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    VkResult result;
 394
 395    if (instance->physicalDeviceCount < 0) {
 396       char path[20];
 397       for (unsigned i = 0; i < 8; i++) {
 398          snprintf(path, sizeof(path), "/dev/dri/renderD%d", 128 + i);
 399          result = anv_physical_device_init(&instance->physicalDevice,
 400                                            instance, path);
 401          if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 402             break;
 403       }
 404
 405       if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
 406          instance->physicalDeviceCount = 0;
 407       } else if (result == VK_SUCCESS) {
 408          instance->physicalDeviceCount = 1;
 409       } else {
 410          return result;
 411       }
 412    }
 413
 414    /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
 415     * otherwise it's an inout parameter.
 416     *
 417     * The Vulkan spec (git aaed022) says:
 418     *
 419     *    pPhysicalDeviceCount is a pointer to an unsigned integer variable
 420     *    that is initialized with the number of devices the application is
 421     *    prepared to receive handles to. pname:pPhysicalDevices is pointer to
 422     *    an array of at least this many VkPhysicalDevice handles [...].
 423     *
 424     *    Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
 425     *    overwrites the contents of the variable pointed to by
 426     *    pPhysicalDeviceCount with the number of physical devices in in the
 427     *    instance; otherwise, vkEnumeratePhysicalDevices overwrites
 428     *    pPhysicalDeviceCount with the number of physical handles written to
 429     *    pPhysicalDevices.
 430     */
 431    if (!pPhysicalDevices) {
 432       *pPhysicalDeviceCount = instance->physicalDeviceCount;
 433    } else if (*pPhysicalDeviceCount >= 1) {
 434       pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
 435       *pPhysicalDeviceCount = 1;
 436    } else if (*pPhysicalDeviceCount < instance->physicalDeviceCount) {
 437       return VK_INCOMPLETE;
 438    } else {
 439       *pPhysicalDeviceCount = 0;
 440    }
 441
 442    return VK_SUCCESS;
 443 }
 444
 445 void anv_GetPhysicalDeviceFeatures(
 446     VkPhysicalDevice                            physicalDevice,
 447     VkPhysicalDeviceFeatures*                   pFeatures)
 448 {
 449    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 450
 451    *pFeatures = (VkPhysicalDeviceFeatures) {
 452       .robustBufferAccess                       = true,
 453       .fullDrawIndexUint32                      = true,
 454       .imageCubeArray                           = true,
 455       .independentBlend                         = true,
 456       .geometryShader                           = true,
 457       .tessellationShader                       = true,
 458       .sampleRateShading                        = true,
 459       .dualSrcBlend                             = true,
 460       .logicOp                                  = true,
 461       .multiDrawIndirect                        = false,
 462       .drawIndirectFirstInstance                = true,
 463       .depthClamp                               = true,
 464       .depthBiasClamp                           = true,
 465       .fillModeNonSolid                         = true,
 466       .depthBounds                              = false,
 467       .wideLines                                = true,
 468       .largePoints                              = true,
 469       .alphaToOne                               = true,
 470       .multiViewport                            = true,
 471       .samplerAnisotropy                        = true,
 472       .textureCompressionETC2                   = pdevice->info.gen >= 8 ||
 473                                                   pdevice->info.is_baytrail,
 474       .textureCompressionASTC_LDR               = pdevice->info.gen >= 9, /* FINISHME CHV */
 475       .textureCompressionBC                     = true,
 476       .occlusionQueryPrecise                    = true,
 477       .pipelineStatisticsQuery                  = false,
 478       .fragmentStoresAndAtomics                 = true,
 479       .shaderTessellationAndGeometryPointSize   = true,
 480       .shaderImageGatherExtended                = true,
 481       .shaderStorageImageExtendedFormats        = true,
 482       .shaderStorageImageMultisample            = false,
 483       .shaderStorageImageReadWithoutFormat      = false,
 484       .shaderStorageImageWriteWithoutFormat     = true,
 485       .shaderUniformBufferArrayDynamicIndexing  = true,
 486       .shaderSampledImageArrayDynamicIndexing   = true,
 487       .shaderStorageBufferArrayDynamicIndexing  = true,
 488       .shaderStorageImageArrayDynamicIndexing   = true,
 489       .shaderClipDistance                       = true,
 490       .shaderCullDistance                       = true,
 491       .shaderFloat64                            = pdevice->info.gen >= 8,
 492       .shaderInt64                              = pdevice->info.gen >= 8,
 493       .shaderInt16                              = false,
 494       .shaderResourceMinLod                     = false,
 495       .variableMultisampleRate                  = false,
 496       .inheritedQueries                         = false,
 497    };
 498
 499    /* We can't do image stores in vec4 shaders */
 500    pFeatures->vertexPipelineStoresAndAtomics =
 501       pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
 502       pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
 503 }
 504
 505 void anv_GetPhysicalDeviceFeatures2KHR(
 506     VkPhysicalDevice                            physicalDevice,
 507     VkPhysicalDeviceFeatures2KHR*               pFeatures)
 508 {
 509    anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 510
 511    vk_foreach_struct(ext, pFeatures->pNext) {
 512       switch (ext->sType) {
 513       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 514          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 515             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 516
 517          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 518          break;
 519       }
 520
 521       default:
 522          anv_debug_ignored_stype(ext->sType);
 523          break;
 524       }
 525    }
 526 }
 527
 528 void anv_GetPhysicalDeviceProperties(
 529     VkPhysicalDevice                            physicalDevice,
 530     VkPhysicalDeviceProperties*                 pProperties)
 531 {
 532    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 533    const struct gen_device_info *devinfo = &pdevice->info;
 534
 535    const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
 536
 537    /* See assertions made when programming the buffer surface state. */
 538    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
 539                                       (1ul << 30) : (1ul << 27);
 540
 541    VkSampleCountFlags sample_counts =
 542       isl_device_get_sample_counts(&pdevice->isl_dev);
 543
 544    VkPhysicalDeviceLimits limits = {
 545       .maxImageDimension1D                      = (1 << 14),
 546       .maxImageDimension2D                      = (1 << 14),
 547       .maxImageDimension3D                      = (1 << 11),
 548       .maxImageDimensionCube                    = (1 << 14),
 549       .maxImageArrayLayers                      = (1 << 11),
 550       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 551       .maxUniformBufferRange                    = (1ul << 27),
 552       .maxStorageBufferRange                    = max_raw_buffer_sz,
 553       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 554       .maxMemoryAllocationCount                 = UINT32_MAX,
 555       .maxSamplerAllocationCount                = 64 * 1024,
 556       .bufferImageGranularity                   = 64, /* A cache line */
 557       .sparseAddressSpaceSize                   = 0,
 558       .maxBoundDescriptorSets                   = MAX_SETS,
 559       .maxPerStageDescriptorSamplers            = 64,
 560       .maxPerStageDescriptorUniformBuffers      = 64,
 561       .maxPerStageDescriptorStorageBuffers      = 64,
 562       .maxPerStageDescriptorSampledImages       = 64,
 563       .maxPerStageDescriptorStorageImages       = 64,
 564       .maxPerStageDescriptorInputAttachments    = 64,
 565       .maxPerStageResources                     = 128,
 566       .maxDescriptorSetSamplers                 = 256,
 567       .maxDescriptorSetUniformBuffers           = 256,
 568       .maxDescriptorSetUniformBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
