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 struct anv_dispatch_table dtable;
  40
  41 static void
  42 compiler_debug_log(void *data, const char *fmt, ...)
  43 { }
  44
  45 static void
  46 compiler_perf_log(void *data, const char *fmt, ...)
  47 {
  48    va_list args;
  49    va_start(args, fmt);
  50
  51    if (unlikely(INTEL_DEBUG & DEBUG_PERF))
  52       vfprintf(stderr, fmt, args);
  53
  54    va_end(args);
  55 }
  56
  57 static bool
  58 anv_device_get_cache_uuid(void *uuid)
  59 {
  60    const struct build_id_note *note = build_id_find_nhdr("libvulkan_intel.so");
  61    if (!note)
  62       return false;
  63
  64    unsigned len = build_id_length(note);
  65    if (len < VK_UUID_SIZE)
  66       return false;
  67
  68    build_id_read(note, uuid, VK_UUID_SIZE);
  69    return true;
  70 }
  71
  72 static VkResult
  73 anv_physical_device_init(struct anv_physical_device *device,
  74                          struct anv_instance *instance,
  75                          const char *path)
  76 {
  77    VkResult result;
  78    int fd;
  79
  80    fd = open(path, O_RDWR | O_CLOEXEC);
  81    if (fd < 0)
  82       return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  83
  84    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
  85    device->instance = instance;
  86
  87    assert(strlen(path) < ARRAY_SIZE(device->path));
  88    strncpy(device->path, path, ARRAY_SIZE(device->path));
  89
  90    device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
  91    if (!device->chipset_id) {
  92       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  93       goto fail;
  94    }
  95
  96    device->name = gen_get_device_name(device->chipset_id);
  97    if (!gen_get_device_info(device->chipset_id, &device->info)) {
  98       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  99       goto fail;
 100    }
 101
 102    if (device->info.is_haswell) {
 103       fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
 104    } else if (device->info.gen == 7 && !device->info.is_baytrail) {
 105       fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
 106    } else if (device->info.gen == 7 && device->info.is_baytrail) {
 107       fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
 108    } else if (device->info.gen >= 8) {
 109       /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
 110        * supported as anything */
 111    } else {
 112       result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 113                          "Vulkan not yet supported on %s", device->name);
 114       goto fail;
 115    }
 116
 117    device->cmd_parser_version = -1;
 118    if (device->info.gen == 7) {
 119       device->cmd_parser_version =
 120          anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
 121       if (device->cmd_parser_version == -1) {
 122          result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 123                             "failed to get command parser version");
 124          goto fail;
 125       }
 126    }
 127
 128    if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
 129       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 130                          "failed to get aperture size: %m");
 131       goto fail;
 132    }
 133
 134    if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
 135       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 136                          "kernel missing gem wait");
 137       goto fail;
 138    }
 139
 140    if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
 141       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 142                          "kernel missing execbuf2");
 143       goto fail;
 144    }
 145
 146    if (!device->info.has_llc &&
 147        anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
 148       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 149                          "kernel missing wc mmap");
 150       goto fail;
 151    }
 152
 153    if (!anv_device_get_cache_uuid(device->uuid)) {
 154       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 155                          "cannot generate UUID");
 156       goto fail;
 157    }
 158    bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
 159
 160    /* GENs prior to 8 do not support EU/Subslice info */
 161    if (device->info.gen >= 8) {
 162       device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
 163       device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
 164
 165       /* Without this information, we cannot get the right Braswell
 166        * brandstrings, and we have to use conservative numbers for GPGPU on
 167        * many platforms, but otherwise, things will just work.
 168        */
 169       if (device->subslice_total < 1 || device->eu_total < 1) {
 170          fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
 171                          " query GPU properties.\n");
 172       }
 173    } else if (device->info.gen == 7) {
 174       device->subslice_total = 1 << (device->info.gt - 1);
 175    }
 176
 177    if (device->info.is_cherryview &&
 178        device->subslice_total > 0 && device->eu_total > 0) {
 179       /* Logical CS threads = EUs per subslice * 7 threads per EU */
 180       uint32_t max_cs_threads = device->eu_total / device->subslice_total * 7;
 181
 182       /* Fuse configurations may give more threads than expected, never less. */
 183       if (max_cs_threads > device->info.max_cs_threads)
 184          device->info.max_cs_threads = max_cs_threads;
 185    }
 186
 187    brw_process_intel_debug_variable();
 188
 189    device->compiler = brw_compiler_create(NULL, &device->info);
 190    if (device->compiler == NULL) {
 191       result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 192       goto fail;
 193    }
 194    device->compiler->shader_debug_log = compiler_debug_log;
 195    device->compiler->shader_perf_log = compiler_perf_log;
 196
 197    result = anv_init_wsi(device);
 198    if (result != VK_SUCCESS) {
 199       ralloc_free(device->compiler);
 200       goto fail;
 201    }
 202
 203    isl_device_init(&device->isl_dev, &device->info, swizzled);
 204
 205    device->local_fd = fd;
 206    return VK_SUCCESS;
 207
 208 fail:
 209    close(fd);
 210    return result;
 211 }
 212
 213 static void
 214 anv_physical_device_finish(struct anv_physical_device *device)
 215 {
 216    anv_finish_wsi(device);
 217    ralloc_free(device->compiler);
 218    close(device->local_fd);
 219 }
 220
 221 static const VkExtensionProperties global_extensions[] = {
 222    {
 223       .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
 224       .specVersion = 25,
 225    },
 226 #ifdef VK_USE_PLATFORM_XCB_KHR
 227    {
 228       .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
 229       .specVersion = 6,
 230    },
 231 #endif
 232 #ifdef VK_USE_PLATFORM_XLIB_KHR
 233    {
 234       .extensionName = VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
 235       .specVersion = 6,
 236    },
 237 #endif
 238 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
 239    {
 240       .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
 241       .specVersion = 5,
 242    },
 243 #endif
 244    {
 245       .extensionName = VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
 246       .specVersion = 1,
 247    },
 248 };
 249
 250 static const VkExtensionProperties device_extensions[] = {
 251    {
 252       .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
 253       .specVersion = 68,
 254    },
 255    {
 256       .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
 257       .specVersion = 1,
 258    },
 259    {
 260       .extensionName = VK_KHR_MAINTENANCE1_EXTENSION_NAME,
 261       .specVersion = 1,
 262    },
 263    {
 264       .extensionName = VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
 265       .specVersion = 1,
 266    }
 267 };
 268
 269 static void *
 270 default_alloc_func(void *pUserData, size_t size, size_t align,
 271                    VkSystemAllocationScope allocationScope)
 272 {
 273    return malloc(size);
 274 }
 275
 276 static void *
 277 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
 278                      size_t align, VkSystemAllocationScope allocationScope)
 279 {
 280    return realloc(pOriginal, size);
 281 }
 282
 283 static void
 284 default_free_func(void *pUserData, void *pMemory)
 285 {
 286    free(pMemory);
 287 }
 288
 289 static const VkAllocationCallbacks default_alloc = {
 290    .pUserData = NULL,
 291    .pfnAllocation = default_alloc_func,
 292    .pfnReallocation = default_realloc_func,
 293    .pfnFree = default_free_func,
 294 };
 295
 296 VkResult anv_CreateInstance(
 297     const VkInstanceCreateInfo*                 pCreateInfo,
 298     const VkAllocationCallbacks*                pAllocator,
 299     VkInstance*                                 pInstance)
 300 {
 301    struct anv_instance *instance;
 302
 303    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 304
 305    uint32_t client_version;
 306    if (pCreateInfo->pApplicationInfo &&
 307        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 308       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 309    } else {
 310       client_version = VK_MAKE_VERSION(1, 0, 0);
 311    }
 312
 313    if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
 314        client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
 315       return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 316                        "Client requested version %d.%d.%d",
 317                        VK_VERSION_MAJOR(client_version),
 318                        VK_VERSION_MINOR(client_version),
 319                        VK_VERSION_PATCH(client_version));
 320    }
 321
 322    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 323       bool found = false;
 324       for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
 325          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 326                     global_extensions[j].extensionName) == 0) {
 327             found = true;
 328             break;
 329          }
 330       }
 331       if (!found)
 332          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 333    }
 334
 335    instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 336                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 337    if (!instance)
 338       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 339
 340    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 341
 342    if (pAllocator)
 343       instance->alloc = *pAllocator;
 344    else
 345       instance->alloc = default_alloc;
 346
 347    instance->apiVersion = client_version;
 348    instance->physicalDeviceCount = -1;
 349
 350    _mesa_locale_init();
 351
 352    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 353
 354    *pInstance = anv_instance_to_handle(instance);
 355
 356    return VK_SUCCESS;
 357 }
 358
 359 void anv_DestroyInstance(
 360     VkInstance                                  _instance,
 361     const VkAllocationCallbacks*                pAllocator)
 362 {
 363    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 364
 365    if (instance->physicalDeviceCount > 0) {
 366       /* We support at most one physical device. */
 367       assert(instance->physicalDeviceCount == 1);
 368       anv_physical_device_finish(&instance->physicalDevice);
 369    }
 370
 371    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 372
 373    _mesa_locale_fini();
 374
 375    vk_free(&instance->alloc, instance);
 376 }
 377
 378 VkResult anv_EnumeratePhysicalDevices(
 379     VkInstance                                  _instance,
 380     uint32_t*                                   pPhysicalDeviceCount,
 381     VkPhysicalDevice*                           pPhysicalDevices)
 382 {
 383    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 384    VkResult result;
 385
 386    if (instance->physicalDeviceCount < 0) {
 387       char path[20];
 388       for (unsigned i = 0; i < 8; i++) {
 389          snprintf(path, sizeof(path), "/dev/dri/renderD%d", 128 + i);
 390          result = anv_physical_device_init(&instance->physicalDevice,
 391                                            instance, path);
 392          if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 393             break;
 394       }
 395
 396       if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
 397          instance->physicalDeviceCount = 0;
 398       } else if (result == VK_SUCCESS) {
 399          instance->physicalDeviceCount = 1;
 400       } else {
 401          return result;
 402       }
 403    }
 404
 405    /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
 406     * otherwise it's an inout parameter.
