &device->instance->enabled_extensions,
1774                                        &device->enabled_extensions);
1775 }
1776
1777 static VkResult
1778 tu_alloc_memory(struct tu_device *device,
1779                 const VkMemoryAllocateInfo *pAllocateInfo,
1780                 const VkAllocationCallbacks *pAllocator,
1781                 VkDeviceMemory *pMem)
1782 {
1783    struct tu_device_memory *mem;
1784    VkResult result;
1785
1786    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1787
1788    if (pAllocateInfo->allocationSize == 0) {
1789       /* Apparently, this is allowed */
1790       *pMem = VK_NULL_HANDLE;
1791       return VK_SUCCESS;
1792    }
1793
1794    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1795                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1796    if (mem == NULL)
1797       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1798
1799    const VkImportMemoryFdInfoKHR *fd_info =
1800       vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
1801    if (fd_info && !fd_info->handleType)
1802       fd_info = NULL;
1803
1804    if (fd_info) {
1805       assert(fd_info->handleType ==
1806                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
1807              fd_info->handleType ==
1808                 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1809
1810       /*
1811        * TODO Importing the same fd twice gives us the same handle without
1812        * reference counting.  We need to maintain a per-instance handle-to-bo
1813        * table and add reference count to tu_bo.
1814        */
1815       result = tu_bo_init_dmabuf(device, &mem->bo,
1816                                  pAllocateInfo->allocationSize, fd_info->fd);
1817       if (result == VK_SUCCESS) {
1818          /* take ownership and close the fd */
1819          close(fd_info->fd);
1820       }
1821    } else {
1822       result =
1823          tu_bo_init_new(device, &mem->bo, pAllocateInfo->allocationSize);
1824    }
1825
1826    if (result != VK_SUCCESS) {
1827       vk_free2(&device->alloc, pAllocator, mem);
1828       return result;
1829    }
1830
1831    mem->size = pAllocateInfo->allocationSize;
1832    mem->type_index = pAllocateInfo->memoryTypeIndex;
1833
1834    mem->map = NULL;
1835    mem->user_ptr = NULL;
1836
1837    *pMem = tu_device_memory_to_handle(mem);
1838
1839    return VK_SUCCESS;
1840 }
1841
1842 VkResult
1843 tu_AllocateMemory(VkDevice _device,
1844                   const VkMemoryAllocateInfo *pAllocateInfo,
1845                   const VkAllocationCallbacks *pAllocator,
1846                   VkDeviceMemory *pMem)
1847 {
1848    TU_FROM_HANDLE(tu_device, device, _device);
1849    return tu_alloc_memory(device, pAllocateInfo, pAllocator, pMem);
1850 }
1851
1852 void
1853 tu_FreeMemory(VkDevice _device,
1854               VkDeviceMemory _mem,
1855               const VkAllocationCallbacks *pAllocator)
1856 {
1857    TU_FROM_HANDLE(tu_device, device, _device);
1858    TU_FROM_HANDLE(tu_device_memory, mem, _mem);
1859
1860    if (mem == NULL)
1861       return;
1862
1863    tu_bo_finish(device, &mem->bo);
1864    vk_free2(&device->alloc, pAllocator, mem);
1865 }
1866
1867 VkResult
1868 tu_MapMemory(VkDevice _device,
1869              VkDeviceMemory _memory,
1870              VkDeviceSize offset,
1871              VkDeviceSize size,
1872              VkMemoryMapFlags flags,
1873              void **ppData)
1874 {
1875    TU_FROM_HANDLE(tu_device, device, _device);
1876    TU_FROM_HANDLE(tu_device_memory, mem, _memory);
1877    VkResult result;
1878
1879    if (mem == NULL) {
1880       *ppData = NULL;
1881       return VK_SUCCESS;
1882    }
1883
1884    if (mem->user_ptr) {
1885       *ppData = mem->user_ptr;
1886    } else if (!mem->map) {
1887       result = tu_bo_map(device, &mem->bo);
1888       if (result != VK_SUCCESS)
1889          return result;
1890       *ppData = mem->map = mem->bo.map;
1891    } else
1892       *ppData = mem->map;
1893
1894    if (*ppData) {
1895       *ppData += offset;
1896       return VK_SUCCESS;
1897    }
1898
1899    return vk_error(device->instance, VK_ERROR_MEMORY_MAP_FAILED);
1900 }
1901
1902 void
1903 tu_UnmapMemory(VkDevice _device, VkDeviceMemory _memory)
1904 {
1905    /* I do not see any unmapping done by the freedreno Gallium driver. */
1906 }
1907
1908 VkResult
1909 tu_FlushMappedMemoryRanges(VkDevice _device,
1910                            uint32_t memoryRangeCount,
1911                            const VkMappedMemoryRange *pMemoryRanges)