 569       .maxDescriptorSetStorageBuffers           = 256,
 570       .maxDescriptorSetStorageBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
 571       .maxDescriptorSetSampledImages            = 256,
 572       .maxDescriptorSetStorageImages            = 256,
 573       .maxDescriptorSetInputAttachments         = 256,
 574       .maxVertexInputAttributes                 = MAX_VBS,
 575       .maxVertexInputBindings                   = MAX_VBS,
 576       .maxVertexInputAttributeOffset            = 2047,
 577       .maxVertexInputBindingStride              = 2048,
 578       .maxVertexOutputComponents                = 128,
 579       .maxTessellationGenerationLevel           = 64,
 580       .maxTessellationPatchSize                 = 32,
 581       .maxTessellationControlPerVertexInputComponents = 128,
 582       .maxTessellationControlPerVertexOutputComponents = 128,
 583       .maxTessellationControlPerPatchOutputComponents = 128,
 584       .maxTessellationControlTotalOutputComponents = 2048,
 585       .maxTessellationEvaluationInputComponents = 128,
 586       .maxTessellationEvaluationOutputComponents = 128,
 587       .maxGeometryShaderInvocations             = 32,
 588       .maxGeometryInputComponents               = 64,
 589       .maxGeometryOutputComponents              = 128,
 590       .maxGeometryOutputVertices                = 256,
 591       .maxGeometryTotalOutputComponents         = 1024,
 592       .maxFragmentInputComponents               = 128,
 593       .maxFragmentOutputAttachments             = 8,
 594       .maxFragmentDualSrcAttachments            = 1,
 595       .maxFragmentCombinedOutputResources       = 8,
 596       .maxComputeSharedMemorySize               = 32768,
 597       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 598       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 599       .maxComputeWorkGroupSize = {
 600          16 * devinfo->max_cs_threads,
 601          16 * devinfo->max_cs_threads,
 602          16 * devinfo->max_cs_threads,
 603       },
 604       .subPixelPrecisionBits                    = 4 /* FIXME */,
 605       .subTexelPrecisionBits                    = 4 /* FIXME */,
 606       .mipmapPrecisionBits                      = 4 /* FIXME */,
 607       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 608       .maxDrawIndirectCount                     = UINT32_MAX,
 609       .maxSamplerLodBias                        = 16,
 610       .maxSamplerAnisotropy                     = 16,
 611       .maxViewports                             = MAX_VIEWPORTS,
 612       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 613       .viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
 614       .viewportSubPixelBits                     = 13, /* We take a float? */
 615       .minMemoryMapAlignment                    = 4096, /* A page */
 616       .minTexelBufferOffsetAlignment            = 1,
 617       .minUniformBufferOffsetAlignment          = 16,
 618       .minStorageBufferOffsetAlignment          = 4,
 619       .minTexelOffset                           = -8,
 620       .maxTexelOffset                           = 7,
 621       .minTexelGatherOffset                     = -32,
 622       .maxTexelGatherOffset                     = 31,
 623       .minInterpolationOffset                   = -0.5,
 624       .maxInterpolationOffset                   = 0.4375,
 625       .subPixelInterpolationOffsetBits          = 4,
 626       .maxFramebufferWidth                      = (1 << 14),
 627       .maxFramebufferHeight                     = (1 << 14),
 628       .maxFramebufferLayers                     = (1 << 11),
 629       .framebufferColorSampleCounts             = sample_counts,
 630       .framebufferDepthSampleCounts             = sample_counts,
 631       .framebufferStencilSampleCounts           = sample_counts,
 632       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 633       .maxColorAttachments                      = MAX_RTS,
 634       .sampledImageColorSampleCounts            = sample_counts,
 635       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 636       .sampledImageDepthSampleCounts            = sample_counts,
 637       .sampledImageStencilSampleCounts          = sample_counts,
 638       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 639       .maxSampleMaskWords                       = 1,
 640       .timestampComputeAndGraphics              = false,
 641       .timestampPeriod                          = time_stamp_base,
 642       .maxClipDistances                         = 8,
 643       .maxCullDistances                         = 8,
 644       .maxCombinedClipAndCullDistances          = 8,
 645       .discreteQueuePriorities                  = 1,
 646       .pointSizeRange                           = { 0.125, 255.875 },
 647       .lineWidthRange                           = { 0.0, 7.9921875 },
 648       .pointSizeGranularity                     = (1.0 / 8.0),
 649       .lineWidthGranularity                     = (1.0 / 128.0),
 650       .strictLines                              = false, /* FINISHME */
 651       .standardSampleLocations                  = true,
 652       .optimalBufferCopyOffsetAlignment         = 128,
 653       .optimalBufferCopyRowPitchAlignment       = 128,
 654       .nonCoherentAtomSize                      = 64,
 655    };
 656
 657    *pProperties = (VkPhysicalDeviceProperties) {
 658       .apiVersion = VK_MAKE_VERSION(1, 0, 42),
 659       .driverVersion = 1,
 660       .vendorID = 0x8086,
 661       .deviceID = pdevice->chipset_id,
 662       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 663       .limits = limits,
 664       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 665    };
 666
 667    strcpy(pProperties->deviceName, pdevice->name);
 668    memcpy(pProperties->pipelineCacheUUID, pdevice->uuid, VK_UUID_SIZE);
 669 }
 670
 671 void anv_GetPhysicalDeviceProperties2KHR(
 672     VkPhysicalDevice                            physicalDevice,
 673     VkPhysicalDeviceProperties2KHR*             pProperties)
 674 {
 675    anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 676
 677    vk_foreach_struct(ext, pProperties->pNext) {
 678       switch (ext->sType) {
 679       default:
 680          anv_debug_ignored_stype(ext->sType);
 681          break;
 682       }
 683    }
 684 }
 685
 686 static void
 687 anv_get_queue_family_properties(struct anv_physical_device *phys_dev,
 688                                 VkQueueFamilyProperties *props)
 689 {
 690    *props = (VkQueueFamilyProperties) {
 691       .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 692                     VK_QUEUE_COMPUTE_BIT |
 693                     VK_QUEUE_TRANSFER_BIT,
 694       .queueCount = 1,
 695       .timestampValidBits = 36, /* XXX: Real value here */
 696       .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
 697    };
 698 }
 699
 700 void anv_GetPhysicalDeviceQueueFamilyProperties(
 701     VkPhysicalDevice                            physicalDevice,
 702     uint32_t*                                   pCount,
 703     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 704 {
 705    ANV_FROM_HANDLE(anv_physical_device, phys_dev, physicalDevice);
 706
 707    if (pQueueFamilyProperties == NULL) {
 708       *pCount = 1;
 709       return;
 710    }
 711
 712    /* The spec implicitly allows the incoming count to be 0. From the Vulkan
 713     * 1.0.38 spec, Section 4.1 Physical Devices:
 714     *
 715     *     If the value referenced by pQueueFamilyPropertyCount is not 0 [then
 716     *     do stuff].
 717     */
 718    if (*pCount == 0)
 719       return;
 720
 721    *pCount = 1;
 722    anv_get_queue_family_properties(phys_dev, pQueueFamilyProperties);
 723 }
 724
 725 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
 726     VkPhysicalDevice                            physicalDevice,
 727     uint32_t*                                   pQueueFamilyPropertyCount,
 728     VkQueueFamilyProperties2KHR*                pQueueFamilyProperties)
 729 {
 730
 731    ANV_FROM_HANDLE(anv_physical_device, phys_dev, physicalDevice);
 732
 733    if (pQueueFamilyProperties == NULL) {
 734       *pQueueFamilyPropertyCount = 1;
 735       return;
 736    }
 737
 738    /* The spec implicitly allows the incoming count to be 0. From the Vulkan
 739     * 1.0.38 spec, Section 4.1 Physical Devices:
 740     *
 741     *     If the value referenced by pQueueFamilyPropertyCount is not 0 [then
 742     *     do stuff].
 743     */
 744    if (*pQueueFamilyPropertyCount == 0)
 745       return;
 746
 747    /* We support exactly one queue family. So need to traverse only the first
 748     * array element's pNext chain.
