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 <unistd.h>
  29 #include <fcntl.h>
  30
  31 #include "anv_private.h"
  32 #include "util/strtod.h"
  33 #include "util/debug.h"
  34 #include "util/build_id.h"
  35 #include "util/vk_util.h"
  36
  37 #include "genxml/gen7_pack.h"
  38
  39 static void
  40 compiler_debug_log(void *data, const char *fmt, ...)
  41 { }
  42
  43 static void
  44 compiler_perf_log(void *data, const char *fmt, ...)
  45 {
  46    va_list args;
  47    va_start(args, fmt);
  48
  49    if (unlikely(INTEL_DEBUG & DEBUG_PERF))
  50       vfprintf(stderr, fmt, args);
  51
  52    va_end(args);
  53 }
  54
  55 static bool
  56 anv_device_get_cache_uuid(void *uuid)
  57 {
  58    const struct build_id_note *note = build_id_find_nhdr("libvulkan_intel.so");
  59    if (!note)
  60       return false;
  61
  62    unsigned len = build_id_length(note);
  63    if (len < VK_UUID_SIZE)
  64       return false;
  65
  66    build_id_read(note, uuid, VK_UUID_SIZE);
  67    return true;
  68 }
  69
  70 static VkResult
  71 anv_physical_device_init(struct anv_physical_device *device,
  72                          struct anv_instance *instance,
  73                          const char *path)
  74 {
  75    VkResult result;
  76    int fd;
  77
  78    fd = open(path, O_RDWR | O_CLOEXEC);
  79    if (fd < 0)
  80       return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  81
  82    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
  83    device->instance = instance;
  84
  85    assert(strlen(path) < ARRAY_SIZE(device->path));
  86    strncpy(device->path, path, ARRAY_SIZE(device->path));
  87
  88    device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
  89    if (!device->chipset_id) {
  90       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  91       goto fail;
  92    }
  93
  94    device->name = gen_get_device_name(device->chipset_id);
  95    if (!gen_get_device_info(device->chipset_id, &device->info)) {
  96       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  97       goto fail;
  98    }
  99
 100    if (device->info.is_haswell) {
 101       fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
 102    } else if (device->info.gen == 7 && !device->info.is_baytrail) {
 103       fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
 104    } else if (device->info.gen == 7 && device->info.is_baytrail) {
 105       fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
 106    } else if (device->info.gen >= 8) {
 107       /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
 108        * supported as anything */
 109    } else {
 110       result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 111                          "Vulkan not yet supported on %s", device->name);
 112       goto fail;
 113    }
 114
 115    device->cmd_parser_version = -1;
 116    if (device->info.gen == 7) {
 117       device->cmd_parser_version =
 118          anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
 119       if (device->cmd_parser_version == -1) {
 120          result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 121                             "failed to get command parser version");
 122          goto fail;
 123       }
 124    }
 125
 126    if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
 127       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 128                          "failed to get aperture size: %m");
 129       goto fail;
 130    }
 131
 132    if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
 133       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 134                          "kernel missing gem wait");
 135       goto fail;
 136    }
 137
 138    if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
 139       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 140                          "kernel missing execbuf2");
 141       goto fail;
 142    }
 143
 144    if (!device->info.has_llc &&
 145        anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
 146       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 147                          "kernel missing wc mmap");
 148       goto fail;
 149    }
 150
 151    if (!anv_device_get_cache_uuid(device->uuid)) {
 152       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 153                          "cannot generate UUID");
 154       goto fail;
 155    }
 156    bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
 157
 158    /* GENs prior to 8 do not support EU/Subslice info */
 159    if (device->info.gen >= 8) {
 160       device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
 161       device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
 162
 163       /* Without this information, we cannot get the right Braswell
 164        * brandstrings, and we have to use conservative numbers for GPGPU on
 165        * many platforms, but otherwise, things will just work.
 166        */
 167       if (device->subslice_total < 1 || device->eu_total < 1) {
 168          fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
 169                          " query GPU properties.\n");
 170       }
 171    } else if (device->info.gen == 7) {
 172       device->subslice_total = 1 << (device->info.gt - 1);
 173    }
 174
 175    if (device->info.is_cherryview &&
 176        device->subslice_total > 0 && device->eu_total > 0) {
 177       /* Logical CS threads = EUs per subslice * 7 threads per EU */
 178       uint32_t max_cs_threads = device->eu_total / device->subslice_total * 7;
 179
 180       /* Fuse configurations may give more threads than expected, never less. */
 181       if (max_cs_threads > device->info.max_cs_threads)
 182          device->info.max_cs_threads = max_cs_threads;
 183    }
 184
 185    brw_process_intel_debug_variable();
 186
 187    device->compiler = brw_compiler_create(NULL, &device->info);
 188    if (device->compiler == NULL) {
 189       result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 190       goto fail;
 191    }
 192    device->compiler->shader_debug_log = compiler_debug_log;
 193    device->compiler->shader_perf_log = compiler_perf_log;
 194
 195    result = anv_init_wsi(device);
 196    if (result != VK_SUCCESS) {
 197       ralloc_free(device->compiler);
 198       goto fail;
 199    }
 200
 201    isl_device_init(&device->isl_dev, &device->info, swizzled);
 202
 203    device->local_fd = fd;
 204    return VK_SUCCESS;
 205
 206 fail:
 207    close(fd);
 208    return result;
 209 }
 210
 211 static void
 212 anv_physical_device_finish(struct anv_physical_device *device)
 213 {
 214    anv_finish_wsi(device);
 215    ralloc_free(device->compiler);
 216    close(device->local_fd);
 217 }
 218
 219 static const VkExtensionProperties global_extensions[] = {
 220    {
 221       .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
 222       .specVersion = 25,
 223    },
 224 #ifdef VK_USE_PLATFORM_XCB_KHR
 225    {
 226       .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
 227       .specVersion = 6,
 228    },
 229 #endif
 230 #ifdef VK_USE_PLATFORM_XLIB_KHR
 231    {
 232       .extensionName = VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
 233       .specVersion = 6,
 234    },
 235 #endif
 236 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
 237    {
 238       .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
 239       .specVersion = 5,
 240    },
 241 #endif
 242    {
 243       .extensionName = VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
 244       .specVersion = 1,
 245    },
 246 };
 247
 248 static const VkExtensionProperties device_extensions[] = {
 249    {
 250       .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
 251       .specVersion = 68,
 252    },
 253    {
 254       .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
 255       .specVersion = 1,
 256    },
 257    {
 258       .extensionName = VK_KHR_MAINTENANCE1_EXTENSION_NAME,
 259       .specVersion = 1,
 260    },
 261    {
 262       .extensionName = VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
 263       .specVersion = 1,
 264    }
 265 };
 266
 267 static void *
 268 default_alloc_func(void *pUserData, size_t size, size_t align,
 269                    VkSystemAllocationScope allocationScope)
 270 {
 271    return malloc(size);
 272 }
 273
 274 static void *
 275 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
 276                      size_t align, VkSystemAllocationScope allocationScope)
 277 {
 278    return realloc(pOriginal, size);
 279 }
 280
 281 static void
 282 default_free_func(void *pUserData, void *pMemory)
 283 {
 284    free(pMemory);
 285 }
 286
 287 static const VkAllocationCallbacks default_alloc = {
 288    .pUserData = NULL,
 289    .pfnAllocation = default_alloc_func,
 290    .pfnReallocation = default_realloc_func,
 291    .pfnFree = default_free_func,
 292 };
 293
 294 VkResult anv_CreateInstance(
 295     const VkInstanceCreateInfo*                 pCreateInfo,
 296     const VkAllocationCallbacks*                pAllocator,
 297     VkInstance*                                 pInstance)
 298 {
 299    struct anv_instance *instance;
 300
 301    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 302
 303    uint32_t client_version;
 304    if (pCreateInfo->pApplicationInfo &&
 305        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 306       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 307    } else {
 308       client_version = VK_MAKE_VERSION(1, 0, 0);
 309    }
 310
 311    if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
 312        client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
 313       return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 314                        "Client requested version %d.%d.%d",
 315                        VK_VERSION_MAJOR(client_version),
 316                        VK_VERSION_MINOR(client_version),
 317                        VK_VERSION_PATCH(client_version));
 318    }
 319
 320    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 321       bool found = false;
 322       for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
 323          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 324                     global_extensions[j].extensionName) == 0) {
 325             found = true;
 326             break;
 327          }
 328       }
 329       if (!found)
 330          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 331    }
 332
 333    instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 334                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 335    if (!instance)
 336       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 337
 338    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 339
 340    if (pAllocator)
 341       instance->alloc = *pAllocator;
 342    else
 343       instance->alloc = default_alloc;
 344
 345    instance->apiVersion = client_version;
 346    instance->physicalDeviceCount = -1;
 347
 348    _mesa_locale_init();
 349
 350    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 351
 352    *pInstance = anv_instance_to_handle(instance);
 353
 354    return VK_SUCCESS;
 355 }
 356
 357 void anv_DestroyInstance(
 358     VkInstance                                  _instance,
 359     const VkAllocationCallbacks*                pAllocator)
 360 {
 361    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 362
 363    if (!instance)
 364       return;
 365
 366    if (instance->physicalDeviceCount > 0) {
 367       /* We support at most one physical device. */
 368       assert(instance->physicalDeviceCount == 1);
 369       anv_physical_device_finish(&instance->physicalDevice);
 370    }
 371
 372    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 373
 374    _mesa_locale_fini();
 375
 376    vk_free(&instance->alloc, instance);
 377 }
 378
 379 VkResult anv_EnumeratePhysicalDevices(
 380     VkInstance                                  _instance,
 381     uint32_t*                                   pPhysicalDeviceCount,
 382     VkPhysicalDevice*                           pPhysicalDevices)
 383 {
 384    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 385    VkResult result;
 386
 387    if (instance->physicalDeviceCount < 0) {
 388       char path[20];
 389       for (unsigned i = 0; i < 8; i++) {
 390          snprintf(path, sizeof(path), "/dev/dri/renderD%d", 128 + i);
 391          result = anv_physical_device_init(&instance->physicalDevice,
 392                                            instance, path);
 393          if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 394             break;
 395       }
 396
 397       if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
 398          instance->physicalDeviceCount = 0;
 399       } else if (result == VK_SUCCESS) {
 400          instance->physicalDeviceCount = 1;
 401       } else {
 402          return result;
 403       }
 404    }
 405
 406    /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
 407     * otherwise it's an inout parameter.
