src/intel/vulkan/anv_device.c

   1 /*
   2  * Copyright © 2015 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <assert.h>
  25 #include <stdbool.h>
  26 #include <string.h>
  27 #include <sys/mman.h>
  28 #include <sys/sysinfo.h>
  29 #include <unistd.h>
  30 #include <fcntl.h>
  31 #include <xf86drm.h>
  32
  33 #include "anv_private.h"
  34 #include "util/strtod.h"
  35 #include "util/debug.h"
  36 #include "util/build_id.h"
  37 #include "util/mesa-sha1.h"
  38 #include "vk_util.h"
  39
  40 #include "genxml/gen7_pack.h"
  41
  42 static void
  43 compiler_debug_log(void *data, const char *fmt, ...)
  44 { }
  45
  46 static void
  47 compiler_perf_log(void *data, const char *fmt, ...)
  48 {
  49    va_list args;
  50    va_start(args, fmt);
  51
  52    if (unlikely(INTEL_DEBUG & DEBUG_PERF))
  53       vfprintf(stderr, fmt, args);
  54
  55    va_end(args);
  56 }
  57
  58 static VkResult
  59 anv_compute_heap_size(int fd, uint64_t *heap_size)
  60 {
  61    uint64_t gtt_size;
  62    if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
  63                                  &gtt_size) == -1) {
  64       /* If, for whatever reason, we can't actually get the GTT size from the
  65        * kernel (too old?) fall back to the aperture size.
  66        */
  67       anv_perf_warn("Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
  68
  69       if (anv_gem_get_aperture(fd, &gtt_size) == -1) {
  70          return vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
  71                           "failed to get aperture size: %m");
  72       }
  73    }
  74
  75    /* Query the total ram from the system */
  76    struct sysinfo info;
  77    sysinfo(&info);
  78
  79    uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
  80
  81    /* We don't want to burn too much ram with the GPU.  If the user has 4GiB
  82     * or less, we use at most half.  If they have more than 4GiB, we use 3/4.
  83     */
  84    uint64_t available_ram;
  85    if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
  86       available_ram = total_ram / 2;
  87    else
  88       available_ram = total_ram * 3 / 4;
  89
  90    /* We also want to leave some padding for things we allocate in the driver,
  91     * so don't go over 3/4 of the GTT either.
  92     */
  93    uint64_t available_gtt = gtt_size * 3 / 4;
  94
  95    *heap_size = MIN2(available_ram, available_gtt);
  96
  97    return VK_SUCCESS;
  98 }
  99
 100 static VkResult
 101 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
 102 {
 103    /* The kernel query only tells us whether or not the kernel supports the
 104     * EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and not whether or not the
 105     * hardware has actual 48bit address support.
 106     */
 107    device->supports_48bit_addresses =
 108       (device->info.gen >= 8) && anv_gem_supports_48b_addresses(fd);
 109
 110    uint64_t heap_size;
 111    VkResult result = anv_compute_heap_size(fd, &heap_size);
 112    if (result != VK_SUCCESS)
 113       return result;
 114
 115    if (heap_size <= 3ull * (1ull << 30)) {
 116       /* In this case, everything fits nicely into the 32-bit address space,
 117        * so there's no need for supporting 48bit addresses on client-allocated
 118        * memory objects.
 119        */
 120       device->memory.heap_count = 1;
 121       device->memory.heaps[0] = (struct anv_memory_heap) {
 122          .size = heap_size,
 123          .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 124          .supports_48bit_addresses = false,
 125       };
 126    } else {
 127       /* Not everything will fit nicely into a 32-bit address space.  In this
 128        * case we need a 64-bit heap.  Advertise a small 32-bit heap and a
 129        * larger 48-bit heap.  If we're in this case, then we have a total heap
 130        * size larger than 3GiB which most likely means they have 8 GiB of
 131        * video memory and so carving off 1 GiB for the 32-bit heap should be
 132        * reasonable.
 133        */
 134       const uint64_t heap_size_32bit = 1ull << 30;
 135       const uint64_t heap_size_48bit = heap_size - heap_size_32bit;
 136
 137       assert(device->supports_48bit_addresses);
 138
 139       device->memory.heap_count = 2;
 140       device->memory.heaps[0] = (struct anv_memory_heap) {
 141          .size = heap_size_48bit,
 142          .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 143          .supports_48bit_addresses = true,
 144       };
 145       device->memory.heaps[1] = (struct anv_memory_heap) {
 146          .size = heap_size_32bit,
 147          .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 148          .supports_48bit_addresses = false,
 149       };
 150    }
 151
 152    uint32_t type_count = 0;
 153    for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
 154       uint32_t valid_buffer_usage = ~0;
 155
 156       /* There appears to be a hardware issue in the VF cache where it only
 157        * considers the bottom 32 bits of memory addresses.  If you happen to
 158        * have two vertex buffers which get placed exactly 4 GiB apart and use
 159        * them in back-to-back draw calls, you can get collisions.  In order to
 160        * solve this problem, we require vertex and index buffers be bound to
 161        * memory allocated out of the 32-bit heap.
 162        */
 163       if (device->memory.heaps[heap].supports_48bit_addresses) {
 164          valid_buffer_usage &= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
 165                                  VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
 166       }
 167
 168       if (device->info.has_llc) {
 169          /* Big core GPUs share LLC with the CPU and thus one memory type can be
 170           * both cached and coherent at the same time.
 171           */
 172          device->memory.types[type_count++] = (struct anv_memory_type) {
 173             .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 174                              VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 175                              VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
 176                              VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 177             .heapIndex = heap,
 178             .valid_buffer_usage = valid_buffer_usage,
 179          };
 180       } else {
 181          /* The spec requires that we expose a host-visible, coherent memory
 182           * type, but Atom GPUs don't share LLC. Thus we offer two memory types
 183           * to give the application a choice between cached, but not coherent and
 184           * coherent but uncached (WC though).
 185           */
 186          device->memory.types[type_count++] = (struct anv_memory_type) {
 187             .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 188                              VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 189                              VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
 190             .heapIndex = heap,
 191             .valid_buffer_usage = valid_buffer_usage,
 192          };
 193          device->memory.types[type_count++] = (struct anv_memory_type) {
 194             .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 195                              VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 196                              VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 197             .heapIndex = heap,
 198             .valid_buffer_usage = valid_buffer_usage,
 199          };
 200       }
 201    }
 202    device->memory.type_count = type_count;
 203
 204    return VK_SUCCESS;
 205 }
 206
 207 static VkResult
 208 anv_physical_device_init_uuids(struct anv_physical_device *device)
 209 {
 210    const struct build_id_note *note = build_id_find_nhdr("libvulkan_intel.so");
 211    if (!note) {
 212       return vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 213                        "Failed to find build-id");
 214    }
 215
 216    unsigned build_id_len = build_id_length(note);
 217    if (build_id_len < 20) {
 218       return vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 219                        "build-id too short.  It needs to be a SHA");
 220    }
 221
 222    struct mesa_sha1 sha1_ctx;
 223    uint8_t sha1[20];
 224    STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
 225
 226    /* The pipeline cache UUID is used for determining when a pipeline cache is
 227     * invalid.  It needs both a driver build and the PCI ID of the device.
 228     */
 229    _mesa_sha1_init(&sha1_ctx);
 230    _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
 231    _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
 232                      sizeof(device->chipset_id));
 233    _mesa_sha1_final(&sha1_ctx, sha1);
 234    memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
 235
 236    /* The driver UUID is used for determining sharability of images and memory
 237     * between two Vulkan instances in separate processes.  People who want to
 238     * share memory need to also check the device UUID (below) so all this
 239     * needs to be is the build-id.
 240     */
 241    memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
 242
 243    /* The device UUID uniquely identifies the given device within the machine.
 244     * Since we never have more than one device, this doesn't need to be a real
 245     * UUID.  However, on the off-chance that someone tries to use this to
 246     * cache pre-tiled images or something of the like, we use the PCI ID and
 247     * some bits of ISL info to ensure that this is safe.
