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