1912 {
1913    return VK_SUCCESS;
1914 }
1915
1916 VkResult
1917 tu_InvalidateMappedMemoryRanges(VkDevice _device,
1918                                 uint32_t memoryRangeCount,
1919                                 const VkMappedMemoryRange *pMemoryRanges)
1920 {
1921    return VK_SUCCESS;
1922 }
1923
1924 void
1925 tu_GetBufferMemoryRequirements(VkDevice _device,
1926                                VkBuffer _buffer,
1927                                VkMemoryRequirements *pMemoryRequirements)
1928 {
1929    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
1930
1931    pMemoryRequirements->memoryTypeBits = 1;
1932    pMemoryRequirements->alignment = 64;
1933    pMemoryRequirements->size =
1934       align64(buffer->size, pMemoryRequirements->alignment);
1935 }
1936
1937 void
1938 tu_GetBufferMemoryRequirements2(
1939    VkDevice device,
1940    const VkBufferMemoryRequirementsInfo2 *pInfo,
1941    VkMemoryRequirements2 *pMemoryRequirements)
1942 {
1943    tu_GetBufferMemoryRequirements(device, pInfo->buffer,
1944                                   &pMemoryRequirements->memoryRequirements);
1945 }
1946
1947 void
1948 tu_GetImageMemoryRequirements(VkDevice _device,
1949                               VkImage _image,
1950                               VkMemoryRequirements *pMemoryRequirements)
1951 {
1952    TU_FROM_HANDLE(tu_image, image, _image);
1953
1954    pMemoryRequirements->memoryTypeBits = 1;
1955    pMemoryRequirements->size = image->layout.size;
1956    pMemoryRequirements->alignment = image->layout.base_align;
1957 }
1958
1959 void
1960 tu_GetImageMemoryRequirements2(VkDevice device,
1961                                const VkImageMemoryRequirementsInfo2 *pInfo,
1962                                VkMemoryRequirements2 *pMemoryRequirements)
1963 {
1964    tu_GetImageMemoryRequirements(device, pInfo->image,
1965                                  &pMemoryRequirements->memoryRequirements);
1966 }
1967
1968 void
1969 tu_GetImageSparseMemoryRequirements(
1970    VkDevice device,
1971    VkImage image,
1972    uint32_t *pSparseMemoryRequirementCount,
1973    VkSparseImageMemoryRequirements *pSparseMemoryRequirements)
1974 {
1975    tu_stub();
1976 }
1977
1978 void
1979 tu_GetImageSparseMemoryRequirements2(
1980    VkDevice device,
1981    const VkImageSparseMemoryRequirementsInfo2 *pInfo,
1982    uint32_t *pSparseMemoryRequirementCount,
1983    VkSparseImageMemoryRequirements2 *pSparseMemoryRequirements)
1984 {
1985    tu_stub();
1986 }
1987
1988 void
1989 tu_GetDeviceMemoryCommitment(VkDevice device,
1990                              VkDeviceMemory memory,
1991                              VkDeviceSize *pCommittedMemoryInBytes)
1992 {
1993    *pCommittedMemoryInBytes = 0;
1994 }
1995
1996 VkResult
1997 tu_BindBufferMemory2(VkDevice device,
1998                      uint32_t bindInfoCount,
1999                      const VkBindBufferMemoryInfo *pBindInfos)
2000 {
2001    for (uint32_t i = 0; i < bindInfoCount; ++i) {
2002       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
2003       TU_FROM_HANDLE(tu_buffer, buffer, pBindInfos[i].buffer);
2004
2005       if (mem) {
2006          buffer->bo = &mem->bo;
2007          buffer->bo_offset = pBindInfos[i].memoryOffset;
2008       } else {
2009          buffer->bo = NULL;
2010       }
2011    }
2012    return VK_SUCCESS;
2013 }
2014
2015 VkResult
2016 tu_BindBufferMemory(VkDevice device,
2017                     VkBuffer buffer,
2018                     VkDeviceMemory memory,
2019                     VkDeviceSize memoryOffset)
2020 {
2021    const VkBindBufferMemoryInfo info = {
2022       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
2023       .buffer = buffer,
2024       .memory = memory,
2025       .memoryOffset = memoryOffset
2026    };
2027
2028    return tu_BindBufferMemory2(device, 1, &info);
2029 }
2030
2031 VkResult
2032 tu_BindImageMemory2(VkDevice device,
2033                     uint32_t bindInfoCount,
2034                     const VkBindImageMemoryInfo *pBindInfos)
2035 {
2036    for (uint32_t i = 0; i < bindInfoCount; ++i) {
2037       TU_FROM_HANDLE(tu_image, image, pBindInfos[i].image);
2038       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
2039
2040       if (mem) {
2041          image->bo = &mem->bo;
2042          image->bo_offset = pBindInfos[i].memoryOffset;
2043       } else {
2044          image->bo = NULL;
2045          image->bo_offset = 0;
2046       }
2047    }
2048
2049    return VK_SUCCESS;
2050 }
2051