 248     */
 249    _mesa_sha1_init(&sha1_ctx);
 250    _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
 251                      sizeof(device->chipset_id));
 252    _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
 253                      sizeof(device->isl_dev.has_bit6_swizzling));
 254    _mesa_sha1_final(&sha1_ctx, sha1);
 255    memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
 256
 257    return VK_SUCCESS;
 258 }
 259
 260 static VkResult
 261 anv_physical_device_init(struct anv_physical_device *device,
 262                          struct anv_instance *instance,
 263                          const char *path)
 264 {
 265    VkResult result;
 266    int fd;
 267
 268    fd = open(path, O_RDWR | O_CLOEXEC);
 269    if (fd < 0)
 270       return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
 271
 272    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 273    device->instance = instance;
 274
 275    assert(strlen(path) < ARRAY_SIZE(device->path));
 276    strncpy(device->path, path, ARRAY_SIZE(device->path));
 277
 278    device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
 279    if (!device->chipset_id) {
 280       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
 281       goto fail;
 282    }
 283
 284    device->name = gen_get_device_name(device->chipset_id);
 285    if (!gen_get_device_info(device->chipset_id, &device->info)) {
 286       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
 287       goto fail;
 288    }
 289
 290    if (device->info.is_haswell) {
 291       fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
 292    } else if (device->info.gen == 7 && !device->info.is_baytrail) {
 293       fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
 294    } else if (device->info.gen == 7 && device->info.is_baytrail) {
 295       fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
 296    } else if (device->info.gen >= 8 && device->info.gen <= 9) {
 297       /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
 298        * supported as anything */
 299    } else {
 300       result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 301                          "Vulkan not yet supported on %s", device->name);
 302       goto fail;
 303    }
 304
 305    device->cmd_parser_version = -1;
 306    if (device->info.gen == 7) {
 307       device->cmd_parser_version =
 308          anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
 309       if (device->cmd_parser_version == -1) {
 310          result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 311                             "failed to get command parser version");
 312          goto fail;
 313       }
 314    }
 315
 316    if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
 317       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 318                          "kernel missing gem wait");
 319       goto fail;
 320    }
 321
 322    if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
 323       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 324                          "kernel missing execbuf2");
 325       goto fail;
 326    }
 327
 328    if (!device->info.has_llc &&
 329        anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
 330       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 331                          "kernel missing wc mmap");
 332       goto fail;
 333    }
 334
 335    result = anv_physical_device_init_heaps(device, fd);
 336    if (result != VK_SUCCESS)
 337       goto fail;
 338
 339    device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
 340
 341    bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
 342
 343    /* GENs prior to 8 do not support EU/Subslice info */
 344    if (device->info.gen >= 8) {
 345       device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
 346       device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
 347
 348       /* Without this information, we cannot get the right Braswell
 349        * brandstrings, and we have to use conservative numbers for GPGPU on
 350        * many platforms, but otherwise, things will just work.
 351        */
 352       if (device->subslice_total < 1 || device->eu_total < 1) {
 353          fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
 354                          " query GPU properties.\n");
 355       }
 356    } else if (device->info.gen == 7) {
 357       device->subslice_total = 1 << (device->info.gt - 1);
 358    }
 359
 360    if (device->info.is_cherryview &&
 361        device->subslice_total > 0 && device->eu_total > 0) {
 362       /* Logical CS threads = EUs per subslice * num threads per EU */
 363       uint32_t max_cs_threads =
 364          device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
 365
 366       /* Fuse configurations may give more threads than expected, never less. */
 367       if (max_cs_threads > device->info.max_cs_threads)
 368          device->info.max_cs_threads = max_cs_threads;
 369    }
 370
 371    brw_process_intel_debug_variable();
 372
 373    device->compiler = brw_compiler_create(NULL, &device->info);
 374    if (device->compiler == NULL) {
 375       result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 376       goto fail;
 377    }
 378    device->compiler->shader_debug_log = compiler_debug_log;
 379    device->compiler->shader_perf_log = compiler_perf_log;
 380
 381    isl_device_init(&device->isl_dev, &device->info, swizzled);
 382
 383    result = anv_physical_device_init_uuids(device);
 384    if (result != VK_SUCCESS)
 385       goto fail;
 386
 387    result = anv_init_wsi(device);
 388    if (result != VK_SUCCESS) {
 389       ralloc_free(device->compiler);
 390       goto fail;
 391    }
 392
 393    device->local_fd = fd;
 394    return VK_SUCCESS;
 395
 396 fail:
 397    close(fd);
 398    return result;
 399 }
 400
 401 static void
 402 anv_physical_device_finish(struct anv_physical_device *device)
 403 {
 404    anv_finish_wsi(device);
 405    ralloc_free(device->compiler);
 406    close(device->local_fd);
 407 }
 408
 409 static void *
 410 default_alloc_func(void *pUserData, size_t size, size_t align,
 411                    VkSystemAllocationScope allocationScope)
 412 {
 413    return malloc(size);
 414 }
 415
 416 static void *
 417 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
 418                      size_t align, VkSystemAllocationScope allocationScope)
 419 {
 420    return realloc(pOriginal, size);
 421 }
 422
 423 static void
 424 default_free_func(void *pUserData, void *pMemory)
 425 {
 426    free(pMemory);
 427 }
 428
 429 static const VkAllocationCallbacks default_alloc = {
 430    .pUserData = NULL,
 431    .pfnAllocation = default_alloc_func,
 432    .pfnReallocation = default_realloc_func,
 433    .pfnFree = default_free_func,
 434 };
 435
 436 VkResult anv_CreateInstance(
 437     const VkInstanceCreateInfo*                 pCreateInfo,
 438     const VkAllocationCallbacks*                pAllocator,
 439     VkInstance*                                 pInstance)
 440 {
 441    struct anv_instance *instance;
 442
 443    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 444
 445    uint32_t client_version;
 446    if (pCreateInfo->pApplicationInfo &&
 447        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 448       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 449    } else {
 450       client_version = VK_MAKE_VERSION(1, 0, 0);
 451    }
 452
 453    if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
 454        client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
 455       return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 456                        "Client requested version %d.%d.%d",
 457                        VK_VERSION_MAJOR(client_version),
 458                        VK_VERSION_MINOR(client_version),
 459                        VK_VERSION_PATCH(client_version));
 460    }
 461
 462    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 463       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
 464       if (!anv_instance_extension_supported(ext_name))
 465          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 466    }
 467
 468    instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 469                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 470    if (!instance)
 471       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 472
 473    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 474
 475    if (pAllocator)
 476       instance->alloc = *pAllocator;
 477    else
 478       instance->alloc = default_alloc;
 479
 480    instance->apiVersion = client_version;
 481    instance->physicalDeviceCount = -1;
 482
 483    _mesa_locale_init();
 484
 485    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 486
 487    *pInstance = anv_instance_to_handle(instance);
 488
 489    return VK_SUCCESS;
 490 }
 491
 492 void anv_DestroyInstance(
 493     VkInstance                                  _instance,
 494     const VkAllocationCallbacks*                pAllocator)
 495 {
 496    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 497
 498    if (!instance)
 499       return;
 500
 501    if (instance->physicalDeviceCount > 0) {
 502       /* We support at most one physical device. */
 503       assert(instance->physicalDeviceCount == 1);
 504       anv_physical_device_finish(&instance->physicalDevice);
 505    }
 506
 507    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 508
 509    _mesa_locale_fini();
 510
 511    vk_free(&instance->alloc, instance);
 512 }
 513
 514 static VkResult
 515 anv_enumerate_devices(struct anv_instance *instance)
 516 {
 517    /* TODO: Check for more devices ? */
 518    drmDevicePtr devices[8];
 519    VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
 520    int max_devices;
 521
 522    instance->physicalDeviceCount = 0;
 523
 524    max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
 525    if (max_devices < 1)
 526       return VK_ERROR_INCOMPATIBLE_DRIVER;
 527
 528    for (unsigned i = 0; i < (unsigned)max_devices; i++) {
 529       if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
 530           devices[i]->bustype == DRM_BUS_PCI &&
 531           devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
 532
 533          result = anv_physical_device_init(&instance->physicalDevice,
 534                         instance,
 535                         devices[i]->nodes[DRM_NODE_RENDER]);
 536          if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 537             break;
 538       }
 539    }
 540    drmFreeDevices(devices, max_devices);
 541
 542    if (result == VK_SUCCESS)
 543       instance->physicalDeviceCount = 1;
 544
 545    return result;
 546 }
 547
 548
 549 VkResult anv_EnumeratePhysicalDevices(
 550     VkInstance                                  _instance,
 551     uint32_t*                                   pPhysicalDeviceCount,
 552     VkPhysicalDevice*                           pPhysicalDevices)
 553 {
 554    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 555    VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
 556    VkResult result;
 557
 558    if (instance->physicalDeviceCount < 0) {
 559       result = anv_enumerate_devices(instance);
 560       if (result != VK_SUCCESS &&
 561           result != VK_ERROR_INCOMPATIBLE_DRIVER)
 562          return result;
 563    }
 564
 565    if (instance->physicalDeviceCount > 0) {
 566       assert(instance->physicalDeviceCount == 1);
 567       vk_outarray_append(&out, i) {
 568          *i = anv_physical_device_to_handle(&instance->physicalDevice);
 569       }
 570    }
 571
 572    return vk_outarray_status(&out);
 573 }
 574
 575 void anv_GetPhysicalDeviceFeatures(
 576     VkPhysicalDevice                            physicalDevice,
 577     VkPhysicalDeviceFeatures*                   pFeatures)
 578 {
 579    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 580
 581    *pFeatures = (VkPhysicalDeviceFeatures) {
