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 <unistd.h>
  28 #include <fcntl.h>
  29
  30 #include "anv_private.h"
  31 #include "anv_timestamp.h"
  32 #include "util/strtod.h"
  33 #include "util/debug.h"
  34
  35 #include "genxml/gen7_pack.h"
  36
  37 struct anv_dispatch_table dtable;
  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 VkResult
  56 anv_physical_device_init(struct anv_physical_device *device,
  57                          struct anv_instance *instance,
  58                          const char *path)
  59 {
  60    VkResult result;
  61    int fd;
  62
  63    fd = open(path, O_RDWR | O_CLOEXEC);
  64    if (fd < 0)
  65       return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  66
  67    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
  68    device->instance = instance;
  69
  70    assert(strlen(path) < ARRAY_SIZE(device->path));
  71    strncpy(device->path, path, ARRAY_SIZE(device->path));
  72
  73    device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
  74    if (!device->chipset_id) {
  75       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  76       goto fail;
  77    }
  78
  79    device->name = gen_get_device_name(device->chipset_id);
  80    if (!gen_get_device_info(device->chipset_id, &device->info)) {
  81       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  82       goto fail;
  83    }
  84
  85    if (device->info.is_haswell) {
  86       fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
  87    } else if (device->info.gen == 7 && !device->info.is_baytrail) {
  88       fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
  89    } else if (device->info.gen == 7 && device->info.is_baytrail) {
  90       fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
  91    } else if (device->info.gen >= 8) {
  92       /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
  93        * supported as anything */
  94    } else {
  95       result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
  96                          "Vulkan not yet supported on %s", device->name);
  97       goto fail;
  98    }
  99
 100    device->cmd_parser_version = -1;
 101    if (device->info.gen == 7) {
 102       device->cmd_parser_version =
 103          anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
 104       if (device->cmd_parser_version == -1) {
 105          result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 106                             "failed to get command parser version");
 107          goto fail;
 108       }
 109    }
 110
 111    if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
 112       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 113                          "failed to get aperture size: %m");
 114       goto fail;
 115    }
 116
 117    if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
 118       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 119                          "kernel missing gem wait");
 120       goto fail;
 121    }
 122
 123    if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
 124       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 125                          "kernel missing execbuf2");
 126       goto fail;
 127    }
 128
 129    if (!device->info.has_llc &&
 130        anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
 131       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 132                          "kernel missing wc mmap");
 133       goto fail;
 134    }
 135
 136    bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
 137
 138    /* GENs prior to 8 do not support EU/Subslice info */
 139    if (device->info.gen >= 8) {
 140       device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
 141       device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
 142
 143       /* Without this information, we cannot get the right Braswell
 144        * brandstrings, and we have to use conservative numbers for GPGPU on
 145        * many platforms, but otherwise, things will just work.
 146        */
 147       if (device->subslice_total < 1 || device->eu_total < 1) {
 148          fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
 149                          " query GPU properties.\n");
 150       }
 151    } else if (device->info.gen == 7) {
 152       device->subslice_total = 1 << (device->info.gt - 1);
 153    }
 154
 155    if (device->info.is_cherryview &&
 156        device->subslice_total > 0 && device->eu_total > 0) {
 157       /* Logical CS threads = EUs per subslice * 7 threads per EU */
 158       uint32_t max_cs_threads = device->eu_total / device->subslice_total * 7;
 159
 160       /* Fuse configurations may give more threads than expected, never less. */
 161       if (max_cs_threads > device->info.max_cs_threads)
 162          device->info.max_cs_threads = max_cs_threads;
 163    }
 164
 165    close(fd);
 166
 167    brw_process_intel_debug_variable();
 168
 169    device->compiler = brw_compiler_create(NULL, &device->info);
 170    if (device->compiler == NULL) {
 171       result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 172       goto fail;
 173    }
 174    device->compiler->shader_debug_log = compiler_debug_log;
 175    device->compiler->shader_perf_log = compiler_perf_log;
 176
 177    result = anv_init_wsi(device);
 178    if (result != VK_SUCCESS)
 179        goto fail;
 180
 181    /* XXX: Actually detect bit6 swizzling */
 182    isl_device_init(&device->isl_dev, &device->info, swizzled);
 183
 184    return VK_SUCCESS;
 185
 186 fail:
 187    close(fd);
 188    return result;
 189 }
 190
 191 static void
 192 anv_physical_device_finish(struct anv_physical_device *device)
 193 {
 194    anv_finish_wsi(device);
 195    ralloc_free(device->compiler);
 196 }
 197