2052 VkResult
2053 tu_BindImageMemory(VkDevice device,
2054                    VkImage image,
2055                    VkDeviceMemory memory,
2056                    VkDeviceSize memoryOffset)
2057 {
2058    const VkBindImageMemoryInfo info = {
2059       .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
2060       .image = image,
2061       .memory = memory,
2062       .memoryOffset = memoryOffset
2063    };
2064
2065    return tu_BindImageMemory2(device, 1, &info);
2066 }
2067
2068 VkResult
2069 tu_QueueBindSparse(VkQueue _queue,
2070                    uint32_t bindInfoCount,
2071                    const VkBindSparseInfo *pBindInfo,
2072                    VkFence _fence)
2073 {
2074    return VK_SUCCESS;
2075 }
2076
2077 // Queue semaphore functions
2078
2079
2080 static void
2081 tu_semaphore_part_destroy(struct tu_device *device,
2082                           struct tu_semaphore_part *part)
2083 {
2084    switch(part->kind) {
2085    case TU_SEMAPHORE_NONE:
2086       break;
2087    case TU_SEMAPHORE_SYNCOBJ:
2088       drmSyncobjDestroy(device->physical_device->local_fd, part->syncobj);
2089       break;
2090    }
2091    part->kind = TU_SEMAPHORE_NONE;
2092 }
2093
2094 static void
2095 tu_semaphore_remove_temp(struct tu_device *device,
2096                          struct tu_semaphore *sem)
2097 {
2098    if (sem->temporary.kind != TU_SEMAPHORE_NONE) {
2099       tu_semaphore_part_destroy(device, &sem->temporary);
2100    }
2101 }
2102
2103 VkResult
2104 tu_CreateSemaphore(VkDevice _device,
2105                    const VkSemaphoreCreateInfo *pCreateInfo,
2106                    const VkAllocationCallbacks *pAllocator,
2107                    VkSemaphore *pSemaphore)
2108 {
2109    TU_FROM_HANDLE(tu_device, device, _device);
2110
2111    struct tu_semaphore *sem =
2112       vk_alloc2(&device->alloc, pAllocator, sizeof(*sem), 8,
2113                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2114    if (!sem)
2115       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2116
2117    const VkExportSemaphoreCreateInfo *export =
2118       vk_find_struct_const(pCreateInfo->pNext, EXPORT_SEMAPHORE_CREATE_INFO);
2119    VkExternalSemaphoreHandleTypeFlags handleTypes =
2120       export ? export->handleTypes : 0;
2121
2122    sem->permanent.kind = TU_SEMAPHORE_NONE;
2123    sem->temporary.kind = TU_SEMAPHORE_NONE;
2124
2125    if (handleTypes) {
2126       if (drmSyncobjCreate(device->physical_device->local_fd, 0, &sem->permanent.syncobj) < 0) {
2127           vk_free2(&device->alloc, pAllocator, sem);
2128           return VK_ERROR_OUT_OF_HOST_MEMORY;
2129       }
2130       sem->permanent.kind = TU_SEMAPHORE_SYNCOBJ;
2131    }
2132    *pSemaphore = tu_semaphore_to_handle(sem);
2133    return VK_SUCCESS;
2134 }
2135
2136 void
2137 tu_DestroySemaphore(VkDevice _device,
2138                     VkSemaphore _semaphore,
2139                     const VkAllocationCallbacks *pAllocator)
2140 {
2141    TU_FROM_HANDLE(tu_device, device, _device);
2142    TU_FROM_HANDLE(tu_semaphore, sem, _semaphore);
2143    if (!_semaphore)
2144       return;
2145
2146    tu_semaphore_part_destroy(device, &sem->permanent);
2147    tu_semaphore_part_destroy(device, &sem->temporary);
2148
2149    vk_free2(&device->alloc, pAllocator, sem);
2150 }
2151
2152 VkResult
2153 tu_CreateEvent(VkDevice _device,
2154                const VkEventCreateInfo *pCreateInfo,
2155                const VkAllocationCallbacks *pAllocator,
2156                VkEvent *pEvent)
2157 {
2158    TU_FROM_HANDLE(tu_device, device, _device);
2159    struct tu_event *event =
2160       vk_alloc2(&device->alloc, pAllocator, sizeof(*event), 8,
2161                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2162
2163    if (!event)
2164       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2165
2166    VkResult result = tu_bo_init_new(device, &event->bo, 0x1000);
2167    if (result != VK_SUCCESS)
2168       goto fail_alloc;
2169
2170    result = tu_bo_map(device, &event->bo);
2171    if (result != VK_SUCCESS)
2172       goto fail_map;
2173
2174    *pEvent = tu_event_to_handle(event);
2175
2176    return VK_SUCCESS;
2177
2178 fail_map:
2179    tu_bo_finish(device, &event->bo);
2180 fail_alloc:
2181    vk_free2(&device->alloc, pAllocator, event);
2182    return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2183 }
2184
2185 void
2186 tu_DestroyEvent(VkDevice _device,
2187                 VkEvent _event,
2188                 const VkAllocationCallbacks *pAllocator)