 582       .robustBufferAccess                       = true,
 583       .fullDrawIndexUint32                      = true,
 584       .imageCubeArray                           = true,
 585       .independentBlend                         = true,
 586       .geometryShader                           = true,
 587       .tessellationShader                       = true,
 588       .sampleRateShading                        = true,
 589       .dualSrcBlend                             = true,
 590       .logicOp                                  = true,
 591       .multiDrawIndirect                        = true,
 592       .drawIndirectFirstInstance                = true,
 593       .depthClamp                               = true,
 594       .depthBiasClamp                           = true,
 595       .fillModeNonSolid                         = true,
 596       .depthBounds                              = false,
 597       .wideLines                                = true,
 598       .largePoints                              = true,
 599       .alphaToOne                               = true,
 600       .multiViewport                            = true,
 601       .samplerAnisotropy                        = true,
 602       .textureCompressionETC2                   = pdevice->info.gen >= 8 ||
 603                                                   pdevice->info.is_baytrail,
 604       .textureCompressionASTC_LDR               = pdevice->info.gen >= 9, /* FINISHME CHV */
 605       .textureCompressionBC                     = true,
 606       .occlusionQueryPrecise                    = true,
 607       .pipelineStatisticsQuery                  = true,
 608       .fragmentStoresAndAtomics                 = true,
 609       .shaderTessellationAndGeometryPointSize   = true,
 610       .shaderImageGatherExtended                = true,
 611       .shaderStorageImageExtendedFormats        = true,
 612       .shaderStorageImageMultisample            = false,
 613       .shaderStorageImageReadWithoutFormat      = false,
 614       .shaderStorageImageWriteWithoutFormat     = true,
 615       .shaderUniformBufferArrayDynamicIndexing  = true,
 616       .shaderSampledImageArrayDynamicIndexing   = true,
 617       .shaderStorageBufferArrayDynamicIndexing  = true,
 618       .shaderStorageImageArrayDynamicIndexing   = true,
 619       .shaderClipDistance                       = true,
 620       .shaderCullDistance                       = true,
 621       .shaderFloat64                            = pdevice->info.gen >= 8,
 622       .shaderInt64                              = pdevice->info.gen >= 8,
 623       .shaderInt16                              = false,
 624       .shaderResourceMinLod                     = false,
 625       .variableMultisampleRate                  = false,
 626       .inheritedQueries                         = true,
 627    };
 628
 629    /* We can't do image stores in vec4 shaders */
 630    pFeatures->vertexPipelineStoresAndAtomics =
 631       pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
 632       pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
 633 }
 634
 635 void anv_GetPhysicalDeviceFeatures2KHR(
 636     VkPhysicalDevice                            physicalDevice,
 637     VkPhysicalDeviceFeatures2KHR*               pFeatures)
 638 {
 639    anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 640
 641    vk_foreach_struct(ext, pFeatures->pNext) {
 642       switch (ext->sType) {
 643       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX: {
 644          VkPhysicalDeviceMultiviewFeaturesKHX *features =
 645             (VkPhysicalDeviceMultiviewFeaturesKHX *)ext;
 646          features->multiview = true;
 647          features->multiviewGeometryShader = true;
 648          features->multiviewTessellationShader = true;
 649          break;
 650       }
 651
 652       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR: {
 653          VkPhysicalDeviceVariablePointerFeaturesKHR *features = (void *)ext;
 654          features->variablePointersStorageBuffer = true;
 655          features->variablePointers = false;
 656          break;
 657       }
 658
 659       default:
 660          anv_debug_ignored_stype(ext->sType);
 661          break;
 662       }
 663    }
 664 }
 665
 666 void anv_GetPhysicalDeviceProperties(
 667     VkPhysicalDevice                            physicalDevice,
 668     VkPhysicalDeviceProperties*                 pProperties)
 669 {
 670    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 671    const struct gen_device_info *devinfo = &pdevice->info;
 672
 673    /* See assertions made when programming the buffer surface state. */
 674    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
 675                                       (1ul << 30) : (1ul << 27);
 676
 677    const uint32_t max_samplers = (devinfo->gen >= 8 || devinfo->is_haswell) ?
 678                                  128 : 16;
 679
 680    VkSampleCountFlags sample_counts =
 681       isl_device_get_sample_counts(&pdevice->isl_dev);
 682
 683    VkPhysicalDeviceLimits limits = {
 684       .maxImageDimension1D                      = (1 << 14),
 685       .maxImageDimension2D                      = (1 << 14),
 686       .maxImageDimension3D                      = (1 << 11),
 687       .maxImageDimensionCube                    = (1 << 14),
 688       .maxImageArrayLayers                      = (1 << 11),
 689       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 690       .maxUniformBufferRange                    = (1ul << 27),
 691       .maxStorageBufferRange                    = max_raw_buffer_sz,
 692       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 693       .maxMemoryAllocationCount                 = UINT32_MAX,
 694       .maxSamplerAllocationCount                = 64 * 1024,
 695       .bufferImageGranularity                   = 64, /* A cache line */
 696       .sparseAddressSpaceSize                   = 0,
 697       .maxBoundDescriptorSets                   = MAX_SETS,
 698       .maxPerStageDescriptorSamplers            = max_samplers,
 699       .maxPerStageDescriptorUniformBuffers      = 64,
 700       .maxPerStageDescriptorStorageBuffers      = 64,
 701       .maxPerStageDescriptorSampledImages       = max_samplers,
 702       .maxPerStageDescriptorStorageImages       = 64,
 703       .maxPerStageDescriptorInputAttachments    = 64,
 704       .maxPerStageResources                     = 250,
 705       .maxDescriptorSetSamplers                 = 256,
 706       .maxDescriptorSetUniformBuffers           = 256,
 707       .maxDescriptorSetUniformBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
 708       .maxDescriptorSetStorageBuffers           = 256,
 709       .maxDescriptorSetStorageBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
 710       .maxDescriptorSetSampledImages            = 256,
 711       .maxDescriptorSetStorageImages            = 256,
 712       .maxDescriptorSetInputAttachments         = 256,
 713       .maxVertexInputAttributes                 = MAX_VBS,
 714       .maxVertexInputBindings                   = MAX_VBS,
 715       .maxVertexInputAttributeOffset            = 2047,
 716       .maxVertexInputBindingStride              = 2048,
 717       .maxVertexOutputComponents                = 128,
 718       .maxTessellationGenerationLevel           = 64,
 719       .maxTessellationPatchSize                 = 32,
 720       .maxTessellationControlPerVertexInputComponents = 128,
 721       .maxTessellationControlPerVertexOutputComponents = 128,
 722       .maxTessellationControlPerPatchOutputComponents = 128,
 723       .maxTessellationControlTotalOutputComponents = 2048,
 724       .maxTessellationEvaluationInputComponents = 128,
 725       .maxTessellationEvaluationOutputComponents = 128,
 726       .maxGeometryShaderInvocations             = 32,
 727       .maxGeometryInputComponents               = 64,
 728       .maxGeometryOutputComponents              = 128,
 729       .maxGeometryOutputVertices                = 256,
 730       .maxGeometryTotalOutputComponents         = 1024,
 731       .maxFragmentInputComponents               = 128,
 732       .maxFragmentOutputAttachments             = 8,
 733       .maxFragmentDualSrcAttachments            = 1,
 734       .maxFragmentCombinedOutputResources       = 8,
 735       .maxComputeSharedMemorySize               = 32768,
 736       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 737       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 738       .maxComputeWorkGroupSize = {
 739          16 * devinfo->max_cs_threads,
 740          16 * devinfo->max_cs_threads,
 741          16 * devinfo->max_cs_threads,
 742       },
 743       .subPixelPrecisionBits                    = 4 /* FIXME */,
 744       .subTexelPrecisionBits                    = 4 /* FIXME */,
 745       .mipmapPrecisionBits                      = 4 /* FIXME */,
 746       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 747       .maxDrawIndirectCount                     = UINT32_MAX,
 748       .maxSamplerLodBias                        = 16,
 749       .maxSamplerAnisotropy                     = 16,
 750       .maxViewports                             = MAX_VIEWPORTS,
 751       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 752       .viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
 753       .viewportSubPixelBits                     = 13, /* We take a float? */
 754       .minMemoryMapAlignment                    = 4096, /* A page */
 755       .minTexelBufferOffsetAlignment            = 1,
 756       .minUniformBufferOffsetAlignment          = 16,
 757       .minStorageBufferOffsetAlignment          = 4,
 758       .minTexelOffset                           = -8,
 759       .maxTexelOffset                           = 7,
 760       .minTexelGatherOffset                     = -32,
 761       .maxTexelGatherOffset                     = 31,
 762       .minInterpolationOffset                   = -0.5,
 763       .maxInterpolationOffset                   = 0.4375,
 764       .subPixelInterpolationOffsetBits          = 4,
 765       .maxFramebufferWidth                      = (1 << 14),
 766       .maxFramebufferHeight                     = (1 << 14),
 767       .maxFramebufferLayers                     = (1 << 11),
 768       .framebufferColorSampleCounts             = sample_counts,
 769       .framebufferDepthSampleCounts             = sample_counts,
 770       .framebufferStencilSampleCounts           = sample_counts,
 771       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 772       .maxColorAttachments                      = MAX_RTS,
 773       .sampledImageColorSampleCounts            = sample_counts,
 774       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 775       .sampledImageDepthSampleCounts            = sample_counts,
 776       .sampledImageStencilSampleCounts          = sample_counts,
 777       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 778       .maxSampleMaskWords                       = 1,
 779       .timestampComputeAndGraphics              = false,
 780       .timestampPeriod                          = 1000000000.0 / devinfo->timestamp_frequency,
 781       .maxClipDistances                         = 8,
 782       .maxCullDistances                         = 8,
 783       .maxCombinedClipAndCullDistances          = 8,
 784       .discreteQueuePriorities                  = 1,
 785       .pointSizeRange                           = { 0.125, 255.875 },
 786       .lineWidthRange                           = { 0.0, 7.9921875 },
 787       .pointSizeGranularity                     = (1.0 / 8.0),
 788       .lineWidthGranularity                     = (1.0 / 128.0),
 789       .strictLines                              = false, /* FINISHME */
 790       .standardSampleLocations                  = true,
 791       .optimalBufferCopyOffsetAlignment         = 128,
 792       .optimalBufferCopyRowPitchAlignment       = 128,
 793       .nonCoherentAtomSize                      = 64,
 794    };
 795
 796    *pProperties = (VkPhysicalDeviceProperties) {