 749     */
 750    *pQueueFamilyPropertyCount = 1;
 751    anv_get_queue_family_properties(phys_dev,
 752          &pQueueFamilyProperties->queueFamilyProperties);
 753
 754    vk_foreach_struct(ext, pQueueFamilyProperties->pNext) {
 755       switch (ext->sType) {
 756       default:
 757          anv_debug_ignored_stype(ext->sType);
 758          break;
 759       }
 760    }
 761 }
 762
 763 void anv_GetPhysicalDeviceMemoryProperties(
 764     VkPhysicalDevice                            physicalDevice,
 765     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
 766 {
 767    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 768    VkDeviceSize heap_size;
 769
 770    /* Reserve some wiggle room for the driver by exposing only 75% of the
 771     * aperture to the heap.
 772     */
 773    heap_size = 3 * physical_device->aperture_size / 4;
 774
 775    if (physical_device->info.has_llc) {
 776       /* Big core GPUs share LLC with the CPU and thus one memory type can be
 777        * both cached and coherent at the same time.
 778        */
 779       pMemoryProperties->memoryTypeCount = 1;
 780       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 781          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 782                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 783                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
 784                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 785          .heapIndex = 0,
 786       };
 787    } else {
 788       /* The spec requires that we expose a host-visible, coherent memory
 789        * type, but Atom GPUs don't share LLC. Thus we offer two memory types
 790        * to give the application a choice between cached, but not coherent and
 791        * coherent but uncached (WC though).
 792        */
 793       pMemoryProperties->memoryTypeCount = 2;
 794       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 795          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 796                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 797                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
 798          .heapIndex = 0,
 799       };
 800       pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
 801          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 802                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 803                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 804          .heapIndex = 0,
 805       };
 806    }
 807
 808    pMemoryProperties->memoryHeapCount = 1;
 809    pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
 810       .size = heap_size,
 811       .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 812    };
 813 }
 814
 815 void anv_GetPhysicalDeviceMemoryProperties2KHR(
 816     VkPhysicalDevice                            physicalDevice,
 817     VkPhysicalDeviceMemoryProperties2KHR*       pMemoryProperties)
 818 {
 819    anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
 820                                          &pMemoryProperties->memoryProperties);
 821
 822    vk_foreach_struct(ext, pMemoryProperties->pNext) {
 823       switch (ext->sType) {
 824       default:
 825          anv_debug_ignored_stype(ext->sType);
 826          break;
 827       }
 828    }
 829 }
 830
 831 PFN_vkVoidFunction anv_GetInstanceProcAddr(
 832     VkInstance                                  instance,
 833     const char*                                 pName)
 834 {
 835    return anv_lookup_entrypoint(NULL, pName);
 836 }
 837
 838 /* With version 1+ of the loader interface the ICD should expose
 839  * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
 840  */
 841 PUBLIC
 842 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 843     VkInstance                                  instance,
 844     const char*                                 pName);
 845
 846 PUBLIC
 847 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 848     VkInstance                                  instance,
 849     const char*                                 pName)
 850 {
 851    return anv_GetInstanceProcAddr(instance, pName);
 852 }
 853
 854 PFN_vkVoidFunction anv_GetDeviceProcAddr(
 855     VkDevice                                    _device,
 856     const char*                                 pName)
 857 {
 858    ANV_FROM_HANDLE(anv_device, device, _device);
 859    return anv_lookup_entrypoint(&device->info, pName);
 860 }
 861
 862 static void
 863 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
 864 {
 865    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 866    queue->device = device;
 867    queue->pool = &device->surface_state_pool;
 868 }
 869
 870 static void
 871 anv_queue_finish(struct anv_queue *queue)
 872 {
 873 }
 874
 875 static struct anv_state
 876 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
 877 {
 878    struct anv_state state;
 879
 880    state = anv_state_pool_alloc(pool, size, align);
 881    memcpy(state.map, p, size);
 882
 883    anv_state_flush(pool->block_pool->device, state);
 884
 885    return state;
 886 }
 887
 888 struct gen8_border_color {
 889    union {
 890       float float32[4];
 891       uint32_t uint32[4];
 892    };
 893    /* Pad out to 64 bytes */
 894    uint32_t _pad[12];
 895 };
 896
 897 static void
 898 anv_device_init_border_colors(struct anv_device *device)
 899 {
 900    static const struct gen8_border_color border_colors[] = {
 901       [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] =  { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
 902       [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] =       { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
 903       [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] =       { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
 904       [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] =    { .uint32 = { 0, 0, 0, 0 } },
 905       [VK_BORDER_COLOR_INT_OPAQUE_BLACK] =         { .uint32 = { 0, 0, 0, 1 } },
 906       [VK_BORDER_COLOR_INT_OPAQUE_WHITE] =         { .uint32 = { 1, 1, 1, 1 } },
 907    };
 908
 909    device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
 910                                                     sizeof(border_colors), 64,
 911                                                     border_colors);
 912 }
 913
 914 VkResult
 915 anv_device_submit_simple_batch(struct anv_device *device,
 916                                struct anv_batch *batch)
 917 {
 918    struct drm_i915_gem_execbuffer2 execbuf;
 919    struct drm_i915_gem_exec_object2 exec2_objects[1];
 920    struct anv_bo bo, *exec_bos[1];
 921    VkResult result = VK_SUCCESS;
 922    uint32_t size;
 923    int64_t timeout;
 924    int ret;
 925
 926    /* Kernel driver requires 8 byte aligned batch length */
 927    size = align_u32(batch->next - batch->start, 8);
 928    result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
 929    if (result != VK_SUCCESS)
 930       return result;
 931
 932    memcpy(bo.map, batch->start, size);
 933    if (!device->info.has_llc)
 934       anv_flush_range(bo.map, size);
 935
 936    exec_bos[0] = &bo;
 937    exec2_objects[0].handle = bo.gem_handle;
 938    exec2_objects[0].relocation_count = 0;
 939    exec2_objects[0].relocs_ptr = 0;
 940    exec2_objects[0].alignment = 0;
 941    exec2_objects[0].offset = bo.offset;
 942    exec2_objects[0].flags = 0;
 943    exec2_objects[0].rsvd1 = 0;
 944    exec2_objects[0].rsvd2 = 0;
 945
 946    execbuf.buffers_ptr = (uintptr_t) exec2_objects;
 947    execbuf.buffer_count = 1;
 948    execbuf.batch_start_offset = 0;
 949    execbuf.batch_len = size;
 950    execbuf.cliprects_ptr = 0;
 951    execbuf.num_cliprects = 0;
 952    execbuf.DR1 = 0;
 953    execbuf.DR4 = 0;
 954
 955    execbuf.flags =
 956       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
 957    execbuf.rsvd1 = device->context_id;
 958    execbuf.rsvd2 = 0;
 959
 960    result = anv_device_execbuf(device, &execbuf, exec_bos);
 961    if (result != VK_SUCCESS)
 962       goto fail;
 963
 964    timeout = INT64_MAX;
 965    ret = anv_gem_wait(device, bo.gem_handle, &timeout);
 966    if (ret != 0) {
 967       /* We don't know the real error. */
 968       result = vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
 969       goto fail;
 970    }
 971
 972  fail:
 973    anv_bo_pool_free(&device->batch_bo_pool, &bo);
 974
 975    return result;
 976 }
 977
 978 VkResult anv_CreateDevice(
 979     VkPhysicalDevice                            physicalDevice,
 980     const VkDeviceCreateInfo*                   pCreateInfo,
 981     const VkAllocationCallbacks*                pAllocator,
 982     VkDevice*                                   pDevice)
 983 {
 984    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 985    VkResult result;
 986    struct anv_device *device;
 987
 988    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
 989
 990    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 991       bool found = false;
 992       for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
 993          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 994                     device_extensions[j].extensionName) == 0) {
 995             found = true;
 996             break;
 997          }
 998       }
 999       if (!found)
1000          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1001    }
1002
1003    device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
1004                        sizeof(*device), 8,
1005                        VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1006    if (!device)
1007       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1008
1009    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1010    device->instance = physical_device->instance;
1011    device->chipset_id = physical_device->chipset_id;
1012
1013    if (pAllocator)
1014       device->alloc = *pAllocator;
1015    else
1016       device->alloc = physical_device->instance->alloc;
1017
1018    /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1019    device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1020    if (device->fd == -1) {
1021       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1022       goto fail_device;
1023    }
1024
1025    device->context_id = anv_gem_create_context(device);
1026    if (device->context_id == -1) {
1027       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1028       goto fail_fd;
1029    }
1030
1031    device->info = physical_device->info;
1032    device->isl_dev = physical_device->isl_dev;
1033
1034    /* On Broadwell and later, we can use batch chaining to more efficiently
1035     * implement growing command buffers.  Prior to Haswell, the kernel
1036     * command parser gets in the way and we have to fall back to growing
1037     * the batch.