 408     *
 409     * The Vulkan spec (git aaed022) says:
 410     *
 411     *    pPhysicalDeviceCount is a pointer to an unsigned integer variable
 412     *    that is initialized with the number of devices the application is
 413     *    prepared to receive handles to. pname:pPhysicalDevices is pointer to
 414     *    an array of at least this many VkPhysicalDevice handles [...].
 415     *
 416     *    Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
 417     *    overwrites the contents of the variable pointed to by
 418     *    pPhysicalDeviceCount with the number of physical devices in in the
 419     *    instance; otherwise, vkEnumeratePhysicalDevices overwrites
 420     *    pPhysicalDeviceCount with the number of physical handles written to
 421     *    pPhysicalDevices.
 422     */
 423    if (!pPhysicalDevices) {
 424       *pPhysicalDeviceCount = instance->physicalDeviceCount;
 425    } else if (*pPhysicalDeviceCount >= 1) {
 426       pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
 427       *pPhysicalDeviceCount = 1;
 428    } else if (*pPhysicalDeviceCount < instance->physicalDeviceCount) {
 429       return VK_INCOMPLETE;
 430    } else {
 431       *pPhysicalDeviceCount = 0;
 432    }
 433
 434    return VK_SUCCESS;
 435 }
 436
 437 void anv_GetPhysicalDeviceFeatures(
 438     VkPhysicalDevice                            physicalDevice,
 439     VkPhysicalDeviceFeatures*                   pFeatures)
 440 {
 441    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 442
 443    *pFeatures = (VkPhysicalDeviceFeatures) {
 444       .robustBufferAccess                       = true,
 445       .fullDrawIndexUint32                      = true,
 446       .imageCubeArray                           = true,
 447       .independentBlend                         = true,
 448       .geometryShader                           = true,
 449       .tessellationShader                       = true,
 450       .sampleRateShading                        = true,
 451       .dualSrcBlend                             = true,
 452       .logicOp                                  = true,
 453       .multiDrawIndirect                        = false,
 454       .drawIndirectFirstInstance                = true,
 455       .depthClamp                               = true,
 456       .depthBiasClamp                           = true,
 457       .fillModeNonSolid                         = true,
 458       .depthBounds                              = false,
 459       .wideLines                                = true,
 460       .largePoints                              = true,
 461       .alphaToOne                               = true,
 462       .multiViewport                            = true,
 463       .samplerAnisotropy                        = true,
 464       .textureCompressionETC2                   = pdevice->info.gen >= 8 ||
 465                                                   pdevice->info.is_baytrail,
 466       .textureCompressionASTC_LDR               = pdevice->info.gen >= 9, /* FINISHME CHV */
 467       .textureCompressionBC                     = true,
 468       .occlusionQueryPrecise                    = true,
 469       .pipelineStatisticsQuery                  = false,
 470       .fragmentStoresAndAtomics                 = true,
 471       .shaderTessellationAndGeometryPointSize   = true,
 472       .shaderImageGatherExtended                = true,
 473       .shaderStorageImageExtendedFormats        = true,
 474       .shaderStorageImageMultisample            = false,
 475       .shaderStorageImageReadWithoutFormat      = false,
 476       .shaderStorageImageWriteWithoutFormat     = true,
 477       .shaderUniformBufferArrayDynamicIndexing  = true,
 478       .shaderSampledImageArrayDynamicIndexing   = true,
 479       .shaderStorageBufferArrayDynamicIndexing  = true,
 480       .shaderStorageImageArrayDynamicIndexing   = true,
 481       .shaderClipDistance                       = true,
 482       .shaderCullDistance                       = true,
 483       .shaderFloat64                            = pdevice->info.gen >= 8,
 484       .shaderInt64                              = false,
 485       .shaderInt16                              = false,
 486       .shaderResourceMinLod                     = false,
 487       .variableMultisampleRate                  = false,
 488       .inheritedQueries                         = false,
 489    };
 490
 491    /* We can't do image stores in vec4 shaders */
 492    pFeatures->vertexPipelineStoresAndAtomics =
 493       pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
 494       pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
 495 }
 496
 497 void anv_GetPhysicalDeviceFeatures2KHR(
 498     VkPhysicalDevice                            physicalDevice,
 499     VkPhysicalDeviceFeatures2KHR*               pFeatures)
 500 {
 501    anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 502
 503    vk_foreach_struct(ext, pFeatures->pNext) {
 504       switch (ext->sType) {
 505       default:
 506          anv_debug_ignored_stype(ext->sType);
 507          break;
 508       }
 509    }
 510 }
 511
 512 void anv_GetPhysicalDeviceProperties(
 513     VkPhysicalDevice                            physicalDevice,
 514     VkPhysicalDeviceProperties*                 pProperties)
 515 {
 516    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 517    const struct gen_device_info *devinfo = &pdevice->info;
 518
 519    const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
 520
 521    /* See assertions made when programming the buffer surface state. */
 522    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
 523                                       (1ul << 30) : (1ul << 27);
 524
 525    VkSampleCountFlags sample_counts =
 526       isl_device_get_sample_counts(&pdevice->isl_dev);
 527
 528    VkPhysicalDeviceLimits limits = {
 529       .maxImageDimension1D                      = (1 << 14),
 530       .maxImageDimension2D                      = (1 << 14),
 531       .maxImageDimension3D                      = (1 << 11),
 532       .maxImageDimensionCube                    = (1 << 14),
 533       .maxImageArrayLayers                      = (1 << 11),
 534       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 535       .maxUniformBufferRange                    = (1ul << 27),
 536       .maxStorageBufferRange                    = max_raw_buffer_sz,
 537       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 538       .maxMemoryAllocationCount                 = UINT32_MAX,
 539       .maxSamplerAllocationCount                = 64 * 1024,
 540       .bufferImageGranularity                   = 64, /* A cache line */
 541       .sparseAddressSpaceSize                   = 0,
 542       .maxBoundDescriptorSets                   = MAX_SETS,
 543       .maxPerStageDescriptorSamplers            = 64,
 544       .maxPerStageDescriptorUniformBuffers      = 64,
 545       .maxPerStageDescriptorStorageBuffers      = 64,
 546       .maxPerStageDescriptorSampledImages       = 64,
 547       .maxPerStageDescriptorStorageImages       = 64,
 548       .maxPerStageDescriptorInputAttachments    = 64,
 549       .maxPerStageResources                     = 128,
 550       .maxDescriptorSetSamplers                 = 256,
 551       .maxDescriptorSetUniformBuffers           = 256,
 552       .maxDescriptorSetUniformBuffersDynamic    = 256,
 553       .maxDescriptorSetStorageBuffers           = 256,
 554       .maxDescriptorSetStorageBuffersDynamic    = 256,
 555       .maxDescriptorSetSampledImages            = 256,
 556       .maxDescriptorSetStorageImages            = 256,
 557       .maxDescriptorSetInputAttachments         = 256,
 558       .maxVertexInputAttributes                 = MAX_VBS,
 559       .maxVertexInputBindings                   = MAX_VBS,
 560       .maxVertexInputAttributeOffset            = 2047,
 561       .maxVertexInputBindingStride              = 2048,
 562       .maxVertexOutputComponents                = 128,
 563       .maxTessellationGenerationLevel           = 64,
 564       .maxTessellationPatchSize                 = 32,
 565       .maxTessellationControlPerVertexInputComponents = 128,
 566       .maxTessellationControlPerVertexOutputComponents = 128,
 567       .maxTessellationControlPerPatchOutputComponents = 128,
 568       .maxTessellationControlTotalOutputComponents = 2048,
 569       .maxTessellationEvaluationInputComponents = 128,
 570       .maxTessellationEvaluationOutputComponents = 128,
 571       .maxGeometryShaderInvocations             = 32,
 572       .maxGeometryInputComponents               = 64,
 573       .maxGeometryOutputComponents              = 128,
 574       .maxGeometryOutputVertices                = 256,
 575       .maxGeometryTotalOutputComponents         = 1024,
 576       .maxFragmentInputComponents               = 128,
 577       .maxFragmentOutputAttachments             = 8,