 198 static const VkExtensionProperties global_extensions[] = {
 199    {
 200       .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
 201       .specVersion = 25,
 202    },
 203 #ifdef VK_USE_PLATFORM_XCB_KHR
 204    {
 205       .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
 206       .specVersion = 5,
 207    },
 208 #endif
 209 #ifdef VK_USE_PLATFORM_XLIB_KHR
 210    {
 211       .extensionName = VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
 212       .specVersion = 5,
 213    },
 214 #endif
 215 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
 216    {
 217       .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
 218       .specVersion = 4,
 219    },
 220 #endif
 221 };
 222
 223 static const VkExtensionProperties device_extensions[] = {
 224    {
 225       .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
 226       .specVersion = 67,
 227    },
 228 };
 229
 230 static void *
 231 default_alloc_func(void *pUserData, size_t size, size_t align,
 232                    VkSystemAllocationScope allocationScope)
 233 {
 234    return malloc(size);
 235 }
 236
 237 static void *
 238 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
 239                      size_t align, VkSystemAllocationScope allocationScope)
 240 {
 241    return realloc(pOriginal, size);
 242 }
 243
 244 static void
 245 default_free_func(void *pUserData, void *pMemory)
 246 {
 247    free(pMemory);
 248 }
 249
 250 static const VkAllocationCallbacks default_alloc = {
 251    .pUserData = NULL,
 252    .pfnAllocation = default_alloc_func,
 253    .pfnReallocation = default_realloc_func,
 254    .pfnFree = default_free_func,
 255 };
 256
 257 VkResult anv_CreateInstance(
 258     const VkInstanceCreateInfo*                 pCreateInfo,
 259     const VkAllocationCallbacks*                pAllocator,
 260     VkInstance*                                 pInstance)
 261 {
 262    struct anv_instance *instance;
 263
 264    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 265
 266    uint32_t client_version;
 267    if (pCreateInfo->pApplicationInfo &&
 268        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 269       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 270    } else {
 271       client_version = VK_MAKE_VERSION(1, 0, 0);
 272    }
 273
 274    if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
 275        client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
 276       return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 277                        "Client requested version %d.%d.%d",
 278                        VK_VERSION_MAJOR(client_version),
 279                        VK_VERSION_MINOR(client_version),
 280                        VK_VERSION_PATCH(client_version));
 281    }
 282
 283    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 284       bool found = false;
 285       for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
 286          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 287                     global_extensions[j].extensionName) == 0) {
 288             found = true;
 289             break;
 290          }
 291       }
 292       if (!found)
 293          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 294    }
 295
 296    instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 297                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 298    if (!instance)
 299       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 300
 301    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 302
 303    if (pAllocator)
 304       instance->alloc = *pAllocator;
 305    else
 306       instance->alloc = default_alloc;
 307
 308    instance->apiVersion = client_version;
 309    instance->physicalDeviceCount = -1;
 310
 311    _mesa_locale_init();
 312
 313    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 314
 315    *pInstance = anv_instance_to_handle(instance);
 316
 317    return VK_SUCCESS;
 318 }
 319
 320 void anv_DestroyInstance(
 321     VkInstance                                  _instance,
 322     const VkAllocationCallbacks*                pAllocator)
 323 {
 324    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 325
 326    if (instance->physicalDeviceCount > 0) {
 327       /* We support at most one physical device. */
 328       assert(instance->physicalDeviceCount == 1);
 329       anv_physical_device_finish(&instance->physicalDevice);
 330    }
 331
 332    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 333
 334    _mesa_locale_fini();
 335
 336    vk_free(&instance->alloc, instance);
 337 }
 338
 339 VkResult anv_EnumeratePhysicalDevices(
 340     VkInstance                                  _instance,
 341     uint32_t*                                   pPhysicalDeviceCount,
 342     VkPhysicalDevice*                           pPhysicalDevices)
 343 {
 344    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 345    VkResult result;
 346
 347    if (instance->physicalDeviceCount < 0) {
 348       char path[20];
 349       for (unsigned i = 0; i < 8; i++) {
 350          snprintf(path, sizeof(path), "/dev/dri/renderD%d", 128 + i);
 351          result = anv_physical_device_init(&instance->physicalDevice,
 352                                            instance, path);
 353          if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 354             break;
 355       }
 356
 357       if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
 358          instance->physicalDeviceCount = 0;
 359       } else if (result == VK_SUCCESS) {
 360          instance->physicalDeviceCount = 1;
 361       } else {
 362          return result;
 363       }
 364    }
 365
 366    /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
 367     * otherwise it's an inout parameter.
