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 <dlfcn.h>
  25 #include <assert.h>
  26 #include <stdbool.h>
  27 #include <string.h>
  28 #include <sys/mman.h>
  29 #include <sys/stat.h>
  30 #include <unistd.h>
  31 #include <fcntl.h>
  32
  33 #include "anv_private.h"
  34 #include "util/strtod.h"
  35 #include "util/debug.h"
  36
  37 #include "genxml/gen7_pack.h"
  38
  39 struct anv_dispatch_table dtable;
  40
  41 static void
  42 compiler_debug_log(void *data, const char *fmt, ...)
  43 { }
  44
  45 static void
  46 compiler_perf_log(void *data, const char *fmt, ...)
  47 {
  48    va_list args;
  49    va_start(args, fmt);
  50
  51    if (unlikely(INTEL_DEBUG & DEBUG_PERF))
  52       vfprintf(stderr, fmt, args);
  53
  54    va_end(args);
  55 }
  56
  57 static bool
  58 anv_get_function_timestamp(void *ptr, uint32_t* timestamp)
  59 {
  60    Dl_info info;
  61    struct stat st;
  62    if (!dladdr(ptr, &info) || !info.dli_fname)
  63       return false;
  64
  65    if (stat(info.dli_fname, &st))
  66       return false;
  67
  68    *timestamp = st.st_mtim.tv_sec;
  69    return true;
  70 }
  71
  72 static bool
  73 anv_device_get_cache_uuid(void *uuid)
  74 {
  75    uint32_t timestamp;
  76
  77    memset(uuid, 0, VK_UUID_SIZE);
  78    if (!anv_get_function_timestamp(anv_device_get_cache_uuid, &timestamp))
  79       return false;
  80
  81    snprintf(uuid, VK_UUID_SIZE, "anv-%d", timestamp);
  82    return true;
  83 }
  84
  85 static VkResult
  86 anv_physical_device_init(struct anv_physical_device *device,
  87                          struct anv_instance *instance,
  88                          const char *path)
  89 {
  90    VkResult result;
  91    int fd;
  92
  93    fd = open(path, O_RDWR | O_CLOEXEC);
  94    if (fd < 0)
  95       return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  96
  97    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
  98    device->instance = instance;
  99
 100    assert(strlen(path) < ARRAY_SIZE(device->path));
 101    strncpy(device->path, path, ARRAY_SIZE(device->path));
 102
 103    device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
 104    if (!device->chipset_id) {
 105       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
 106       goto fail;
 107    }
 108
 109    device->name = gen_get_device_name(device->chipset_id);
 110    if (!gen_get_device_info(device->chipset_id, &device->info)) {
 111       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
 112       goto fail;
 113    }
 114
 115    if (device->info.is_haswell) {
 116       fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
 117    } else if (device->info.gen == 7 && !device->info.is_baytrail) {
 118       fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
 119    } else if (device->info.gen == 7 && device->info.is_baytrail) {
 120       fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
 121    } else if (device->info.gen >= 8) {
 122       /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
 123        * supported as anything */
 124    } else {
 125       result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 126                          "Vulkan not yet supported on %s", device->name);
 127       goto fail;
 128    }
 129
 130    device->cmd_parser_version = -1;
 131    if (device->info.gen == 7) {
 132       device->cmd_parser_version =
 133          anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
 134       if (device->cmd_parser_version == -1) {
 135          result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 136                             "failed to get command parser version");
 137          goto fail;
 138       }
 139    }
 140
 141    if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
 142       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 143                          "failed to get aperture size: %m");
 144       goto fail;
 145    }
 146
 147    if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
 148       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 149                          "kernel missing gem wait");
 150       goto fail;
 151    }
 152
 153    if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
 154       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 155                          "kernel missing execbuf2");
 156       goto fail;
 157    }
 158
 159    if (!device->info.has_llc &&
 160        anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
 161       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 162                          "kernel missing wc mmap");
 163       goto fail;
 164    }
 165
 166    if (!anv_device_get_cache_uuid(device->uuid)) {
 167       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 168                          "cannot generate UUID");
 169       goto fail;
 170    }
 171    bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
 172
 173    /* GENs prior to 8 do not support EU/Subslice info */
 174    if (device->info.gen >= 8) {
 175       device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
 176       device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
 177
 178       /* Without this information, we cannot get the right Braswell
 179        * brandstrings, and we have to use conservative numbers for GPGPU on
 180        * many platforms, but otherwise, things will just work.
 181        */
 182       if (device->subslice_total < 1 || device->eu_total < 1) {
 183          fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
 184                          " query GPU properties.\n");
 185       }
 186    } else if (device->info.gen == 7) {
 187       device->subslice_total = 1 << (device->info.gt - 1);
 188    }
 189
 190    if (device->info.is_cherryview &&
 191        device->subslice_total > 0 && device->eu_total > 0) {
 192       /* Logical CS threads = EUs per subslice * 7 threads per EU */
 193       uint32_t max_cs_threads = device->eu_total / device->subslice_total * 7;
 194
 195       /* Fuse configurations may give more threads than expected, never less. */
 196       if (max_cs_threads > device->info.max_cs_threads)
 197          device->info.max_cs_threads = max_cs_threads;
 198    }
 199
 200    brw_process_intel_debug_variable();
 201
 202    device->compiler = brw_compiler_create(NULL, &device->info);
 203    if (device->compiler == NULL) {
 204       result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 205       goto fail;
 206    }
 207    device->compiler->shader_debug_log = compiler_debug_log;
 208    device->compiler->shader_perf_log = compiler_perf_log;
 209
 210    result = anv_init_wsi(device);
 211    if (result != VK_SUCCESS) {
 212       ralloc_free(device->compiler);
 213       goto fail;
 214    }
 215
 216    isl_device_init(&device->isl_dev, &device->info, swizzled);
 217
 218    close(fd);
 219    return VK_SUCCESS;
 220
 221 fail:
 222    close(fd);
 223    return result;
 224 }
 225
 226 static void
 227 anv_physical_device_finish(struct anv_physical_device *device)
 228 {
 229    anv_finish_wsi(device);
 230    ralloc_free(device->compiler);
 231 }
 232
 233 static const VkExtensionProperties global_extensions[] = {
 234    {
 235       .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
 236       .specVersion = 25,
 237    },
 238 #ifdef VK_USE_PLATFORM_XCB_KHR
 239    {
 240       .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
 241       .specVersion = 6,
 242    },
 243 #endif
 244 #ifdef VK_USE_PLATFORM_XLIB_KHR
 245    {
 246       .extensionName = VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
 247       .specVersion = 6,
 248    },
 249 #endif
 250 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
 251    {
 252       .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
 253       .specVersion = 5,
 254    },
 255 #endif
 256    {
 257       .extensionName = VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
 258       .specVersion = 1,
 259    },
 260 };
 261
 262 static const VkExtensionProperties device_extensions[] = {
 263    {
 264       .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
 265       .specVersion = 68,
 266    },
 267    {
 268       .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
 269       .specVersion = 1,
 270    },
 271    {
 272       .extensionName = VK_KHR_MAINTENANCE1_EXTENSION_NAME,
 273       .specVersion = 1,
 274    },
 275 };
 276
 277 static void *
 278 default_alloc_func(void *pUserData, size_t size, size_t align,
 279                    VkSystemAllocationScope allocationScope)
 280 {
 281    return malloc(size);
 282 }
 283
 284 static void *
 285 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
 286                      size_t align, VkSystemAllocationScope allocationScope)
 287 {
 288    return realloc(pOriginal, size);
 289 }
 290
 291 static void
 292 default_free_func(void *pUserData, void *pMemory)
 293 {
 294    free(pMemory);
 295 }
 296
 297 static const VkAllocationCallbacks default_alloc = {
 298    .pUserData = NULL,
 299    .pfnAllocation = default_alloc_func,
 300    .pfnReallocation = default_realloc_func,
 301    .pfnFree = default_free_func,
 302 };
 303
 304 VkResult anv_CreateInstance(
 305     const VkInstanceCreateInfo*                 pCreateInfo,
 306     const VkAllocationCallbacks*                pAllocator,
 307     VkInstance*                                 pInstance)
 308 {
 309    struct anv_instance *instance;
 310
 311    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 312
 313    uint32_t client_version;
 314    if (pCreateInfo->pApplicationInfo &&
 315        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 316       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 317    } else {
 318       client_version = VK_MAKE_VERSION(1, 0, 0);
 319    }
 320
 321    if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
 322        client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
 323       return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 324                        "Client requested version %d.%d.%d",
 325                        VK_VERSION_MAJOR(client_version),
 326                        VK_VERSION_MINOR(client_version),
 327                        VK_VERSION_PATCH(client_version));
 328    }
 329
 330    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 331       bool found = false;
 332       for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
 333          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 334                     global_extensions[j].extensionName) == 0) {
 335             found = true;
 336             break;
 337          }
 338       }
 339       if (!found)
 340          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 341    }
 342
 343    instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 344                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 345    if (!instance)
 346       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 347
 348    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 349
 350    if (pAllocator)
 351       instance->alloc = *pAllocator;
 352    else
 353       instance->alloc = default_alloc;
 354
 355    instance->apiVersion = client_version;
 356    instance->physicalDeviceCount = -1;
 357
 358    _mesa_locale_init();
 359
 360    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 361
 362    *pInstance = anv_instance_to_handle(instance);
 363
 364    return VK_SUCCESS;
 365 }
 366
 367 void anv_DestroyInstance(
 368     VkInstance                                  _instance,
 369     const VkAllocationCallbacks*                pAllocator)
 370 {
 371    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 372
 373    if (instance->physicalDeviceCount > 0) {
 374       /* We support at most one physical device. */
 375       assert(instance->physicalDeviceCount == 1);
 376       anv_physical_device_finish(&instance->physicalDevice);
 377    }
 378
 379    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 380
 381    _mesa_locale_fini();
 382
 383    vk_free(&instance->alloc, instance);
 384 }
 385
 386 VkResult anv_EnumeratePhysicalDevices(
 387     VkInstance                                  _instance,
 388     uint32_t*                                   pPhysicalDeviceCount,
 389     VkPhysicalDevice*                           pPhysicalDevices)
 390 {
 391    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 392    VkResult result;
 393
 394    if (instance->physicalDeviceCount < 0) {
 395       char path[20];
 396       for (unsigned i = 0; i < 8; i++) {
 397          snprintf(path, sizeof(path), "/dev/dri/renderD%d", 128 + i);
 398          result = anv_physical_device_init(&instance->physicalDevice,
 399                                            instance, path);
 400          if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 401             break;
 402       }
 403
 404       if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
 405          instance->physicalDeviceCount = 0;
 406       } else if (result == VK_SUCCESS) {
 407          instance->physicalDeviceCount = 1;
 408       } else {
 409          return result;
 410       }
 411    }
 412
 413    /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
 414     * otherwise it's an inout parameter.