 578       .maxFragmentDualSrcAttachments            = 1,
 579       .maxFragmentCombinedOutputResources       = 8,
 580       .maxComputeSharedMemorySize               = 32768,
 581       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 582       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 583       .maxComputeWorkGroupSize = {
 584          16 * devinfo->max_cs_threads,
 585          16 * devinfo->max_cs_threads,
 586          16 * devinfo->max_cs_threads,
 587       },
 588       .subPixelPrecisionBits                    = 4 /* FIXME */,
 589       .subTexelPrecisionBits                    = 4 /* FIXME */,
 590       .mipmapPrecisionBits                      = 4 /* FIXME */,
 591       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 592       .maxDrawIndirectCount                     = UINT32_MAX,
 593       .maxSamplerLodBias                        = 16,
 594       .maxSamplerAnisotropy                     = 16,
 595       .maxViewports                             = MAX_VIEWPORTS,
 596       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 597       .viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
 598       .viewportSubPixelBits                     = 13, /* We take a float? */
 599       .minMemoryMapAlignment                    = 4096, /* A page */
 600       .minTexelBufferOffsetAlignment            = 1,
 601       .minUniformBufferOffsetAlignment          = 16,
 602       .minStorageBufferOffsetAlignment          = 4,
 603       .minTexelOffset                           = -8,
 604       .maxTexelOffset                           = 7,
 605       .minTexelGatherOffset                     = -32,
 606       .maxTexelGatherOffset                     = 31,
 607       .minInterpolationOffset                   = -0.5,
 608       .maxInterpolationOffset                   = 0.4375,
 609       .subPixelInterpolationOffsetBits          = 4,
 610       .maxFramebufferWidth                      = (1 << 14),
 611       .maxFramebufferHeight                     = (1 << 14),
 612       .maxFramebufferLayers                     = (1 << 11),
 613       .framebufferColorSampleCounts             = sample_counts,
 614       .framebufferDepthSampleCounts             = sample_counts,
 615       .framebufferStencilSampleCounts           = sample_counts,
 616       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 617       .maxColorAttachments                      = MAX_RTS,
 618       .sampledImageColorSampleCounts            = sample_counts,
 619       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 620       .sampledImageDepthSampleCounts            = sample_counts,
 621       .sampledImageStencilSampleCounts          = sample_counts,
 622       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 623       .maxSampleMaskWords                       = 1,
 624       .timestampComputeAndGraphics              = false,
 625       .timestampPeriod                          = time_stamp_base,
 626       .maxClipDistances                         = 8,
 627       .maxCullDistances                         = 8,
 628       .maxCombinedClipAndCullDistances          = 8,
 629       .discreteQueuePriorities                  = 1,
 630       .pointSizeRange                           = { 0.125, 255.875 },
 631       .lineWidthRange                           = { 0.0, 7.9921875 },
 632       .pointSizeGranularity                     = (1.0 / 8.0),
 633       .lineWidthGranularity                     = (1.0 / 128.0),
 634       .strictLines                              = false, /* FINISHME */
 635       .standardSampleLocations                  = true,
 636       .optimalBufferCopyOffsetAlignment         = 128,
 637       .optimalBufferCopyRowPitchAlignment       = 128,
 638       .nonCoherentAtomSize                      = 64,
 639    };
 640
 641    *pProperties = (VkPhysicalDeviceProperties) {
 642       .apiVersion = VK_MAKE_VERSION(1, 0, 42),
 643       .driverVersion = 1,
 644       .vendorID = 0x8086,
 645       .deviceID = pdevice->chipset_id,
 646       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 647       .limits = limits,
 648       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 649    };
 650
 651    strcpy(pProperties->deviceName, pdevice->name);
 652    memcpy(pProperties->pipelineCacheUUID, pdevice->uuid, VK_UUID_SIZE);
 653 }
 654
 655 void anv_GetPhysicalDeviceProperties2KHR(
 656     VkPhysicalDevice                            physicalDevice,
 657     VkPhysicalDeviceProperties2KHR*             pProperties)
 658 {
 659    anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 660
 661    vk_foreach_struct(ext, pProperties->pNext) {
 662       switch (ext->sType) {
 663       default:
 664          anv_debug_ignored_stype(ext->sType);
 665          break;
 666       }
 667    }
 668 }
 669
 670 static void
 671 anv_get_queue_family_properties(struct anv_physical_device *phys_dev,
 672                                 VkQueueFamilyProperties *props)
 673 {
 674    *props = (VkQueueFamilyProperties) {
 675       .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 676                     VK_QUEUE_COMPUTE_BIT |
 677                     VK_QUEUE_TRANSFER_BIT,
 678       .queueCount = 1,
 679       .timestampValidBits = 36, /* XXX: Real value here */
 680       .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
 681    };
 682 }
 683
 684 void anv_GetPhysicalDeviceQueueFamilyProperties(
 685     VkPhysicalDevice                            physicalDevice,
 686     uint32_t*                                   pCount,
 687     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 688 {
 689    ANV_FROM_HANDLE(anv_physical_device, phys_dev, physicalDevice);
 690
 691    if (pQueueFamilyProperties == NULL) {
 692       *pCount = 1;
 693       return;
 694    }
 695
 696    /* The spec implicitly allows the incoming count to be 0. From the Vulkan
 697     * 1.0.38 spec, Section 4.1 Physical Devices:
 698     *
 699     *     If the value referenced by pQueueFamilyPropertyCount is not 0 [then
 700     *     do stuff].
 701     */
 702    if (*pCount == 0)
 703       return;
 704
 705    *pCount = 1;
 706    anv_get_queue_family_properties(phys_dev, pQueueFamilyProperties);
 707 }
 708
 709 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
 710     VkPhysicalDevice                            physicalDevice,
 711     uint32_t*                                   pQueueFamilyPropertyCount,
 712     VkQueueFamilyProperties2KHR*                pQueueFamilyProperties)
 713 {
 714
 715    ANV_FROM_HANDLE(anv_physical_device, phys_dev, physicalDevice);
 716
 717    if (pQueueFamilyProperties == NULL) {
 718       *pQueueFamilyPropertyCount = 1;
 719       return;
 720    }
 721
 722    /* The spec implicitly allows the incoming count to be 0. From the Vulkan
 723     * 1.0.38 spec, Section 4.1 Physical Devices:
 724     *
 725     *     If the value referenced by pQueueFamilyPropertyCount is not 0 [then
 726     *     do stuff].
 727     */
 728    if (*pQueueFamilyPropertyCount == 0)
 729       return;
 730
 731    /* We support exactly one queue family. So need to traverse only the first
 732     * array element's pNext chain.
 733     */
 734    *pQueueFamilyPropertyCount = 1;
 735    anv_get_queue_family_properties(phys_dev,
 736          &pQueueFamilyProperties->queueFamilyProperties);
 737
 738    vk_foreach_struct(ext, pQueueFamilyProperties->pNext) {
 739       switch (ext->sType) {
 740       default:
 741          anv_debug_ignored_stype(ext->sType);
 742          break;
 743       }
 744    }
 745 }
 746
 747 void anv_GetPhysicalDeviceMemoryProperties(
 748     VkPhysicalDevice                            physicalDevice,
 749     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
 750 {
 751    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 752    VkDeviceSize heap_size;
 753
 754    /* Reserve some wiggle room for the driver by exposing only 75% of the
 755     * aperture to the heap.
 756     */
 757    heap_size = 3 * physical_device->aperture_size / 4;
 758
 759    if (physical_device->info.has_llc) {
 760       /* Big core GPUs share LLC with the CPU and thus one memory type can be
 761        * both cached and coherent at the same time.
 762        */
 763       pMemoryProperties->memoryTypeCount = 1;
 764       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 765          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 766                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 767                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
 768                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 769          .heapIndex = 0,
 770       };
 771    } else {
 772       /* The spec requires that we expose a host-visible, coherent memory
 773        * type, but Atom GPUs don't share LLC. Thus we offer two memory types
 774        * to give the application a choice between cached, but not coherent and
 775        * coherent but uncached (WC though).