2189 {
2190    TU_FROM_HANDLE(tu_device, device, _device);
2191    TU_FROM_HANDLE(tu_event, event, _event);
2192
2193    if (!event)
2194       return;
2195
2196    tu_bo_finish(device, &event->bo);
2197    vk_free2(&device->alloc, pAllocator, event);
2198 }
2199
2200 VkResult
2201 tu_GetEventStatus(VkDevice _device, VkEvent _event)
2202 {
2203    TU_FROM_HANDLE(tu_event, event, _event);
2204
2205    if (*(uint64_t*) event->bo.map == 1)
2206       return VK_EVENT_SET;
2207    return VK_EVENT_RESET;
2208 }
2209
2210 VkResult
2211 tu_SetEvent(VkDevice _device, VkEvent _event)
2212 {
2213    TU_FROM_HANDLE(tu_event, event, _event);
2214    *(uint64_t*) event->bo.map = 1;
2215
2216    return VK_SUCCESS;
2217 }
2218
2219 VkResult
2220 tu_ResetEvent(VkDevice _device, VkEvent _event)
2221 {
2222    TU_FROM_HANDLE(tu_event, event, _event);
2223    *(uint64_t*) event->bo.map = 0;
2224
2225    return VK_SUCCESS;
2226 }
2227
2228 VkResult
2229 tu_CreateBuffer(VkDevice _device,
2230                 const VkBufferCreateInfo *pCreateInfo,
2231                 const VkAllocationCallbacks *pAllocator,
2232                 VkBuffer *pBuffer)
2233 {
2234    TU_FROM_HANDLE(tu_device, device, _device);
2235    struct tu_buffer *buffer;
2236
2237    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2238
2239    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2240                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2241    if (buffer == NULL)
2242       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2243
2244    buffer->size = pCreateInfo->size;
2245    buffer->usage = pCreateInfo->usage;
2246    buffer->flags = pCreateInfo->flags;
2247
2248    *pBuffer = tu_buffer_to_handle(buffer);
2249
2250    return VK_SUCCESS;
2251 }
2252
2253 void
2254 tu_DestroyBuffer(VkDevice _device,
2255                  VkBuffer _buffer,
2256                  const VkAllocationCallbacks *pAllocator)
2257 {
2258    TU_FROM_HANDLE(tu_device, device, _device);
2259    TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
2260
2261    if (!buffer)
2262       return;
2263
2264    vk_free2(&device->alloc, pAllocator, buffer);
2265 }
2266
2267 VkResult
2268 tu_CreateFramebuffer(VkDevice _device,
2269                      const VkFramebufferCreateInfo *pCreateInfo,
2270                      const VkAllocationCallbacks *pAllocator,
2271                      VkFramebuffer *pFramebuffer)
2272 {
2273    TU_FROM_HANDLE(tu_device, device, _device);
2274    struct tu_framebuffer *framebuffer;
2275
2276    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2277
2278    size_t size = sizeof(*framebuffer) + sizeof(struct tu_attachment_info) *
2279                                            pCreateInfo->attachmentCount;
2280    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2281                            VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2282    if (framebuffer == NULL)
2283       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2284
2285    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2286    framebuffer->width = pCreateInfo->width;
2287    framebuffer->height = pCreateInfo->height;
2288    framebuffer->layers = pCreateInfo->layers;
2289    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2290       VkImageView _iview = pCreateInfo->pAttachments[i];
2291       struct tu_image_view *iview = tu_image_view_from_handle(_iview);
2292       framebuffer->attachments[i].attachment = iview;
2293    }
2294
2295    *pFramebuffer = tu_framebuffer_to_handle(framebuffer);
2296    return VK_SUCCESS;
2297 }
2298
2299 void
2300 tu_DestroyFramebuffer(VkDevice _device,
2301                       VkFramebuffer _fb,
2302                       const VkAllocationCallbacks *pAllocator)
2303 {
2304    TU_FROM_HANDLE(tu_device, device, _device);
2305    TU_FROM_HANDLE(tu_framebuffer, fb, _fb);
2306
2307    if (!fb)
2308       return;
2309    vk_free2(&device->alloc, pAllocator, fb);
2310 }
2311
2312 static void
2313 tu_init_sampler(struct tu_device *device,
2314                 struct tu_sampler *sampler,
2315                 const VkSamplerCreateInfo *pCreateInfo)
2316 {
2317    const struct VkSamplerReductionModeCreateInfo *reduction =
2318       vk_find_struct_const(pCreateInfo->pNext, SAMPLER_REDUCTION_MODE_CREATE_INFO);
2319    const struct VkSamplerYcbcrConversionInfo *ycbcr_conversion =
2320       vk_find_struct_const(pCreateInfo->pNext,  SAMPLER_YCBCR_CONVERSION_INFO);
2321
2322    unsigned aniso = pCreateInfo->anisotropyEnable ?