 407     *
 408     * The Vulkan spec (git aaed022) says:
 409     *
 410     *    pPhysicalDeviceCount is a pointer to an unsigned integer variable
 411     *    that is initialized with the number of devices the application is
 412     *    prepared to receive handles to. pname:pPhysicalDevices is pointer to
 413     *    an array of at least this many VkPhysicalDevice handles [...].
 414     *
 415     *    Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
 416     *    overwrites the contents of the variable pointed to by
 417     *    pPhysicalDeviceCount with the number of physical devices in in the
 418     *    instance; otherwise, vkEnumeratePhysicalDevices overwrites
 419     *    pPhysicalDeviceCount with the number of physical handles written to
 420     *    pPhysicalDevices.
 421     */
 422    if (!pPhysicalDevices) {
 423       *pPhysicalDeviceCount = instance->physicalDeviceCount;
 424    } else if (*pPhysicalDeviceCount >= 1) {
 425       pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
 426       *pPhysicalDeviceCount = 1;
 427    } else if (*pPhysicalDeviceCount < instance->physicalDeviceCount) {
 428       return VK_INCOMPLETE;
 429    } else {
 430       *pPhysicalDeviceCount = 0;
 431    }
 432
 433    return VK_SUCCESS;
 434 }
 435
 436 void anv_GetPhysicalDeviceFeatures(
 437     VkPhysicalDevice                            physicalDevice,
 438     VkPhysicalDeviceFeatures*                   pFeatures)
 439 {
 440    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 441
 442    *pFeatures = (VkPhysicalDeviceFeatures) {
 443       .robustBufferAccess                       = true,
 444       .fullDrawIndexUint32                      = true,
 445       .imageCubeArray                           = true,
 446       .independentBlend                         = true,
 447       .geometryShader                           = true,
 448       .tessellationShader                       = true,
 449       .sampleRateShading                        = true,
 450       .dualSrcBlend                             = true,
 451       .logicOp                                  = true,
 452       .multiDrawIndirect                        = false,
 453       .drawIndirectFirstInstance                = true,
 454       .depthClamp                               = true,
 455       .depthBiasClamp                           = true,
 456       .fillModeNonSolid                         = true,
 457       .depthBounds                              = false,
 458       .wideLines                                = true,
 459       .largePoints                              = true,
 460       .alphaToOne                               = true,
 461       .multiViewport                            = true,
 462       .samplerAnisotropy                        = true,
 463       .textureCompressionETC2                   = pdevice->info.gen >= 8 ||
 464                                                   pdevice->info.is_baytrail,
 465       .textureCompressionASTC_LDR               = pdevice->info.gen >= 9, /* FINISHME CHV */
 466       .textureCompressionBC                     = true,
 467       .occlusionQueryPrecise                    = true,
 468       .pipelineStatisticsQuery                  = false,
 469       .fragmentStoresAndAtomics                 = true,
 470       .shaderTessellationAndGeometryPointSize   = true,
 471       .shaderImageGatherExtended                = true,
 472       .shaderStorageImageExtendedFormats        = true,
 473       .shaderStorageImageMultisample            = false,
 474       .shaderStorageImageReadWithoutFormat      = false,
 475       .shaderStorageImageWriteWithoutFormat     = true,
 476       .shaderUniformBufferArrayDynamicIndexing  = true,
 477       .shaderSampledImageArrayDynamicIndexing   = true,
 478       .shaderStorageBufferArrayDynamicIndexing  = true,
 479       .shaderStorageImageArrayDynamicIndexing   = true,
 480       .shaderClipDistance                       = true,
 481       .shaderCullDistance                       = true,
 482       .shaderFloat64                            = pdevice->info.gen >= 8,
 483       .shaderInt64                              = false,
 484       .shaderInt16                              = false,
 485       .shaderResourceMinLod                     = false,
 486       .variableMultisampleRate                  = false,
 487       .inheritedQueries                         = false,
 488    };
 489
 490    /* We can't do image stores in vec4 shaders */
 491    pFeatures->vertexPipelineStoresAndAtomics =
 492       pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
 493       pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
 494 }
 495
 496 void anv_GetPhysicalDeviceFeatures2KHR(
 497     VkPhysicalDevice                            physicalDevice,
 498     VkPhysicalDeviceFeatures2KHR*               pFeatures)
 499 {
 500    anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 501
 502    vk_foreach_struct(ext, pFeatures->pNext) {
 503       switch (ext->sType) {
 504       default:
 505          anv_debug_ignored_stype(ext->sType);
 506          break;
 507       }
 508    }
 509 }
 510
 511 void anv_GetPhysicalDeviceProperties(
 512     VkPhysicalDevice                            physicalDevice,
 513     VkPhysicalDeviceProperties*                 pProperties)
 514 {
 515    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 516    const struct gen_device_info *devinfo = &pdevice->info;
 517
 518    const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
 519
 520    /* See assertions made when programming the buffer surface state. */
 521    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
 522                                       (1ul << 30) : (1ul << 27);
 523
 524    VkSampleCountFlags sample_counts =
 525       isl_device_get_sample_counts(&pdevice->isl_dev);
 526
 527    VkPhysicalDeviceLimits limits = {
 528       .maxImageDimension1D                      = (1 << 14),
 529       .maxImageDimension2D                      = (1 << 14),
 530       .maxImageDimension3D                      = (1 << 11),
 531       .maxImageDimensionCube                    = (1 << 14),
 532       .maxImageArrayLayers                      = (1 << 11),
 533       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 534       .maxUniformBufferRange                    = (1ul << 27),
 535       .maxStorageBufferRange                    = max_raw_buffer_sz,
 536       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 537       .maxMemoryAllocationCount                 = UINT32_MAX,
 538       .maxSamplerAllocationCount                = 64 * 1024,
 539       .bufferImageGranularity                   = 64, /* A cache line */
 540       .sparseAddressSpaceSize                   = 0,
 541       .maxBoundDescriptorSets                   = MAX_SETS,
 542       .maxPerStageDescriptorSamplers            = 64,
 543       .maxPerStageDescriptorUniformBuffers      = 64,
 544       .maxPerStageDescriptorStorageBuffers      = 64,
 545       .maxPerStageDescriptorSampledImages       = 64,
 546       .maxPerStageDescriptorStorageImages       = 64,
 547       .maxPerStageDescriptorInputAttachments    = 64,
 548       .maxPerStageResources                     = 128,
 549       .maxDescriptorSetSamplers                 = 256,
 550       .maxDescriptorSetUniformBuffers           = 256,
 551       .maxDescriptorSetUniformBuffersDynamic    = 256,
 552       .maxDescriptorSetStorageBuffers           = 256,
 553       .maxDescriptorSetStorageBuffersDynamic    = 256,
 554       .maxDescriptorSetSampledImages            = 256,
 555       .maxDescriptorSetStorageImages            = 256,
 556       .maxDescriptorSetInputAttachments         = 256,
 557       .maxVertexInputAttributes                 = MAX_VBS,
 558       .maxVertexInputBindings                   = MAX_VBS,
 559       .maxVertexInputAttributeOffset            = 2047,
 560       .maxVertexInputBindingStride              = 2048,