 797       .apiVersion = anv_physical_device_api_version(pdevice),
 798       .driverVersion = vk_get_driver_version(),
 799       .vendorID = 0x8086,
 800       .deviceID = pdevice->chipset_id,
 801       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 802       .limits = limits,
 803       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 804    };
 805
 806    snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
 807             "%s", pdevice->name);
 808    memcpy(pProperties->pipelineCacheUUID,
 809           pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
 810 }
 811
 812 void anv_GetPhysicalDeviceProperties2KHR(
 813     VkPhysicalDevice                            physicalDevice,
 814     VkPhysicalDeviceProperties2KHR*             pProperties)
 815 {
 816    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 817
 818    anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 819
 820    vk_foreach_struct(ext, pProperties->pNext) {
 821       switch (ext->sType) {
 822       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 823          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 824             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 825
 826          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 827          break;
 828       }
 829
 830       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR: {
 831          VkPhysicalDeviceIDPropertiesKHR *id_props =
 832             (VkPhysicalDeviceIDPropertiesKHR *)ext;
 833          memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
 834          memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
 835          /* The LUID is for Windows. */
 836          id_props->deviceLUIDValid = false;
 837          break;
 838       }
 839
 840       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX: {
 841          VkPhysicalDeviceMultiviewPropertiesKHX *properties =
 842             (VkPhysicalDeviceMultiviewPropertiesKHX *)ext;
 843          properties->maxMultiviewViewCount = 16;
 844          properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
 845          break;
 846       }
 847
 848       default:
 849          anv_debug_ignored_stype(ext->sType);
 850          break;
 851       }
 852    }
 853 }
 854
 855 /* We support exactly one queue family. */
 856 static const VkQueueFamilyProperties
 857 anv_queue_family_properties = {
 858    .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 859                  VK_QUEUE_COMPUTE_BIT |
 860                  VK_QUEUE_TRANSFER_BIT,
 861    .queueCount = 1,
 862    .timestampValidBits = 36, /* XXX: Real value here */
 863    .minImageTransferGranularity = { 1, 1, 1 },
 864 };
 865
 866 void anv_GetPhysicalDeviceQueueFamilyProperties(
 867     VkPhysicalDevice                            physicalDevice,
 868     uint32_t*                                   pCount,
 869     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 870 {
 871    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
 872
 873    vk_outarray_append(&out, p) {
 874       *p = anv_queue_family_properties;
 875    }
 876 }
 877
 878 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
 879     VkPhysicalDevice                            physicalDevice,
 880     uint32_t*                                   pQueueFamilyPropertyCount,
 881     VkQueueFamilyProperties2KHR*                pQueueFamilyProperties)
 882 {
 883
 884    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 885
 886    vk_outarray_append(&out, p) {
 887       p->queueFamilyProperties = anv_queue_family_properties;
 888
 889       vk_foreach_struct(s, p->pNext) {
 890          anv_debug_ignored_stype(s->sType);
 891       }
 892    }
 893 }
 894
 895 void anv_GetPhysicalDeviceMemoryProperties(
 896     VkPhysicalDevice                            physicalDevice,
 897     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
 898 {
 899    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 900
 901    pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
 902    for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
 903       pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
 904          .propertyFlags = physical_device->memory.types[i].propertyFlags,
 905          .heapIndex     = physical_device->memory.types[i].heapIndex,
 906       };
 907    }
 908
 909    pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
 910    for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
 911       pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
 912          .size    = physical_device->memory.heaps[i].size,
 913          .flags   = physical_device->memory.heaps[i].flags,
 914       };
 915    }
 916 }
 917
 918 void anv_GetPhysicalDeviceMemoryProperties2KHR(
 919     VkPhysicalDevice                            physicalDevice,
 920     VkPhysicalDeviceMemoryProperties2KHR*       pMemoryProperties)
 921 {
 922    anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
 923                                          &pMemoryProperties->memoryProperties);
 924
 925    vk_foreach_struct(ext, pMemoryProperties->pNext) {
 926       switch (ext->sType) {
 927       default:
 928          anv_debug_ignored_stype(ext->sType);
 929          break;
 930       }
 931    }
 932 }
 933
 934 PFN_vkVoidFunction anv_GetInstanceProcAddr(
 935     VkInstance                                  instance,
 936     const char*                                 pName)
 937 {
 938    return anv_lookup_entrypoint(NULL, pName);
 939 }
 940
 941 /* With version 1+ of the loader interface the ICD should expose
 942  * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
 943  */
 944 PUBLIC
 945 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 946     VkInstance                                  instance,
 947     const char*                                 pName);
 948
 949 PUBLIC
 950 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 951     VkInstance                                  instance,
 952     const char*                                 pName)
 953 {
 954    return anv_GetInstanceProcAddr(instance, pName);
 955 }
 956
 957 PFN_vkVoidFunction anv_GetDeviceProcAddr(
 958     VkDevice                                    _device,
 959     const char*                                 pName)
 960 {
 961    ANV_FROM_HANDLE(anv_device, device, _device);
 962    return anv_lookup_entrypoint(&device->info, pName);
 963 }
 964
 965 static void
 966 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
 967 {
 968    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 969    queue->device = device;
 970    queue->pool = &device->surface_state_pool;
 971 }
 972
 973 static void
 974 anv_queue_finish(struct anv_queue *queue)
 975 {
 976 }
 977
 978 static struct anv_state
 979 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
 980 {
 981    struct anv_state state;
 982
 983    state = anv_state_pool_alloc(pool, size, align);
 984    memcpy(state.map, p, size);
 985
 986    anv_state_flush(pool->block_pool.device, state);
 987
 988    return state;
 989 }
 990
 991 struct gen8_border_color {
 992    union {
 993       float float32[4];
 994       uint32_t uint32[4];
 995    };
 996    /* Pad out to 64 bytes */
 997    uint32_t _pad[12];
 998 };
 999
1000 static void
1001 anv_device_init_border_colors(struct anv_device *device)
1002 {
1003    static const struct gen8_border_color border_colors[] = {
1004       [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] =  { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
1005       [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] =       { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
1006       [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] =       { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
1007       [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] =    { .uint32 = { 0, 0, 0, 0 } },
1008       [VK_BORDER_COLOR_INT_OPAQUE_BLACK] =         { .uint32 = { 0, 0, 0, 1 } },
1009       [VK_BORDER_COLOR_INT_OPAQUE_WHITE] =         { .uint32 = { 1, 1, 1, 1 } },
1010    };
1011
1012    device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
1013                                                     sizeof(border_colors), 64,
1014                                                     border_colors);
1015 }
1016
1017 static void
1018 anv_device_init_trivial_batch(struct anv_device *device)
1019 {
1020    anv_bo_init_new(&device->trivial_batch_bo, device, 4096);
1021
1022    if (device->instance->physicalDevice.has_exec_async)
1023       device->trivial_batch_bo.flags |= EXEC_OBJECT_ASYNC;
1024
1025    void *map = anv_gem_mmap(device, device->trivial_batch_bo.gem_handle,
1026                             0, 4096, 0);
1027
1028    struct anv_batch batch = {
1029       .start = map,
1030       .next = map,
1031       .end = map + 4096,
1032    };
1033
1034    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1035    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1036
1037    if (!device->info.has_llc)
1038       gen_clflush_range(map, batch.next - map);
1039
1040    anv_gem_munmap(map, device->trivial_batch_bo.size);
1041 }
1042
1043 VkResult anv_CreateDevice(
1044     VkPhysicalDevice                            physicalDevice,
1045     const VkDeviceCreateInfo*                   pCreateInfo,
1046     const VkAllocationCallbacks*                pAllocator,
1047     VkDevice*                                   pDevice)
1048 {
1049    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1050    VkResult result;
1051    struct anv_device *device;
1052
1053    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
1054
1055    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1056       const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
1057       if (!anv_physical_device_extension_supported(physical_device, ext_name))
1058          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1059    }
1060
1061    /* Check enabled features */
1062    if (pCreateInfo->pEnabledFeatures) {
1063       VkPhysicalDeviceFeatures supported_features;
1064       anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1065       VkBool32 *supported_feature = (VkBool32 *)&supported_features;
1066       VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
1067       unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1068       for (uint32_t i = 0; i < num_features; i++) {
1069          if (enabled_feature[i] && !supported_feature[i])
1070             return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1071       }
1072    }
1073
1074    device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
1075                        sizeof(*device), 8,
1076                        VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1077    if (!device)
1078       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1079
1080    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1081    device->instance = physical_device->instance;
1082    device->chipset_id = physical_device->chipset_id;
1083    device->lost = false;
1084
1085    if (pAllocator)
1086       device->alloc = *pAllocator;
1087    else
1088       device->alloc = physical_device->instance->alloc;
1089
1090    /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1091    device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1092    if (device->fd == -1) {
1093       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1094       goto fail_device;
1095    }
1096
1097    device->context_id = anv_gem_create_context(device);
1098    if (device->context_id == -1) {
1099       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1100       goto fail_fd;
1101    }
1102
1103    device->info = physical_device->info;
1104    device->isl_dev = physical_device->isl_dev;
1105
1106    /* On Broadwell and later, we can use batch chaining to more efficiently
1107     * implement growing command buffers.  Prior to Haswell, the kernel
1108     * command parser gets in the way and we have to fall back to growing
1109     * the batch.