2323       util_last_bit(MIN2((uint32_t)pCreateInfo->maxAnisotropy >> 1, 8)) : 0;
2324    bool miplinear = (pCreateInfo->mipmapMode == VK_SAMPLER_MIPMAP_MODE_LINEAR);
2325    float min_lod = CLAMP(pCreateInfo->minLod, 0.0f, 4095.0f / 256.0f);
2326    float max_lod = CLAMP(pCreateInfo->maxLod, 0.0f, 4095.0f / 256.0f);
2327
2328    sampler->descriptor[0] =
2329       COND(miplinear, A6XX_TEX_SAMP_0_MIPFILTER_LINEAR_NEAR) |
2330       A6XX_TEX_SAMP_0_XY_MAG(tu6_tex_filter(pCreateInfo->magFilter, aniso)) |
2331       A6XX_TEX_SAMP_0_XY_MIN(tu6_tex_filter(pCreateInfo->minFilter, aniso)) |
2332       A6XX_TEX_SAMP_0_ANISO(aniso) |
2333       A6XX_TEX_SAMP_0_WRAP_S(tu6_tex_wrap(pCreateInfo->addressModeU)) |
2334       A6XX_TEX_SAMP_0_WRAP_T(tu6_tex_wrap(pCreateInfo->addressModeV)) |
2335       A6XX_TEX_SAMP_0_WRAP_R(tu6_tex_wrap(pCreateInfo->addressModeW)) |
2336       A6XX_TEX_SAMP_0_LOD_BIAS(pCreateInfo->mipLodBias);
2337    sampler->descriptor[1] =
2338       /* COND(!cso->seamless_cube_map, A6XX_TEX_SAMP_1_CUBEMAPSEAMLESSFILTOFF) | */
2339       COND(pCreateInfo->unnormalizedCoordinates, A6XX_TEX_SAMP_1_UNNORM_COORDS) |
2340       A6XX_TEX_SAMP_1_MIN_LOD(min_lod) |
2341       A6XX_TEX_SAMP_1_MAX_LOD(max_lod) |
2342       COND(pCreateInfo->compareEnable,
2343            A6XX_TEX_SAMP_1_COMPARE_FUNC(tu6_compare_func(pCreateInfo->compareOp)));
2344    /* This is an offset into the border_color BO, which we fill with all the
2345     * possible Vulkan border colors in the correct order, so we can just use
2346     * the Vulkan enum with no translation necessary.
2347     */
2348    sampler->descriptor[2] =
2349       A6XX_TEX_SAMP_2_BCOLOR_OFFSET((unsigned) pCreateInfo->borderColor *
2350                                     sizeof(struct bcolor_entry));
2351    sampler->descriptor[3] = 0;
2352
2353    if (reduction) {
2354       sampler->descriptor[2] |= A6XX_TEX_SAMP_2_REDUCTION_MODE(
2355          tu6_reduction_mode(reduction->reductionMode));
2356    }
2357
2358    sampler->ycbcr_sampler = ycbcr_conversion ?
2359       tu_sampler_ycbcr_conversion_from_handle(ycbcr_conversion->conversion) : NULL;
2360
2361    if (sampler->ycbcr_sampler &&
2362        sampler->ycbcr_sampler->chroma_filter == VK_FILTER_LINEAR) {
2363       sampler->descriptor[2] |= A6XX_TEX_SAMP_2_CHROMA_LINEAR;
2364    }
2365
2366    /* TODO:
2367     * A6XX_TEX_SAMP_1_MIPFILTER_LINEAR_FAR disables mipmapping, but vk has no NONE mipfilter?