 415     *
 416     * The Vulkan spec (git aaed022) says:
 417     *
 418     *    pPhysicalDeviceCount is a pointer to an unsigned integer variable
 419     *    that is initialized with the number of devices the application is
 420     *    prepared to receive handles to. pname:pPhysicalDevices is pointer to
 421     *    an array of at least this many VkPhysicalDevice handles [...].
 422     *
 423     *    Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
 424     *    overwrites the contents of the variable pointed to by
 425     *    pPhysicalDeviceCount with the number of physical devices in in the
 426     *    instance; otherwise, vkEnumeratePhysicalDevices overwrites
 427     *    pPhysicalDeviceCount with the number of physical handles written to
 428     *    pPhysicalDevices.
 429     */
 430    if (!pPhysicalDevices) {
 431       *pPhysicalDeviceCount = instance->physicalDeviceCount;
 432    } else if (*pPhysicalDeviceCount >= 1) {
 433       pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
 434       *pPhysicalDeviceCount = 1;
 435    } else if (*pPhysicalDeviceCount < instance->physicalDeviceCount) {
 436       return VK_INCOMPLETE;
 437    } else {
 438       *pPhysicalDeviceCount = 0;
 439    }
 440
 441    return VK_SUCCESS;
 442 }
 443
 444 void anv_GetPhysicalDeviceFeatures(
 445     VkPhysicalDevice                            physicalDevice,
 446     VkPhysicalDeviceFeatures*                   pFeatures)
 447 {
 448    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 449
 450    *pFeatures = (VkPhysicalDeviceFeatures) {
 451       .robustBufferAccess                       = true,
 452       .fullDrawIndexUint32                      = true,
 453       .imageCubeArray                           = true,
 454       .independentBlend                         = true,
 455       .geometryShader                           = true,
 456       .tessellationShader                       = true,
 457       .sampleRateShading                        = true,
 458       .dualSrcBlend                             = true,
 459       .logicOp                                  = true,
 460       .multiDrawIndirect                        = false,
 461       .drawIndirectFirstInstance                = true,
 462       .depthClamp                               = true,
 463       .depthBiasClamp                           = true,
 464       .fillModeNonSolid                         = true,
 465       .depthBounds                              = false,
 466       .wideLines                                = true,
 467       .largePoints                              = true,
 468       .alphaToOne                               = true,
 469       .multiViewport                            = true,
 470       .samplerAnisotropy                        = true,
 471       .textureCompressionETC2                   = pdevice->info.gen >= 8 ||
 472                                                   pdevice->info.is_baytrail,
 473       .textureCompressionASTC_LDR               = pdevice->info.gen >= 9, /* FINISHME CHV */
 474       .textureCompressionBC                     = true,
 475       .occlusionQueryPrecise                    = true,
 476       .pipelineStatisticsQuery                  = false,
 477       .fragmentStoresAndAtomics                 = true,
 478       .shaderTessellationAndGeometryPointSize   = true,
 479       .shaderImageGatherExtended                = true,
 480       .shaderStorageImageExtendedFormats        = true,
 481       .shaderStorageImageMultisample            = false,
 482       .shaderStorageImageReadWithoutFormat      = false,
 483       .shaderStorageImageWriteWithoutFormat     = false,
 484       .shaderUniformBufferArrayDynamicIndexing  = true,
 485       .shaderSampledImageArrayDynamicIndexing   = true,
 486       .shaderStorageBufferArrayDynamicIndexing  = true,
 487       .shaderStorageImageArrayDynamicIndexing   = true,
 488       .shaderClipDistance                       = true,
 489       .shaderCullDistance                       = true,
 490       .shaderFloat64                            = pdevice->info.gen >= 8,
 491       .shaderInt64                              = false,
 492       .shaderInt16                              = false,
 493       .shaderResourceMinLod                     = false,
 494       .variableMultisampleRate                  = false,
 495       .inheritedQueries                         = false,
 496    };
 497
 498    /* We can't do image stores in vec4 shaders */
 499    pFeatures->vertexPipelineStoresAndAtomics =
 500       pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
 501       pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
 502 }
 503
 504 void anv_GetPhysicalDeviceFeatures2KHR(
 505     VkPhysicalDevice                            physicalDevice,
 506     VkPhysicalDeviceFeatures2KHR*               pFeatures)
 507 {
 508    anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 509
 510    for (struct anv_common *c = pFeatures->pNext; c != NULL; c = c->pNext) {
 511       switch (c->sType) {
 512       default:
 513          anv_debug_ignored_stype(c->sType);
 514          break;
 515       }
 516    }
 517 }
 518
 519 void anv_GetPhysicalDeviceProperties(
 520     VkPhysicalDevice                            physicalDevice,
 521     VkPhysicalDeviceProperties*                 pProperties)
 522 {
 523    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 524    const struct gen_device_info *devinfo = &pdevice->info;
 525
 526    const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
 527
 528    /* See assertions made when programming the buffer surface state. */
 529    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
 530                                       (1ul << 30) : (1ul << 27);
 531
 532    VkSampleCountFlags sample_counts =
 533       isl_device_get_sample_counts(&pdevice->isl_dev);
 534
 535    VkPhysicalDeviceLimits limits = {
 536       .maxImageDimension1D                      = (1 << 14),
 537       .maxImageDimension2D                      = (1 << 14),
 538       .maxImageDimension3D                      = (1 << 11),
 539       .maxImageDimensionCube                    = (1 << 14),
 540       .maxImageArrayLayers                      = (1 << 11),
 541       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 542       .maxUniformBufferRange                    = (1ul << 27),
 543       .maxStorageBufferRange                    = max_raw_buffer_sz,
 544       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 545       .maxMemoryAllocationCount                 = UINT32_MAX,
 546       .maxSamplerAllocationCount                = 64 * 1024,
 547       .bufferImageGranularity                   = 64, /* A cache line */
 548       .sparseAddressSpaceSize                   = 0,
 549       .maxBoundDescriptorSets                   = MAX_SETS,
 550       .maxPerStageDescriptorSamplers            = 64,
 551       .maxPerStageDescriptorUniformBuffers      = 64,
 552       .maxPerStageDescriptorStorageBuffers      = 64,
 553       .maxPerStageDescriptorSampledImages       = 64,
 554       .maxPerStageDescriptorStorageImages       = 64,
 555       .maxPerStageDescriptorInputAttachments    = 64,
 556       .maxPerStageResources                     = 128,
 557       .maxDescriptorSetSamplers                 = 256,
 558       .maxDescriptorSetUniformBuffers           = 256,
 559       .maxDescriptorSetUniformBuffersDynamic    = 256,
 560       .maxDescriptorSetStorageBuffers           = 256,
 561       .maxDescriptorSetStorageBuffersDynamic    = 256,
 562       .maxDescriptorSetSampledImages            = 256,
 563       .maxDescriptorSetStorageImages            = 256,
 564       .maxDescriptorSetInputAttachments         = 256,
 565       .maxVertexInputAttributes                 = MAX_VBS,
 566       .maxVertexInputBindings                   = MAX_VBS,
 567       .maxVertexInputAttributeOffset            = 2047,
 568       .maxVertexInputBindingStride              = 2048,
 569       .maxVertexOutputComponents                = 128,
 570       .maxTessellationGenerationLevel           = 64,
 571       .maxTessellationPatchSize                 = 32,
 572       .maxTessellationControlPerVertexInputComponents = 128,
 573       .maxTessellationControlPerVertexOutputComponents = 128,
 574       .maxTessellationControlPerPatchOutputComponents = 128,
 575       .maxTessellationControlTotalOutputComponents = 2048,
 576       .maxTessellationEvaluationInputComponents = 128,
 577       .maxTessellationEvaluationOutputComponents = 128,
 578       .maxGeometryShaderInvocations             = 32,
 579       .maxGeometryInputComponents               = 64,
 580       .maxGeometryOutputComponents              = 128,
 581       .maxGeometryOutputVertices                = 256,
 582       .maxGeometryTotalOutputComponents         = 1024,
 583       .maxFragmentInputComponents               = 128,
 584       .maxFragmentOutputAttachments             = 8,
 585       .maxFragmentDualSrcAttachments            = 1,