 561       .maxVertexOutputComponents                = 128,
 562       .maxTessellationGenerationLevel           = 64,
 563       .maxTessellationPatchSize                 = 32,
 564       .maxTessellationControlPerVertexInputComponents = 128,
 565       .maxTessellationControlPerVertexOutputComponents = 128,
 566       .maxTessellationControlPerPatchOutputComponents = 128,
 567       .maxTessellationControlTotalOutputComponents = 2048,
 568       .maxTessellationEvaluationInputComponents = 128,
 569       .maxTessellationEvaluationOutputComponents = 128,
 570       .maxGeometryShaderInvocations             = 32,
 571       .maxGeometryInputComponents               = 64,
 572       .maxGeometryOutputComponents              = 128,
 573       .maxGeometryOutputVertices                = 256,
 574       .maxGeometryTotalOutputComponents         = 1024,
 575       .maxFragmentInputComponents               = 128,
 576       .maxFragmentOutputAttachments             = 8,
 577       .maxFragmentDualSrcAttachments            = 1,
 578       .maxFragmentCombinedOutputResources       = 8,
 579       .maxComputeSharedMemorySize               = 32768,
 580       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 581       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 582       .maxComputeWorkGroupSize = {
 583          16 * devinfo->max_cs_threads,
 584          16 * devinfo->max_cs_threads,
 585          16 * devinfo->max_cs_threads,
 586       },
 587       .subPixelPrecisionBits                    = 4 /* FIXME */,
 588       .subTexelPrecisionBits                    = 4 /* FIXME */,
 589       .mipmapPrecisionBits                      = 4 /* FIXME */,
 590       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 591       .maxDrawIndirectCount                     = UINT32_MAX,
 592       .maxSamplerLodBias                        = 16,
 593       .maxSamplerAnisotropy                     = 16,
 594       .maxViewports                             = MAX_VIEWPORTS,
 595       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 596       .viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
 597       .viewportSubPixelBits                     = 13, /* We take a float? */
 598       .minMemoryMapAlignment                    = 4096, /* A page */
 599       .minTexelBufferOffsetAlignment            = 1,
 600       .minUniformBufferOffsetAlignment          = 16,
 601       .minStorageBufferOffsetAlignment          = 4,
 602       .minTexelOffset                           = -8,
 603       .maxTexelOffset                           = 7,
 604       .minTexelGatherOffset                     = -32,
 605       .maxTexelGatherOffset                     = 31,
 606       .minInterpolationOffset                   = -0.5,
 607       .maxInterpolationOffset                   = 0.4375,
 608       .subPixelInterpolationOffsetBits          = 4,
 609       .maxFramebufferWidth                      = (1 << 14),
 610       .maxFramebufferHeight                     = (1 << 14),
 611       .maxFramebufferLayers                     = (1 << 11),
 612       .framebufferColorSampleCounts             = sample_counts,
 613       .framebufferDepthSampleCounts             = sample_counts,
 614       .framebufferStencilSampleCounts           = sample_counts,
 615       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 616       .maxColorAttachments                      = MAX_RTS,
 617       .sampledImageColorSampleCounts            = sample_counts,
 618       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 619       .sampledImageDepthSampleCounts            = sample_counts,
 620       .sampledImageStencilSampleCounts          = sample_counts,
 621       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 622       .maxSampleMaskWords                       = 1,
 623       .timestampComputeAndGraphics              = false,
 624       .timestampPeriod                          = time_stamp_base,
 625       .maxClipDistances                         = 8,
 626       .maxCullDistances                         = 8,
 627       .maxCombinedClipAndCullDistances          = 8,
 628       .discreteQueuePriorities                  = 1,
 629       .pointSizeRange                           = { 0.125, 255.875 },
 630       .lineWidthRange                           = { 0.0, 7.9921875 },
 631       .pointSizeGranularity                     = (1.0 / 8.0),
 632       .lineWidthGranularity                     = (1.0 / 128.0),
 633       .strictLines                              = false, /* FINISHME */
 634       .standardSampleLocations                  = true,
 635       .optimalBufferCopyOffsetAlignment         = 128,
 636       .optimalBufferCopyRowPitchAlignment       = 128,
 637       .nonCoherentAtomSize                      = 64,
 638    };
 639
 640    *pProperties = (VkPhysicalDeviceProperties) {
 641       .apiVersion = VK_MAKE_VERSION(1, 0, 39),
 642       .driverVersion = 1,
 643       .vendorID = 0x8086,
 644       .deviceID = pdevice->chipset_id,
 645       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 646       .limits = limits,
 647       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 648    };
 649
 650    strcpy(pProperties->deviceName, pdevice->name);
 651    memcpy(pProperties->pipelineCacheUUID, pdevice->uuid, VK_UUID_SIZE);
 652 }
 653
 654 void anv_GetPhysicalDeviceProperties2KHR(
 655     VkPhysicalDevice                            physicalDevice,
 656     VkPhysicalDeviceProperties2KHR*             pProperties)
 657 {
 658    anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 659
 660    vk_foreach_struct(ext, pProperties->pNext) {
 661       switch (ext->sType) {
 662       default:
 663          anv_debug_ignored_stype(ext->sType);
 664          break;
 665       }
 666    }
 667 }
 668
 669 static void
 670 anv_get_queue_family_properties(struct anv_physical_device *phys_dev,
 671                                 VkQueueFamilyProperties *props)
 672 {
 673    *props = (VkQueueFamilyProperties) {
 674       .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 675                     VK_QUEUE_COMPUTE_BIT |
 676                     VK_QUEUE_TRANSFER_BIT,
 677       .queueCount = 1,
 678       .timestampValidBits = 36, /* XXX: Real value here */
 679       .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
 680    };
 681 }
 682
 683 void anv_GetPhysicalDeviceQueueFamilyProperties(
 684     VkPhysicalDevice                            physicalDevice,
 685     uint32_t*                                   pCount,
 686     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 687 {
 688    ANV_FROM_HANDLE(anv_physical_device, phys_dev, physicalDevice);
 689
 690    if (pQueueFamilyProperties == NULL) {
 691       *pCount = 1;
 692       return;
 693    }
 694
 695    /* The spec implicitly allows the incoming count to be 0. From the Vulkan
 696     * 1.0.38 spec, Section 4.1 Physical Devices:
 697     *
 698     *     If the value referenced by pQueueFamilyPropertyCount is not 0 [then
 699     *     do stuff].
 700     */
 701    if (*pCount == 0)
 702       return;
 703
 704    *pCount = 1;
 705    anv_get_queue_family_properties(phys_dev, pQueueFamilyProperties);
 706 }
 707
 708 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
 709     VkPhysicalDevice                            physicalDevice,
 710     uint32_t*                                   pQueueFamilyPropertyCount,
 711     VkQueueFamilyProperties2KHR*                pQueueFamilyProperties)
 712 {
 713
 714    ANV_FROM_HANDLE(anv_physical_device, phys_dev, physicalDevice);
 715
 716    if (pQueueFamilyProperties == NULL) {
 717       *pQueueFamilyPropertyCount = 1;
 718       return;
 719    }
 720
 721    /* The spec implicitly allows the incoming count to be 0. From the Vulkan
 722     * 1.0.38 spec, Section 4.1 Physical Devices:
 723     *
 724     *     If the value referenced by pQueueFamilyPropertyCount is not 0 [then
 725     *     do stuff].