2368     */
2369 }
2370
2371 VkResult
2372 tu_CreateSampler(VkDevice _device,
2373                  const VkSamplerCreateInfo *pCreateInfo,
2374                  const VkAllocationCallbacks *pAllocator,
2375                  VkSampler *pSampler)
2376 {
2377    TU_FROM_HANDLE(tu_device, device, _device);
2378    struct tu_sampler *sampler;
2379
2380    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO);
2381
2382    sampler = vk_alloc2(&device->alloc, pAllocator, sizeof(*sampler), 8,
2383                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2384    if (!sampler)
2385       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2386
2387    tu_init_sampler(device, sampler, pCreateInfo);
2388    *pSampler = tu_sampler_to_handle(sampler);
2389
2390    return VK_SUCCESS;
2391 }
2392
2393 void
2394 tu_DestroySampler(VkDevice _device,
2395                   VkSampler _sampler,
2396                   const VkAllocationCallbacks *pAllocator)
2397 {
2398    TU_FROM_HANDLE(tu_device, device, _device);
2399    TU_FROM_HANDLE(tu_sampler, sampler, _sampler);
2400
2401    if (!sampler)
2402       return;
2403    vk_free2(&device->alloc, pAllocator, sampler);
2404 }
2405
2406 /* vk_icd.h does not declare this function, so we declare it here to
2407  * suppress Wmissing-prototypes.
2408  */
2409 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2410 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion);
2411
2412 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2413 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion)
2414 {
2415    /* For the full details on loader interface versioning, see
2416     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2417     * What follows is a condensed summary, to help you navigate the large and
2418     * confusing official doc.
2419     *
2420     *   - Loader interface v0 is incompatible with later versions. We don't
2421     *     support it.
2422     *
2423     *   - In loader interface v1:
2424     *       - The first ICD entrypoint called by the loader is
2425     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2426     *         entrypoint.
2427     *       - The ICD must statically expose no other Vulkan symbol unless it
2428     * is linked with -Bsymbolic.
2429     *       - Each dispatchable Vulkan handle created by the ICD must be
2430     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2431     *         ICD must initialize VK_LOADER_DATA.loadMagic to
2432     * ICD_LOADER_MAGIC.
2433     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2434     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2435     *         such loader-managed surfaces.
2436     *
2437     *    - Loader interface v2 differs from v1 in:
2438     *       - The first ICD entrypoint called by the loader is
2439     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2440     *         statically expose this entrypoint.
2441     *
2442     *    - Loader interface v3 differs from v2 in:
2443     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2444     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2445     *          because the loader no longer does so.
2446     */
2447    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2448    return VK_SUCCESS;
2449 }
2450
2451 VkResult
2452 tu_GetMemoryFdKHR(VkDevice _device,
2453                   const VkMemoryGetFdInfoKHR *pGetFdInfo,
2454                   int *pFd)
2455 {
2456    TU_FROM_HANDLE(tu_device, device, _device);
2457    TU_FROM_HANDLE(tu_device_memory, memory, pGetFdInfo->memory);
2458
2459    assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2460
2461    /* At the moment, we support only the below handle types. */
2462    assert(pGetFdInfo->handleType ==
2463              VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2464           pGetFdInfo->handleType ==
2465              VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2466
2467    int prime_fd = tu_bo_export_dmabuf(device, &memory->bo);
2468    if (prime_fd < 0)
2469       return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
2470
2471    *pFd = prime_fd;
2472    return VK_SUCCESS;
2473 }
2474
2475 VkResult
2476 tu_GetMemoryFdPropertiesKHR(VkDevice _device,
2477                             VkExternalMemoryHandleTypeFlagBits handleType,
2478                             int fd,
2479                             VkMemoryFdPropertiesKHR *pMemoryFdProperties)
2480 {
2481    assert(handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2482    pMemoryFdProperties->memoryTypeBits = 1;
2483    return VK_SUCCESS;