 586       .maxFragmentCombinedOutputResources       = 8,
 587       .maxComputeSharedMemorySize               = 32768,
 588       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 589       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 590       .maxComputeWorkGroupSize = {
 591          16 * devinfo->max_cs_threads,
 592          16 * devinfo->max_cs_threads,
 593          16 * devinfo->max_cs_threads,
 594       },
 595       .subPixelPrecisionBits                    = 4 /* FIXME */,
 596       .subTexelPrecisionBits                    = 4 /* FIXME */,
 597       .mipmapPrecisionBits                      = 4 /* FIXME */,
 598       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 599       .maxDrawIndirectCount                     = UINT32_MAX,
 600       .maxSamplerLodBias                        = 16,
 601       .maxSamplerAnisotropy                     = 16,
 602       .maxViewports                             = MAX_VIEWPORTS,
 603       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 604       .viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
 605       .viewportSubPixelBits                     = 13, /* We take a float? */
 606       .minMemoryMapAlignment                    = 4096, /* A page */
 607       .minTexelBufferOffsetAlignment            = 1,
 608       .minUniformBufferOffsetAlignment          = 16,
 609       .minStorageBufferOffsetAlignment          = 4,
 610       .minTexelOffset                           = -8,
 611       .maxTexelOffset                           = 7,
 612       .minTexelGatherOffset                     = -32,
 613       .maxTexelGatherOffset                     = 31,
 614       .minInterpolationOffset                   = -0.5,
 615       .maxInterpolationOffset                   = 0.4375,
 616       .subPixelInterpolationOffsetBits          = 4,
 617       .maxFramebufferWidth                      = (1 << 14),
 618       .maxFramebufferHeight                     = (1 << 14),
 619       .maxFramebufferLayers                     = (1 << 11),
 620       .framebufferColorSampleCounts             = sample_counts,
 621       .framebufferDepthSampleCounts             = sample_counts,
 622       .framebufferStencilSampleCounts           = sample_counts,
 623       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 624       .maxColorAttachments                      = MAX_RTS,
 625       .sampledImageColorSampleCounts            = sample_counts,
 626       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 627       .sampledImageDepthSampleCounts            = sample_counts,
 628       .sampledImageStencilSampleCounts          = sample_counts,
 629       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 630       .maxSampleMaskWords                       = 1,
 631       .timestampComputeAndGraphics              = false,
 632       .timestampPeriod                          = time_stamp_base,
 633       .maxClipDistances                         = 8,
 634       .maxCullDistances                         = 8,
 635       .maxCombinedClipAndCullDistances          = 8,
 636       .discreteQueuePriorities                  = 1,
 637       .pointSizeRange                           = { 0.125, 255.875 },
 638       .lineWidthRange                           = { 0.0, 7.9921875 },
 639       .pointSizeGranularity                     = (1.0 / 8.0),
 640       .lineWidthGranularity                     = (1.0 / 128.0),
 641       .strictLines                              = false, /* FINISHME */
 642       .standardSampleLocations                  = true,
 643       .optimalBufferCopyOffsetAlignment         = 128,
 644       .optimalBufferCopyRowPitchAlignment       = 128,
 645       .nonCoherentAtomSize                      = 64,
 646    };
 647
 648    *pProperties = (VkPhysicalDeviceProperties) {
 649       .apiVersion = VK_MAKE_VERSION(1, 0, 39),
 650       .driverVersion = 1,
 651       .vendorID = 0x8086,
 652       .deviceID = pdevice->chipset_id,
 653       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 654       .limits = limits,
 655       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 656    };
 657
 658    strcpy(pProperties->deviceName, pdevice->name);
 659    memcpy(pProperties->pipelineCacheUUID, pdevice->uuid, VK_UUID_SIZE);
 660 }
 661
 662 void anv_GetPhysicalDeviceProperties2KHR(
 663     VkPhysicalDevice                            physicalDevice,
 664     VkPhysicalDeviceProperties2KHR*             pProperties)
 665 {
 666    anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 667
 668    for (struct anv_common *c = pProperties->pNext; c != NULL; c = c->pNext) {
 669       switch (c->sType) {
 670       default:
 671          anv_debug_ignored_stype(c->sType);
 672          break;
 673       }
 674    }
 675 }
 676
 677 static void
 678 anv_get_queue_family_properties(struct anv_physical_device *phys_dev,
 679                                 VkQueueFamilyProperties *props)
 680 {
 681    *props = (VkQueueFamilyProperties) {
 682       .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 683                     VK_QUEUE_COMPUTE_BIT |
 684                     VK_QUEUE_TRANSFER_BIT,
 685       .queueCount = 1,
 686       .timestampValidBits = 36, /* XXX: Real value here */
 687       .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
 688    };
 689 }
 690
 691 void anv_GetPhysicalDeviceQueueFamilyProperties(
 692     VkPhysicalDevice                            physicalDevice,
 693     uint32_t*                                   pCount,
 694     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 695 {
 696    ANV_FROM_HANDLE(anv_physical_device, phys_dev, physicalDevice);
 697
 698    if (pQueueFamilyProperties == NULL) {
 699       *pCount = 1;
 700       return;
 701    }
 702
 703    /* The spec implicitly allows the incoming count to be 0. From the Vulkan
 704     * 1.0.38 spec, Section 4.1 Physical Devices:
 705     *
 706     *     If the value referenced by pQueueFamilyPropertyCount is not 0 [then
 707     *     do stuff].
 708     */
 709    if (*pCount == 0)
 710       return;
 711
 712    *pCount = 1;
 713    anv_get_queue_family_properties(phys_dev, pQueueFamilyProperties);
 714 }
 715
 716 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
 717     VkPhysicalDevice                            physicalDevice,
 718     uint32_t*                                   pQueueFamilyPropertyCount,
 719     VkQueueFamilyProperties2KHR*                pQueueFamilyProperties)
 720 {
 721
 722    ANV_FROM_HANDLE(anv_physical_device, phys_dev, physicalDevice);
 723
 724    if (pQueueFamilyProperties == NULL) {
 725       *pQueueFamilyPropertyCount = 1;
 726       return;
 727    }
 728
 729    /* The spec implicitly allows the incoming count to be 0. From the Vulkan
 730     * 1.0.38 spec, Section 4.1 Physical Devices:
 731     *
 732     *     If the value referenced by pQueueFamilyPropertyCount is not 0 [then
 733     *     do stuff].
 734     */
 735    if (*pQueueFamilyPropertyCount == 0)
 736       return;
 737
 738    /* We support exactly one queue family. So need to traverse only the first
 739     * array element's pNext chain.
 740     */
 741    *pQueueFamilyPropertyCount = 1;
 742    anv_get_queue_family_properties(phys_dev,
 743          &pQueueFamilyProperties->queueFamilyProperties);
 744
 745    for (struct anv_common *c = pQueueFamilyProperties->pNext;
 746         c != NULL; c = c->pNext) {
 747       switch (c->sType) {
 748       default:
 749          anv_debug_ignored_stype(c->sType);
 750          break;
 751       }
 752    }
 753 }
 754
 755 void anv_GetPhysicalDeviceMemoryProperties(
 756     VkPhysicalDevice                            physicalDevice,
 757     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
 758 {
 759    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 760    VkDeviceSize heap_size;
 761
 762    /* Reserve some wiggle room for the driver by exposing only 75% of the
 763     * aperture to the heap.
 764     */
 765    heap_size = 3 * physical_device->aperture_size / 4;
 766
 767    if (physical_device->info.has_llc) {
 768       /* Big core GPUs share LLC with the CPU and thus one memory type can be
 769        * both cached and coherent at the same time.
 770        */
 771       pMemoryProperties->memoryTypeCount = 1;
 772       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 773          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 774                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 775                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
 776                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 777          .heapIndex = 0,
 778       };
 779    } else {
 780       /* The spec requires that we expose a host-visible, coherent memory
 781        * type, but Atom GPUs don't share LLC. Thus we offer two memory types
 782        * to give the application a choice between cached, but not coherent and
 783        * coherent but uncached (WC though).