 726     */
 727    if (*pQueueFamilyPropertyCount == 0)
 728       return;
 729
 730    /* We support exactly one queue family. So need to traverse only the first
 731     * array element's pNext chain.
 732     */
 733    *pQueueFamilyPropertyCount = 1;
 734    anv_get_queue_family_properties(phys_dev,
 735          &pQueueFamilyProperties->queueFamilyProperties);
 736
 737    vk_foreach_struct(ext, pQueueFamilyProperties->pNext) {
 738       switch (ext->sType) {
 739       default:
 740          anv_debug_ignored_stype(ext->sType);
 741          break;
 742       }
 743    }
 744 }
 745
 746 void anv_GetPhysicalDeviceMemoryProperties(
 747     VkPhysicalDevice                            physicalDevice,
 748     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
 749 {
 750    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 751    VkDeviceSize heap_size;
 752
 753    /* Reserve some wiggle room for the driver by exposing only 75% of the
 754     * aperture to the heap.
 755     */
 756    heap_size = 3 * physical_device->aperture_size / 4;
 757
 758    if (physical_device->info.has_llc) {
 759       /* Big core GPUs share LLC with the CPU and thus one memory type can be
 760        * both cached and coherent at the same time.
 761        */
 762       pMemoryProperties->memoryTypeCount = 1;
 763       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 764          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 765                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 766                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
 767                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 768          .heapIndex = 0,
 769       };
 770    } else {
 771       /* The spec requires that we expose a host-visible, coherent memory
 772        * type, but Atom GPUs don't share LLC. Thus we offer two memory types
 773        * to give the application a choice between cached, but not coherent and
 774        * coherent but uncached (WC though).
 775        */
 776       pMemoryProperties->memoryTypeCount = 2;
 777       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 778          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 779                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 780                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
 781          .heapIndex = 0,
 782       };
 783       pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
 784          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 785                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 786                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 787          .heapIndex = 0,
 788       };
 789    }
 790
 791    pMemoryProperties->memoryHeapCount = 1;
 792    pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
 793       .size = heap_size,
 794       .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 795    };
 796 }
 797
 798 void anv_GetPhysicalDeviceMemoryProperties2KHR(
 799     VkPhysicalDevice                            physicalDevice,
 800     VkPhysicalDeviceMemoryProperties2KHR*       pMemoryProperties)
 801 {
 802    anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
 803                                          &pMemoryProperties->memoryProperties);
 804
 805    vk_foreach_struct(ext, pMemoryProperties->pNext) {
 806       switch (ext->sType) {
 807       default:
 808          anv_debug_ignored_stype(ext->sType);
 809          break;
 810       }
 811    }
 812 }
 813
 814 PFN_vkVoidFunction anv_GetInstanceProcAddr(
 815     VkInstance                                  instance,
 816     const char*                                 pName)
 817 {
 818    return anv_lookup_entrypoint(NULL, pName);
 819 }
 820
 821 /* With version 1+ of the loader interface the ICD should expose
 822  * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
 823  */
 824 PUBLIC
 825 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 826     VkInstance                                  instance,
 827     const char*                                 pName);
 828
 829 PUBLIC
 830 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 831     VkInstance                                  instance,
 832     const char*                                 pName)
 833 {
 834    return anv_GetInstanceProcAddr(instance, pName);
 835 }
 836
 837 PFN_vkVoidFunction anv_GetDeviceProcAddr(
 838     VkDevice                                    _device,
 839     const char*                                 pName)
 840 {
 841    ANV_FROM_HANDLE(anv_device, device, _device);
 842    return anv_lookup_entrypoint(&device->info, pName);
 843 }
 844
 845 static void
 846 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
 847 {
 848    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 849    queue->device = device;
 850    queue->pool = &device->surface_state_pool;
 851 }
 852
 853 static void
 854 anv_queue_finish(struct anv_queue *queue)
 855 {
 856 }
 857
 858 static struct anv_state
 859 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
 860 {
 861    struct anv_state state;
 862
 863    state = anv_state_pool_alloc(pool, size, align);
 864    memcpy(state.map, p, size);
 865
 866    if (!pool->block_pool->device->info.has_llc)
 867       anv_state_clflush(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_clflush_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    anv_device_finish_blorp(device);
1102
1103    anv_queue_finish(&device->queue);
1104
1105 #ifdef HAVE_VALGRIND
1106    /* We only need to free these to prevent valgrind errors.  The backing
1107     * BO will go away in a couple of lines so we don't actually leak.
1108     */
1109    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1110 #endif
1111
1112    anv_scratch_pool_finish(device, &device->scratch_pool);
1113
1114    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1115    anv_gem_close(device, device->workaround_bo.gem_handle);
1116
1117    anv_state_pool_finish(&device->surface_state_pool);
1118    anv_block_pool_finish(&device->surface_state_block_pool);
1119    anv_state_pool_finish(&device->instruction_state_pool);
1120    anv_block_pool_finish(&device->instruction_block_pool);
1121    anv_state_pool_finish(&device->dynamic_state_pool);
1122    anv_block_pool_finish(&device->dynamic_state_block_pool);
1123
1124    anv_bo_pool_finish(&device->batch_bo_pool);
1125
1126    pthread_cond_destroy(&device->queue_submit);
1127    pthread_mutex_destroy(&device->mutex);
1128
1129    anv_gem_destroy_context(device, device->context_id);
1130
1131    close(device->fd);
1132
1133    vk_free(&device->alloc, device);
1134 }
1135
1136 VkResult anv_EnumerateInstanceExtensionProperties(
1137     const char*                                 pLayerName,
1138     uint32_t*                                   pPropertyCount,
1139     VkExtensionProperties*                      pProperties)
1140 {
1141    if (pProperties == NULL) {
1142       *pPropertyCount = ARRAY_SIZE(global_extensions);
1143       return VK_SUCCESS;
1144    }
1145
1146    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(global_extensions));
1147    typed_memcpy(pProperties, global_extensions, *pPropertyCount);
1148
1149    if (*pPropertyCount < ARRAY_SIZE(global_extensions))
1150       return VK_INCOMPLETE;
1151
1152    return VK_SUCCESS;
1153 }
1154
1155 VkResult anv_EnumerateDeviceExtensionProperties(
1156     VkPhysicalDevice                            physicalDevice,
1157     const char*                                 pLayerName,
1158     uint32_t*                                   pPropertyCount,
1159     VkExtensionProperties*                      pProperties)
1160 {
1161    if (pProperties == NULL) {
1162       *pPropertyCount = ARRAY_SIZE(device_extensions);
1163       return VK_SUCCESS;
1164    }
1165
1166    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(device_extensions));
1167    typed_memcpy(pProperties, device_extensions, *pPropertyCount);
1168
1169    if (*pPropertyCount < ARRAY_SIZE(device_extensions))
1170       return VK_INCOMPLETE;
1171
1172    return VK_SUCCESS;
1173 }
1174
1175 VkResult anv_EnumerateInstanceLayerProperties(
1176     uint32_t*                                   pPropertyCount,
1177     VkLayerProperties*                          pProperties)
1178 {
1179    if (pProperties == NULL) {
1180       *pPropertyCount = 0;
1181       return VK_SUCCESS;
1182    }
1183
1184    /* None supported at this time */
1185    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1186 }
1187
1188 VkResult anv_EnumerateDeviceLayerProperties(
1189     VkPhysicalDevice                            physicalDevice,
1190     uint32_t*                                   pPropertyCount,
1191     VkLayerProperties*                          pProperties)
1192 {
1193    if (pProperties == NULL) {
1194       *pPropertyCount = 0;
1195       return VK_SUCCESS;
1196    }
1197
1198    /* None supported at this time */
1199    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1200 }
1201
1202 void anv_GetDeviceQueue(
1203     VkDevice                                    _device,
1204     uint32_t                                    queueNodeIndex,
1205     uint32_t                                    queueIndex,
1206     VkQueue*                                    pQueue)
1207 {
1208    ANV_FROM_HANDLE(anv_device, device, _device);
1209
1210    assert(queueIndex == 0);
1211
1212    *pQueue = anv_queue_to_handle(&device->queue);
1213 }
1214
1215 VkResult
1216 anv_device_execbuf(struct anv_device *device,
1217                    struct drm_i915_gem_execbuffer2 *execbuf,
1218                    struct anv_bo **execbuf_bos)
1219 {
1220    int ret = anv_gem_execbuffer(device, execbuf);
1221    if (ret != 0) {
1222       /* We don't know the real error. */
1223       return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
1224    }
1225
1226    struct drm_i915_gem_exec_object2 *objects =
1227       (void *)(uintptr_t)execbuf->buffers_ptr;
1228    for (uint32_t k = 0; k < execbuf->buffer_count; k++)
1229       execbuf_bos[k]->offset = objects[k].offset;
1230
1231    return VK_SUCCESS;
1232 }
1233
1234 VkResult anv_QueueSubmit(
1235     VkQueue                                     _queue,
1236     uint32_t                                    submitCount,
1237     const VkSubmitInfo*                         pSubmits,
1238     VkFence                                     _fence)
1239 {
1240    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1241    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1242    struct anv_device *device = queue->device;
1243    VkResult result = VK_SUCCESS;
1244
1245    /* We lock around QueueSubmit for three main reasons:
1246     *
1247     *  1) When a block pool is resized, we create a new gem handle with a
1248     *     different size and, in the case of surface states, possibly a
1249     *     different center offset but we re-use the same anv_bo struct when
1250     *     we do so.  If this happens in the middle of setting up an execbuf,
1251     *     we could end up with our list of BOs out of sync with our list of
1252     *     gem handles.