 776        */
 777       pMemoryProperties->memoryTypeCount = 2;
 778       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 779          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 780                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 781                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
 782          .heapIndex = 0,
 783       };
 784       pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
 785          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 786                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 787                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 788          .heapIndex = 0,
 789       };
 790    }
 791
 792    pMemoryProperties->memoryHeapCount = 1;
 793    pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
 794       .size = heap_size,
 795       .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 796    };
 797 }
 798
 799 void anv_GetPhysicalDeviceMemoryProperties2KHR(
 800     VkPhysicalDevice                            physicalDevice,
 801     VkPhysicalDeviceMemoryProperties2KHR*       pMemoryProperties)
 802 {
 803    anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
 804                                          &pMemoryProperties->memoryProperties);
 805
 806    vk_foreach_struct(ext, pMemoryProperties->pNext) {
 807       switch (ext->sType) {
 808       default:
 809          anv_debug_ignored_stype(ext->sType);
 810          break;
 811       }
 812    }
 813 }
 814
 815 PFN_vkVoidFunction anv_GetInstanceProcAddr(
 816     VkInstance                                  instance,
 817     const char*                                 pName)
 818 {
 819    return anv_lookup_entrypoint(NULL, pName);
 820 }
 821
 822 /* With version 1+ of the loader interface the ICD should expose
 823  * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
 824  */
 825 PUBLIC
 826 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 827     VkInstance                                  instance,
 828     const char*                                 pName);
 829
 830 PUBLIC
 831 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 832     VkInstance                                  instance,
 833     const char*                                 pName)
 834 {
 835    return anv_GetInstanceProcAddr(instance, pName);
 836 }
 837
 838 PFN_vkVoidFunction anv_GetDeviceProcAddr(
 839     VkDevice                                    _device,
 840     const char*                                 pName)
 841 {
 842    ANV_FROM_HANDLE(anv_device, device, _device);
 843    return anv_lookup_entrypoint(&device->info, pName);
 844 }
 845
 846 static void
 847 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
 848 {
 849    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 850    queue->device = device;
 851    queue->pool = &device->surface_state_pool;
 852 }
 853
 854 static void
 855 anv_queue_finish(struct anv_queue *queue)
 856 {
 857 }
 858
 859 static struct anv_state
 860 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
 861 {
 862    struct anv_state state;
 863
 864    state = anv_state_pool_alloc(pool, size, align);
 865    memcpy(state.map, p, size);
 866
 867    anv_state_flush(pool->block_pool->device, state);
 868
 869    return state;
 870 }
 871
 872 struct gen8_border_color {
 873    union {
 874       float float32[4];
 875       uint32_t uint32[4];
 876    };
 877    /* Pad out to 64 bytes */
 878    uint32_t _pad[12];
 879 };
 880
 881 static void
 882 anv_device_init_border_colors(struct anv_device *device)
 883 {
 884    static const struct gen8_border_color border_colors[] = {
 885       [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] =  { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
 886       [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] =       { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
 887       [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] =       { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
 888       [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] =    { .uint32 = { 0, 0, 0, 0 } },
 889       [VK_BORDER_COLOR_INT_OPAQUE_BLACK] =         { .uint32 = { 0, 0, 0, 1 } },
 890       [VK_BORDER_COLOR_INT_OPAQUE_WHITE] =         { .uint32 = { 1, 1, 1, 1 } },
 891    };
 892
 893    device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
 894                                                     sizeof(border_colors), 64,
 895                                                     border_colors);
 896 }
 897
 898 VkResult
 899 anv_device_submit_simple_batch(struct anv_device *device,
 900                                struct anv_batch *batch)
 901 {
 902    struct drm_i915_gem_execbuffer2 execbuf;
 903    struct drm_i915_gem_exec_object2 exec2_objects[1];
 904    struct anv_bo bo, *exec_bos[1];
 905    VkResult result = VK_SUCCESS;
 906    uint32_t size;
 907    int64_t timeout;
 908    int ret;
 909
 910    /* Kernel driver requires 8 byte aligned batch length */
 911    size = align_u32(batch->next - batch->start, 8);
 912    result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
 913    if (result != VK_SUCCESS)
 914       return result;
 915
 916    memcpy(bo.map, batch->start, size);
 917    if (!device->info.has_llc)
 918       anv_flush_range(bo.map, size);
 919
 920    exec_bos[0] = &bo;
 921    exec2_objects[0].handle = bo.gem_handle;
 922    exec2_objects[0].relocation_count = 0;
 923    exec2_objects[0].relocs_ptr = 0;
 924    exec2_objects[0].alignment = 0;
 925    exec2_objects[0].offset = bo.offset;
 926    exec2_objects[0].flags = 0;
 927    exec2_objects[0].rsvd1 = 0;
 928    exec2_objects[0].rsvd2 = 0;
 929
 930    execbuf.buffers_ptr = (uintptr_t) exec2_objects;
 931    execbuf.buffer_count = 1;
 932    execbuf.batch_start_offset = 0;
 933    execbuf.batch_len = size;
 934    execbuf.cliprects_ptr = 0;
 935    execbuf.num_cliprects = 0;
 936    execbuf.DR1 = 0;
 937    execbuf.DR4 = 0;
 938
 939    execbuf.flags =
 940       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
 941    execbuf.rsvd1 = device->context_id;
 942    execbuf.rsvd2 = 0;
 943
 944    result = anv_device_execbuf(device, &execbuf, exec_bos);
 945    if (result != VK_SUCCESS)
 946       goto fail;
 947
 948    timeout = INT64_MAX;
 949    ret = anv_gem_wait(device, bo.gem_handle, &timeout);
 950    if (ret != 0) {
 951       /* We don't know the real error. */
 952       result = vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
 953       goto fail;
 954    }
 955
 956  fail:
 957    anv_bo_pool_free(&device->batch_bo_pool, &bo);
 958
 959    return result;
 960 }
 961
 962 VkResult anv_CreateDevice(
 963     VkPhysicalDevice                            physicalDevice,
 964     const VkDeviceCreateInfo*                   pCreateInfo,
 965     const VkAllocationCallbacks*                pAllocator,
 966     VkDevice*                                   pDevice)
 967 {
 968    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 969    VkResult result;
 970    struct anv_device *device;
 971
 972    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
 973
 974    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 975       bool found = false;
 976       for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
 977          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 978                     device_extensions[j].extensionName) == 0) {
 979             found = true;
 980             break;
 981          }
 982       }
 983       if (!found)
 984          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 985    }
 986
 987    device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
 988                        sizeof(*device), 8,
 989                        VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
 990    if (!device)
 991       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 992
 993    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 994    device->instance = physical_device->instance;
 995    device->chipset_id = physical_device->chipset_id;
 996
 997    if (pAllocator)
 998       device->alloc = *pAllocator;
 999    else
1000       device->alloc = physical_device->instance->alloc;
1001
1002    /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1003    device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1004    if (device->fd == -1) {
1005       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1006       goto fail_device;
1007    }
1008
1009    device->context_id = anv_gem_create_context(device);
1010    if (device->context_id == -1) {
1011       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1012       goto fail_fd;
1013    }
1014
1015    device->info = physical_device->info;
1016    device->isl_dev = physical_device->isl_dev;
1017
1018    /* On Broadwell and later, we can use batch chaining to more efficiently
1019     * implement growing command buffers.  Prior to Haswell, the kernel
1020     * command parser gets in the way and we have to fall back to growing
1021     * the batch.
1022     */
1023    device->can_chain_batches = device->info.gen >= 8;
1024
1025    device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1026       pCreateInfo->pEnabledFeatures->robustBufferAccess;
1027
1028    pthread_mutex_init(&device->mutex, NULL);
1029
1030    pthread_condattr_t condattr;
1031    pthread_condattr_init(&condattr);
1032    pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC);
1033    pthread_cond_init(&device->queue_submit, NULL);
1034    pthread_condattr_destroy(&condattr);
1035
1036    anv_bo_pool_init(&device->batch_bo_pool, device);
1037
1038    anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
1039
1040    anv_state_pool_init(&device->dynamic_state_pool,
1041                        &device->dynamic_state_block_pool);
1042
1043    anv_block_pool_init(&device->instruction_block_pool, device, 1024 * 1024);
1044    anv_state_pool_init(&device->instruction_state_pool,
1045                        &device->instruction_block_pool);
1046
1047    anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
1048
1049    anv_state_pool_init(&device->surface_state_pool,
1050                        &device->surface_state_block_pool);
1051
1052    anv_bo_init_new(&device->workaround_bo, device, 1024);
1053
1054    anv_scratch_pool_init(device, &device->scratch_pool);
1055
1056    anv_queue_init(device, &device->queue);
1057
1058    switch (device->info.gen) {
1059    case 7:
1060       if (!device->info.is_haswell)
1061          result = gen7_init_device_state(device);
1062       else
1063          result = gen75_init_device_state(device);
1064       break;
1065    case 8:
1066       result = gen8_init_device_state(device);
1067       break;
1068    case 9:
1069       result = gen9_init_device_state(device);
1070       break;
1071    default:
1072       /* Shouldn't get here as we don't create physical devices for any other
1073        * gens. */
1074       unreachable("unhandled gen");
1075    }
1076    if (result != VK_SUCCESS)
1077       goto fail_fd;
1078
1079    anv_device_init_blorp(device);
1080
1081    anv_device_init_border_colors(device);
1082
1083    *pDevice = anv_device_to_handle(device);
1084
1085    return VK_SUCCESS;
1086
1087  fail_fd:
1088    close(device->fd);
1089  fail_device:
1090    vk_free(&device->alloc, device);
1091
1092    return result;
1093 }
1094
1095 void anv_DestroyDevice(
1096     VkDevice                                    _device,
1097     const VkAllocationCallbacks*                pAllocator)
1098 {
1099    ANV_FROM_HANDLE(anv_device, device, _device);
1100
1101    if (!device)
1102       return;
1103
1104    anv_device_finish_blorp(device);
1105
1106    anv_queue_finish(&device->queue);
1107
1108 #ifdef HAVE_VALGRIND
1109    /* We only need to free these to prevent valgrind errors.  The backing
1110     * BO will go away in a couple of lines so we don't actually leak.