1110     */
1111    device->can_chain_batches = device->info.gen >= 8;
1112
1113    device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1114       pCreateInfo->pEnabledFeatures->robustBufferAccess;
1115
1116    if (pthread_mutex_init(&device->mutex, NULL) != 0) {
1117       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1118       goto fail_context_id;
1119    }
1120
1121    pthread_condattr_t condattr;
1122    if (pthread_condattr_init(&condattr) != 0) {
1123       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1124       goto fail_mutex;
1125    }
1126    if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
1127       pthread_condattr_destroy(&condattr);
1128       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1129       goto fail_mutex;
1130    }
1131    if (pthread_cond_init(&device->queue_submit, NULL) != 0) {
1132       pthread_condattr_destroy(&condattr);
1133       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1134       goto fail_mutex;
1135    }
1136    pthread_condattr_destroy(&condattr);
1137
1138    anv_bo_pool_init(&device->batch_bo_pool, device);
1139
1140    result = anv_bo_cache_init(&device->bo_cache);
1141    if (result != VK_SUCCESS)
1142       goto fail_batch_bo_pool;
1143
1144    result = anv_state_pool_init(&device->dynamic_state_pool, device, 16384);
1145    if (result != VK_SUCCESS)
1146       goto fail_bo_cache;
1147
1148    result = anv_state_pool_init(&device->instruction_state_pool, device, 16384);
1149    if (result != VK_SUCCESS)
1150       goto fail_dynamic_state_pool;
1151
1152    result = anv_state_pool_init(&device->surface_state_pool, device, 4096);
1153    if (result != VK_SUCCESS)
1154       goto fail_instruction_state_pool;
1155
1156    result = anv_bo_init_new(&device->workaround_bo, device, 1024);
1157    if (result != VK_SUCCESS)
1158       goto fail_surface_state_pool;
1159
1160    anv_device_init_trivial_batch(device);
1161
1162    anv_scratch_pool_init(device, &device->scratch_pool);
1163
1164    anv_queue_init(device, &device->queue);
1165
1166    switch (device->info.gen) {
1167    case 7:
1168       if (!device->info.is_haswell)
1169          result = gen7_init_device_state(device);
1170       else
1171          result = gen75_init_device_state(device);
1172       break;
1173    case 8:
1174       result = gen8_init_device_state(device);
1175       break;
1176    case 9:
1177       result = gen9_init_device_state(device);
1178       break;
1179    case 10:
1180       result = gen10_init_device_state(device);
1181       break;
1182    default:
1183       /* Shouldn't get here as we don't create physical devices for any other
1184        * gens. */
1185       unreachable("unhandled gen");
1186    }
1187    if (result != VK_SUCCESS)
1188       goto fail_workaround_bo;
1189
1190    anv_device_init_blorp(device);
1191
1192    anv_device_init_border_colors(device);
1193
1194    *pDevice = anv_device_to_handle(device);
1195
1196    return VK_SUCCESS;
1197
1198  fail_workaround_bo:
1199    anv_queue_finish(&device->queue);
1200    anv_scratch_pool_finish(device, &device->scratch_pool);
1201    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1202    anv_gem_close(device, device->workaround_bo.gem_handle);
1203  fail_surface_state_pool:
1204    anv_state_pool_finish(&device->surface_state_pool);
1205  fail_instruction_state_pool:
1206    anv_state_pool_finish(&device->instruction_state_pool);
1207  fail_dynamic_state_pool:
1208    anv_state_pool_finish(&device->dynamic_state_pool);
1209  fail_bo_cache:
1210    anv_bo_cache_finish(&device->bo_cache);
1211  fail_batch_bo_pool:
1212    anv_bo_pool_finish(&device->batch_bo_pool);
1213    pthread_cond_destroy(&device->queue_submit);
1214  fail_mutex:
1215    pthread_mutex_destroy(&device->mutex);
1216  fail_context_id:
1217    anv_gem_destroy_context(device, device->context_id);
1218  fail_fd:
1219    close(device->fd);
1220  fail_device:
1221    vk_free(&device->alloc, device);
1222
1223    return result;
1224 }
1225
1226 void anv_DestroyDevice(
1227     VkDevice                                    _device,
1228     const VkAllocationCallbacks*                pAllocator)
1229 {
1230    ANV_FROM_HANDLE(anv_device, device, _device);
1231
1232    if (!device)
1233       return;
1234
1235    anv_device_finish_blorp(device);
1236
1237    anv_queue_finish(&device->queue);
1238
1239 #ifdef HAVE_VALGRIND
1240    /* We only need to free these to prevent valgrind errors.  The backing
1241     * BO will go away in a couple of lines so we don't actually leak.
1242     */
1243    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1244 #endif
1245
1246    anv_scratch_pool_finish(device, &device->scratch_pool);
1247
1248    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1249    anv_gem_close(device, device->workaround_bo.gem_handle);
1250
1251    anv_gem_close(device, device->trivial_batch_bo.gem_handle);
1252
1253    anv_state_pool_finish(&device->surface_state_pool);
1254    anv_state_pool_finish(&device->instruction_state_pool);
1255    anv_state_pool_finish(&device->dynamic_state_pool);
1256
1257    anv_bo_cache_finish(&device->bo_cache);
1258
1259    anv_bo_pool_finish(&device->batch_bo_pool);
1260
1261    pthread_cond_destroy(&device->queue_submit);
1262    pthread_mutex_destroy(&device->mutex);
1263
1264    anv_gem_destroy_context(device, device->context_id);
1265
1266    close(device->fd);
1267
1268    vk_free(&device->alloc, device);
1269 }
1270
1271 VkResult anv_EnumerateInstanceLayerProperties(
1272     uint32_t*                                   pPropertyCount,
1273     VkLayerProperties*                          pProperties)
1274 {
1275    if (pProperties == NULL) {
1276       *pPropertyCount = 0;
1277       return VK_SUCCESS;
1278    }
1279
1280    /* None supported at this time */
1281    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1282 }
1283
1284 VkResult anv_EnumerateDeviceLayerProperties(
1285     VkPhysicalDevice                            physicalDevice,
1286     uint32_t*                                   pPropertyCount,
1287     VkLayerProperties*                          pProperties)
1288 {
1289    if (pProperties == NULL) {
1290       *pPropertyCount = 0;
1291       return VK_SUCCESS;
1292    }
1293
1294    /* None supported at this time */
1295    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1296 }
1297
1298 void anv_GetDeviceQueue(
1299     VkDevice                                    _device,
1300     uint32_t                                    queueNodeIndex,
1301     uint32_t                                    queueIndex,
1302     VkQueue*                                    pQueue)
1303 {
1304    ANV_FROM_HANDLE(anv_device, device, _device);
1305
1306    assert(queueIndex == 0);
1307
1308    *pQueue = anv_queue_to_handle(&device->queue);
1309 }
1310
1311 VkResult
1312 anv_device_query_status(struct anv_device *device)
1313 {
1314    /* This isn't likely as most of the callers of this function already check
1315     * for it.  However, it doesn't hurt to check and it potentially lets us
1316     * avoid an ioctl.