1038     */
1039    device->can_chain_batches = device->info.gen >= 8;
1040
1041    device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1042       pCreateInfo->pEnabledFeatures->robustBufferAccess;
1043
1044    pthread_mutex_init(&device->mutex, NULL);
1045
1046    pthread_condattr_t condattr;
1047    pthread_condattr_init(&condattr);
1048    pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC);
1049    pthread_cond_init(&device->queue_submit, NULL);
1050    pthread_condattr_destroy(&condattr);
1051
1052    anv_bo_pool_init(&device->batch_bo_pool, device);
1053
1054    anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
1055
1056    anv_state_pool_init(&device->dynamic_state_pool,
1057                        &device->dynamic_state_block_pool);
1058
1059    anv_block_pool_init(&device->instruction_block_pool, device, 1024 * 1024);
1060    anv_state_pool_init(&device->instruction_state_pool,
1061                        &device->instruction_block_pool);
1062
1063    anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
1064
1065    anv_state_pool_init(&device->surface_state_pool,
1066                        &device->surface_state_block_pool);
1067
1068    anv_bo_init_new(&device->workaround_bo, device, 1024);
1069
1070    anv_scratch_pool_init(device, &device->scratch_pool);
1071
1072    anv_queue_init(device, &device->queue);
1073
1074    switch (device->info.gen) {
1075    case 7:
1076       if (!device->info.is_haswell)
1077          result = gen7_init_device_state(device);
1078       else
1079          result = gen75_init_device_state(device);
1080       break;
1081    case 8:
1082       result = gen8_init_device_state(device);
1083       break;
1084    case 9:
1085       result = gen9_init_device_state(device);
1086       break;
1087    default:
1088       /* Shouldn't get here as we don't create physical devices for any other
1089        * gens. */
1090       unreachable("unhandled gen");
1091    }
1092    if (result != VK_SUCCESS)
1093       goto fail_fd;
1094
1095    anv_device_init_blorp(device);
1096
1097    anv_device_init_border_colors(device);
1098
1099    *pDevice = anv_device_to_handle(device);
1100
1101    return VK_SUCCESS;
1102
1103  fail_fd:
1104    close(device->fd);
1105  fail_device:
1106    vk_free(&device->alloc, device);
1107
1108    return result;
1109 }
1110
1111 void anv_DestroyDevice(
1112     VkDevice                                    _device,
1113     const VkAllocationCallbacks*                pAllocator)
1114 {
1115    ANV_FROM_HANDLE(anv_device, device, _device);
1116
1117    if (!device)
1118       return;
1119
1120    anv_device_finish_blorp(device);
1121
1122    anv_queue_finish(&device->queue);
1123
1124 #ifdef HAVE_VALGRIND
1125    /* We only need to free these to prevent valgrind errors.  The backing
1126     * BO will go away in a couple of lines so we don't actually leak.
1127     */
1128    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1129 #endif
1130
1131    anv_scratch_pool_finish(device, &device->scratch_pool);
1132
1133    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1134    anv_gem_close(device, device->workaround_bo.gem_handle);
1135
1136    anv_state_pool_finish(&device->surface_state_pool);
1137    anv_block_pool_finish(&device->surface_state_block_pool);
1138    anv_state_pool_finish(&device->instruction_state_pool);
1139    anv_block_pool_finish(&device->instruction_block_pool);
1140    anv_state_pool_finish(&device->dynamic_state_pool);
1141    anv_block_pool_finish(&device->dynamic_state_block_pool);
1142
1143    anv_bo_pool_finish(&device->batch_bo_pool);
1144
1145    pthread_cond_destroy(&device->queue_submit);
1146    pthread_mutex_destroy(&device->mutex);
1147
1148    anv_gem_destroy_context(device, device->context_id);
1149
1150    close(device->fd);
1151
1152    vk_free(&device->alloc, device);
1153 }
1154
1155 VkResult anv_EnumerateInstanceExtensionProperties(
1156     const char*                                 pLayerName,
1157     uint32_t*                                   pPropertyCount,
1158     VkExtensionProperties*                      pProperties)
1159 {
1160    if (pProperties == NULL) {
1161       *pPropertyCount = ARRAY_SIZE(global_extensions);
1162       return VK_SUCCESS;
1163    }
1164
1165    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(global_extensions));
1166    typed_memcpy(pProperties, global_extensions, *pPropertyCount);
1167
1168    if (*pPropertyCount < ARRAY_SIZE(global_extensions))
1169       return VK_INCOMPLETE;
1170
1171    return VK_SUCCESS;
1172 }
1173
1174 VkResult anv_EnumerateDeviceExtensionProperties(
1175     VkPhysicalDevice                            physicalDevice,
1176     const char*                                 pLayerName,
1177     uint32_t*                                   pPropertyCount,
1178     VkExtensionProperties*                      pProperties)
1179 {
1180    if (pProperties == NULL) {
1181       *pPropertyCount = ARRAY_SIZE(device_extensions);
1182       return VK_SUCCESS;
1183    }
1184
1185    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(device_extensions));
1186    typed_memcpy(pProperties, device_extensions, *pPropertyCount);
1187
1188    if (*pPropertyCount < ARRAY_SIZE(device_extensions))
1189       return VK_INCOMPLETE;
1190
1191    return VK_SUCCESS;
1192 }
1193
1194 VkResult anv_EnumerateInstanceLayerProperties(
1195     uint32_t*                                   pPropertyCount,
1196     VkLayerProperties*                          pProperties)
1197 {
1198    if (pProperties == NULL) {
1199       *pPropertyCount = 0;
1200       return VK_SUCCESS;
1201    }
1202
1203    /* None supported at this time */
1204    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1205 }
1206
1207 VkResult anv_EnumerateDeviceLayerProperties(
1208     VkPhysicalDevice                            physicalDevice,
1209     uint32_t*                                   pPropertyCount,
1210     VkLayerProperties*                          pProperties)
1211 {
1212    if (pProperties == NULL) {
1213       *pPropertyCount = 0;
1214       return VK_SUCCESS;
1215    }
1216
1217    /* None supported at this time */
1218    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1219 }
1220
1221 void anv_GetDeviceQueue(
1222     VkDevice                                    _device,
1223     uint32_t                                    queueNodeIndex,
1224     uint32_t                                    queueIndex,
1225     VkQueue*                                    pQueue)
1226 {
1227    ANV_FROM_HANDLE(anv_device, device, _device);
1228
1229    assert(queueIndex == 0);
1230
1231    *pQueue = anv_queue_to_handle(&device->queue);
1232 }
1233
1234 VkResult
1235 anv_device_execbuf(struct anv_device *device,
1236                    struct drm_i915_gem_execbuffer2 *execbuf,
1237                    struct anv_bo **execbuf_bos)
1238 {
1239    int ret = anv_gem_execbuffer(device, execbuf);
1240    if (ret != 0) {
1241       /* We don't know the real error. */
1242       return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
1243    }
1244
1245    struct drm_i915_gem_exec_object2 *objects =
1246       (void *)(uintptr_t)execbuf->buffers_ptr;
1247    for (uint32_t k = 0; k < execbuf->buffer_count; k++)
1248       execbuf_bos[k]->offset = objects[k].offset;
1249
1250    return VK_SUCCESS;
1251 }
1252
1253 VkResult anv_QueueSubmit(
1254     VkQueue                                     _queue,
1255     uint32_t                                    submitCount,
1256     const VkSubmitInfo*                         pSubmits,
1257     VkFence                                     _fence)
1258 {
1259    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1260    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1261    struct anv_device *device = queue->device;
1262    VkResult result = VK_SUCCESS;
1263
1264    /* We lock around QueueSubmit for three main reasons:
1265     *
1266     *  1) When a block pool is resized, we create a new gem handle with a
1267     *     different size and, in the case of surface states, possibly a
1268     *     different center offset but we re-use the same anv_bo struct when
1269     *     we do so.  If this happens in the middle of setting up an execbuf,
1270     *     we could end up with our list of BOs out of sync with our list of
1271     *     gem handles.