2484 }
2485
2486 VkResult
2487 tu_ImportFenceFdKHR(VkDevice _device,
2488                     const VkImportFenceFdInfoKHR *pImportFenceFdInfo)
2489 {
2490    tu_stub();
2491
2492    return VK_SUCCESS;
2493 }
2494
2495 VkResult
2496 tu_GetFenceFdKHR(VkDevice _device,
2497                  const VkFenceGetFdInfoKHR *pGetFdInfo,
2498                  int *pFd)
2499 {
2500    tu_stub();
2501
2502    return VK_SUCCESS;
2503 }
2504
2505 VkResult
2506 tu_ImportSemaphoreFdKHR(VkDevice _device,
2507                         const VkImportSemaphoreFdInfoKHR *pImportSemaphoreFdInfo)
2508 {
2509    TU_FROM_HANDLE(tu_device, device, _device);
2510    TU_FROM_HANDLE(tu_semaphore, sem, pImportSemaphoreFdInfo->semaphore);
2511    int ret;
2512    struct tu_semaphore_part *dst = NULL;
2513
2514    if (pImportSemaphoreFdInfo->flags & VK_SEMAPHORE_IMPORT_TEMPORARY_BIT) {
2515       dst = &sem->temporary;
2516    } else {
2517       dst = &sem->permanent;
2518    }
2519
2520    uint32_t syncobj = dst->kind == TU_SEMAPHORE_SYNCOBJ ? dst->syncobj : 0;
2521
2522    switch(pImportSemaphoreFdInfo->handleType) {
2523       case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT: {
2524          uint32_t old_syncobj = syncobj;
2525          ret = drmSyncobjFDToHandle(device->physical_device->local_fd, pImportSemaphoreFdInfo->fd, &syncobj);
2526          if (ret == 0) {
2527             close(pImportSemaphoreFdInfo->fd);
2528             if (old_syncobj)
2529                drmSyncobjDestroy(device->physical_device->local_fd, old_syncobj);
2530          }
2531          break;
2532       }
2533       case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT: {
2534          if (!syncobj) {
2535             ret = drmSyncobjCreate(device->physical_device->local_fd, 0, &syncobj);
2536             if (ret)
2537                break;
2538          }
2539          if (pImportSemaphoreFdInfo->fd == -1) {
2540             ret = drmSyncobjSignal(device->physical_device->local_fd, &syncobj, 1);
2541          } else {
2542             ret = drmSyncobjImportSyncFile(device->physical_device->local_fd, syncobj, pImportSemaphoreFdInfo->fd);
2543          }
2544          if (!ret)
2545             close(pImportSemaphoreFdInfo->fd);
2546          break;
2547       }
2548       default:
2549          unreachable("Unhandled semaphore handle type");
2550    }
2551
2552    if (ret) {
2553       return VK_ERROR_INVALID_EXTERNAL_HANDLE;
2554    }
2555    dst->syncobj = syncobj;
2556    dst->kind = TU_SEMAPHORE_SYNCOBJ;
2557
2558    return VK_SUCCESS;
2559 }
2560
2561 VkResult
2562 tu_GetSemaphoreFdKHR(VkDevice _device,
2563                      const VkSemaphoreGetFdInfoKHR *pGetFdInfo,
2564                      int *pFd)
2565 {
2566    TU_FROM_HANDLE(tu_device, device, _device);
2567    TU_FROM_HANDLE(tu_semaphore, sem, pGetFdInfo->semaphore);
2568    int ret;
2569    uint32_t syncobj_handle;
2570
2571    if (sem->temporary.kind != TU_SEMAPHORE_NONE) {
2572       assert(sem->temporary.kind == TU_SEMAPHORE_SYNCOBJ);
2573       syncobj_handle = sem->temporary.syncobj;
2574    } else {
2575       assert(sem->permanent.kind == TU_SEMAPHORE_SYNCOBJ);
2576       syncobj_handle = sem->permanent.syncobj;
2577    }
2578
2579    switch(pGetFdInfo->handleType) {
2580    case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT:
2581       ret = drmSyncobjHandleToFD(device->physical_device->local_fd, syncobj_handle, pFd);
2582       break;
2583    case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT:
2584       ret = drmSyncobjExportSyncFile(device->physical_device->local_fd, syncobj_handle, pFd);
2585       if (!ret) {
2586          if (sem->temporary.kind != TU_SEMAPHORE_NONE) {
2587             tu_semaphore_part_destroy(device, &sem->temporary);
2588          } else {
2589             drmSyncobjReset(device->physical_device->local_fd, &syncobj_handle, 1);
2590          }
2591       }
2592       break;
2593    default:
2594       unreachable("Unhandled semaphore handle type");
2595    }
2596
2597    if (ret)
2598       return vk_error(device->instance, VK_ERROR_INVALID_EXTERNAL_HANDLE);
2599    return VK_SUCCESS;
2600 }
2601
2602
2603 static bool tu_has_syncobj(struct tu_physical_device *pdev)
2604 {
2605    uint64_t value;
2606    if (drmGetCap(pdev->local_fd, DRM_CAP_SYNCOBJ, &value))
2607       return false;
2608    return value && pdev->msm_major_version == 1 && pdev->msm_minor_version >= 6;
2609 }
2610
2611 void
2612 tu_GetPhysicalDeviceExternalSemaphoreProperties(
2613    VkPhysicalDevice physicalDevice,