1317     */
1318    if (unlikely(device->lost))
1319       return VK_ERROR_DEVICE_LOST;
1320
1321    uint32_t active, pending;
1322    int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
1323    if (ret == -1) {
1324       /* We don't know the real error. */
1325       device->lost = true;
1326       return vk_errorf(VK_ERROR_DEVICE_LOST, "get_reset_stats failed: %m");
1327    }
1328
1329    if (active) {
1330       device->lost = true;
1331       return vk_errorf(VK_ERROR_DEVICE_LOST,
1332                        "GPU hung on one of our command buffers");
1333    } else if (pending) {
1334       device->lost = true;
1335       return vk_errorf(VK_ERROR_DEVICE_LOST,
1336                        "GPU hung with commands in-flight");
1337    }
1338
1339    return VK_SUCCESS;
1340 }
1341
1342 VkResult
1343 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
1344 {
1345    /* Note:  This only returns whether or not the BO is in use by an i915 GPU.
1346     * Other usages of the BO (such as on different hardware) will not be
1347     * flagged as "busy" by this ioctl.  Use with care.
1348     */
1349    int ret = anv_gem_busy(device, bo->gem_handle);
1350    if (ret == 1) {
1351       return VK_NOT_READY;
1352    } else if (ret == -1) {
1353       /* We don't know the real error. */
1354       device->lost = true;
1355       return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1356    }
1357
1358    /* Query for device status after the busy call.  If the BO we're checking
1359     * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1360     * client because it clearly doesn't have valid data.  Yes, this most
1361     * likely means an ioctl, but we just did an ioctl to query the busy status
1362     * so it's no great loss.
1363     */
1364    return anv_device_query_status(device);
1365 }
1366
1367 VkResult
1368 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
1369                 int64_t timeout)
1370 {
1371    int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
1372    if (ret == -1 && errno == ETIME) {
1373       return VK_TIMEOUT;
1374    } else if (ret == -1) {
1375       /* We don't know the real error. */
1376       device->lost = true;
1377       return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1378    }
1379
1380    /* Query for device status after the wait.  If the BO we're waiting on got
1381     * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1382     * because it clearly doesn't have valid data.  Yes, this most likely means
1383     * an ioctl, but we just did an ioctl to wait so it's no great loss.
1384     */
1385    return anv_device_query_status(device);
1386 }
1387
1388 VkResult anv_DeviceWaitIdle(
1389     VkDevice                                    _device)
1390 {
1391    ANV_FROM_HANDLE(anv_device, device, _device);
1392    if (unlikely(device->lost))
1393       return VK_ERROR_DEVICE_LOST;
1394
1395    struct anv_batch batch;
1396
1397    uint32_t cmds[8];
1398    batch.start = batch.next = cmds;
1399    batch.end = (void *) cmds + sizeof(cmds);
1400
1401    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1402    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1403
1404    return anv_device_submit_simple_batch(device, &batch);
1405 }
1406
1407 VkResult
1408 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1409 {
1410    uint32_t gem_handle = anv_gem_create(device, size);
1411    if (!gem_handle)
1412       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1413
1414    anv_bo_init(bo, gem_handle, size);
1415
1416    return VK_SUCCESS;
1417 }
1418
1419 VkResult anv_AllocateMemory(
1420     VkDevice                                    _device,
1421     const VkMemoryAllocateInfo*                 pAllocateInfo,
1422     const VkAllocationCallbacks*                pAllocator,
1423     VkDeviceMemory*                             pMem)
1424 {
1425    ANV_FROM_HANDLE(anv_device, device, _device);
1426    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
1427    struct anv_device_memory *mem;
1428    VkResult result = VK_SUCCESS;
1429
1430    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1431
1432    /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1433    assert(pAllocateInfo->allocationSize > 0);
1434
1435    /* The kernel relocation API has a limitation of a 32-bit delta value
1436     * applied to the address before it is written which, in spite of it being
1437     * unsigned, is treated as signed .  Because of the way that this maps to
1438     * the Vulkan API, we cannot handle an offset into a buffer that does not
1439     * fit into a signed 32 bits.  The only mechanism we have for dealing with
1440     * this at the moment is to limit all VkDeviceMemory objects to a maximum
1441     * of 2GB each.  The Vulkan spec allows us to do this:
1442     *
1443     *    "Some platforms may have a limit on the maximum size of a single
1444     *    allocation. For example, certain systems may fail to create
1445     *    allocations with a size greater than or equal to 4GB. Such a limit is
1446     *    implementation-dependent, and if such a failure occurs then the error
1447     *    VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1448     *
1449     * We don't use vk_error here because it's not an error so much as an
1450     * indication to the application that the allocation is too large.
1451     */
1452    if (pAllocateInfo->allocationSize > (1ull << 31))
1453       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
1454
1455    /* FINISHME: Fail if allocation request exceeds heap size. */
1456
1457    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1458                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1459    if (mem == NULL)
1460       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1461
1462    assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
1463    mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
1464    mem->map = NULL;
1465    mem->map_size = 0;
1466
1467    const VkImportMemoryFdInfoKHR *fd_info =
1468       vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
1469
1470    /* The Vulkan spec permits handleType to be 0, in which case the struct is
1471     * ignored.
1472     */
1473    if (fd_info && fd_info->handleType) {
1474       /* At the moment, we only support the OPAQUE_FD memory type which is
1475        * just a GEM buffer.
1476        */
1477       assert(fd_info->handleType ==
1478              VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR);
1479
1480       result = anv_bo_cache_import(device, &device->bo_cache,
1481                                    fd_info->fd, pAllocateInfo->allocationSize,
1482                                    &mem->bo);
1483       if (result != VK_SUCCESS)
1484          goto fail;
1485    } else {
1486       result = anv_bo_cache_alloc(device, &device->bo_cache,
1487                                   pAllocateInfo->allocationSize,
1488                                   &mem->bo);
1489       if (result != VK_SUCCESS)
1490          goto fail;
1491    }
1492
1493    assert(mem->type->heapIndex < pdevice->memory.heap_count);
1494    if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
1495       mem->bo->flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
1496
1497    if (pdevice->has_exec_async)
1498       mem->bo->flags |= EXEC_OBJECT_ASYNC;
1499
1500    *pMem = anv_device_memory_to_handle(mem);
1501
1502    return VK_SUCCESS;
1503
1504  fail:
1505    vk_free2(&device->alloc, pAllocator, mem);
1506
1507    return result;
1508 }
1509
1510 VkResult anv_GetMemoryFdKHR(
1511     VkDevice                                    device_h,
1512     const VkMemoryGetFdInfoKHR*                 pGetFdInfo,
1513     int*                                        pFd)
1514 {
1515    ANV_FROM_HANDLE(anv_device, dev, device_h);
1516    ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
1517
1518    assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
1519
1520    /* We support only one handle type. */
1521    assert(pGetFdInfo->handleType ==
1522           VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR);
1523
1524    return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
1525 }
1526
1527 VkResult anv_GetMemoryFdPropertiesKHR(
1528     VkDevice                                    device_h,
1529     VkExternalMemoryHandleTypeFlagBitsKHR       handleType,
1530     int                                         fd,
1531     VkMemoryFdPropertiesKHR*                    pMemoryFdProperties)
1532 {
1533    /* The valid usage section for this function says:
1534     *
1535     *    "handleType must not be one of the handle types defined as opaque."
1536     *
1537     * Since we only handle opaque handles for now, there are no FD properties.