1272     *
1273     *  2) The algorithm we use for building the list of unique buffers isn't
1274     *     thread-safe.  While the client is supposed to syncronize around
1275     *     QueueSubmit, this would be extremely difficult to debug if it ever
1276     *     came up in the wild due to a broken app.  It's better to play it
1277     *     safe and just lock around QueueSubmit.
1278     *
1279     *  3)  The anv_cmd_buffer_execbuf function may perform relocations in
1280     *      userspace.  Due to the fact that the surface state buffer is shared
1281     *      between batches, we can't afford to have that happen from multiple
1282     *      threads at the same time.  Even though the user is supposed to
1283     *      ensure this doesn't happen, we play it safe as in (2) above.
1284     *
1285     * Since the only other things that ever take the device lock such as block
1286     * pool resize only rarely happen, this will almost never be contended so
1287     * taking a lock isn't really an expensive operation in this case.
1288     */
1289    pthread_mutex_lock(&device->mutex);
1290
1291    for (uint32_t i = 0; i < submitCount; i++) {
1292       for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1293          ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1294                          pSubmits[i].pCommandBuffers[j]);
1295          assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1296
1297          result = anv_cmd_buffer_execbuf(device, cmd_buffer);
1298          if (result != VK_SUCCESS)
1299             goto out;
1300       }
1301    }
1302
1303    if (fence) {
1304       struct anv_bo *fence_bo = &fence->bo;
1305       result = anv_device_execbuf(device, &fence->execbuf, &fence_bo);
1306       if (result != VK_SUCCESS)
1307          goto out;
1308
1309       /* Update the fence and wake up any waiters */
1310       assert(fence->state == ANV_FENCE_STATE_RESET);
1311       fence->state = ANV_FENCE_STATE_SUBMITTED;
1312       pthread_cond_broadcast(&device->queue_submit);
1313    }
1314
1315 out:
1316    pthread_mutex_unlock(&device->mutex);
1317
1318    return result;
1319 }
1320
1321 VkResult anv_QueueWaitIdle(
1322     VkQueue                                     _queue)
1323 {
1324    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1325
1326    return anv_DeviceWaitIdle(anv_device_to_handle(queue->device));
1327 }
1328
1329 VkResult anv_DeviceWaitIdle(
1330     VkDevice                                    _device)
1331 {
1332    ANV_FROM_HANDLE(anv_device, device, _device);
1333    struct anv_batch batch;
1334
1335    uint32_t cmds[8];
1336    batch.start = batch.next = cmds;
1337    batch.end = (void *) cmds + sizeof(cmds);
1338
1339    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1340    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1341
1342    return anv_device_submit_simple_batch(device, &batch);
1343 }
1344
1345 VkResult
1346 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1347 {
1348    uint32_t gem_handle = anv_gem_create(device, size);
1349    if (!gem_handle)
1350       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1351
1352    anv_bo_init(bo, gem_handle, size);
1353
1354    return VK_SUCCESS;
1355 }
1356
1357 VkResult anv_AllocateMemory(
1358     VkDevice                                    _device,
1359     const VkMemoryAllocateInfo*                 pAllocateInfo,
1360     const VkAllocationCallbacks*                pAllocator,
1361     VkDeviceMemory*                             pMem)
1362 {
1363    ANV_FROM_HANDLE(anv_device, device, _device);
1364    struct anv_device_memory *mem;
1365    VkResult result;
1366
1367    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1368
1369    /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1370    assert(pAllocateInfo->allocationSize > 0);
1371
1372    /* We support exactly one memory heap. */
1373    assert(pAllocateInfo->memoryTypeIndex == 0 ||
1374           (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1375
1376    /* FINISHME: Fail if allocation request exceeds heap size. */
1377
1378    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1379                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1380    if (mem == NULL)
1381       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1382
1383    /* The kernel is going to give us whole pages anyway */
1384    uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1385
1386    result = anv_bo_init_new(&mem->bo, device, alloc_size);
1387    if (result != VK_SUCCESS)
1388       goto fail;
1389
1390    mem->type_index = pAllocateInfo->memoryTypeIndex;
1391
1392    mem->map = NULL;
1393    mem->map_size = 0;
1394
1395    *pMem = anv_device_memory_to_handle(mem);
1396
1397    return VK_SUCCESS;
1398
1399  fail:
1400    vk_free2(&device->alloc, pAllocator, mem);
1401
1402    return result;
1403 }
1404
1405 void anv_FreeMemory(
1406     VkDevice                                    _device,
1407     VkDeviceMemory                              _mem,
1408     const VkAllocationCallbacks*                pAllocator)
1409 {
1410    ANV_FROM_HANDLE(anv_device, device, _device);
1411    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1412
1413    if (mem == NULL)
1414       return;
1415
1416    if (mem->map)
1417       anv_UnmapMemory(_device, _mem);
1418
1419    if (mem->bo.map)
1420       anv_gem_munmap(mem->bo.map, mem->bo.size);
1421
1422    if (mem->bo.gem_handle != 0)
1423       anv_gem_close(device, mem->bo.gem_handle);
1424
1425    vk_free2(&device->alloc, pAllocator, mem);
1426 }
1427
1428 VkResult anv_MapMemory(
1429     VkDevice                                    _device,
1430     VkDeviceMemory                              _memory,
1431     VkDeviceSize                                offset,
1432     VkDeviceSize                                size,
1433     VkMemoryMapFlags                            flags,
1434     void**                                      ppData)
1435 {
1436    ANV_FROM_HANDLE(anv_device, device, _device);
1437    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1438
1439    if (mem == NULL) {
1440       *ppData = NULL;
1441       return VK_SUCCESS;
1442    }
1443
1444    if (size == VK_WHOLE_SIZE)
1445       size = mem->bo.size - offset;
1446
1447    /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1448     *
1449     *  * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1450     *    assert(size != 0);
1451     *  * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1452     *    equal to the size of the memory minus offset
1453     */
1454    assert(size > 0);
1455    assert(offset + size <= mem->bo.size);
1456
1457    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1458     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1459     * at a time is valid. We could just mmap up front and return an offset
1460     * pointer here, but that may exhaust virtual memory on 32 bit
1461     * userspace. */
1462
1463    uint32_t gem_flags = 0;
1464    if (!device->info.has_llc && mem->type_index == 0)
1465       gem_flags |= I915_MMAP_WC;
1466
1467    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1468    uint64_t map_offset = offset & ~4095ull;
1469    assert(offset >= map_offset);
1470    uint64_t map_size = (offset + size) - map_offset;
1471
1472    /* Let's map whole pages */
1473    map_size = align_u64(map_size, 4096);
1474
1475    void *map = anv_gem_mmap(device, mem->bo.gem_handle,
1476                             map_offset, map_size, gem_flags);
1477    if (map == MAP_FAILED)
1478       return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
1479
1480    mem->map = map;
1481    mem->map_size = map_size;
1482
1483    *ppData = mem->map + (offset - map_offset);
1484
1485    return VK_SUCCESS;
1486 }
1487
1488 void anv_UnmapMemory(
1489     VkDevice                                    _device,
1490     VkDeviceMemory                              _memory)
1491 {
1492    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1493
1494    if (mem == NULL)
1495       return;
1496
1497    anv_gem_munmap(mem->map, mem->map_size);
1498