 784        */
 785       pMemoryProperties->memoryTypeCount = 2;
 786       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 787          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 788                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 789                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
 790          .heapIndex = 0,
 791       };
 792       pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
 793          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 794                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 795                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 796          .heapIndex = 0,
 797       };
 798    }
 799
 800    pMemoryProperties->memoryHeapCount = 1;
 801    pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
 802       .size = heap_size,
 803       .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 804    };
 805 }
 806
 807 void anv_GetPhysicalDeviceMemoryProperties2KHR(
 808     VkPhysicalDevice                            physicalDevice,
 809     VkPhysicalDeviceMemoryProperties2KHR*       pMemoryProperties)
 810 {
 811    anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
 812                                          &pMemoryProperties->memoryProperties);
 813
 814    for (struct anv_common *c = pMemoryProperties->pNext;
 815         c != NULL; c = c->pNext) {
 816       switch (c->sType) {
 817       default:
 818          anv_debug_ignored_stype(c->sType);
 819          break;
 820       }
 821    }
 822 }
 823
 824 PFN_vkVoidFunction anv_GetInstanceProcAddr(
 825     VkInstance                                  instance,
 826     const char*                                 pName)
 827 {
 828    return anv_lookup_entrypoint(NULL, pName);
 829 }
 830
 831 /* With version 1+ of the loader interface the ICD should expose
 832  * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
 833  */
 834 PUBLIC
 835 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 836     VkInstance                                  instance,
 837     const char*                                 pName);
 838
 839 PUBLIC
 840 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 841     VkInstance                                  instance,
 842     const char*                                 pName)
 843 {
 844    return anv_GetInstanceProcAddr(instance, pName);
 845 }
 846
 847 PFN_vkVoidFunction anv_GetDeviceProcAddr(
 848     VkDevice                                    _device,
 849     const char*                                 pName)
 850 {
 851    ANV_FROM_HANDLE(anv_device, device, _device);
 852    return anv_lookup_entrypoint(&device->info, pName);
 853 }
 854
 855 static void
 856 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
 857 {
 858    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 859    queue->device = device;
 860    queue->pool = &device->surface_state_pool;
 861 }
 862
 863 static void
 864 anv_queue_finish(struct anv_queue *queue)
 865 {
 866 }
 867
 868 static struct anv_state
 869 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
 870 {
 871    struct anv_state state;
 872
 873    state = anv_state_pool_alloc(pool, size, align);
 874    memcpy(state.map, p, size);
 875
 876    if (!pool->block_pool->device->info.has_llc)
 877       anv_state_clflush(state);
 878
 879    return state;
 880 }
 881
 882 struct gen8_border_color {
 883    union {
 884       float float32[4];
 885       uint32_t uint32[4];
 886    };
 887    /* Pad out to 64 bytes */
 888    uint32_t _pad[12];
 889 };
 890
 891 static void
 892 anv_device_init_border_colors(struct anv_device *device)
 893 {
 894    static const struct gen8_border_color border_colors[] = {
 895       [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] =  { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
 896       [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] =       { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
 897       [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] =       { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
 898       [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] =    { .uint32 = { 0, 0, 0, 0 } },
 899       [VK_BORDER_COLOR_INT_OPAQUE_BLACK] =         { .uint32 = { 0, 0, 0, 1 } },
 900       [VK_BORDER_COLOR_INT_OPAQUE_WHITE] =         { .uint32 = { 1, 1, 1, 1 } },
 901    };
 902
 903    device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
 904                                                     sizeof(border_colors), 64,
 905                                                     border_colors);
 906 }
 907
 908 VkResult
 909 anv_device_submit_simple_batch(struct anv_device *device,
 910                                struct anv_batch *batch)
 911 {
 912    struct drm_i915_gem_execbuffer2 execbuf;
 913    struct drm_i915_gem_exec_object2 exec2_objects[1];
 914    struct anv_bo bo, *exec_bos[1];
 915    VkResult result = VK_SUCCESS;
 916    uint32_t size;
 917    int64_t timeout;
 918    int ret;
 919
 920    /* Kernel driver requires 8 byte aligned batch length */
 921    size = align_u32(batch->next - batch->start, 8);
 922    result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
 923    if (result != VK_SUCCESS)
 924       return result;
 925
 926    memcpy(bo.map, batch->start, size);
 927    if (!device->info.has_llc)
 928       anv_clflush_range(bo.map, size);
 929
 930    exec_bos[0] = &bo;
 931    exec2_objects[0].handle = bo.gem_handle;
 932    exec2_objects[0].relocation_count = 0;
 933    exec2_objects[0].relocs_ptr = 0;
 934    exec2_objects[0].alignment = 0;
 935    exec2_objects[0].offset = bo.offset;
 936    exec2_objects[0].flags = 0;
 937    exec2_objects[0].rsvd1 = 0;
 938    exec2_objects[0].rsvd2 = 0;
 939
 940    execbuf.buffers_ptr = (uintptr_t) exec2_objects;
 941    execbuf.buffer_count = 1;
 942    execbuf.batch_start_offset = 0;
 943    execbuf.batch_len = size;
 944    execbuf.cliprects_ptr = 0;
 945    execbuf.num_cliprects = 0;
 946    execbuf.DR1 = 0;
 947    execbuf.DR4 = 0;
 948
 949    execbuf.flags =
 950       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
 951    execbuf.rsvd1 = device->context_id;
 952    execbuf.rsvd2 = 0;
 953
 954    result = anv_device_execbuf(device, &execbuf, exec_bos);
 955    if (result != VK_SUCCESS)
 956       goto fail;
 957
 958    timeout = INT64_MAX;
 959    ret = anv_gem_wait(device, bo.gem_handle, &timeout);
 960    if (ret != 0) {
 961       /* We don't know the real error. */
 962       result = vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
 963       goto fail;
 964    }
 965
 966  fail:
 967    anv_bo_pool_free(&device->batch_bo_pool, &bo);
 968
 969    return result;
 970 }
 971
 972 VkResult anv_CreateDevice(
 973     VkPhysicalDevice                            physicalDevice,
 974     const VkDeviceCreateInfo*                   pCreateInfo,
 975     const VkAllocationCallbacks*                pAllocator,
 976     VkDevice*                                   pDevice)
 977 {
 978    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 979    VkResult result;
 980    struct anv_device *device;
 981
 982    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
 983
 984    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 985       bool found = false;
 986       for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
 987          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 988                     device_extensions[j].extensionName) == 0) {
 989             found = true;
 990             break;
 991          }
 992       }
 993       if (!found)
 994          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 995    }
 996
 997    device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
 998                        sizeof(*device), 8,
 999                        VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1000    if (!device)
1001       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1002
1003    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1004    device->instance = physical_device->instance;
1005    device->chipset_id = physical_device->chipset_id;
1006
1007    if (pAllocator)
1008       device->alloc = *pAllocator;
1009    else
1010       device->alloc = physical_device->instance->alloc;
1011
1012    /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1013    device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1014    if (device->fd == -1) {
1015       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1016       goto fail_device;
1017    }
1018
1019    device->context_id = anv_gem_create_context(device);
1020    if (device->context_id == -1) {
1021       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1022       goto fail_fd;
1023    }
1024
1025    device->info = physical_device->info;
1026    device->isl_dev = physical_device->isl_dev;
1027
1028    /* On Broadwell and later, we can use batch chaining to more efficiently
1029     * implement growing command buffers.  Prior to Haswell, the kernel
1030     * command parser gets in the way and we have to fall back to growing
1031     * the batch.
1032     */
1033    device->can_chain_batches = device->info.gen >= 8;
1034
1035    device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1036       pCreateInfo->pEnabledFeatures->robustBufferAccess;
1037
1038    pthread_mutex_init(&device->mutex, NULL);
1039
1040    pthread_condattr_t condattr;
1041    pthread_condattr_init(&condattr);
1042    pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC);
1043    pthread_cond_init(&device->queue_submit, NULL);
1044    pthread_condattr_destroy(&condattr);
1045
1046    anv_bo_pool_init(&device->batch_bo_pool, device);
1047
1048    anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
1049
1050    anv_state_pool_init(&device->dynamic_state_pool,
1051                        &device->dynamic_state_block_pool);
1052
1053    anv_block_pool_init(&device->instruction_block_pool, device, 1024 * 1024);
1054    anv_state_pool_init(&device->instruction_state_pool,
1055                        &device->instruction_block_pool);
1056
1057    anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
1058
1059    anv_state_pool_init(&device->surface_state_pool,
1060                        &device->surface_state_block_pool);
1061
1062    anv_bo_init_new(&device->workaround_bo, device, 1024);
1063
1064    anv_scratch_pool_init(device, &device->scratch_pool);
1065
1066    anv_queue_init(device, &device->queue);
1067
1068    switch (device->info.gen) {
1069    case 7:
1070       if (!device->info.is_haswell)
1071          result = gen7_init_device_state(device);
1072       else
1073          result = gen75_init_device_state(device);
1074       break;
1075    case 8:
1076       result = gen8_init_device_state(device);
1077       break;
1078    case 9:
1079       result = gen9_init_device_state(device);
1080       break;
1081    default:
1082       /* Shouldn't get here as we don't create physical devices for any other
1083        * gens. */
1084       unreachable("unhandled gen");
1085    }
1086    if (result != VK_SUCCESS)
1087       goto fail_fd;
1088
1089    anv_device_init_blorp(device);
1090
1091    anv_device_init_border_colors(device);
1092
1093    *pDevice = anv_device_to_handle(device);
1094
1095    return VK_SUCCESS;
1096
1097  fail_fd:
1098    close(device->fd);
1099  fail_device:
1100    vk_free(&device->alloc, device);
1101
1102    return result;
1103 }
1104
1105 void anv_DestroyDevice(
1106     VkDevice                                    _device,
1107     const VkAllocationCallbacks*                pAllocator)
1108 {
1109    ANV_FROM_HANDLE(anv_device, device, _device);
1110
1111    anv_device_finish_blorp(device);
1112
1113    anv_queue_finish(&device->queue);
1114
1115 #ifdef HAVE_VALGRIND
1116    /* We only need to free these to prevent valgrind errors.  The backing
1117     * BO will go away in a couple of lines so we don't actually leak.