 368     *
 369     * The Vulkan spec (git aaed022) says:
 370     *
 371     *    pPhysicalDeviceCount is a pointer to an unsigned integer variable
 372     *    that is initialized with the number of devices the application is
 373     *    prepared to receive handles to. pname:pPhysicalDevices is pointer to
 374     *    an array of at least this many VkPhysicalDevice handles [...].
 375     *
 376     *    Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
 377     *    overwrites the contents of the variable pointed to by
 378     *    pPhysicalDeviceCount with the number of physical devices in in the
 379     *    instance; otherwise, vkEnumeratePhysicalDevices overwrites
 380     *    pPhysicalDeviceCount with the number of physical handles written to
 381     *    pPhysicalDevices.
 382     */
 383    if (!pPhysicalDevices) {
 384       *pPhysicalDeviceCount = instance->physicalDeviceCount;
 385    } else if (*pPhysicalDeviceCount >= 1) {
 386       pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
 387       *pPhysicalDeviceCount = 1;
 388    } else if (*pPhysicalDeviceCount < instance->physicalDeviceCount) {
 389       return VK_INCOMPLETE;
 390    } else {
 391       *pPhysicalDeviceCount = 0;
 392    }
 393
 394    return VK_SUCCESS;
 395 }
 396
 397 void anv_GetPhysicalDeviceFeatures(
 398     VkPhysicalDevice                            physicalDevice,
 399     VkPhysicalDeviceFeatures*                   pFeatures)
 400 {
 401    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 402
 403    *pFeatures = (VkPhysicalDeviceFeatures) {
 404       .robustBufferAccess                       = true,
 405       .fullDrawIndexUint32                      = true,
 406       .imageCubeArray                           = false,
 407       .independentBlend                         = true,
 408       .geometryShader                           = true,
 409       .tessellationShader                       = false,
 410       .sampleRateShading                        = true,
 411       .dualSrcBlend                             = true,
 412       .logicOp                                  = true,
 413       .multiDrawIndirect                        = false,
 414       .drawIndirectFirstInstance                = false,
 415       .depthClamp                               = true,
 416       .depthBiasClamp                           = false,
 417       .fillModeNonSolid                         = true,
 418       .depthBounds                              = false,
 419       .wideLines                                = true,
 420       .largePoints                              = true,
 421       .alphaToOne                               = true,
 422       .multiViewport                            = true,
 423       .samplerAnisotropy                        = true,
 424       .textureCompressionETC2                   = pdevice->info.gen >= 8 ||
 425                                                   pdevice->info.is_baytrail,
 426       .textureCompressionASTC_LDR               = pdevice->info.gen >= 9, /* FINISHME CHV */
 427       .textureCompressionBC                     = true,
 428       .occlusionQueryPrecise                    = true,
 429       .pipelineStatisticsQuery                  = false,
 430       .fragmentStoresAndAtomics                 = true,
 431       .shaderTessellationAndGeometryPointSize   = true,
 432       .shaderImageGatherExtended                = false,
 433       .shaderStorageImageExtendedFormats        = false,
 434       .shaderStorageImageMultisample            = false,
 435       .shaderUniformBufferArrayDynamicIndexing  = true,
 436       .shaderSampledImageArrayDynamicIndexing   = true,
 437       .shaderStorageBufferArrayDynamicIndexing  = true,
 438       .shaderStorageImageArrayDynamicIndexing   = true,
 439       .shaderStorageImageReadWithoutFormat      = false,
 440       .shaderStorageImageWriteWithoutFormat     = true,
 441       .shaderClipDistance                       = false,
 442       .shaderCullDistance                       = false,
 443       .shaderFloat64                            = false,
 444       .shaderInt64                              = false,
 445       .shaderInt16                              = false,
 446       .alphaToOne                               = true,
 447       .variableMultisampleRate                  = false,