1253     *
1254     *  2) The algorithm we use for building the list of unique buffers isn't
1255     *     thread-safe.  While the client is supposed to syncronize around
1256     *     QueueSubmit, this would be extremely difficult to debug if it ever
1257     *     came up in the wild due to a broken app.  It's better to play it
1258     *     safe and just lock around QueueSubmit.
1259     *
1260     *  3)  The anv_cmd_buffer_execbuf function may perform relocations in
1261     *      userspace.  Due to the fact that the surface state buffer is shared
1262     *      between batches, we can't afford to have that happen from multiple
1263     *      threads at the same time.  Even though the user is supposed to
1264     *      ensure this doesn't happen, we play it safe as in (2) above.
1265     *
1266     * Since the only other things that ever take the device lock such as block
1267     * pool resize only rarely happen, this will almost never be contended so
1268     * taking a lock isn't really an expensive operation in this case.
1269     */
1270    pthread_mutex_lock(&device->mutex);
1271
1272    for (uint32_t i = 0; i < submitCount; i++) {
1273       for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1274          ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1275                          pSubmits[i].pCommandBuffers[j]);
1276          assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1277
1278          result = anv_cmd_buffer_execbuf(device, cmd_buffer);
1279          if (result != VK_SUCCESS)
1280             goto out;
1281       }
1282    }
1283
1284    if (fence) {
1285       struct anv_bo *fence_bo = &fence->bo;
1286       result = anv_device_execbuf(device, &fence->execbuf, &fence_bo);
1287       if (result != VK_SUCCESS)
1288          goto out;
1289
1290       /* Update the fence and wake up any waiters */
1291       assert(fence->state == ANV_FENCE_STATE_RESET);
1292       fence->state = ANV_FENCE_STATE_SUBMITTED;
1293       pthread_cond_broadcast(&device->queue_submit);
1294    }
1295
1296 out:
1297    pthread_mutex_unlock(&device->mutex);
1298
1299    return result;
1300 }
1301
1302 VkResult anv_QueueWaitIdle(
1303     VkQueue                                     _queue)
1304 {
1305    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1306
1307    return anv_DeviceWaitIdle(anv_device_to_handle(queue->device));
1308 }
1309
1310 VkResult anv_DeviceWaitIdle(
1311     VkDevice                                    _device)
1312 {
1313    ANV_FROM_HANDLE(anv_device, device, _device);
1314    struct anv_batch batch;
1315
1316    uint32_t cmds[8];
1317    batch.start = batch.next = cmds;
1318    batch.end = (void *) cmds + sizeof(cmds);
1319
1320    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1321    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1322
1323    return anv_device_submit_simple_batch(device, &batch);
1324 }
1325
1326 VkResult
1327 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1328 {
1329    uint32_t gem_handle = anv_gem_create(device, size);
1330    if (!gem_handle)
1331       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1332
1333    anv_bo_init(bo, gem_handle, size);
1334
1335    return VK_SUCCESS;
1336 }
1337
1338 VkResult anv_AllocateMemory(
1339     VkDevice                                    _device,
1340     const VkMemoryAllocateInfo*                 pAllocateInfo,
1341     const VkAllocationCallbacks*                pAllocator,
1342     VkDeviceMemory*                             pMem)
1343 {
1344    ANV_FROM_HANDLE(anv_device, device, _device);
1345    struct anv_device_memory *mem;
1346    VkResult result;
1347
1348    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1349
1350    /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1351    assert(pAllocateInfo->allocationSize > 0);
1352
1353    /* We support exactly one memory heap. */
1354    assert(pAllocateInfo->memoryTypeIndex == 0 ||
1355           (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1356
1357    /* FINISHME: Fail if allocation request exceeds heap size. */
1358
1359    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1360                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1361    if (mem == NULL)
1362       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1363
1364    /* The kernel is going to give us whole pages anyway */
1365    uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1366
1367    result = anv_bo_init_new(&mem->bo, device, alloc_size);
1368    if (result != VK_SUCCESS)
1369       goto fail;
1370
1371    mem->type_index = pAllocateInfo->memoryTypeIndex;
1372
1373    mem->map = NULL;
1374    mem->map_size = 0;
1375
1376    *pMem = anv_device_memory_to_handle(mem);
1377
1378    return VK_SUCCESS;
1379
1380  fail:
1381    vk_free2(&device->alloc, pAllocator, mem);
1382
1383    return result;
1384 }
1385
1386 void anv_FreeMemory(
1387     VkDevice                                    _device,
1388     VkDeviceMemory                              _mem,
1389     const VkAllocationCallbacks*                pAllocator)
1390 {
1391    ANV_FROM_HANDLE(anv_device, device, _device);
1392    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1393
1394    if (mem == NULL)
1395       return;
1396
1397    if (mem->map)
1398       anv_UnmapMemory(_device, _mem);
1399
1400    if (mem->bo.map)
1401       anv_gem_munmap(mem->bo.map, mem->bo.size);
1402
1403    if (mem->bo.gem_handle != 0)
1404       anv_gem_close(device, mem->bo.gem_handle);
1405
1406    vk_free2(&device->alloc, pAllocator, mem);
1407 }
1408
1409 VkResult anv_MapMemory(
1410     VkDevice                                    _device,
1411     VkDeviceMemory                              _memory,
1412     VkDeviceSize                                offset,
1413     VkDeviceSize                                size,
1414     VkMemoryMapFlags                            flags,
1415     void**                                      ppData)
1416 {
1417    ANV_FROM_HANDLE(anv_device, device, _device);
1418    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1419
1420    if (mem == NULL) {
1421       *ppData = NULL;
1422       return VK_SUCCESS;
1423    }
1424
1425    if (size == VK_WHOLE_SIZE)
1426       size = mem->bo.size - offset;
1427
1428    /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1429     *
1430     *  * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1431     *    assert(size != 0);
1432     *  * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1433     *    equal to the size of the memory minus offset
1434     */
1435    assert(size > 0);
1436    assert(offset + size <= mem->bo.size);
1437
1438    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1439     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1440     * at a time is valid. We could just mmap up front and return an offset
1441     * pointer here, but that may exhaust virtual memory on 32 bit
1442     * userspace. */
1443
1444    uint32_t gem_flags = 0;
1445    if (!device->info.has_llc && mem->type_index == 0)
1446       gem_flags |= I915_MMAP_WC;
1447
1448    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1449    uint64_t map_offset = offset & ~4095ull;
1450    assert(offset >= map_offset);
1451    uint64_t map_size = (offset + size) - map_offset;
1452
1453    /* Let's map whole pages */
1454    map_size = align_u64(map_size, 4096);
1455
1456    void *map = anv_gem_mmap(device, mem->bo.gem_handle,
1457                             map_offset, map_size, gem_flags);
1458    if (map == MAP_FAILED)
1459       return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
1460
1461    mem->map = map;
1462    mem->map_size = map_size;
1463
1464    *ppData = mem->map + (offset - map_offset);
1465
1466    return VK_SUCCESS;
1467 }
1468
1469 void anv_UnmapMemory(
1470     VkDevice                                    _device,
1471     VkDeviceMemory                              _memory)
1472 {
1473    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1474
1475    if (mem == NULL)
1476       return;
1477
1478    anv_gem_munmap(mem->map, mem->map_size);
1479
1480    mem->map = NULL;
1481    mem->map_size = 0;