1499    mem->map = NULL;
1500    mem->map_size = 0;
1501 }
1502
1503 static void
1504 clflush_mapped_ranges(struct anv_device         *device,
1505                       uint32_t                   count,
1506                       const VkMappedMemoryRange *ranges)
1507 {
1508    for (uint32_t i = 0; i < count; i++) {
1509       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1510       if (ranges[i].offset >= mem->map_size)
1511          continue;
1512
1513       anv_clflush_range(mem->map + ranges[i].offset,
1514                         MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
1515    }
1516 }
1517
1518 VkResult anv_FlushMappedMemoryRanges(
1519     VkDevice                                    _device,
1520     uint32_t                                    memoryRangeCount,
1521     const VkMappedMemoryRange*                  pMemoryRanges)
1522 {
1523    ANV_FROM_HANDLE(anv_device, device, _device);
1524
1525    if (device->info.has_llc)
1526       return VK_SUCCESS;
1527
1528    /* Make sure the writes we're flushing have landed. */
1529    __builtin_ia32_mfence();
1530
1531    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1532
1533    return VK_SUCCESS;
1534 }
1535
1536 VkResult anv_InvalidateMappedMemoryRanges(
1537     VkDevice                                    _device,
1538     uint32_t                                    memoryRangeCount,
1539     const VkMappedMemoryRange*                  pMemoryRanges)
1540 {
1541    ANV_FROM_HANDLE(anv_device, device, _device);
1542
1543    if (device->info.has_llc)
1544       return VK_SUCCESS;
1545
1546    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1547
1548    /* Make sure no reads get moved up above the invalidate. */
1549    __builtin_ia32_mfence();
1550
1551    return VK_SUCCESS;
1552 }
1553
1554 void anv_GetBufferMemoryRequirements(
1555     VkDevice                                    _device,
1556     VkBuffer                                    _buffer,
1557     VkMemoryRequirements*                       pMemoryRequirements)
1558 {
1559    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1560    ANV_FROM_HANDLE(anv_device, device, _device);
1561
1562    /* The Vulkan spec (git aaed022) says:
1563     *
1564     *    memoryTypeBits is a bitfield and contains one bit set for every
1565     *    supported memory type for the resource. The bit `1<<i` is set if and
1566     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1567     *    structure for the physical device is supported.
1568     *
1569     * We support exactly one memory type on LLC, two on non-LLC.
1570     */
1571    pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1572
1573    pMemoryRequirements->size = buffer->size;
1574    pMemoryRequirements->alignment = 16;
1575 }
1576
1577 void anv_GetImageMemoryRequirements(
1578     VkDevice                                    _device,
1579     VkImage                                     _image,
1580     VkMemoryRequirements*                       pMemoryRequirements)
1581 {
1582    ANV_FROM_HANDLE(anv_image, image, _image);
1583    ANV_FROM_HANDLE(anv_device, device, _device);
1584
1585    /* The Vulkan spec (git aaed022) says:
1586     *
1587     *    memoryTypeBits is a bitfield and contains one bit set for every
1588     *    supported memory type for the resource. The bit `1<<i` is set if and
1589     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1590     *    structure for the physical device is supported.
1591     *
1592     * We support exactly one memory type on LLC, two on non-LLC.
1593     */
1594    pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1595
1596    pMemoryRequirements->size = image->size;
1597    pMemoryRequirements->alignment = image->alignment;
1598 }
1599
1600 void anv_GetImageSparseMemoryRequirements(
1601     VkDevice                                    device,
1602     VkImage                                     image,
1603     uint32_t*                                   pSparseMemoryRequirementCount,
1604     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1605 {
1606    *pSparseMemoryRequirementCount = 0;
1607 }
1608
1609 void anv_GetDeviceMemoryCommitment(
1610     VkDevice                                    device,
1611     VkDeviceMemory                              memory,
1612     VkDeviceSize*                               pCommittedMemoryInBytes)
1613 {
1614    *pCommittedMemoryInBytes = 0;
1615 }
1616
1617 VkResult anv_BindBufferMemory(
1618     VkDevice                                    device,
1619     VkBuffer                                    _buffer,
1620     VkDeviceMemory                              _memory,
1621     VkDeviceSize                                memoryOffset)
1622 {
1623    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1624    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1625
1626    if (mem) {
1627       buffer->bo = &mem->bo;
1628       buffer->offset = memoryOffset;
1629    } else {
1630       buffer->bo = NULL;
1631       buffer->offset = 0;
1632    }
1633
1634    return VK_SUCCESS;
1635 }
1636
1637 VkResult anv_QueueBindSparse(
1638     VkQueue                                     queue,
1639     uint32_t                                    bindInfoCount,
1640     const VkBindSparseInfo*                     pBindInfo,
1641     VkFence                                     fence)
1642 {
1643    return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1644 }
1645
1646 VkResult anv_CreateFence(
1647     VkDevice                                    _device,
1648     const VkFenceCreateInfo*                    pCreateInfo,
1649     const VkAllocationCallbacks*                pAllocator,
1650     VkFence*                                    pFence)
1651 {
1652    ANV_FROM_HANDLE(anv_device, device, _device);
1653    struct anv_bo fence_bo;
1654    struct anv_fence *fence;
1655    struct anv_batch batch;
1656    VkResult result;
1657
1658    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1659
1660    result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1661    if (result != VK_SUCCESS)
1662       return result;
1663
1664    /* Fences are small.  Just store the CPU data structure in the BO. */
1665    fence = fence_bo.map;
1666    fence->bo = fence_bo;
1667
1668    /* Place the batch after the CPU data but on its own cache line. */
1669    const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1670    batch.next = batch.start = fence->bo.map + batch_offset;
1671    batch.end = fence->bo.map + fence->bo.size;
1672    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1673    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1674
1675    if (!device->info.has_llc) {
1676       assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1677       assert(batch.next - batch.start <= CACHELINE_SIZE);
1678       __builtin_ia32_mfence();
1679       __builtin_ia32_clflush(batch.start);
1680    }
1681
1682    fence->exec2_objects[0].handle = fence->bo.gem_handle;
1683    fence->exec2_objects[0].relocation_count = 0;
1684    fence->exec2_objects[0].relocs_ptr = 0;
1685    fence->exec2_objects[0].alignment = 0;
1686    fence->exec2_objects[0].offset = fence->bo.offset;
1687    fence->exec2_objects[0].flags = 0;
1688    fence->exec2_objects[0].rsvd1 = 0;
1689    fence->exec2_objects[0].rsvd2 = 0;
1690
1691    fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1692    fence->execbuf.buffer_count = 1;
1693    fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1694    fence->execbuf.batch_len = batch.next - batch.start;
1695    fence->execbuf.cliprects_ptr = 0;
1696    fence->execbuf.num_cliprects = 0;
1697    fence->execbuf.DR1 = 0;
1698    fence->execbuf.DR4 = 0;
1699
1700    fence->execbuf.flags =
1701       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1702    fence->execbuf.rsvd1 = device->context_id;
1703    fence->execbuf.rsvd2 = 0;
1704
1705    if (pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT) {
1706       fence->state = ANV_FENCE_STATE_SIGNALED;
1707    } else {
1708       fence->state = ANV_FENCE_STATE_RESET;
1709    }
1710
1711    *pFence = anv_fence_to_handle(fence);