1111     */
1112    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1113 #endif
1114
1115    anv_scratch_pool_finish(device, &device->scratch_pool);
1116
1117    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1118    anv_gem_close(device, device->workaround_bo.gem_handle);
1119
1120    anv_state_pool_finish(&device->surface_state_pool);
1121    anv_block_pool_finish(&device->surface_state_block_pool);
1122    anv_state_pool_finish(&device->instruction_state_pool);
1123    anv_block_pool_finish(&device->instruction_block_pool);
1124    anv_state_pool_finish(&device->dynamic_state_pool);
1125    anv_block_pool_finish(&device->dynamic_state_block_pool);
1126
1127    anv_bo_pool_finish(&device->batch_bo_pool);
1128
1129    pthread_cond_destroy(&device->queue_submit);
1130    pthread_mutex_destroy(&device->mutex);
1131
1132    anv_gem_destroy_context(device, device->context_id);
1133
1134    close(device->fd);
1135
1136    vk_free(&device->alloc, device);
1137 }
1138
1139 VkResult anv_EnumerateInstanceExtensionProperties(
1140     const char*                                 pLayerName,
1141     uint32_t*                                   pPropertyCount,
1142     VkExtensionProperties*                      pProperties)
1143 {
1144    if (pProperties == NULL) {
1145       *pPropertyCount = ARRAY_SIZE(global_extensions);
1146       return VK_SUCCESS;
1147    }
1148
1149    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(global_extensions));
1150    typed_memcpy(pProperties, global_extensions, *pPropertyCount);
1151
1152    if (*pPropertyCount < ARRAY_SIZE(global_extensions))
1153       return VK_INCOMPLETE;
1154
1155    return VK_SUCCESS;
1156 }
1157
1158 VkResult anv_EnumerateDeviceExtensionProperties(
1159     VkPhysicalDevice                            physicalDevice,
1160     const char*                                 pLayerName,
1161     uint32_t*                                   pPropertyCount,
1162     VkExtensionProperties*                      pProperties)
1163 {
1164    if (pProperties == NULL) {
1165       *pPropertyCount = ARRAY_SIZE(device_extensions);
1166       return VK_SUCCESS;
1167    }
1168
1169    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(device_extensions));
1170    typed_memcpy(pProperties, device_extensions, *pPropertyCount);
1171
1172    if (*pPropertyCount < ARRAY_SIZE(device_extensions))
1173       return VK_INCOMPLETE;
1174
1175    return VK_SUCCESS;
1176 }
1177
1178 VkResult anv_EnumerateInstanceLayerProperties(
1179     uint32_t*                                   pPropertyCount,
1180     VkLayerProperties*                          pProperties)
1181 {
1182    if (pProperties == NULL) {
1183       *pPropertyCount = 0;
1184       return VK_SUCCESS;
1185    }
1186
1187    /* None supported at this time */
1188    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1189 }
1190
1191 VkResult anv_EnumerateDeviceLayerProperties(
1192     VkPhysicalDevice                            physicalDevice,
1193     uint32_t*                                   pPropertyCount,
1194     VkLayerProperties*                          pProperties)
1195 {
1196    if (pProperties == NULL) {
1197       *pPropertyCount = 0;
1198       return VK_SUCCESS;
1199    }
1200
1201    /* None supported at this time */
1202    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1203 }
1204
1205 void anv_GetDeviceQueue(
1206     VkDevice                                    _device,
1207     uint32_t                                    queueNodeIndex,
1208     uint32_t                                    queueIndex,
1209     VkQueue*                                    pQueue)
1210 {
1211    ANV_FROM_HANDLE(anv_device, device, _device);
1212
1213    assert(queueIndex == 0);
1214
1215    *pQueue = anv_queue_to_handle(&device->queue);
1216 }
1217
1218 VkResult
1219 anv_device_execbuf(struct anv_device *device,
1220                    struct drm_i915_gem_execbuffer2 *execbuf,
1221                    struct anv_bo **execbuf_bos)
1222 {
1223    int ret = anv_gem_execbuffer(device, execbuf);
1224    if (ret != 0) {
1225       /* We don't know the real error. */
1226       return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
1227    }
1228
1229    struct drm_i915_gem_exec_object2 *objects =
1230       (void *)(uintptr_t)execbuf->buffers_ptr;
1231    for (uint32_t k = 0; k < execbuf->buffer_count; k++)
1232       execbuf_bos[k]->offset = objects[k].offset;
1233
1234    return VK_SUCCESS;
1235 }
1236
1237 VkResult anv_QueueSubmit(
1238     VkQueue                                     _queue,
1239     uint32_t                                    submitCount,
1240     const VkSubmitInfo*                         pSubmits,
1241     VkFence                                     _fence)
1242 {
1243    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1244    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1245    struct anv_device *device = queue->device;
1246    VkResult result = VK_SUCCESS;
1247
1248    /* We lock around QueueSubmit for three main reasons:
1249     *
1250     *  1) When a block pool is resized, we create a new gem handle with a
1251     *     different size and, in the case of surface states, possibly a
1252     *     different center offset but we re-use the same anv_bo struct when
1253     *     we do so.  If this happens in the middle of setting up an execbuf,
1254     *     we could end up with our list of BOs out of sync with our list of
1255     *     gem handles.
1256     *
1257     *  2) The algorithm we use for building the list of unique buffers isn't
1258     *     thread-safe.  While the client is supposed to syncronize around
1259     *     QueueSubmit, this would be extremely difficult to debug if it ever
1260     *     came up in the wild due to a broken app.  It's better to play it
1261     *     safe and just lock around QueueSubmit.
1262     *
1263     *  3)  The anv_cmd_buffer_execbuf function may perform relocations in
1264     *      userspace.  Due to the fact that the surface state buffer is shared
1265     *      between batches, we can't afford to have that happen from multiple
1266     *      threads at the same time.  Even though the user is supposed to
1267     *      ensure this doesn't happen, we play it safe as in (2) above.
1268     *
1269     * Since the only other things that ever take the device lock such as block
1270     * pool resize only rarely happen, this will almost never be contended so
1271     * taking a lock isn't really an expensive operation in this case.