2614    const VkPhysicalDeviceExternalSemaphoreInfo *pExternalSemaphoreInfo,
2615    VkExternalSemaphoreProperties *pExternalSemaphoreProperties)
2616 {
2617    TU_FROM_HANDLE(tu_physical_device, pdev, physicalDevice);
2618
2619    if (tu_has_syncobj(pdev) &&
2620        (pExternalSemaphoreInfo->handleType == VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT ||
2621         pExternalSemaphoreInfo->handleType == VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT)) {
2622       pExternalSemaphoreProperties->exportFromImportedHandleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT | VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT;
2623       pExternalSemaphoreProperties->compatibleHandleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT | VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT;
2624       pExternalSemaphoreProperties->externalSemaphoreFeatures = VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT |
2625          VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT;
2626    } else {
2627       pExternalSemaphoreProperties->exportFromImportedHandleTypes = 0;
2628       pExternalSemaphoreProperties->compatibleHandleTypes = 0;
2629       pExternalSemaphoreProperties->externalSemaphoreFeatures = 0;
2630    }
2631 }
2632
2633 void
2634 tu_GetPhysicalDeviceExternalFenceProperties(
2635    VkPhysicalDevice physicalDevice,
2636    const VkPhysicalDeviceExternalFenceInfo *pExternalFenceInfo,
2637    VkExternalFenceProperties *pExternalFenceProperties)
2638 {
2639    pExternalFenceProperties->exportFromImportedHandleTypes = 0;
2640    pExternalFenceProperties->compatibleHandleTypes = 0;
2641    pExternalFenceProperties->externalFenceFeatures = 0;
2642 }
2643
2644 VkResult
2645 tu_CreateDebugReportCallbackEXT(
2646    VkInstance _instance,
2647    const VkDebugReportCallbackCreateInfoEXT *pCreateInfo,
2648    const VkAllocationCallbacks *pAllocator,
2649    VkDebugReportCallbackEXT *pCallback)
2650 {
2651    TU_FROM_HANDLE(tu_instance, instance, _instance);
2652    return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2653                                           pCreateInfo, pAllocator,
2654                                           &instance->alloc, pCallback);
2655 }
2656
2657 void
2658 tu_DestroyDebugReportCallbackEXT(VkInstance _instance,
2659                                  VkDebugReportCallbackEXT _callback,
2660                                  const VkAllocationCallbacks *pAllocator)
2661 {
2662    TU_FROM_HANDLE(tu_instance, instance, _instance);
2663    vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2664                                     _callback, pAllocator, &instance->alloc);
2665 }
2666
2667 void
2668 tu_DebugReportMessageEXT(VkInstance _instance,
2669                          VkDebugReportFlagsEXT flags,
2670                          VkDebugReportObjectTypeEXT objectType,
2671                          uint64_t object,
2672                          size_t location,
2673                          int32_t messageCode,
2674                          const char *pLayerPrefix,
2675                          const char *pMessage)
2676 {
2677    TU_FROM_HANDLE(tu_instance, instance, _instance);
2678    vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2679                    object, location, messageCode, pLayerPrefix, pMessage);
2680 }
2681
2682 void
2683 tu_GetDeviceGroupPeerMemoryFeatures(
2684    VkDevice device,
2685    uint32_t heapIndex,
2686    uint32_t localDeviceIndex,
2687    uint32_t remoteDeviceIndex,
2688    VkPeerMemoryFeatureFlags *pPeerMemoryFeatures)
2689 {
2690    assert(localDeviceIndex == remoteDeviceIndex);
2691
2692    *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2693                           VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2694                           VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2695                           VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2696 }
2697
2698 void tu_GetPhysicalDeviceMultisamplePropertiesEXT(
2699    VkPhysicalDevice                            physicalDevice,
2700    VkSampleCountFlagBits                       samples,
2701    VkMultisamplePropertiesEXT*                 pMultisampleProperties)
2702 {
2703    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
2704
2705    if (samples <= VK_SAMPLE_COUNT_4_BIT && pdevice->supported_extensions.EXT_sample_locations)
2706       pMultisampleProperties->maxSampleLocationGridSize = (VkExtent2D){ 1, 1 };
2707    else
2708       pMultisampleProperties->maxSampleLocationGridSize = (VkExtent2D){ 0, 0 };
2709 }