1712
1713    return VK_SUCCESS;
1714 }
1715
1716 void anv_DestroyFence(
1717     VkDevice                                    _device,
1718     VkFence                                     _fence,
1719     const VkAllocationCallbacks*                pAllocator)
1720 {
1721    ANV_FROM_HANDLE(anv_device, device, _device);
1722    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1723
1724    if (!fence)
1725       return;
1726
1727    assert(fence->bo.map == fence);
1728    anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1729 }
1730
1731 VkResult anv_ResetFences(
1732     VkDevice                                    _device,
1733     uint32_t                                    fenceCount,
1734     const VkFence*                              pFences)
1735 {
1736    for (uint32_t i = 0; i < fenceCount; i++) {
1737       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1738       fence->state = ANV_FENCE_STATE_RESET;
1739    }
1740
1741    return VK_SUCCESS;
1742 }
1743
1744 VkResult anv_GetFenceStatus(
1745     VkDevice                                    _device,
1746     VkFence                                     _fence)
1747 {
1748    ANV_FROM_HANDLE(anv_device, device, _device);
1749    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1750    int64_t t = 0;
1751    int ret;
1752
1753    switch (fence->state) {
1754    case ANV_FENCE_STATE_RESET:
1755       /* If it hasn't even been sent off to the GPU yet, it's not ready */
1756       return VK_NOT_READY;
1757
1758    case ANV_FENCE_STATE_SIGNALED:
1759       /* It's been signaled, return success */
1760       return VK_SUCCESS;
1761
1762    case ANV_FENCE_STATE_SUBMITTED:
1763       /* It's been submitted to the GPU but we don't know if it's done yet. */
1764       ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1765       if (ret == 0) {
1766          fence->state = ANV_FENCE_STATE_SIGNALED;
1767          return VK_SUCCESS;
1768       } else {
1769          return VK_NOT_READY;
1770       }
1771    default:
1772       unreachable("Invalid fence status");
1773    }
1774 }
1775
1776 #define NSEC_PER_SEC 1000000000
1777 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
1778
1779 VkResult anv_WaitForFences(
1780     VkDevice                                    _device,
1781     uint32_t                                    fenceCount,
1782     const VkFence*                              pFences,
1783     VkBool32                                    waitAll,
1784     uint64_t                                    _timeout)
1785 {
1786    ANV_FROM_HANDLE(anv_device, device, _device);
1787    int ret;
1788
1789    /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1790     * to block indefinitely timeouts <= 0.  Unfortunately, this was broken
1791     * for a couple of kernel releases.  Since there's no way to know
1792     * whether or not the kernel we're using is one of the broken ones, the
1793     * best we can do is to clamp the timeout to INT64_MAX.  This limits the
1794     * maximum timeout from 584 years to 292 years - likely not a big deal.
1795     */
1796    int64_t timeout = MIN2(_timeout, INT64_MAX);
1797
1798    uint32_t pending_fences = fenceCount;
1799    while (pending_fences) {
1800       pending_fences = 0;
1801       bool signaled_fences = false;
1802       for (uint32_t i = 0; i < fenceCount; i++) {
1803          ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1804          switch (fence->state) {
1805          case ANV_FENCE_STATE_RESET:
1806             /* This fence hasn't been submitted yet, we'll catch it the next
1807              * time around.  Yes, this may mean we dead-loop but, short of
1808              * lots of locking and a condition variable, there's not much that
1809              * we can do about that.
1810              */
1811             pending_fences++;
1812             continue;
1813
1814          case ANV_FENCE_STATE_SIGNALED:
1815             /* This fence is not pending.  If waitAll isn't set, we can return
1816              * early.  Otherwise, we have to keep going.
1817              */
1818             if (!waitAll)
1819                return VK_SUCCESS;
1820             continue;
1821
1822          case ANV_FENCE_STATE_SUBMITTED:
1823             /* These are the fences we really care about.  Go ahead and wait
1824              * on it until we hit a timeout.
1825              */
1826             ret = anv_gem_wait(device, fence->bo.gem_handle, &timeout);
1827             if (ret == -1 && errno == ETIME) {
1828                return VK_TIMEOUT;
1829             } else if (ret == -1) {
1830                /* We don't know the real error. */
1831                return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1832             } else {
1833                fence->state = ANV_FENCE_STATE_SIGNALED;
1834                signaled_fences = true;
1835                if (!waitAll)
1836                   return VK_SUCCESS;
1837                continue;
1838             }
1839          }
1840       }
1841
1842       if (pending_fences && !signaled_fences) {
1843          /* If we've hit this then someone decided to vkWaitForFences before
1844           * they've actually submitted any of them to a queue.  This is a
1845           * fairly pessimal case, so it's ok to lock here and use a standard
1846           * pthreads condition variable.
1847           */
1848          pthread_mutex_lock(&device->mutex);
1849
1850          /* It's possible that some of the fences have changed state since the
1851           * last time we checked.  Now that we have the lock, check for
1852           * pending fences again and don't wait if it's changed.
1853           */
1854          uint32_t now_pending_fences = 0;
1855          for (uint32_t i = 0; i < fenceCount; i++) {
1856             ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1857             if (fence->state == ANV_FENCE_STATE_RESET)
1858                now_pending_fences++;
1859          }
1860          assert(now_pending_fences <= pending_fences);
1861
1862          if (now_pending_fences == pending_fences) {
1863             struct timespec before;
1864             clock_gettime(CLOCK_MONOTONIC, &before);
1865
1866             uint32_t abs_nsec = before.tv_nsec + timeout % NSEC_PER_SEC;
1867             uint64_t abs_sec = before.tv_sec + (abs_nsec / NSEC_PER_SEC) +
1868                                (timeout / NSEC_PER_SEC);
1869             abs_nsec %= NSEC_PER_SEC;
1870
1871             /* Avoid roll-over in tv_sec on 32-bit systems if the user
1872              * provided timeout is UINT64_MAX
1873              */
1874             struct timespec abstime;
1875             abstime.tv_nsec = abs_nsec;
1876             abstime.tv_sec = MIN2(abs_sec, INT_TYPE_MAX(abstime.tv_sec));
1877
1878             ret = pthread_cond_timedwait(&device->queue_submit,
1879                                          &device->mutex, &abstime);
1880             assert(ret != EINVAL);
1881
1882             struct timespec after;
1883             clock_gettime(CLOCK_MONOTONIC, &after);
1884             uint64_t time_elapsed =
1885                ((uint64_t)after.tv_sec * NSEC_PER_SEC + after.tv_nsec) -
1886                ((uint64_t)before.tv_sec * NSEC_PER_SEC + before.tv_nsec);
1887
1888             if (time_elapsed >= timeout) {
1889                pthread_mutex_unlock(&device->mutex);
1890                return VK_TIMEOUT;
1891             }
1892
1893             timeout -= time_elapsed;
1894          }
1895
1896          pthread_mutex_unlock(&device->mutex);
1897       }
1898    }
1899
1900    return VK_SUCCESS;
1901 }
1902
1903 // Queue semaphore functions
1904
1905 VkResult anv_CreateSemaphore(
1906     VkDevice                                    device,
1907     const VkSemaphoreCreateInfo*                pCreateInfo,
1908     const VkAllocationCallbacks*                pAllocator,
1909     VkSemaphore*                                pSemaphore)
1910 {
1911    /* The DRM execbuffer ioctl always execute in-oder, even between different
1912     * rings. As such, there's nothing to do for the user space semaphore.