1272     */
1273    pthread_mutex_lock(&device->mutex);
1274
1275    for (uint32_t i = 0; i < submitCount; i++) {
1276       for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1277          ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1278                          pSubmits[i].pCommandBuffers[j]);
1279          assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1280
1281          result = anv_cmd_buffer_execbuf(device, cmd_buffer);
1282          if (result != VK_SUCCESS)
1283             goto out;
1284       }
1285    }
1286
1287    if (fence) {
1288       struct anv_bo *fence_bo = &fence->bo;
1289       result = anv_device_execbuf(device, &fence->execbuf, &fence_bo);
1290       if (result != VK_SUCCESS)
1291          goto out;
1292
1293       /* Update the fence and wake up any waiters */
1294       assert(fence->state == ANV_FENCE_STATE_RESET);
1295       fence->state = ANV_FENCE_STATE_SUBMITTED;
1296       pthread_cond_broadcast(&device->queue_submit);
1297    }
1298
1299 out:
1300    pthread_mutex_unlock(&device->mutex);
1301
1302    return result;
1303 }
1304
1305 VkResult anv_QueueWaitIdle(
1306     VkQueue                                     _queue)
1307 {
1308    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1309
1310    return anv_DeviceWaitIdle(anv_device_to_handle(queue->device));
1311 }
1312
1313 VkResult anv_DeviceWaitIdle(
1314     VkDevice                                    _device)
1315 {
1316    ANV_FROM_HANDLE(anv_device, device, _device);
1317    struct anv_batch batch;
1318
1319    uint32_t cmds[8];
1320    batch.start = batch.next = cmds;
1321    batch.end = (void *) cmds + sizeof(cmds);
1322
1323    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1324    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1325
1326    return anv_device_submit_simple_batch(device, &batch);
1327 }
1328
1329 VkResult
1330 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1331 {
1332    uint32_t gem_handle = anv_gem_create(device, size);
1333    if (!gem_handle)
1334       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1335
1336    anv_bo_init(bo, gem_handle, size);
1337
1338    return VK_SUCCESS;
1339 }
1340
1341 VkResult anv_AllocateMemory(
1342     VkDevice                                    _device,
1343     const VkMemoryAllocateInfo*                 pAllocateInfo,
1344     const VkAllocationCallbacks*                pAllocator,
1345     VkDeviceMemory*                             pMem)
1346 {
1347    ANV_FROM_HANDLE(anv_device, device, _device);
1348    struct anv_device_memory *mem;
1349    VkResult result;
1350
1351    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1352
1353    /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1354    assert(pAllocateInfo->allocationSize > 0);
1355
1356    /* We support exactly one memory heap. */
1357    assert(pAllocateInfo->memoryTypeIndex == 0 ||
1358           (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1359
1360    /* FINISHME: Fail if allocation request exceeds heap size. */
1361
1362    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1363                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1364    if (mem == NULL)
1365       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1366
1367    /* The kernel is going to give us whole pages anyway */
1368    uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1369
1370    result = anv_bo_init_new(&mem->bo, device, alloc_size);
1371    if (result != VK_SUCCESS)
1372       goto fail;
1373
1374    mem->type_index = pAllocateInfo->memoryTypeIndex;
1375
1376    mem->map = NULL;
1377    mem->map_size = 0;
1378
1379    *pMem = anv_device_memory_to_handle(mem);
1380
1381    return VK_SUCCESS;
1382
1383  fail:
1384    vk_free2(&device->alloc, pAllocator, mem);
1385
1386    return result;
1387 }
1388
1389 void anv_FreeMemory(
1390     VkDevice                                    _device,
1391     VkDeviceMemory                              _mem,
1392     const VkAllocationCallbacks*                pAllocator)
1393 {
1394    ANV_FROM_HANDLE(anv_device, device, _device);
1395    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1396
1397    if (mem == NULL)
1398       return;
1399
1400    if (mem->map)
1401       anv_UnmapMemory(_device, _mem);
1402
1403    if (mem->bo.map)
1404       anv_gem_munmap(mem->bo.map, mem->bo.size);
1405
1406    if (mem->bo.gem_handle != 0)
1407       anv_gem_close(device, mem->bo.gem_handle);
1408
1409    vk_free2(&device->alloc, pAllocator, mem);
1410 }
1411
1412 VkResult anv_MapMemory(
1413     VkDevice                                    _device,
1414     VkDeviceMemory                              _memory,
1415     VkDeviceSize                                offset,
1416     VkDeviceSize                                size,
1417     VkMemoryMapFlags                            flags,
1418     void**                                      ppData)
1419 {
1420    ANV_FROM_HANDLE(anv_device, device, _device);
1421    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1422
1423    if (mem == NULL) {
1424       *ppData = NULL;
1425       return VK_SUCCESS;
1426    }
1427
1428    if (size == VK_WHOLE_SIZE)
1429       size = mem->bo.size - offset;
1430
1431    /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1432     *
1433     *  * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1434     *    assert(size != 0);
1435     *  * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1436     *    equal to the size of the memory minus offset
1437     */
1438    assert(size > 0);
1439    assert(offset + size <= mem->bo.size);
1440
1441    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1442     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1443     * at a time is valid. We could just mmap up front and return an offset
1444     * pointer here, but that may exhaust virtual memory on 32 bit
1445     * userspace. */
1446
1447    uint32_t gem_flags = 0;
1448    if (!device->info.has_llc && mem->type_index == 0)
1449       gem_flags |= I915_MMAP_WC;
1450
1451    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1452    uint64_t map_offset = offset & ~4095ull;
1453    assert(offset >= map_offset);
1454    uint64_t map_size = (offset + size) - map_offset;
1455
1456    /* Let's map whole pages */
1457    map_size = align_u64(map_size, 4096);
1458
1459    void *map = anv_gem_mmap(device, mem->bo.gem_handle,
1460                             map_offset, map_size, gem_flags);
1461    if (map == MAP_FAILED)
1462       return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
1463
1464    mem->map = map;
1465    mem->map_size = map_size;
1466
1467    *ppData = mem->map + (offset - map_offset);
1468
1469    return VK_SUCCESS;
1470 }
1471
1472 void anv_UnmapMemory(
1473     VkDevice                                    _device,
1474     VkDeviceMemory                              _memory)
1475 {
1476    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1477
1478    if (mem == NULL)
1479       return;
1480
1481    anv_gem_munmap(mem->map, mem->map_size);
1482
1483    mem->map = NULL;
1484    mem->map_size = 0;
1485 }
1486
1487 static void
1488 clflush_mapped_ranges(struct anv_device         *device,
1489                       uint32_t                   count,
1490                       const VkMappedMemoryRange *ranges)
1491 {
1492    for (uint32_t i = 0; i < count; i++) {
1493       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1494       if (ranges[i].offset >= mem->map_size)
1495          continue;
1496
1497       anv_clflush_range(mem->map + ranges[i].offset,
1498                         MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
1499    }
1500 }
1501
1502 VkResult anv_FlushMappedMemoryRanges(
1503     VkDevice                                    _device,
1504     uint32_t                                    memoryRangeCount,
1505     const VkMappedMemoryRange*                  pMemoryRanges)
1506 {
1507    ANV_FROM_HANDLE(anv_device, device, _device);
1508
1509    if (device->info.has_llc)
1510       return VK_SUCCESS;
1511
1512    /* Make sure the writes we're flushing have landed. */
1513    __builtin_ia32_mfence();
1514
1515    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1516
1517    return VK_SUCCESS;
1518 }
1519
1520 VkResult anv_InvalidateMappedMemoryRanges(
1521     VkDevice                                    _device,
1522     uint32_t                                    memoryRangeCount,
1523     const VkMappedMemoryRange*                  pMemoryRanges)
1524 {
1525    ANV_FROM_HANDLE(anv_device, device, _device);
1526
1527    if (device->info.has_llc)
1528       return VK_SUCCESS;
1529
1530    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1531
1532    /* Make sure no reads get moved up above the invalidate. */
1533    __builtin_ia32_mfence();
1534
1535    return VK_SUCCESS;
1536 }
1537
1538 void anv_GetBufferMemoryRequirements(
1539     VkDevice                                    _device,
1540     VkBuffer                                    _buffer,
1541     VkMemoryRequirements*                       pMemoryRequirements)
1542 {
1543    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1544    ANV_FROM_HANDLE(anv_device, device, _device);
1545
1546    /* The Vulkan spec (git aaed022) says:
1547     *
1548     *    memoryTypeBits is a bitfield and contains one bit set for every
1549     *    supported memory type for the resource. The bit `1<<i` is set if and
1550     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1551     *    structure for the physical device is supported.
1552     *
1553     * We support exactly one memory type on LLC, two on non-LLC.
1554     */
1555    pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1556
1557    pMemoryRequirements->size = buffer->size;
1558    pMemoryRequirements->alignment = 16;
1559 }
1560
1561 void anv_GetImageMemoryRequirements(
1562     VkDevice                                    _device,
1563     VkImage                                     _image,
1564     VkMemoryRequirements*                       pMemoryRequirements)
1565 {
1566    ANV_FROM_HANDLE(anv_image, image, _image);
1567    ANV_FROM_HANDLE(anv_device, device, _device);
1568
1569    /* The Vulkan spec (git aaed022) says:
1570     *
1571     *    memoryTypeBits is a bitfield and contains one bit set for every
1572     *    supported memory type for the resource. The bit `1<<i` is set if and
1573     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1574     *    structure for the physical device is supported.
1575     *
1576     * We support exactly one memory type on LLC, two on non-LLC.