1118     */
1119    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1120 #endif
1121
1122    anv_scratch_pool_finish(device, &device->scratch_pool);
1123
1124    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1125    anv_gem_close(device, device->workaround_bo.gem_handle);
1126
1127    anv_state_pool_finish(&device->surface_state_pool);
1128    anv_block_pool_finish(&device->surface_state_block_pool);
1129    anv_state_pool_finish(&device->instruction_state_pool);
1130    anv_block_pool_finish(&device->instruction_block_pool);
1131    anv_state_pool_finish(&device->dynamic_state_pool);
1132    anv_block_pool_finish(&device->dynamic_state_block_pool);
1133
1134    anv_bo_pool_finish(&device->batch_bo_pool);
1135
1136    pthread_cond_destroy(&device->queue_submit);
1137    pthread_mutex_destroy(&device->mutex);
1138
1139    anv_gem_destroy_context(device, device->context_id);
1140
1141    close(device->fd);
1142
1143    vk_free(&device->alloc, device);
1144 }
1145
1146 VkResult anv_EnumerateInstanceExtensionProperties(
1147     const char*                                 pLayerName,
1148     uint32_t*                                   pPropertyCount,
1149     VkExtensionProperties*                      pProperties)
1150 {
1151    if (pProperties == NULL) {
1152       *pPropertyCount = ARRAY_SIZE(global_extensions);
1153       return VK_SUCCESS;
1154    }
1155
1156    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(global_extensions));
1157    typed_memcpy(pProperties, global_extensions, *pPropertyCount);
1158
1159    if (*pPropertyCount < ARRAY_SIZE(global_extensions))
1160       return VK_INCOMPLETE;
1161
1162    return VK_SUCCESS;
1163 }
1164
1165 VkResult anv_EnumerateDeviceExtensionProperties(
1166     VkPhysicalDevice                            physicalDevice,
1167     const char*                                 pLayerName,
1168     uint32_t*                                   pPropertyCount,
1169     VkExtensionProperties*                      pProperties)
1170 {
1171    if (pProperties == NULL) {
1172       *pPropertyCount = ARRAY_SIZE(device_extensions);
1173       return VK_SUCCESS;
1174    }
1175
1176    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(device_extensions));
1177    typed_memcpy(pProperties, device_extensions, *pPropertyCount);
1178
1179    if (*pPropertyCount < ARRAY_SIZE(device_extensions))
1180       return VK_INCOMPLETE;
1181
1182    return VK_SUCCESS;
1183 }
1184
1185 VkResult anv_EnumerateInstanceLayerProperties(
1186     uint32_t*                                   pPropertyCount,
1187     VkLayerProperties*                          pProperties)
1188 {
1189    if (pProperties == NULL) {
1190       *pPropertyCount = 0;
1191       return VK_SUCCESS;
1192    }
1193
1194    /* None supported at this time */
1195    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1196 }
1197
1198 VkResult anv_EnumerateDeviceLayerProperties(
1199     VkPhysicalDevice                            physicalDevice,
1200     uint32_t*                                   pPropertyCount,
1201     VkLayerProperties*                          pProperties)
1202 {
1203    if (pProperties == NULL) {
1204       *pPropertyCount = 0;
1205       return VK_SUCCESS;
1206    }
1207
1208    /* None supported at this time */
1209    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1210 }
1211
1212 void anv_GetDeviceQueue(
1213     VkDevice                                    _device,
1214     uint32_t                                    queueNodeIndex,
1215     uint32_t                                    queueIndex,
1216     VkQueue*                                    pQueue)
1217 {
1218    ANV_FROM_HANDLE(anv_device, device, _device);
1219
1220    assert(queueIndex == 0);
1221
1222    *pQueue = anv_queue_to_handle(&device->queue);
1223 }
1224
1225 VkResult
1226 anv_device_execbuf(struct anv_device *device,
1227                    struct drm_i915_gem_execbuffer2 *execbuf,
1228                    struct anv_bo **execbuf_bos)
1229 {
1230    int ret = anv_gem_execbuffer(device, execbuf);
1231    if (ret != 0) {
1232       /* We don't know the real error. */
1233       return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
1234    }
1235
1236    struct drm_i915_gem_exec_object2 *objects =
1237       (void *)(uintptr_t)execbuf->buffers_ptr;
1238    for (uint32_t k = 0; k < execbuf->buffer_count; k++)
1239       execbuf_bos[k]->offset = objects[k].offset;
1240
1241    return VK_SUCCESS;
1242 }
1243
1244 VkResult anv_QueueSubmit(
1245     VkQueue                                     _queue,
1246     uint32_t                                    submitCount,
1247     const VkSubmitInfo*                         pSubmits,
1248     VkFence                                     _fence)
1249 {
1250    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1251    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1252    struct anv_device *device = queue->device;
1253    VkResult result = VK_SUCCESS;
1254
1255    /* We lock around QueueSubmit for three main reasons:
1256     *
1257     *  1) When a block pool is resized, we create a new gem handle with a
1258     *     different size and, in the case of surface states, possibly a
1259     *     different center offset but we re-use the same anv_bo struct when
1260     *     we do so.  If this happens in the middle of setting up an execbuf,
1261     *     we could end up with our list of BOs out of sync with our list of
1262     *     gem handles.
1263     *
1264     *  2) The algorithm we use for building the list of unique buffers isn't
1265     *     thread-safe.  While the client is supposed to syncronize around
1266     *     QueueSubmit, this would be extremely difficult to debug if it ever
1267     *     came up in the wild due to a broken app.  It's better to play it
1268     *     safe and just lock around QueueSubmit.
1269     *
1270     *  3)  The anv_cmd_buffer_execbuf function may perform relocations in
1271     *      userspace.  Due to the fact that the surface state buffer is shared
1272     *      between batches, we can't afford to have that happen from multiple
1273     *      threads at the same time.  Even though the user is supposed to
1274     *      ensure this doesn't happen, we play it safe as in (2) above.
1275     *
1276     * Since the only other things that ever take the device lock such as block
1277     * pool resize only rarely happen, this will almost never be contended so
1278     * taking a lock isn't really an expensive operation in this case.