1538     */
1539    return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR;
1540 }
1541
1542 void anv_FreeMemory(
1543     VkDevice                                    _device,
1544     VkDeviceMemory                              _mem,
1545     const VkAllocationCallbacks*                pAllocator)
1546 {
1547    ANV_FROM_HANDLE(anv_device, device, _device);
1548    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1549
1550    if (mem == NULL)
1551       return;
1552
1553    if (mem->map)
1554       anv_UnmapMemory(_device, _mem);
1555
1556    anv_bo_cache_release(device, &device->bo_cache, mem->bo);
1557
1558    vk_free2(&device->alloc, pAllocator, mem);
1559 }
1560
1561 VkResult anv_MapMemory(
1562     VkDevice                                    _device,
1563     VkDeviceMemory                              _memory,
1564     VkDeviceSize                                offset,
1565     VkDeviceSize                                size,
1566     VkMemoryMapFlags                            flags,
1567     void**                                      ppData)
1568 {
1569    ANV_FROM_HANDLE(anv_device, device, _device);
1570    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1571
1572    if (mem == NULL) {
1573       *ppData = NULL;
1574       return VK_SUCCESS;
1575    }
1576
1577    if (size == VK_WHOLE_SIZE)
1578       size = mem->bo->size - offset;
1579
1580    /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1581     *
1582     *  * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1583     *    assert(size != 0);
1584     *  * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1585     *    equal to the size of the memory minus offset
1586     */
1587    assert(size > 0);
1588    assert(offset + size <= mem->bo->size);
1589
1590    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1591     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1592     * at a time is valid. We could just mmap up front and return an offset
1593     * pointer here, but that may exhaust virtual memory on 32 bit
1594     * userspace. */
1595
1596    uint32_t gem_flags = 0;
1597
1598    if (!device->info.has_llc &&
1599        (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
1600       gem_flags |= I915_MMAP_WC;
1601
1602    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1603    uint64_t map_offset = offset & ~4095ull;
1604    assert(offset >= map_offset);
1605    uint64_t map_size = (offset + size) - map_offset;
1606
1607    /* Let's map whole pages */
1608    map_size = align_u64(map_size, 4096);
1609
1610    void *map = anv_gem_mmap(device, mem->bo->gem_handle,
1611                             map_offset, map_size, gem_flags);
1612    if (map == MAP_FAILED)
1613       return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
1614
1615    mem->map = map;
1616    mem->map_size = map_size;
1617
1618    *ppData = mem->map + (offset - map_offset);
1619
1620    return VK_SUCCESS;
1621 }
1622
1623 void anv_UnmapMemory(
1624     VkDevice                                    _device,
1625     VkDeviceMemory                              _memory)
1626 {
1627    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1628
1629    if (mem == NULL)
1630       return;
1631
1632    anv_gem_munmap(mem->map, mem->map_size);
1633
1634    mem->map = NULL;
1635    mem->map_size = 0;
1636 }
1637
1638 static void
1639 clflush_mapped_ranges(struct anv_device         *device,
1640                       uint32_t                   count,
1641                       const VkMappedMemoryRange *ranges)
1642 {
1643    for (uint32_t i = 0; i < count; i++) {
1644       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1645       if (ranges[i].offset >= mem->map_size)
1646          continue;
1647
1648       gen_clflush_range(mem->map + ranges[i].offset,
1649                         MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
1650    }
1651 }
1652
1653 VkResult anv_FlushMappedMemoryRanges(
1654     VkDevice                                    _device,
1655     uint32_t                                    memoryRangeCount,
1656     const VkMappedMemoryRange*                  pMemoryRanges)
1657 {
1658    ANV_FROM_HANDLE(anv_device, device, _device);
1659
1660    if (device->info.has_llc)
1661       return VK_SUCCESS;
1662
1663    /* Make sure the writes we're flushing have landed. */
1664    __builtin_ia32_mfence();
1665
1666    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1667
1668    return VK_SUCCESS;
1669 }
1670
1671 VkResult anv_InvalidateMappedMemoryRanges(
1672     VkDevice                                    _device,
1673     uint32_t                                    memoryRangeCount,
1674     const VkMappedMemoryRange*                  pMemoryRanges)
1675 {
1676    ANV_FROM_HANDLE(anv_device, device, _device);
1677
1678    if (device->info.has_llc)
1679       return VK_SUCCESS;
1680
1681    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1682
1683    /* Make sure no reads get moved up above the invalidate. */
1684    __builtin_ia32_mfence();
1685
1686    return VK_SUCCESS;
1687 }
1688
1689 void anv_GetBufferMemoryRequirements(
1690     VkDevice                                    _device,
1691     VkBuffer                                    _buffer,
1692     VkMemoryRequirements*                       pMemoryRequirements)
1693 {
1694    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1695    ANV_FROM_HANDLE(anv_device, device, _device);
1696    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
1697
1698    /* The Vulkan spec (git aaed022) says:
1699     *
1700     *    memoryTypeBits is a bitfield and contains one bit set for every
1701     *    supported memory type for the resource. The bit `1<<i` is set if and
1702     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1703     *    structure for the physical device is supported.
1704     */
1705    uint32_t memory_types = 0;
1706    for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
1707       uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
1708       if ((valid_usage & buffer->usage) == buffer->usage)
1709          memory_types |= (1u << i);
1710    }
1711
1712    pMemoryRequirements->size = buffer->size;
1713    pMemoryRequirements->alignment = 16;
1714    pMemoryRequirements->memoryTypeBits = memory_types;
1715 }
1716
1717 void anv_GetBufferMemoryRequirements2KHR(
1718     VkDevice                                    _device,
1719     const VkBufferMemoryRequirementsInfo2KHR*   pInfo,
1720     VkMemoryRequirements2KHR*                   pMemoryRequirements)
1721 {
1722    anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
1723                                    &pMemoryRequirements->memoryRequirements);
1724
1725    vk_foreach_struct(ext, pMemoryRequirements->pNext) {
1726       switch (ext->sType) {
1727       case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR: {
1728          VkMemoryDedicatedRequirementsKHR *requirements = (void *)ext;
1729          requirements->prefersDedicatedAllocation = VK_FALSE;
1730          requirements->requiresDedicatedAllocation = VK_FALSE;
1731          break;
1732       }
1733
1734       default:
1735          anv_debug_ignored_stype(ext->sType);
1736          break;
1737       }
1738    }
1739 }
1740
1741 void anv_GetImageMemoryRequirements(
1742     VkDevice                                    _device,
1743     VkImage                                     _image,
1744     VkMemoryRequirements*                       pMemoryRequirements)
1745 {
1746    ANV_FROM_HANDLE(anv_image, image, _image);
1747    ANV_FROM_HANDLE(anv_device, device, _device);
1748    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
1749
1750    /* The Vulkan spec (git aaed022) says:
1751     *
1752     *    memoryTypeBits is a bitfield and contains one bit set for every
1753     *    supported memory type for the resource. The bit `1<<i` is set if and
1754     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1755     *    structure for the physical device is supported.
1756     *
1757     * All types are currently supported for images.
