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