1279     */
1280    pthread_mutex_lock(&device->mutex);
1281
1282    for (uint32_t i = 0; i < submitCount; i++) {
1283       for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1284          ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1285                          pSubmits[i].pCommandBuffers[j]);
1286          assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1287
1288          result = anv_cmd_buffer_execbuf(device, cmd_buffer);
1289          if (result != VK_SUCCESS)
1290             goto out;
1291       }
1292    }
1293
1294    if (fence) {
1295       struct anv_bo *fence_bo = &fence->bo;
1296       result = anv_device_execbuf(device, &fence->execbuf, &fence_bo);
1297       if (result != VK_SUCCESS)
1298          goto out;
1299
1300       /* Update the fence and wake up any waiters */
1301       assert(fence->state == ANV_FENCE_STATE_RESET);
1302       fence->state = ANV_FENCE_STATE_SUBMITTED;
1303       pthread_cond_broadcast(&device->queue_submit);
1304    }
1305
1306 out:
1307    pthread_mutex_unlock(&device->mutex);
1308
1309    return result;
1310 }
1311
1312 VkResult anv_QueueWaitIdle(
1313     VkQueue                                     _queue)
1314 {
1315    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1316
1317    return anv_DeviceWaitIdle(anv_device_to_handle(queue->device));
1318 }
1319
1320 VkResult anv_DeviceWaitIdle(
1321     VkDevice                                    _device)
1322 {
1323    ANV_FROM_HANDLE(anv_device, device, _device);
1324    struct anv_batch batch;
1325
1326    uint32_t cmds[8];
1327    batch.start = batch.next = cmds;
1328    batch.end = (void *) cmds + sizeof(cmds);
1329
1330    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1331    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1332
1333    return anv_device_submit_simple_batch(device, &batch);
1334 }
1335
1336 VkResult
1337 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1338 {
1339    uint32_t gem_handle = anv_gem_create(device, size);
1340    if (!gem_handle)
1341       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1342
1343    anv_bo_init(bo, gem_handle, size);
1344
1345    return VK_SUCCESS;
1346 }
1347
1348 VkResult anv_AllocateMemory(
1349     VkDevice                                    _device,
1350     const VkMemoryAllocateInfo*                 pAllocateInfo,
1351     const VkAllocationCallbacks*                pAllocator,
1352     VkDeviceMemory*                             pMem)
1353 {
1354    ANV_FROM_HANDLE(anv_device, device, _device);
1355    struct anv_device_memory *mem;
1356    VkResult result;
1357
1358    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1359
1360    /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1361    assert(pAllocateInfo->allocationSize > 0);
1362
1363    /* We support exactly one memory heap. */
1364    assert(pAllocateInfo->memoryTypeIndex == 0 ||
1365           (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1366
1367    /* FINISHME: Fail if allocation request exceeds heap size. */
1368
1369    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1370                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1371    if (mem == NULL)
1372       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1373
1374    /* The kernel is going to give us whole pages anyway */
1375    uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1376
1377    result = anv_bo_init_new(&mem->bo, device, alloc_size);
1378    if (result != VK_SUCCESS)
1379       goto fail;
1380
1381    mem->type_index = pAllocateInfo->memoryTypeIndex;
1382
1383    mem->map = NULL;
1384    mem->map_size = 0;
1385
1386    *pMem = anv_device_memory_to_handle(mem);
1387
1388    return VK_SUCCESS;
1389
1390  fail:
1391    vk_free2(&device->alloc, pAllocator, mem);
1392
1393    return result;
1394 }
1395
1396 void anv_FreeMemory(
1397     VkDevice                                    _device,
1398     VkDeviceMemory                              _mem,
1399     const VkAllocationCallbacks*                pAllocator)
1400 {
1401    ANV_FROM_HANDLE(anv_device, device, _device);
1402    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1403
1404    if (mem == NULL)
1405       return;
1406
1407    if (mem->map)
1408       anv_UnmapMemory(_device, _mem);
1409
1410    if (mem->bo.map)
1411       anv_gem_munmap(mem->bo.map, mem->bo.size);
1412
1413    if (mem->bo.gem_handle != 0)
1414       anv_gem_close(device, mem->bo.gem_handle);
1415
1416    vk_free2(&device->alloc, pAllocator, mem);
1417 }
1418
1419 VkResult anv_MapMemory(
1420     VkDevice                                    _device,
1421     VkDeviceMemory                              _memory,
1422     VkDeviceSize                                offset,
1423     VkDeviceSize                                size,
1424     VkMemoryMapFlags                            flags,
1425     void**                                      ppData)
1426 {
1427    ANV_FROM_HANDLE(anv_device, device, _device);
1428    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1429
1430    if (mem == NULL) {
1431       *ppData = NULL;
1432       return VK_SUCCESS;
1433    }
1434
1435    if (size == VK_WHOLE_SIZE)
1436       size = mem->bo.size - offset;
1437
1438    /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1439     *
1440     *  * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1441     *    assert(size != 0);
1442     *  * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1443     *    equal to the size of the memory minus offset
1444     */
1445    assert(size > 0);
1446    assert(offset + size <= mem->bo.size);
1447
1448    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1449     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1450     * at a time is valid. We could just mmap up front and return an offset
1451     * pointer here, but that may exhaust virtual memory on 32 bit
1452     * userspace. */
1453
1454    uint32_t gem_flags = 0;
1455    if (!device->info.has_llc && mem->type_index == 0)
1456       gem_flags |= I915_MMAP_WC;
1457
1458    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1459    uint64_t map_offset = offset & ~4095ull;
1460    assert(offset >= map_offset);
1461    uint64_t map_size = (offset + size) - map_offset;
1462
1463    /* Let's map whole pages */
1464    map_size = align_u64(map_size, 4096);
1465
1466    void *map = anv_gem_mmap(device, mem->bo.gem_handle,
1467                             map_offset, map_size, gem_flags);
1468    if (map == MAP_FAILED)
1469       return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
1470
1471    mem->map = map;
1472    mem->map_size = map_size;
1473
1474    *ppData = mem->map + (offset - map_offset);
1475
1476    return VK_SUCCESS;
1477 }
1478
1479 void anv_UnmapMemory(
1480     VkDevice                                    _device,
1481     VkDeviceMemory                              _memory)
1482 {
1483    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1484
1485    if (mem == NULL)
1486       return;
1487
1488    anv_gem_munmap(mem->map, mem->map_size);
1489
1490    mem->map = NULL;
1491    mem->map_size = 0;
1492 }
1493
1494 static void
1495 clflush_mapped_ranges(struct anv_device         *device,
1496                       uint32_t                   count,
1497                       const VkMappedMemoryRange *ranges)
1498 {
1499    for (uint32_t i = 0; i < count; i++) {
1500       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1501       void *p = mem->map + (ranges[i].offset & ~CACHELINE_MASK);
1502       void *end;
1503
1504       if (ranges[i].offset + ranges[i].size > mem->map_size)
1505          end = mem->map + mem->map_size;
1506       else
1507          end = mem->map + ranges[i].offset + ranges[i].size;
1508
1509       while (p < end) {
1510          __builtin_ia32_clflush(p);
1511          p += CACHELINE_SIZE;
1512       }
1513    }
1514 }
1515
1516 VkResult anv_FlushMappedMemoryRanges(
1517     VkDevice                                    _device,
1518     uint32_t                                    memoryRangeCount,
1519     const VkMappedMemoryRange*                  pMemoryRanges)
1520 {
1521    ANV_FROM_HANDLE(anv_device, device, _device);
1522
1523    if (device->info.has_llc)
1524       return VK_SUCCESS;
1525
1526    /* Make sure the writes we're flushing have landed. */
1527    __builtin_ia32_mfence();
1528
1529    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1530
1531    return VK_SUCCESS;
1532 }
1533
1534 VkResult anv_InvalidateMappedMemoryRanges(
1535     VkDevice                                    _device,
1536     uint32_t                                    memoryRangeCount,
1537     const VkMappedMemoryRange*                  pMemoryRanges)
1538 {
1539    ANV_FROM_HANDLE(anv_device, device, _device);
1540
1541    if (device->info.has_llc)
1542       return VK_SUCCESS;
1543
1544    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1545
1546    /* Make sure no reads get moved up above the invalidate. */
1547    __builtin_ia32_mfence();
1548
1549    return VK_SUCCESS;
1550 }
1551
1552 void anv_GetBufferMemoryRequirements(
1553     VkDevice                                    device,
1554     VkBuffer                                    _buffer,
1555     VkMemoryRequirements*                       pMemoryRequirements)
1556 {
1557    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1558
1559    /* The Vulkan spec (git aaed022) says:
1560     *
1561     *    memoryTypeBits is a bitfield and contains one bit set for every
1562     *    supported memory type for the resource. The bit `1<<i` is set if and
1563     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1564     *    structure for the physical device is supported.
1565     *
1566     * We support exactly one memory type.
1567     */
1568    pMemoryRequirements->memoryTypeBits = 1;
1569
1570    pMemoryRequirements->size = buffer->size;
1571    pMemoryRequirements->alignment = 16;
1572 }
1573
1574 void anv_GetImageMemoryRequirements(
1575     VkDevice                                    device,
1576     VkImage                                     _image,
1577     VkMemoryRequirements*                       pMemoryRequirements)
1578 {
1579    ANV_FROM_HANDLE(anv_image, image, _image);
1580
1581    /* The Vulkan spec (git aaed022) says:
1582     *
1583     *    memoryTypeBits is a bitfield and contains one bit set for every
1584     *    supported memory type for the resource. The bit `1<<i` is set if and
1585     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1586     *    structure for the physical device is supported.
1587     *
1588     * We support exactly one memory type.