1482 }
1483
1484 static void
1485 clflush_mapped_ranges(struct anv_device         *device,
1486                       uint32_t                   count,
1487                       const VkMappedMemoryRange *ranges)
1488 {
1489    for (uint32_t i = 0; i < count; i++) {
1490       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1491       void *p = mem->map + (ranges[i].offset & ~CACHELINE_MASK);
1492       void *end;
1493
1494       if (ranges[i].offset + ranges[i].size > mem->map_size)
1495          end = mem->map + mem->map_size;
1496       else
1497          end = mem->map + ranges[i].offset + ranges[i].size;
1498
1499       while (p < end) {
1500          __builtin_ia32_clflush(p);
1501          p += CACHELINE_SIZE;
1502       }
1503    }
1504 }
1505
1506 VkResult anv_FlushMappedMemoryRanges(
1507     VkDevice                                    _device,
1508     uint32_t                                    memoryRangeCount,
1509     const VkMappedMemoryRange*                  pMemoryRanges)
1510 {
1511    ANV_FROM_HANDLE(anv_device, device, _device);
1512
1513    if (device->info.has_llc)
1514       return VK_SUCCESS;
1515
1516    /* Make sure the writes we're flushing have landed. */
1517    __builtin_ia32_mfence();
1518
1519    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1520
1521    return VK_SUCCESS;
1522 }
1523
1524 VkResult anv_InvalidateMappedMemoryRanges(
1525     VkDevice                                    _device,
1526     uint32_t                                    memoryRangeCount,
1527     const VkMappedMemoryRange*                  pMemoryRanges)
1528 {
1529    ANV_FROM_HANDLE(anv_device, device, _device);
1530
1531    if (device->info.has_llc)
1532       return VK_SUCCESS;
1533
1534    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1535
1536    /* Make sure no reads get moved up above the invalidate. */
1537    __builtin_ia32_mfence();
1538
1539    return VK_SUCCESS;
1540 }
1541
1542 void anv_GetBufferMemoryRequirements(
1543     VkDevice                                    _device,
1544     VkBuffer                                    _buffer,
1545     VkMemoryRequirements*                       pMemoryRequirements)
1546 {
1547    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1548    ANV_FROM_HANDLE(anv_device, device, _device);
1549
1550    /* The Vulkan spec (git aaed022) says:
1551     *
1552     *    memoryTypeBits is a bitfield and contains one bit set for every
1553     *    supported memory type for the resource. The bit `1<<i` is set if and
1554     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1555     *    structure for the physical device is supported.
1556     *
1557     * We support exactly one memory type on LLC, two on non-LLC.
1558     */
1559    pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1560
1561    pMemoryRequirements->size = buffer->size;
1562    pMemoryRequirements->alignment = 16;
1563 }
1564
1565 void anv_GetImageMemoryRequirements(
1566     VkDevice                                    _device,
1567     VkImage                                     _image,
1568     VkMemoryRequirements*                       pMemoryRequirements)
1569 {
1570    ANV_FROM_HANDLE(anv_image, image, _image);
1571    ANV_FROM_HANDLE(anv_device, device, _device);
1572
1573    /* The Vulkan spec (git aaed022) says:
1574     *
1575     *    memoryTypeBits is a bitfield and contains one bit set for every
1576     *    supported memory type for the resource. The bit `1<<i` is set if and
1577     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1578     *    structure for the physical device is supported.
1579     *
1580     * We support exactly one memory type on LLC, two on non-LLC.
1581     */
1582    pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1583
1584    pMemoryRequirements->size = image->size;
1585    pMemoryRequirements->alignment = image->alignment;
1586 }
1587
1588 void anv_GetImageSparseMemoryRequirements(
1589     VkDevice                                    device,
1590     VkImage                                     image,
1591     uint32_t*                                   pSparseMemoryRequirementCount,
1592     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1593 {
1594    stub();
1595 }
1596
1597 void anv_GetDeviceMemoryCommitment(
1598     VkDevice                                    device,
1599     VkDeviceMemory                              memory,
1600     VkDeviceSize*                               pCommittedMemoryInBytes)
1601 {
1602    *pCommittedMemoryInBytes = 0;
1603 }
1604
1605 VkResult anv_BindBufferMemory(
1606     VkDevice                                    device,
1607     VkBuffer                                    _buffer,
1608     VkDeviceMemory                              _memory,
1609     VkDeviceSize                                memoryOffset)
1610 {
1611    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1612    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1613
1614    if (mem) {
1615       buffer->bo = &mem->bo;
1616       buffer->offset = memoryOffset;
1617    } else {
1618       buffer->bo = NULL;
1619       buffer->offset = 0;
1620    }
1621
1622    return VK_SUCCESS;
1623 }
1624
1625 VkResult anv_QueueBindSparse(
1626     VkQueue                                     queue,
1627     uint32_t                                    bindInfoCount,
1628     const VkBindSparseInfo*                     pBindInfo,
1629     VkFence                                     fence)
1630 {
1631    stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1632 }
1633
1634 VkResult anv_CreateFence(
1635     VkDevice                                    _device,
1636     const VkFenceCreateInfo*                    pCreateInfo,
1637     const VkAllocationCallbacks*                pAllocator,
1638     VkFence*                                    pFence)
1639 {
1640    ANV_FROM_HANDLE(anv_device, device, _device);
1641    struct anv_bo fence_bo;
1642    struct anv_fence *fence;
1643    struct anv_batch batch;
1644    VkResult result;
1645
1646    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1647
1648    result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1649    if (result != VK_SUCCESS)
1650       return result;
1651
1652    /* Fences are small.  Just store the CPU data structure in the BO. */
1653    fence = fence_bo.map;
1654    fence->bo = fence_bo;
1655
1656    /* Place the batch after the CPU data but on its own cache line. */
1657    const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1658    batch.next = batch.start = fence->bo.map + batch_offset;
1659    batch.end = fence->bo.map + fence->bo.size;
1660    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1661    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1662
1663    if (!device->info.has_llc) {
1664       assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1665       assert(batch.next - batch.start <= CACHELINE_SIZE);
1666       __builtin_ia32_mfence();
1667       __builtin_ia32_clflush(batch.start);
1668    }
1669
1670    fence->exec2_objects[0].handle = fence->bo.gem_handle;
1671    fence->exec2_objects[0].relocation_count = 0;
1672    fence->exec2_objects[0].relocs_ptr = 0;
1673    fence->exec2_objects[0].alignment = 0;
1674    fence->exec2_objects[0].offset = fence->bo.offset;
1675    fence->exec2_objects[0].flags = 0;
1676    fence->exec2_objects[0].rsvd1 = 0;
1677    fence->exec2_objects[0].rsvd2 = 0;
1678
1679    fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1680    fence->execbuf.buffer_count = 1;
1681    fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1682    fence->execbuf.batch_len = batch.next - batch.start;
1683    fence->execbuf.cliprects_ptr = 0;
1684    fence->execbuf.num_cliprects = 0;
1685    fence->execbuf.DR1 = 0;
1686    fence->execbuf.DR4 = 0;
1687
1688    fence->execbuf.flags =
1689       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1690    fence->execbuf.rsvd1 = device->context_id;
1691    fence->execbuf.rsvd2 = 0;
1692
1693    if (pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT) {
1694       fence->state = ANV_FENCE_STATE_SIGNALED;
1695    } else {
1696       fence->state = ANV_FENCE_STATE_RESET;