1758     */
1759    uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
1760
1761    pMemoryRequirements->size = image->size;
1762    pMemoryRequirements->alignment = image->alignment;
1763    pMemoryRequirements->memoryTypeBits = memory_types;
1764 }
1765
1766 void anv_GetImageMemoryRequirements2KHR(
1767     VkDevice                                    _device,
1768     const VkImageMemoryRequirementsInfo2KHR*    pInfo,
1769     VkMemoryRequirements2KHR*                   pMemoryRequirements)
1770 {
1771    anv_GetImageMemoryRequirements(_device, pInfo->image,
1772                                   &pMemoryRequirements->memoryRequirements);
1773
1774    vk_foreach_struct(ext, pMemoryRequirements->pNext) {
1775       switch (ext->sType) {
1776       case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR: {
1777          VkMemoryDedicatedRequirementsKHR *requirements = (void *)ext;
1778          requirements->prefersDedicatedAllocation = VK_FALSE;
1779          requirements->requiresDedicatedAllocation = VK_FALSE;
1780          break;
1781       }
1782
1783       default:
1784          anv_debug_ignored_stype(ext->sType);
1785          break;
1786       }
1787    }
1788 }
1789
1790 void anv_GetImageSparseMemoryRequirements(
1791     VkDevice                                    device,
1792     VkImage                                     image,
1793     uint32_t*                                   pSparseMemoryRequirementCount,
1794     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1795 {
1796    *pSparseMemoryRequirementCount = 0;
1797 }
1798
1799 void anv_GetImageSparseMemoryRequirements2KHR(
1800     VkDevice                                    device,
1801     const VkImageSparseMemoryRequirementsInfo2KHR* pInfo,
1802     uint32_t*                                   pSparseMemoryRequirementCount,
1803     VkSparseImageMemoryRequirements2KHR*        pSparseMemoryRequirements)
1804 {
1805    *pSparseMemoryRequirementCount = 0;
1806 }
1807
1808 void anv_GetDeviceMemoryCommitment(
1809     VkDevice                                    device,
1810     VkDeviceMemory                              memory,
1811     VkDeviceSize*                               pCommittedMemoryInBytes)
1812 {
1813    *pCommittedMemoryInBytes = 0;
1814 }
1815
1816 VkResult anv_BindBufferMemory(
1817     VkDevice                                    device,
1818     VkBuffer                                    _buffer,
1819     VkDeviceMemory                              _memory,
1820     VkDeviceSize                                memoryOffset)
1821 {
1822    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1823    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1824
1825    if (mem) {
1826       assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
1827       buffer->bo = mem->bo;
1828       buffer->offset = memoryOffset;
1829    } else {
1830       buffer->bo = NULL;
1831       buffer->offset = 0;
1832    }
1833
1834    return VK_SUCCESS;
1835 }
1836
1837 VkResult anv_QueueBindSparse(
1838     VkQueue                                     _queue,
1839     uint32_t                                    bindInfoCount,
1840     const VkBindSparseInfo*                     pBindInfo,
1841     VkFence                                     fence)
1842 {
1843    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1844    if (unlikely(queue->device->lost))
1845       return VK_ERROR_DEVICE_LOST;
1846
1847    return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1848 }
1849
1850 // Event functions
1851
1852 VkResult anv_CreateEvent(
1853     VkDevice                                    _device,
1854     const VkEventCreateInfo*                    pCreateInfo,
1855     const VkAllocationCallbacks*                pAllocator,
1856     VkEvent*                                    pEvent)
1857 {
1858    ANV_FROM_HANDLE(anv_device, device, _device);
1859    struct anv_state state;
1860    struct anv_event *event;
1861
1862    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1863
1864    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1865                                 sizeof(*event), 8);
1866    event = state.map;
1867    event->state = state;
1868    event->semaphore = VK_EVENT_RESET;
1869
1870    if (!device->info.has_llc) {
1871       /* Make sure the writes we're flushing have landed. */
1872       __builtin_ia32_mfence();
1873       __builtin_ia32_clflush(event);
1874    }
1875
1876    *pEvent = anv_event_to_handle(event);
1877
1878    return VK_SUCCESS;
1879 }
1880
1881 void anv_DestroyEvent(
1882     VkDevice                                    _device,
1883     VkEvent                                     _event,
1884     const VkAllocationCallbacks*                pAllocator)
1885 {
1886    ANV_FROM_HANDLE(anv_device, device, _device);
1887    ANV_FROM_HANDLE(anv_event, event, _event);
1888
1889    if (!event)
1890       return;
1891
1892    anv_state_pool_free(&device->dynamic_state_pool, event->state);
1893 }
1894
1895 VkResult anv_GetEventStatus(
1896     VkDevice                                    _device,
1897     VkEvent                                     _event)
1898 {
1899    ANV_FROM_HANDLE(anv_device, device, _device);
1900    ANV_FROM_HANDLE(anv_event, event, _event);
1901
1902    if (unlikely(device->lost))
1903       return VK_ERROR_DEVICE_LOST;
1904
1905    if (!device->info.has_llc) {
1906       /* Invalidate read cache before reading event written by GPU. */
1907       __builtin_ia32_clflush(event);
1908       __builtin_ia32_mfence();
1909
1910    }
1911
1912    return event->semaphore;
1913 }
1914
1915 VkResult anv_SetEvent(
1916     VkDevice                                    _device,
1917     VkEvent                                     _event)
1918 {
1919    ANV_FROM_HANDLE(anv_device, device, _device);
1920    ANV_FROM_HANDLE(anv_event, event, _event);
1921
1922    event->semaphore = VK_EVENT_SET;
1923
1924    if (!device->info.has_llc) {
1925       /* Make sure the writes we're flushing have landed. */
1926       __builtin_ia32_mfence();
1927       __builtin_ia32_clflush(event);
1928    }
1929
1930    return VK_SUCCESS;
1931 }
1932
1933 VkResult anv_ResetEvent(
1934     VkDevice                                    _device,
1935     VkEvent                                     _event)
1936 {
1937    ANV_FROM_HANDLE(anv_device, device, _device);
1938    ANV_FROM_HANDLE(anv_event, event, _event);
1939
1940    event->semaphore = VK_EVENT_RESET;
1941
1942    if (!device->info.has_llc) {
1943       /* Make sure the writes we're flushing have landed. */
1944       __builtin_ia32_mfence();
1945       __builtin_ia32_clflush(event);
1946    }
1947
1948    return VK_SUCCESS;
1949 }
1950
1951 // Buffer functions
1952
1953 VkResult anv_CreateBuffer(
1954     VkDevice                                    _device,
1955     const VkBufferCreateInfo*                   pCreateInfo,
1956     const VkAllocationCallbacks*                pAllocator,
1957     VkBuffer*                                   pBuffer)
1958 {
1959    ANV_FROM_HANDLE(anv_device, device, _device);
1960    struct anv_buffer *buffer;
1961
1962    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1963
1964    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1965                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1966    if (buffer == NULL)
1967       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1968
1969    buffer->size = pCreateInfo->size;
1970    buffer->usage = pCreateInfo->usage;
1971    buffer->bo = NULL;
1972    buffer->offset = 0;
1973
1974    *pBuffer = anv_buffer_to_handle(buffer);
1975
1976    return VK_SUCCESS;
1977 }
1978
1979 void anv_DestroyBuffer(
1980     VkDevice                                    _device,
1981     VkBuffer                                    _buffer,
1982     const VkAllocationCallbacks*                pAllocator)
1983 {
1984    ANV_FROM_HANDLE(anv_device, device, _device);
1985    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1986
1987    if (!buffer)
1988       return;
1989
1990    vk_free2(&device->alloc, pAllocator, buffer);
1991 }
1992
1993 void
1994 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
1995                               enum isl_format format,
1996                               uint32_t offset, uint32_t range, uint32_t stride)
1997 {
1998    isl_buffer_fill_state(&device->isl_dev, state.map,
1999                          .address = offset,
2000                          .mocs = device->default_mocs,
2001                          .size = range,
2002                          .format = format,
2003                          .stride = stride);
2004
2005    anv_state_flush(device, state);
2006 }
2007
2008 void anv_DestroySampler(
2009     VkDevice                                    _device,
2010     VkSampler                                   _sampler,
2011     const VkAllocationCallbacks*                pAllocator)
2012 {
2013    ANV_FROM_HANDLE(anv_device, device, _device);
2014    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
2015
2016    if (!sampler)
2017       return;
2018
2019    vk_free2(&device->alloc, pAllocator, sampler);
2020 }
2021
2022 VkResult anv_CreateFramebuffer(
2023     VkDevice                                    _device,
2024     const VkFramebufferCreateInfo*              pCreateInfo,
2025     const VkAllocationCallbacks*                pAllocator,
2026     VkFramebuffer*                              pFramebuffer)
2027 {
2028    ANV_FROM_HANDLE(anv_device, device, _device);
2029    struct anv_framebuffer *framebuffer;
2030
2031    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2032
2033    size_t size = sizeof(*framebuffer) +
2034                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2035    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2036                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2037    if (framebuffer == NULL)
2038       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2039
2040    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2041    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2042       VkImageView _iview = pCreateInfo->pAttachments[i];
2043       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2044    }
2045
2046    framebuffer->width = pCreateInfo->width;
2047    framebuffer->height = pCreateInfo->height;
2048    framebuffer->layers = pCreateInfo->layers;
2049
2050    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2051
2052    return VK_SUCCESS;
2053 }
2054
2055 void anv_DestroyFramebuffer(
2056     VkDevice                                    _device,
2057     VkFramebuffer                               _fb,
2058     const VkAllocationCallbacks*                pAllocator)
2059 {
2060    ANV_FROM_HANDLE(anv_device, device, _device);
2061    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2062
2063    if (!fb)
2064       return;
2065
2066    vk_free2(&device->alloc, pAllocator, fb);
2067 }
2068
2069 /* vk_icd.h does not declare this function, so we declare it here to
2070  * suppress Wmissing-prototypes.
2071  */
2072 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2073 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2074
2075 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2076 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2077 {
2078    /* For the full details on loader interface versioning, see
2079     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2080     * What follows is a condensed summary, to help you navigate the large and
2081     * confusing official doc.
2082     *
2083     *   - Loader interface v0 is incompatible with later versions. We don't
2084     *     support it.
2085     *
2086     *   - In loader interface v1:
2087     *       - The first ICD entrypoint called by the loader is
2088     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2089     *         entrypoint.
2090     *       - The ICD must statically expose no other Vulkan symbol unless it is
2091     *         linked with -Bsymbolic.
2092     *       - Each dispatchable Vulkan handle created by the ICD must be
2093     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2094     *         ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2095     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2096     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2097     *         such loader-managed surfaces.
2098     *
2099     *    - Loader interface v2 differs from v1 in:
2100     *       - The first ICD entrypoint called by the loader is
2101     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2102     *         statically expose this entrypoint.
2103     *
2104     *    - Loader interface v3 differs from v2 in:
2105     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2106     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2107     *          because the loader no longer does so.
2108     */
2109    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2110    return VK_SUCCESS;
2111 }