1589     */
1590    pMemoryRequirements->memoryTypeBits = 1;
1591
1592    pMemoryRequirements->size = image->size;
1593    pMemoryRequirements->alignment = image->alignment;
1594 }
1595
1596 void anv_GetImageSparseMemoryRequirements(
1597     VkDevice                                    device,
1598     VkImage                                     image,
1599     uint32_t*                                   pSparseMemoryRequirementCount,
1600     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1601 {
1602    stub();
1603 }
1604
1605 void anv_GetDeviceMemoryCommitment(
1606     VkDevice                                    device,
1607     VkDeviceMemory                              memory,
1608     VkDeviceSize*                               pCommittedMemoryInBytes)
1609 {
1610    *pCommittedMemoryInBytes = 0;
1611 }
1612
1613 VkResult anv_BindBufferMemory(
1614     VkDevice                                    device,
1615     VkBuffer                                    _buffer,
1616     VkDeviceMemory                              _memory,
1617     VkDeviceSize                                memoryOffset)
1618 {
1619    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1620    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1621
1622    if (mem) {
1623       buffer->bo = &mem->bo;
1624       buffer->offset = memoryOffset;
1625    } else {
1626       buffer->bo = NULL;
1627       buffer->offset = 0;
1628    }
1629
1630    return VK_SUCCESS;
1631 }
1632
1633 VkResult anv_QueueBindSparse(
1634     VkQueue                                     queue,
1635     uint32_t                                    bindInfoCount,
1636     const VkBindSparseInfo*                     pBindInfo,
1637     VkFence                                     fence)
1638 {
1639    stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1640 }
1641
1642 VkResult anv_CreateFence(
1643     VkDevice                                    _device,
1644     const VkFenceCreateInfo*                    pCreateInfo,
1645     const VkAllocationCallbacks*                pAllocator,
1646     VkFence*                                    pFence)
1647 {
1648    ANV_FROM_HANDLE(anv_device, device, _device);
1649    struct anv_bo fence_bo;
1650    struct anv_fence *fence;
1651    struct anv_batch batch;
1652    VkResult result;
1653
1654    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1655
1656    result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1657    if (result != VK_SUCCESS)
1658       return result;
1659
1660    /* Fences are small.  Just store the CPU data structure in the BO. */
1661    fence = fence_bo.map;
1662    fence->bo = fence_bo;
1663
1664    /* Place the batch after the CPU data but on its own cache line. */
1665    const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1666    batch.next = batch.start = fence->bo.map + batch_offset;
1667    batch.end = fence->bo.map + fence->bo.size;
1668    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1669    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1670
1671    if (!device->info.has_llc) {
1672       assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1673       assert(batch.next - batch.start <= CACHELINE_SIZE);
1674       __builtin_ia32_mfence();
1675       __builtin_ia32_clflush(batch.start);
1676    }
1677
1678    fence->exec2_objects[0].handle = fence->bo.gem_handle;
1679    fence->exec2_objects[0].relocation_count = 0;
1680    fence->exec2_objects[0].relocs_ptr = 0;
1681    fence->exec2_objects[0].alignment = 0;
1682    fence->exec2_objects[0].offset = fence->bo.offset;
1683    fence->exec2_objects[0].flags = 0;
1684    fence->exec2_objects[0].rsvd1 = 0;
1685    fence->exec2_objects[0].rsvd2 = 0;
1686
1687    fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1688    fence->execbuf.buffer_count = 1;
1689    fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1690    fence->execbuf.batch_len = batch.next - batch.start;
1691    fence->execbuf.cliprects_ptr = 0;
1692    fence->execbuf.num_cliprects = 0;
1693    fence->execbuf.DR1 = 0;
1694    fence->execbuf.DR4 = 0;
1695
1696    fence->execbuf.flags =
1697       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1698    fence->execbuf.rsvd1 = device->context_id;
1699    fence->execbuf.rsvd2 = 0;
1700
1701    if (pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT) {
1702       fence->state = ANV_FENCE_STATE_SIGNALED;
1703    } else {
1704       fence->state = ANV_FENCE_STATE_RESET;
1705    }
1706
1707    *pFence = anv_fence_to_handle(fence);
1708
1709    return VK_SUCCESS;
1710 }
1711
1712 void anv_DestroyFence(
1713     VkDevice                                    _device,
1714     VkFence                                     _fence,
1715     const VkAllocationCallbacks*                pAllocator)
1716 {
1717    ANV_FROM_HANDLE(anv_device, device, _device);
1718    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1719
1720    if (!fence)
1721       return;
1722
1723    assert(fence->bo.map == fence);
1724    anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1725 }
1726
1727 VkResult anv_ResetFences(
1728     VkDevice                                    _device,
1729     uint32_t                                    fenceCount,
1730     const VkFence*                              pFences)
1731 {
1732    for (uint32_t i = 0; i < fenceCount; i++) {
1733       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1734       fence->state = ANV_FENCE_STATE_RESET;
1735    }
1736
1737    return VK_SUCCESS;
1738 }
1739
1740 VkResult anv_GetFenceStatus(
1741     VkDevice                                    _device,
1742     VkFence                                     _fence)
1743 {
1744    ANV_FROM_HANDLE(anv_device, device, _device);
1745    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1746    int64_t t = 0;
1747    int ret;
1748
1749    switch (fence->state) {
1750    case ANV_FENCE_STATE_RESET:
1751       /* If it hasn't even been sent off to the GPU yet, it's not ready */
1752       return VK_NOT_READY;
1753
1754    case ANV_FENCE_STATE_SIGNALED:
1755       /* It's been signaled, return success */
1756       return VK_SUCCESS;
1757
1758    case ANV_FENCE_STATE_SUBMITTED:
1759       /* It's been submitted to the GPU but we don't know if it's done yet. */
1760       ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1761       if (ret == 0) {
1762          fence->state = ANV_FENCE_STATE_SIGNALED;
1763          return VK_SUCCESS;
1764       } else {
1765          return VK_NOT_READY;
1766       }
1767    default:
1768       unreachable("Invalid fence status");
1769    }
1770 }
1771
1772 #define NSEC_PER_SEC 1000000000
1773 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
1774
1775 VkResult anv_WaitForFences(
1776     VkDevice                                    _device,
1777     uint32_t                                    fenceCount,
1778     const VkFence*                              pFences,
1779     VkBool32                                    waitAll,
1780     uint64_t                                    _timeout)
1781 {
1782    ANV_FROM_HANDLE(anv_device, device, _device);
1783    int ret;
1784
1785    /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1786     * to block indefinitely timeouts <= 0.  Unfortunately, this was broken
1787     * for a couple of kernel releases.  Since there's no way to know
1788     * whether or not the kernel we're using is one of the broken ones, the
1789     * best we can do is to clamp the timeout to INT64_MAX.  This limits the
1790     * maximum timeout from 584 years to 292 years - likely not a big deal.
1791     */
1792    int64_t timeout = MIN2(_timeout, INT64_MAX);
1793
1794    uint32_t pending_fences = fenceCount;
1795    while (pending_fences) {
1796       pending_fences = 0;
1797       bool signaled_fences = false;
1798       for (uint32_t i = 0; i < fenceCount; i++) {
1799          ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1800          switch (fence->state) {
1801          case ANV_FENCE_STATE_RESET:
1802             /* This fence hasn't been submitted yet, we'll catch it the next
1803              * time around.  Yes, this may mean we dead-loop but, short of
1804              * lots of locking and a condition variable, there's not much that
1805              * we can do about that.
1806              */
1807             pending_fences++;
1808             continue;
1809
1810          case ANV_FENCE_STATE_SIGNALED:
1811             /* This fence is not pending.  If waitAll isn't set, we can return
1812              * early.  Otherwise, we have to keep going.
1813              */
1814             if (!waitAll)
1815                return VK_SUCCESS;
1816             continue;
1817
1818          case ANV_FENCE_STATE_SUBMITTED:
1819             /* These are the fences we really care about.  Go ahead and wait
1820              * on it until we hit a timeout.
1821              */
1822             ret = anv_gem_wait(device, fence->bo.gem_handle, &timeout);
1823             if (ret == -1 && errno == ETIME) {
1824                return VK_TIMEOUT;
1825             } else if (ret == -1) {
1826                /* We don't know the real error. */
1827                return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1828             } else {
1829                fence->state = ANV_FENCE_STATE_SIGNALED;
1830                signaled_fences = true;
1831                if (!waitAll)
1832                   return VK_SUCCESS;
1833                continue;
1834             }
1835          }
1836       }
1837
1838       if (pending_fences && !signaled_fences) {
1839          /* If we've hit this then someone decided to vkWaitForFences before
1840           * they've actually submitted any of them to a queue.  This is a
1841           * fairly pessimal case, so it's ok to lock here and use a standard
1842           * pthreads condition variable.
1843           */
1844          pthread_mutex_lock(&device->mutex);
1845
1846          /* It's possible that some of the fences have changed state since the
1847           * last time we checked.  Now that we have the lock, check for
1848           * pending fences again and don't wait if it's changed.
1849           */
1850          uint32_t now_pending_fences = 0;
1851          for (uint32_t i = 0; i < fenceCount; i++) {
1852             ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1853             if (fence->state == ANV_FENCE_STATE_RESET)
1854                now_pending_fences++;
1855          }
1856          assert(now_pending_fences <= pending_fences);
1857
1858          if (now_pending_fences == pending_fences) {
1859             struct timespec before;
1860             clock_gettime(CLOCK_MONOTONIC, &before);
1861
1862             uint32_t abs_nsec = before.tv_nsec + timeout % NSEC_PER_SEC;
1863             uint64_t abs_sec = before.tv_sec + (abs_nsec / NSEC_PER_SEC) +
1864                                (timeout / NSEC_PER_SEC);
1865             abs_nsec %= NSEC_PER_SEC;
1866
1867             /* Avoid roll-over in tv_sec on 32-bit systems if the user
1868              * provided timeout is UINT64_MAX
1869              */
1870             struct timespec abstime;
1871             abstime.tv_nsec = abs_nsec;
1872             abstime.tv_sec = MIN2(abs_sec, INT_TYPE_MAX(abstime.tv_sec));
1873
1874             ret = pthread_cond_timedwait(&device->queue_submit,
1875                                          &device->mutex, &abstime);
1876             assert(ret != EINVAL);
1877
1878             struct timespec after;
1879             clock_gettime(CLOCK_MONOTONIC, &after);
1880             uint64_t time_elapsed =
1881                ((uint64_t)after.tv_sec * NSEC_PER_SEC + after.tv_nsec) -
1882                ((uint64_t)before.tv_sec * NSEC_PER_SEC + before.tv_nsec);
1883
1884             if (time_elapsed >= timeout) {
1885                pthread_mutex_unlock(&device->mutex);
1886                return VK_TIMEOUT;
1887             }
1888
1889             timeout -= time_elapsed;
1890          }
1891
1892          pthread_mutex_unlock(&device->mutex);
1893       }
1894    }
1895
1896    return VK_SUCCESS;
1897 }
1898
1899 // Queue semaphore functions
1900
1901 VkResult anv_CreateSemaphore(
1902     VkDevice                                    device,
1903     const VkSemaphoreCreateInfo*                pCreateInfo,
1904     const VkAllocationCallbacks*                pAllocator,
1905     VkSemaphore*                                pSemaphore)
1906 {
1907    /* The DRM execbuffer ioctl always execute in-oder, even between different
1908     * rings. As such, there's nothing to do for the user space semaphore.