1697    }
1698
1699    *pFence = anv_fence_to_handle(fence);
1700
1701    return VK_SUCCESS;
1702 }
1703
1704 void anv_DestroyFence(
1705     VkDevice                                    _device,
1706     VkFence                                     _fence,
1707     const VkAllocationCallbacks*                pAllocator)
1708 {
1709    ANV_FROM_HANDLE(anv_device, device, _device);
1710    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1711
1712    if (!fence)
1713       return;
1714
1715    assert(fence->bo.map == fence);
1716    anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1717 }
1718
1719 VkResult anv_ResetFences(
1720     VkDevice                                    _device,
1721     uint32_t                                    fenceCount,
1722     const VkFence*                              pFences)
1723 {
1724    for (uint32_t i = 0; i < fenceCount; i++) {
1725       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1726       fence->state = ANV_FENCE_STATE_RESET;
1727    }
1728
1729    return VK_SUCCESS;
1730 }
1731
1732 VkResult anv_GetFenceStatus(
1733     VkDevice                                    _device,
1734     VkFence                                     _fence)
1735 {
1736    ANV_FROM_HANDLE(anv_device, device, _device);
1737    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1738    int64_t t = 0;
1739    int ret;
1740
1741    switch (fence->state) {
1742    case ANV_FENCE_STATE_RESET:
1743       /* If it hasn't even been sent off to the GPU yet, it's not ready */
1744       return VK_NOT_READY;
1745
1746    case ANV_FENCE_STATE_SIGNALED:
1747       /* It's been signaled, return success */
1748       return VK_SUCCESS;
1749
1750    case ANV_FENCE_STATE_SUBMITTED:
1751       /* It's been submitted to the GPU but we don't know if it's done yet. */
1752       ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1753       if (ret == 0) {
1754          fence->state = ANV_FENCE_STATE_SIGNALED;
1755          return VK_SUCCESS;
1756       } else {
1757          return VK_NOT_READY;
1758       }
1759    default:
1760       unreachable("Invalid fence status");
1761    }
1762 }
1763
1764 #define NSEC_PER_SEC 1000000000
1765 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
1766
1767 VkResult anv_WaitForFences(
1768     VkDevice                                    _device,
1769     uint32_t                                    fenceCount,
1770     const VkFence*                              pFences,
1771     VkBool32                                    waitAll,
1772     uint64_t                                    _timeout)
1773 {
1774    ANV_FROM_HANDLE(anv_device, device, _device);
1775    int ret;
1776
1777    /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1778     * to block indefinitely timeouts <= 0.  Unfortunately, this was broken
1779     * for a couple of kernel releases.  Since there's no way to know
1780     * whether or not the kernel we're using is one of the broken ones, the
1781     * best we can do is to clamp the timeout to INT64_MAX.  This limits the
1782     * maximum timeout from 584 years to 292 years - likely not a big deal.
1783     */
1784    int64_t timeout = MIN2(_timeout, INT64_MAX);
1785
1786    uint32_t pending_fences = fenceCount;
1787    while (pending_fences) {
1788       pending_fences = 0;
1789       bool signaled_fences = false;
1790       for (uint32_t i = 0; i < fenceCount; i++) {
1791          ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1792          switch (fence->state) {
1793          case ANV_FENCE_STATE_RESET:
1794             /* This fence hasn't been submitted yet, we'll catch it the next
1795              * time around.  Yes, this may mean we dead-loop but, short of
1796              * lots of locking and a condition variable, there's not much that
1797              * we can do about that.
1798              */
1799             pending_fences++;
1800             continue;
1801
1802          case ANV_FENCE_STATE_SIGNALED:
1803             /* This fence is not pending.  If waitAll isn't set, we can return
1804              * early.  Otherwise, we have to keep going.
1805              */
1806             if (!waitAll)
1807                return VK_SUCCESS;
1808             continue;
1809
1810          case ANV_FENCE_STATE_SUBMITTED:
1811             /* These are the fences we really care about.  Go ahead and wait
1812              * on it until we hit a timeout.
1813              */
1814             ret = anv_gem_wait(device, fence->bo.gem_handle, &timeout);
1815             if (ret == -1 && errno == ETIME) {
1816                return VK_TIMEOUT;
1817             } else if (ret == -1) {
1818                /* We don't know the real error. */
1819                return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1820             } else {
1821                fence->state = ANV_FENCE_STATE_SIGNALED;
1822                signaled_fences = true;
1823                if (!waitAll)
1824                   return VK_SUCCESS;
1825                continue;
1826             }
1827          }
1828       }
1829
1830       if (pending_fences && !signaled_fences) {
1831          /* If we've hit this then someone decided to vkWaitForFences before
1832           * they've actually submitted any of them to a queue.  This is a
1833           * fairly pessimal case, so it's ok to lock here and use a standard
1834           * pthreads condition variable.
1835           */
1836          pthread_mutex_lock(&device->mutex);
1837
1838          /* It's possible that some of the fences have changed state since the
1839           * last time we checked.  Now that we have the lock, check for
1840           * pending fences again and don't wait if it's changed.
1841           */
1842          uint32_t now_pending_fences = 0;
1843          for (uint32_t i = 0; i < fenceCount; i++) {
1844             ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1845             if (fence->state == ANV_FENCE_STATE_RESET)
1846                now_pending_fences++;
1847          }
1848          assert(now_pending_fences <= pending_fences);
1849
1850          if (now_pending_fences == pending_fences) {
1851             struct timespec before;
1852             clock_gettime(CLOCK_MONOTONIC, &before);
1853
1854             uint32_t abs_nsec = before.tv_nsec + timeout % NSEC_PER_SEC;
1855             uint64_t abs_sec = before.tv_sec + (abs_nsec / NSEC_PER_SEC) +
1856                                (timeout / NSEC_PER_SEC);
1857             abs_nsec %= NSEC_PER_SEC;
1858
1859             /* Avoid roll-over in tv_sec on 32-bit systems if the user
1860              * provided timeout is UINT64_MAX
1861              */
1862             struct timespec abstime;
1863             abstime.tv_nsec = abs_nsec;
1864             abstime.tv_sec = MIN2(abs_sec, INT_TYPE_MAX(abstime.tv_sec));
1865
1866             ret = pthread_cond_timedwait(&device->queue_submit,
1867                                          &device->mutex, &abstime);
1868             assert(ret != EINVAL);
1869
1870             struct timespec after;
1871             clock_gettime(CLOCK_MONOTONIC, &after);
1872             uint64_t time_elapsed =
1873                ((uint64_t)after.tv_sec * NSEC_PER_SEC + after.tv_nsec) -
1874                ((uint64_t)before.tv_sec * NSEC_PER_SEC + before.tv_nsec);
1875
1876             if (time_elapsed >= timeout) {
1877                pthread_mutex_unlock(&device->mutex);
1878                return VK_TIMEOUT;
1879             }
1880
1881             timeout -= time_elapsed;
1882          }
1883
1884          pthread_mutex_unlock(&device->mutex);
1885       }
1886    }
1887
1888    return VK_SUCCESS;
1889 }
1890
1891 // Queue semaphore functions
1892
1893 VkResult anv_CreateSemaphore(
1894     VkDevice                                    device,
1895     const VkSemaphoreCreateInfo*                pCreateInfo,
1896     const VkAllocationCallbacks*                pAllocator,
1897     VkSemaphore*                                pSemaphore)
1898 {
1899    /* The DRM execbuffer ioctl always execute in-oder, even between different
1900     * rings. As such, there's nothing to do for the user space semaphore.