1577     */
1578    pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1579
1580    pMemoryRequirements->size = image->size;
1581    pMemoryRequirements->alignment = image->alignment;
1582 }
1583
1584 void anv_GetImageSparseMemoryRequirements(
1585     VkDevice                                    device,
1586     VkImage                                     image,
1587     uint32_t*                                   pSparseMemoryRequirementCount,
1588     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1589 {
1590    stub();
1591 }
1592
1593 void anv_GetDeviceMemoryCommitment(
1594     VkDevice                                    device,
1595     VkDeviceMemory                              memory,
1596     VkDeviceSize*                               pCommittedMemoryInBytes)
1597 {
1598    *pCommittedMemoryInBytes = 0;
1599 }
1600
1601 VkResult anv_BindBufferMemory(
1602     VkDevice                                    device,
1603     VkBuffer                                    _buffer,
1604     VkDeviceMemory                              _memory,
1605     VkDeviceSize                                memoryOffset)
1606 {
1607    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1608    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1609
1610    if (mem) {
1611       buffer->bo = &mem->bo;
1612       buffer->offset = memoryOffset;
1613    } else {
1614       buffer->bo = NULL;
1615       buffer->offset = 0;
1616    }
1617
1618    return VK_SUCCESS;
1619 }
1620
1621 VkResult anv_QueueBindSparse(
1622     VkQueue                                     queue,
1623     uint32_t                                    bindInfoCount,
1624     const VkBindSparseInfo*                     pBindInfo,
1625     VkFence                                     fence)
1626 {
1627    stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1628 }
1629
1630 VkResult anv_CreateFence(
1631     VkDevice                                    _device,
1632     const VkFenceCreateInfo*                    pCreateInfo,
1633     const VkAllocationCallbacks*                pAllocator,
1634     VkFence*                                    pFence)
1635 {
1636    ANV_FROM_HANDLE(anv_device, device, _device);
1637    struct anv_bo fence_bo;
1638    struct anv_fence *fence;
1639    struct anv_batch batch;
1640    VkResult result;
1641
1642    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1643
1644    result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1645    if (result != VK_SUCCESS)
1646       return result;
1647
1648    /* Fences are small.  Just store the CPU data structure in the BO. */
1649    fence = fence_bo.map;
1650    fence->bo = fence_bo;
1651
1652    /* Place the batch after the CPU data but on its own cache line. */
1653    const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1654    batch.next = batch.start = fence->bo.map + batch_offset;
1655    batch.end = fence->bo.map + fence->bo.size;
1656    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1657    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1658
1659    if (!device->info.has_llc) {
1660       assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1661       assert(batch.next - batch.start <= CACHELINE_SIZE);
1662       __builtin_ia32_mfence();
1663       __builtin_ia32_clflush(batch.start);
1664    }
1665
1666    fence->exec2_objects[0].handle = fence->bo.gem_handle;
1667    fence->exec2_objects[0].relocation_count = 0;
1668    fence->exec2_objects[0].relocs_ptr = 0;
1669    fence->exec2_objects[0].alignment = 0;
1670    fence->exec2_objects[0].offset = fence->bo.offset;
1671    fence->exec2_objects[0].flags = 0;
1672    fence->exec2_objects[0].rsvd1 = 0;
1673    fence->exec2_objects[0].rsvd2 = 0;
1674
1675    fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1676    fence->execbuf.buffer_count = 1;
1677    fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1678    fence->execbuf.batch_len = batch.next - batch.start;
1679    fence->execbuf.cliprects_ptr = 0;
1680    fence->execbuf.num_cliprects = 0;
1681    fence->execbuf.DR1 = 0;
1682    fence->execbuf.DR4 = 0;
1683
1684    fence->execbuf.flags =
1685       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1686    fence->execbuf.rsvd1 = device->context_id;
1687    fence->execbuf.rsvd2 = 0;
1688
1689    if (pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT) {
1690       fence->state = ANV_FENCE_STATE_SIGNALED;
1691    } else {
1692       fence->state = ANV_FENCE_STATE_RESET;
1693    }
1694
1695    *pFence = anv_fence_to_handle(fence);
1696
1697    return VK_SUCCESS;
1698 }
1699
1700 void anv_DestroyFence(
1701     VkDevice                                    _device,
1702     VkFence                                     _fence,
1703     const VkAllocationCallbacks*                pAllocator)
1704 {
1705    ANV_FROM_HANDLE(anv_device, device, _device);
1706    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1707
1708    if (!fence)
1709       return;
1710
1711    assert(fence->bo.map == fence);
1712    anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1713 }
1714
1715 VkResult anv_ResetFences(
1716     VkDevice                                    _device,
1717     uint32_t                                    fenceCount,
1718     const VkFence*                              pFences)
1719 {
1720    for (uint32_t i = 0; i < fenceCount; i++) {
1721       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1722       fence->state = ANV_FENCE_STATE_RESET;
1723    }
1724
1725    return VK_SUCCESS;
1726 }
1727
1728 VkResult anv_GetFenceStatus(
1729     VkDevice                                    _device,
1730     VkFence                                     _fence)
1731 {
1732    ANV_FROM_HANDLE(anv_device, device, _device);
1733    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1734    int64_t t = 0;
1735    int ret;
1736
1737    switch (fence->state) {
1738    case ANV_FENCE_STATE_RESET:
1739       /* If it hasn't even been sent off to the GPU yet, it's not ready */
1740       return VK_NOT_READY;
1741
1742    case ANV_FENCE_STATE_SIGNALED:
1743       /* It's been signaled, return success */
1744       return VK_SUCCESS;
1745
1746    case ANV_FENCE_STATE_SUBMITTED:
1747       /* It's been submitted to the GPU but we don't know if it's done yet. */
1748       ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1749       if (ret == 0) {
1750          fence->state = ANV_FENCE_STATE_SIGNALED;
1751          return VK_SUCCESS;
1752       } else {
1753          return VK_NOT_READY;
1754       }
1755    default:
1756       unreachable("Invalid fence status");
1757    }
1758 }
1759
1760 #define NSEC_PER_SEC 1000000000
1761 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
1762
1763 VkResult anv_WaitForFences(
1764     VkDevice                                    _device,
1765     uint32_t                                    fenceCount,
1766     const VkFence*                              pFences,
1767     VkBool32                                    waitAll,
1768     uint64_t                                    _timeout)
1769 {
1770    ANV_FROM_HANDLE(anv_device, device, _device);
1771    int ret;
1772
1773    /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1774     * to block indefinitely timeouts <= 0.  Unfortunately, this was broken
1775     * for a couple of kernel releases.  Since there's no way to know
1776     * whether or not the kernel we're using is one of the broken ones, the
1777     * best we can do is to clamp the timeout to INT64_MAX.  This limits the
1778     * maximum timeout from 584 years to 292 years - likely not a big deal.
1779     */
1780    int64_t timeout = MIN2(_timeout, INT64_MAX);
1781
1782    uint32_t pending_fences = fenceCount;
1783    while (pending_fences) {
1784       pending_fences = 0;
1785       bool signaled_fences = false;
1786       for (uint32_t i = 0; i < fenceCount; i++) {
1787          ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1788          switch (fence->state) {
1789          case ANV_FENCE_STATE_RESET:
1790             /* This fence hasn't been submitted yet, we'll catch it the next
1791              * time around.  Yes, this may mean we dead-loop but, short of
1792              * lots of locking and a condition variable, there's not much that
1793              * we can do about that.
1794              */
1795             pending_fences++;
1796             continue;
1797
1798          case ANV_FENCE_STATE_SIGNALED:
1799             /* This fence is not pending.  If waitAll isn't set, we can return
1800              * early.  Otherwise, we have to keep going.
1801              */
1802             if (!waitAll)
1803                return VK_SUCCESS;
1804             continue;
1805
1806          case ANV_FENCE_STATE_SUBMITTED:
1807             /* These are the fences we really care about.  Go ahead and wait
1808              * on it until we hit a timeout.
1809              */
1810             ret = anv_gem_wait(device, fence->bo.gem_handle, &timeout);
1811             if (ret == -1 && errno == ETIME) {
1812                return VK_TIMEOUT;
1813             } else if (ret == -1) {
1814                /* We don't know the real error. */
1815                return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1816             } else {
1817                fence->state = ANV_FENCE_STATE_SIGNALED;
1818                signaled_fences = true;
1819                if (!waitAll)
1820                   return VK_SUCCESS;
1821                continue;
1822             }
1823          }
1824       }
1825
1826       if (pending_fences && !signaled_fences) {
1827          /* If we've hit this then someone decided to vkWaitForFences before
1828           * they've actually submitted any of them to a queue.  This is a
1829           * fairly pessimal case, so it's ok to lock here and use a standard
1830           * pthreads condition variable.
1831           */
1832          pthread_mutex_lock(&device->mutex);
1833
1834          /* It's possible that some of the fences have changed state since the
1835           * last time we checked.  Now that we have the lock, check for
1836           * pending fences again and don't wait if it's changed.
1837           */
1838          uint32_t now_pending_fences = 0;
1839          for (uint32_t i = 0; i < fenceCount; i++) {
1840             ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1841             if (fence->state == ANV_FENCE_STATE_RESET)
1842                now_pending_fences++;
1843          }
1844          assert(now_pending_fences <= pending_fences);
1845
1846          if (now_pending_fences == pending_fences) {
1847             struct timespec before;
1848             clock_gettime(CLOCK_MONOTONIC, &before);
1849
1850             uint32_t abs_nsec = before.tv_nsec + timeout % NSEC_PER_SEC;
1851             uint64_t abs_sec = before.tv_sec + (abs_nsec / NSEC_PER_SEC) +
1852                                (timeout / NSEC_PER_SEC);
1853             abs_nsec %= NSEC_PER_SEC;
1854
1855             /* Avoid roll-over in tv_sec on 32-bit systems if the user
1856              * provided timeout is UINT64_MAX
1857              */
1858             struct timespec abstime;
1859             abstime.tv_nsec = abs_nsec;
1860             abstime.tv_sec = MIN2(abs_sec, INT_TYPE_MAX(abstime.tv_sec));
1861
1862             ret = pthread_cond_timedwait(&device->queue_submit,
1863                                          &device->mutex, &abstime);
1864             assert(ret != EINVAL);
1865
1866             struct timespec after;
1867             clock_gettime(CLOCK_MONOTONIC, &after);
1868             uint64_t time_elapsed =
1869                ((uint64_t)after.tv_sec * NSEC_PER_SEC + after.tv_nsec) -
1870                ((uint64_t)before.tv_sec * NSEC_PER_SEC + before.tv_nsec);
1871
1872             if (time_elapsed >= timeout) {
1873                pthread_mutex_unlock(&device->mutex);
1874                return VK_TIMEOUT;
1875             }
1876
1877             timeout -= time_elapsed;
1878          }
1879
1880          pthread_mutex_unlock(&device->mutex);
1881       }
1882    }
1883
1884    return VK_SUCCESS;
1885 }
1886
1887 // Queue semaphore functions
1888
1889 VkResult anv_CreateSemaphore(
1890     VkDevice                                    device,
1891     const VkSemaphoreCreateInfo*                pCreateInfo,
1892     const VkAllocationCallbacks*                pAllocator,
1893     VkSemaphore*                                pSemaphore)
1894 {
1895    /* The DRM execbuffer ioctl always execute in-oder, even between different
1896     * rings. As such, there's nothing to do for the user space semaphore.