1913     */
1914
1915    *pSemaphore = (VkSemaphore)1;
1916
1917    return VK_SUCCESS;
1918 }
1919
1920 void anv_DestroySemaphore(
1921     VkDevice                                    device,
1922     VkSemaphore                                 semaphore,
1923     const VkAllocationCallbacks*                pAllocator)
1924 {
1925 }
1926
1927 // Event functions
1928
1929 VkResult anv_CreateEvent(
1930     VkDevice                                    _device,
1931     const VkEventCreateInfo*                    pCreateInfo,
1932     const VkAllocationCallbacks*                pAllocator,
1933     VkEvent*                                    pEvent)
1934 {
1935    ANV_FROM_HANDLE(anv_device, device, _device);
1936    struct anv_state state;
1937    struct anv_event *event;
1938
1939    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1940
1941    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1942                                 sizeof(*event), 8);
1943    event = state.map;
1944    event->state = state;
1945    event->semaphore = VK_EVENT_RESET;
1946
1947    if (!device->info.has_llc) {
1948       /* Make sure the writes we're flushing have landed. */
1949       __builtin_ia32_mfence();
1950       __builtin_ia32_clflush(event);
1951    }
1952
1953    *pEvent = anv_event_to_handle(event);
1954
1955    return VK_SUCCESS;
1956 }
1957
1958 void anv_DestroyEvent(
1959     VkDevice                                    _device,
1960     VkEvent                                     _event,
1961     const VkAllocationCallbacks*                pAllocator)
1962 {
1963    ANV_FROM_HANDLE(anv_device, device, _device);
1964    ANV_FROM_HANDLE(anv_event, event, _event);
1965
1966    if (!event)
1967       return;
1968
1969    anv_state_pool_free(&device->dynamic_state_pool, event->state);
1970 }
1971
1972 VkResult anv_GetEventStatus(
1973     VkDevice                                    _device,
1974     VkEvent                                     _event)
1975 {
1976    ANV_FROM_HANDLE(anv_device, device, _device);
1977    ANV_FROM_HANDLE(anv_event, event, _event);
1978
1979    if (!device->info.has_llc) {
1980       /* Invalidate read cache before reading event written by GPU. */
1981       __builtin_ia32_clflush(event);
1982       __builtin_ia32_mfence();
1983
1984    }
1985
1986    return event->semaphore;
1987 }
1988
1989 VkResult anv_SetEvent(
1990     VkDevice                                    _device,
1991     VkEvent                                     _event)
1992 {
1993    ANV_FROM_HANDLE(anv_device, device, _device);
1994    ANV_FROM_HANDLE(anv_event, event, _event);
1995
1996    event->semaphore = VK_EVENT_SET;
1997
1998    if (!device->info.has_llc) {
1999       /* Make sure the writes we're flushing have landed. */
2000       __builtin_ia32_mfence();
2001       __builtin_ia32_clflush(event);
2002    }
2003
2004    return VK_SUCCESS;
2005 }
2006
2007 VkResult anv_ResetEvent(
2008     VkDevice                                    _device,
2009     VkEvent                                     _event)
2010 {
2011    ANV_FROM_HANDLE(anv_device, device, _device);
2012    ANV_FROM_HANDLE(anv_event, event, _event);
2013
2014    event->semaphore = VK_EVENT_RESET;
2015
2016    if (!device->info.has_llc) {
2017       /* Make sure the writes we're flushing have landed. */
2018       __builtin_ia32_mfence();
2019       __builtin_ia32_clflush(event);
2020    }
2021
2022    return VK_SUCCESS;
2023 }
2024
2025 // Buffer functions
2026
2027 VkResult anv_CreateBuffer(
2028     VkDevice                                    _device,
2029     const VkBufferCreateInfo*                   pCreateInfo,
2030     const VkAllocationCallbacks*                pAllocator,
2031     VkBuffer*                                   pBuffer)
2032 {
2033    ANV_FROM_HANDLE(anv_device, device, _device);
2034    struct anv_buffer *buffer;
2035
2036    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2037
2038    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2039                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2040    if (buffer == NULL)
2041       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2042
2043    buffer->size = pCreateInfo->size;
2044    buffer->usage = pCreateInfo->usage;
2045    buffer->bo = NULL;
2046    buffer->offset = 0;
2047
2048    *pBuffer = anv_buffer_to_handle(buffer);
2049
2050    return VK_SUCCESS;
2051 }
2052
2053 void anv_DestroyBuffer(
2054     VkDevice                                    _device,
2055     VkBuffer                                    _buffer,
2056     const VkAllocationCallbacks*                pAllocator)
2057 {
2058    ANV_FROM_HANDLE(anv_device, device, _device);
2059    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2060
2061    if (!buffer)
2062       return;
2063
2064    vk_free2(&device->alloc, pAllocator, buffer);
2065 }
2066
2067 void
2068 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
2069                               enum isl_format format,
2070                               uint32_t offset, uint32_t range, uint32_t stride)
2071 {
2072    isl_buffer_fill_state(&device->isl_dev, state.map,
2073                          .address = offset,
2074                          .mocs = device->default_mocs,
2075                          .size = range,
2076                          .format = format,
2077                          .stride = stride);
2078
2079    anv_state_flush(device, state);
2080 }
2081
2082 void anv_DestroySampler(
2083     VkDevice                                    _device,
2084     VkSampler                                   _sampler,
2085     const VkAllocationCallbacks*                pAllocator)
2086 {
2087    ANV_FROM_HANDLE(anv_device, device, _device);
2088    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
2089
2090    if (!sampler)
2091       return;
2092
2093    vk_free2(&device->alloc, pAllocator, sampler);
2094 }
2095
2096 VkResult anv_CreateFramebuffer(
2097     VkDevice                                    _device,
2098     const VkFramebufferCreateInfo*              pCreateInfo,
2099     const VkAllocationCallbacks*                pAllocator,
2100     VkFramebuffer*                              pFramebuffer)
2101 {
2102    ANV_FROM_HANDLE(anv_device, device, _device);
2103    struct anv_framebuffer *framebuffer;
2104
2105    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2106
2107    size_t size = sizeof(*framebuffer) +
2108                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2109    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2110                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2111    if (framebuffer == NULL)
2112       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2113
2114    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2115    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2116       VkImageView _iview = pCreateInfo->pAttachments[i];
2117       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2118    }
2119
2120    framebuffer->width = pCreateInfo->width;
2121    framebuffer->height = pCreateInfo->height;
2122    framebuffer->layers = pCreateInfo->layers;
2123
2124    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2125
2126    return VK_SUCCESS;
2127 }
2128
2129 void anv_DestroyFramebuffer(
2130     VkDevice                                    _device,
2131     VkFramebuffer                               _fb,
2132     const VkAllocationCallbacks*                pAllocator)
2133 {
2134    ANV_FROM_HANDLE(anv_device, device, _device);
2135    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2136
2137    if (!fb)
2138       return;
2139
2140    vk_free2(&device->alloc, pAllocator, fb);
2141 }
2142
2143 /* vk_icd.h does not declare this function, so we declare it here to
2144  * suppress Wmissing-prototypes.
2145  */
2146 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2147 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2148
2149 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2150 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2151 {
2152    /* For the full details on loader interface versioning, see
2153     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2154     * What follows is a condensed summary, to help you navigate the large and
2155     * confusing official doc.
2156     *
2157     *   - Loader interface v0 is incompatible with later versions. We don't
2158     *     support it.
2159     *
2160     *   - In loader interface v1:
2161     *       - The first ICD entrypoint called by the loader is
2162     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2163     *         entrypoint.
2164     *       - The ICD must statically expose no other Vulkan symbol unless it is
2165     *         linked with -Bsymbolic.
2166     *       - Each dispatchable Vulkan handle created by the ICD must be
2167     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2168     *         ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2169     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2170     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2171     *         such loader-managed surfaces.
2172     *
2173     *    - Loader interface v2 differs from v1 in:
2174     *       - The first ICD entrypoint called by the loader is
2175     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2176     *         statically expose this entrypoint.
2177     *
2178     *    - Loader interface v3 differs from v2 in:
2179     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2180     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2181     *          because the loader no longer does so.
2182     */
2183    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2184    return VK_SUCCESS;
2185 }