1901     */
1902
1903    *pSemaphore = (VkSemaphore)1;
1904
1905    return VK_SUCCESS;
1906 }
1907
1908 void anv_DestroySemaphore(
1909     VkDevice                                    device,
1910     VkSemaphore                                 semaphore,
1911     const VkAllocationCallbacks*                pAllocator)
1912 {
1913 }
1914
1915 // Event functions
1916
1917 VkResult anv_CreateEvent(
1918     VkDevice                                    _device,
1919     const VkEventCreateInfo*                    pCreateInfo,
1920     const VkAllocationCallbacks*                pAllocator,
1921     VkEvent*                                    pEvent)
1922 {
1923    ANV_FROM_HANDLE(anv_device, device, _device);
1924    struct anv_state state;
1925    struct anv_event *event;
1926
1927    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1928
1929    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1930                                 sizeof(*event), 8);
1931    event = state.map;
1932    event->state = state;
1933    event->semaphore = VK_EVENT_RESET;
1934
1935    if (!device->info.has_llc) {
1936       /* Make sure the writes we're flushing have landed. */
1937       __builtin_ia32_mfence();
1938       __builtin_ia32_clflush(event);
1939    }
1940
1941    *pEvent = anv_event_to_handle(event);
1942
1943    return VK_SUCCESS;
1944 }
1945
1946 void anv_DestroyEvent(
1947     VkDevice                                    _device,
1948     VkEvent                                     _event,
1949     const VkAllocationCallbacks*                pAllocator)
1950 {
1951    ANV_FROM_HANDLE(anv_device, device, _device);
1952    ANV_FROM_HANDLE(anv_event, event, _event);
1953
1954    if (!event)
1955       return;
1956
1957    anv_state_pool_free(&device->dynamic_state_pool, event->state);
1958 }
1959
1960 VkResult anv_GetEventStatus(
1961     VkDevice                                    _device,
1962     VkEvent                                     _event)
1963 {
1964    ANV_FROM_HANDLE(anv_device, device, _device);
1965    ANV_FROM_HANDLE(anv_event, event, _event);
1966
1967    if (!device->info.has_llc) {
1968       /* Invalidate read cache before reading event written by GPU. */
1969       __builtin_ia32_clflush(event);
1970       __builtin_ia32_mfence();
1971
1972    }
1973
1974    return event->semaphore;
1975 }
1976
1977 VkResult anv_SetEvent(
1978     VkDevice                                    _device,
1979     VkEvent                                     _event)
1980 {
1981    ANV_FROM_HANDLE(anv_device, device, _device);
1982    ANV_FROM_HANDLE(anv_event, event, _event);
1983
1984    event->semaphore = VK_EVENT_SET;
1985
1986    if (!device->info.has_llc) {
1987       /* Make sure the writes we're flushing have landed. */
1988       __builtin_ia32_mfence();
1989       __builtin_ia32_clflush(event);
1990    }
1991
1992    return VK_SUCCESS;
1993 }
1994
1995 VkResult anv_ResetEvent(
1996     VkDevice                                    _device,
1997     VkEvent                                     _event)
1998 {
1999    ANV_FROM_HANDLE(anv_device, device, _device);
2000    ANV_FROM_HANDLE(anv_event, event, _event);
2001
2002    event->semaphore = VK_EVENT_RESET;
2003
2004    if (!device->info.has_llc) {
2005       /* Make sure the writes we're flushing have landed. */
2006       __builtin_ia32_mfence();
2007       __builtin_ia32_clflush(event);
2008    }
2009
2010    return VK_SUCCESS;
2011 }
2012
2013 // Buffer functions
2014
2015 VkResult anv_CreateBuffer(
2016     VkDevice                                    _device,
2017     const VkBufferCreateInfo*                   pCreateInfo,
2018     const VkAllocationCallbacks*                pAllocator,
2019     VkBuffer*                                   pBuffer)
2020 {
2021    ANV_FROM_HANDLE(anv_device, device, _device);
2022    struct anv_buffer *buffer;
2023
2024    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2025
2026    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2027                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2028    if (buffer == NULL)
2029       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2030
2031    buffer->size = pCreateInfo->size;
2032    buffer->usage = pCreateInfo->usage;
2033    buffer->bo = NULL;
2034    buffer->offset = 0;
2035
2036    *pBuffer = anv_buffer_to_handle(buffer);
2037
2038    return VK_SUCCESS;
2039 }
2040
2041 void anv_DestroyBuffer(
2042     VkDevice                                    _device,
2043     VkBuffer                                    _buffer,
2044     const VkAllocationCallbacks*                pAllocator)
2045 {
2046    ANV_FROM_HANDLE(anv_device, device, _device);
2047    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2048
2049    if (!buffer)
2050       return;
2051
2052    vk_free2(&device->alloc, pAllocator, buffer);
2053 }
2054
2055 void
2056 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
2057                               enum isl_format format,
2058                               uint32_t offset, uint32_t range, uint32_t stride)
2059 {
2060    isl_buffer_fill_state(&device->isl_dev, state.map,
2061                          .address = offset,
2062                          .mocs = device->default_mocs,
2063                          .size = range,
2064                          .format = format,
2065                          .stride = stride);
2066
2067    if (!device->info.has_llc)
2068       anv_state_clflush(state);
2069 }
2070
2071 void anv_DestroySampler(
2072     VkDevice                                    _device,
2073     VkSampler                                   _sampler,
2074     const VkAllocationCallbacks*                pAllocator)
2075 {
2076    ANV_FROM_HANDLE(anv_device, device, _device);
2077    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
2078
2079    if (!sampler)
2080       return;
2081
2082    vk_free2(&device->alloc, pAllocator, sampler);
2083 }
2084
2085 VkResult anv_CreateFramebuffer(
2086     VkDevice                                    _device,
2087     const VkFramebufferCreateInfo*              pCreateInfo,
2088     const VkAllocationCallbacks*                pAllocator,
2089     VkFramebuffer*                              pFramebuffer)
2090 {
2091    ANV_FROM_HANDLE(anv_device, device, _device);
2092    struct anv_framebuffer *framebuffer;
2093
2094    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2095
2096    size_t size = sizeof(*framebuffer) +
2097                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2098    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2099                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2100    if (framebuffer == NULL)
2101       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2102
2103    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2104    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2105       VkImageView _iview = pCreateInfo->pAttachments[i];
2106       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2107    }
2108
2109    framebuffer->width = pCreateInfo->width;
2110    framebuffer->height = pCreateInfo->height;
2111    framebuffer->layers = pCreateInfo->layers;
2112
2113    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2114
2115    return VK_SUCCESS;
2116 }
2117
2118 void anv_DestroyFramebuffer(
2119     VkDevice                                    _device,
2120     VkFramebuffer                               _fb,
2121     const VkAllocationCallbacks*                pAllocator)
2122 {
2123    ANV_FROM_HANDLE(anv_device, device, _device);
2124    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2125
2126    if (!fb)
2127       return;
2128
2129    vk_free2(&device->alloc, pAllocator, fb);
2130 }
2131
2132 /* vk_icd.h does not declare this function, so we declare it here to
2133  * suppress Wmissing-prototypes.
2134  */
2135 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2136 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2137
2138 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2139 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2140 {
2141    /* For the full details on loader interface versioning, see
2142     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2143     * What follows is a condensed summary, to help you navigate the large and
2144     * confusing official doc.
2145     *
2146     *   - Loader interface v0 is incompatible with later versions. We don't
2147     *     support it.
2148     *
2149     *   - In loader interface v1:
2150     *       - The first ICD entrypoint called by the loader is
2151     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2152     *         entrypoint.
2153     *       - The ICD must statically expose no other Vulkan symbol unless it is
2154     *         linked with -Bsymbolic.
2155     *       - Each dispatchable Vulkan handle created by the ICD must be
2156     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2157     *         ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2158     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2159     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2160     *         such loader-managed surfaces.
2161     *
2162     *    - Loader interface v2 differs from v1 in:
2163     *       - The first ICD entrypoint called by the loader is
2164     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2165     *         statically expose this entrypoint.
2166     *
2167     *    - Loader interface v3 differs from v2 in:
2168     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2169     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2170     *          because the loader no longer does so.
2171     */
2172    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2173    return VK_SUCCESS;
2174 }