1897     */
1898
1899    *pSemaphore = (VkSemaphore)1;
1900
1901    return VK_SUCCESS;
1902 }
1903
1904 void anv_DestroySemaphore(
1905     VkDevice                                    device,
1906     VkSemaphore                                 semaphore,
1907     const VkAllocationCallbacks*                pAllocator)
1908 {
1909 }
1910
1911 // Event functions
1912
1913 VkResult anv_CreateEvent(
1914     VkDevice                                    _device,
1915     const VkEventCreateInfo*                    pCreateInfo,
1916     const VkAllocationCallbacks*                pAllocator,
1917     VkEvent*                                    pEvent)
1918 {
1919    ANV_FROM_HANDLE(anv_device, device, _device);
1920    struct anv_state state;
1921    struct anv_event *event;
1922
1923    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1924
1925    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1926                                 sizeof(*event), 8);
1927    event = state.map;
1928    event->state = state;
1929    event->semaphore = VK_EVENT_RESET;
1930
1931    if (!device->info.has_llc) {
1932       /* Make sure the writes we're flushing have landed. */
1933       __builtin_ia32_mfence();
1934       __builtin_ia32_clflush(event);
1935    }
1936
1937    *pEvent = anv_event_to_handle(event);
1938
1939    return VK_SUCCESS;
1940 }
1941
1942 void anv_DestroyEvent(
1943     VkDevice                                    _device,
1944     VkEvent                                     _event,
1945     const VkAllocationCallbacks*                pAllocator)
1946 {
1947    ANV_FROM_HANDLE(anv_device, device, _device);
1948    ANV_FROM_HANDLE(anv_event, event, _event);
1949
1950    if (!event)
1951       return;
1952
1953    anv_state_pool_free(&device->dynamic_state_pool, event->state);
1954 }
1955
1956 VkResult anv_GetEventStatus(
1957     VkDevice                                    _device,
1958     VkEvent                                     _event)
1959 {
1960    ANV_FROM_HANDLE(anv_device, device, _device);
1961    ANV_FROM_HANDLE(anv_event, event, _event);
1962
1963    if (!device->info.has_llc) {
1964       /* Invalidate read cache before reading event written by GPU. */
1965       __builtin_ia32_clflush(event);
1966       __builtin_ia32_mfence();
1967
1968    }
1969
1970    return event->semaphore;
1971 }
1972
1973 VkResult anv_SetEvent(
1974     VkDevice                                    _device,
1975     VkEvent                                     _event)
1976 {
1977    ANV_FROM_HANDLE(anv_device, device, _device);
1978    ANV_FROM_HANDLE(anv_event, event, _event);
1979
1980    event->semaphore = VK_EVENT_SET;
1981
1982    if (!device->info.has_llc) {
1983       /* Make sure the writes we're flushing have landed. */
1984       __builtin_ia32_mfence();
1985       __builtin_ia32_clflush(event);
1986    }
1987
1988    return VK_SUCCESS;
1989 }
1990
1991 VkResult anv_ResetEvent(
1992     VkDevice                                    _device,
1993     VkEvent                                     _event)
1994 {
1995    ANV_FROM_HANDLE(anv_device, device, _device);
1996    ANV_FROM_HANDLE(anv_event, event, _event);
1997
1998    event->semaphore = VK_EVENT_RESET;
1999
2000    if (!device->info.has_llc) {
2001       /* Make sure the writes we're flushing have landed. */
2002       __builtin_ia32_mfence();
2003       __builtin_ia32_clflush(event);
2004    }
2005
2006    return VK_SUCCESS;
2007 }
2008
2009 // Buffer functions
2010
2011 VkResult anv_CreateBuffer(
2012     VkDevice                                    _device,
2013     const VkBufferCreateInfo*                   pCreateInfo,
2014     const VkAllocationCallbacks*                pAllocator,
2015     VkBuffer*                                   pBuffer)
2016 {
2017    ANV_FROM_HANDLE(anv_device, device, _device);
2018    struct anv_buffer *buffer;
2019
2020    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2021
2022    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2023                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2024    if (buffer == NULL)
2025       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2026
2027    buffer->size = pCreateInfo->size;
2028    buffer->usage = pCreateInfo->usage;
2029    buffer->bo = NULL;
2030    buffer->offset = 0;
2031
2032    *pBuffer = anv_buffer_to_handle(buffer);
2033
2034    return VK_SUCCESS;
2035 }
2036
2037 void anv_DestroyBuffer(
2038     VkDevice                                    _device,
2039     VkBuffer                                    _buffer,
2040     const VkAllocationCallbacks*                pAllocator)
2041 {
2042    ANV_FROM_HANDLE(anv_device, device, _device);
2043    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2044
2045    if (!buffer)
2046       return;
2047
2048    vk_free2(&device->alloc, pAllocator, buffer);
2049 }
2050
2051 void
2052 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
2053                               enum isl_format format,
2054                               uint32_t offset, uint32_t range, uint32_t stride)
2055 {
2056    isl_buffer_fill_state(&device->isl_dev, state.map,
2057                          .address = offset,
2058                          .mocs = device->default_mocs,
2059                          .size = range,
2060                          .format = format,
2061                          .stride = stride);
2062
2063    anv_state_flush(device, state);
2064 }
2065
2066 void anv_DestroySampler(
2067     VkDevice                                    _device,
2068     VkSampler                                   _sampler,
2069     const VkAllocationCallbacks*                pAllocator)
2070 {
2071    ANV_FROM_HANDLE(anv_device, device, _device);
2072    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
2073
2074    if (!sampler)
2075       return;
2076
2077    vk_free2(&device->alloc, pAllocator, sampler);
2078 }
2079
2080 VkResult anv_CreateFramebuffer(
2081     VkDevice                                    _device,
2082     const VkFramebufferCreateInfo*              pCreateInfo,
2083     const VkAllocationCallbacks*                pAllocator,
2084     VkFramebuffer*                              pFramebuffer)
2085 {
2086    ANV_FROM_HANDLE(anv_device, device, _device);
2087    struct anv_framebuffer *framebuffer;
2088
2089    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2090
2091    size_t size = sizeof(*framebuffer) +
2092                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2093    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2094                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2095    if (framebuffer == NULL)
2096       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2097
2098    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2099    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2100       VkImageView _iview = pCreateInfo->pAttachments[i];
2101       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2102    }
2103
2104    framebuffer->width = pCreateInfo->width;
2105    framebuffer->height = pCreateInfo->height;
2106    framebuffer->layers = pCreateInfo->layers;
2107
2108    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2109
2110    return VK_SUCCESS;
2111 }
2112
2113 void anv_DestroyFramebuffer(
2114     VkDevice                                    _device,
2115     VkFramebuffer                               _fb,
2116     const VkAllocationCallbacks*                pAllocator)
2117 {
2118    ANV_FROM_HANDLE(anv_device, device, _device);
2119    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2120
2121    if (!fb)
2122       return;
2123
2124    vk_free2(&device->alloc, pAllocator, fb);
2125 }
2126
2127 /* vk_icd.h does not declare this function, so we declare it here to
2128  * suppress Wmissing-prototypes.
2129  */
2130 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2131 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2132
2133 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2134 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2135 {
2136    /* For the full details on loader interface versioning, see
2137     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2138     * What follows is a condensed summary, to help you navigate the large and
2139     * confusing official doc.
2140     *
2141     *   - Loader interface v0 is incompatible with later versions. We don't
2142     *     support it.
2143     *
2144     *   - In loader interface v1:
2145     *       - The first ICD entrypoint called by the loader is
2146     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2147     *         entrypoint.
2148     *       - The ICD must statically expose no other Vulkan symbol unless it is
2149     *         linked with -Bsymbolic.
2150     *       - Each dispatchable Vulkan handle created by the ICD must be
2151     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2152     *         ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2153     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2154     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2155     *         such loader-managed surfaces.
2156     *
2157     *    - Loader interface v2 differs from v1 in:
2158     *       - The first ICD entrypoint called by the loader is
2159     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2160     *         statically expose this entrypoint.
2161     *
2162     *    - Loader interface v3 differs from v2 in:
2163     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2164     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2165     *          because the loader no longer does so.
2166     */
2167    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2168    return VK_SUCCESS;
2169 }