1909     */
1910
1911    *pSemaphore = (VkSemaphore)1;
1912
1913    return VK_SUCCESS;
1914 }
1915
1916 void anv_DestroySemaphore(
1917     VkDevice                                    device,
1918     VkSemaphore                                 semaphore,
1919     const VkAllocationCallbacks*                pAllocator)
1920 {
1921 }
1922
1923 // Event functions
1924
1925 VkResult anv_CreateEvent(
1926     VkDevice                                    _device,
1927     const VkEventCreateInfo*                    pCreateInfo,
1928     const VkAllocationCallbacks*                pAllocator,
1929     VkEvent*                                    pEvent)
1930 {
1931    ANV_FROM_HANDLE(anv_device, device, _device);
1932    struct anv_state state;
1933    struct anv_event *event;
1934
1935    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1936
1937    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1938                                 sizeof(*event), 8);
1939    event = state.map;
1940    event->state = state;
1941    event->semaphore = VK_EVENT_RESET;
1942
1943    if (!device->info.has_llc) {
1944       /* Make sure the writes we're flushing have landed. */
1945       __builtin_ia32_mfence();
1946       __builtin_ia32_clflush(event);
1947    }
1948
1949    *pEvent = anv_event_to_handle(event);
1950
1951    return VK_SUCCESS;
1952 }
1953
1954 void anv_DestroyEvent(
1955     VkDevice                                    _device,
1956     VkEvent                                     _event,
1957     const VkAllocationCallbacks*                pAllocator)
1958 {
1959    ANV_FROM_HANDLE(anv_device, device, _device);
1960    ANV_FROM_HANDLE(anv_event, event, _event);
1961
1962    if (!event)
1963       return;
1964
1965    anv_state_pool_free(&device->dynamic_state_pool, event->state);
1966 }
1967
1968 VkResult anv_GetEventStatus(
1969     VkDevice                                    _device,
1970     VkEvent                                     _event)
1971 {
1972    ANV_FROM_HANDLE(anv_device, device, _device);
1973    ANV_FROM_HANDLE(anv_event, event, _event);
1974
1975    if (!device->info.has_llc) {
1976       /* Invalidate read cache before reading event written by GPU. */
1977       __builtin_ia32_clflush(event);
1978       __builtin_ia32_mfence();
1979
1980    }
1981
1982    return event->semaphore;
1983 }
1984
1985 VkResult anv_SetEvent(
1986     VkDevice                                    _device,
1987     VkEvent                                     _event)
1988 {
1989    ANV_FROM_HANDLE(anv_device, device, _device);
1990    ANV_FROM_HANDLE(anv_event, event, _event);
1991
1992    event->semaphore = VK_EVENT_SET;
1993
1994    if (!device->info.has_llc) {
1995       /* Make sure the writes we're flushing have landed. */
1996       __builtin_ia32_mfence();
1997       __builtin_ia32_clflush(event);
1998    }
1999
2000    return VK_SUCCESS;
2001 }
2002
2003 VkResult anv_ResetEvent(
2004     VkDevice                                    _device,
2005     VkEvent                                     _event)
2006 {
2007    ANV_FROM_HANDLE(anv_device, device, _device);
2008    ANV_FROM_HANDLE(anv_event, event, _event);
2009
2010    event->semaphore = VK_EVENT_RESET;
2011
2012    if (!device->info.has_llc) {
2013       /* Make sure the writes we're flushing have landed. */
2014       __builtin_ia32_mfence();
2015       __builtin_ia32_clflush(event);
2016    }
2017
2018    return VK_SUCCESS;
2019 }
2020
2021 // Buffer functions
2022
2023 VkResult anv_CreateBuffer(
2024     VkDevice                                    _device,
2025     const VkBufferCreateInfo*                   pCreateInfo,
2026     const VkAllocationCallbacks*                pAllocator,
2027     VkBuffer*                                   pBuffer)
2028 {
2029    ANV_FROM_HANDLE(anv_device, device, _device);
2030    struct anv_buffer *buffer;
2031
2032    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2033
2034    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2035                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2036    if (buffer == NULL)
2037       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2038
2039    buffer->size = pCreateInfo->size;
2040    buffer->usage = pCreateInfo->usage;
2041    buffer->bo = NULL;
2042    buffer->offset = 0;
2043
2044    *pBuffer = anv_buffer_to_handle(buffer);
2045
2046    return VK_SUCCESS;
2047 }
2048
2049 void anv_DestroyBuffer(
2050     VkDevice                                    _device,
2051     VkBuffer                                    _buffer,
2052     const VkAllocationCallbacks*                pAllocator)
2053 {
2054    ANV_FROM_HANDLE(anv_device, device, _device);
2055    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2056
2057    if (!buffer)
2058       return;
2059
2060    vk_free2(&device->alloc, pAllocator, buffer);
2061 }
2062
2063 void
2064 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
2065                               enum isl_format format,
2066                               uint32_t offset, uint32_t range, uint32_t stride)
2067 {
2068    isl_buffer_fill_state(&device->isl_dev, state.map,
2069                          .address = offset,
2070                          .mocs = device->default_mocs,
2071                          .size = range,
2072                          .format = format,
2073                          .stride = stride);
2074
2075    if (!device->info.has_llc)
2076       anv_state_clflush(state);
2077 }
2078
2079 void anv_DestroySampler(
2080     VkDevice                                    _device,
2081     VkSampler                                   _sampler,
2082     const VkAllocationCallbacks*                pAllocator)
2083 {
2084    ANV_FROM_HANDLE(anv_device, device, _device);
2085    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
2086
2087    if (!sampler)
2088       return;
2089
2090    vk_free2(&device->alloc, pAllocator, sampler);
2091 }
2092
2093 VkResult anv_CreateFramebuffer(
2094     VkDevice                                    _device,
2095     const VkFramebufferCreateInfo*              pCreateInfo,
2096     const VkAllocationCallbacks*                pAllocator,
2097     VkFramebuffer*                              pFramebuffer)
2098 {
2099    ANV_FROM_HANDLE(anv_device, device, _device);
2100    struct anv_framebuffer *framebuffer;
2101
2102    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2103
2104    size_t size = sizeof(*framebuffer) +
2105                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2106    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2107                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2108    if (framebuffer == NULL)
2109       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2110
2111    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2112    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2113       VkImageView _iview = pCreateInfo->pAttachments[i];
2114       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2115    }
2116
2117    framebuffer->width = pCreateInfo->width;
2118    framebuffer->height = pCreateInfo->height;
2119    framebuffer->layers = pCreateInfo->layers;
2120
2121    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2122
2123    return VK_SUCCESS;
2124 }
2125
2126 void anv_DestroyFramebuffer(
2127     VkDevice                                    _device,
2128     VkFramebuffer                               _fb,
2129     const VkAllocationCallbacks*                pAllocator)
2130 {
2131    ANV_FROM_HANDLE(anv_device, device, _device);
2132    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2133
2134    if (!fb)
2135       return;
2136
2137    vk_free2(&device->alloc, pAllocator, fb);
2138 }
2139
2140 /* vk_icd.h does not declare this function, so we declare it here to
2141  * suppress Wmissing-prototypes.
2142  */
2143 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2144 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2145
2146 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2147 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2148 {
2149    /* For the full details on loader interface versioning, see
2150     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2151     * What follows is a condensed summary, to help you navigate the large and
2152     * confusing official doc.
2153     *
2154     *   - Loader interface v0 is incompatible with later versions. We don't
2155     *     support it.
2156     *
2157     *   - In loader interface v1:
2158     *       - The first ICD entrypoint called by the loader is
2159     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2160     *         entrypoint.
2161     *       - The ICD must statically expose no other Vulkan symbol unless it is
2162     *         linked with -Bsymbolic.
2163     *       - Each dispatchable Vulkan handle created by the ICD must be
2164     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2165     *         ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2166     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2167     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2168     *         such loader-managed surfaces.
2169     *
2170     *    - Loader interface v2 differs from v1 in:
2171     *       - The first ICD entrypoint called by the loader is
2172     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2173     *         statically expose this entrypoint.
2174     *
2175     *    - Loader interface v3 differs from v2 in:
2176     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2177     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2178     *          because the loader no longer does so.
2179     */
2180    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2181    return VK_SUCCESS;
2182 }