 448       .inheritedQueries                         = false,
 449    };
 450
 451    /* We can't do image stores in vec4 shaders */
 452    pFeatures->vertexPipelineStoresAndAtomics =
 453       pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
 454       pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
 455 }
 456
 457 void
 458 anv_device_get_cache_uuid(void *uuid)
 459 {
 460    memset(uuid, 0, VK_UUID_SIZE);
 461    snprintf(uuid, VK_UUID_SIZE, "anv-%s", ANV_TIMESTAMP);
 462 }
 463
 464 void anv_GetPhysicalDeviceProperties(
 465     VkPhysicalDevice                            physicalDevice,
 466     VkPhysicalDeviceProperties*                 pProperties)
 467 {
 468    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 469    const struct gen_device_info *devinfo = &pdevice->info;
 470
 471    const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
 472
 473    /* See assertions made when programming the buffer surface state. */
 474    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
 475                                       (1ul << 30) : (1ul << 27);
 476
 477    VkSampleCountFlags sample_counts =
 478       isl_device_get_sample_counts(&pdevice->isl_dev);
 479
 480    VkPhysicalDeviceLimits limits = {
 481       .maxImageDimension1D                      = (1 << 14),
 482       .maxImageDimension2D                      = (1 << 14),
 483       .maxImageDimension3D                      = (1 << 11),
 484       .maxImageDimensionCube                    = (1 << 14),
 485       .maxImageArrayLayers                      = (1 << 11),
 486       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 487       .maxUniformBufferRange                    = (1ul << 27),
 488       .maxStorageBufferRange                    = max_raw_buffer_sz,
 489       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 490       .maxMemoryAllocationCount                 = UINT32_MAX,
 491       .maxSamplerAllocationCount                = 64 * 1024,
 492       .bufferImageGranularity                   = 64, /* A cache line */
 493       .sparseAddressSpaceSize                   = 0,
 494       .maxBoundDescriptorSets                   = MAX_SETS,
 495       .maxPerStageDescriptorSamplers            = 64,
 496       .maxPerStageDescriptorUniformBuffers      = 64,
 497       .maxPerStageDescriptorStorageBuffers      = 64,
 498       .maxPerStageDescriptorSampledImages       = 64,
 499       .maxPerStageDescriptorStorageImages       = 64,
 500       .maxPerStageDescriptorInputAttachments    = 64,
 501       .maxPerStageResources                     = 128,
 502       .maxDescriptorSetSamplers                 = 256,
 503       .maxDescriptorSetUniformBuffers           = 256,
 504       .maxDescriptorSetUniformBuffersDynamic    = 256,
 505       .maxDescriptorSetStorageBuffers           = 256,
 506       .maxDescriptorSetStorageBuffersDynamic    = 256,
 507       .maxDescriptorSetSampledImages            = 256,
 508       .maxDescriptorSetStorageImages            = 256,
 509       .maxDescriptorSetInputAttachments         = 256,
 510       .maxVertexInputAttributes                 = 32,
 511       .maxVertexInputBindings                   = 32,
 512       .maxVertexInputAttributeOffset            = 2047,
 513       .maxVertexInputBindingStride              = 2048,
 514       .maxVertexOutputComponents                = 128,
 515       .maxTessellationGenerationLevel           = 0,
 516       .maxTessellationPatchSize                 = 0,
 517       .maxTessellationControlPerVertexInputComponents = 0,
 518       .maxTessellationControlPerVertexOutputComponents = 0,
 519       .maxTessellationControlPerPatchOutputComponents = 0,
 520       .maxTessellationControlTotalOutputComponents = 0,
 521       .maxTessellationEvaluationInputComponents = 0,
 522       .maxTessellationEvaluationOutputComponents = 0,
 523       .maxGeometryShaderInvocations             = 32,
 524       .maxGeometryInputComponents               = 64,
 525       .maxGeometryOutputComponents              = 128,
 526       .maxGeometryOutputVertices                = 256,
 527       .maxGeometryTotalOutputComponents         = 1024,
 528       .maxFragmentInputComponents               = 128,
 529       .maxFragmentOutputAttachments             = 8,
 530       .maxFragmentDualSrcAttachments            = 2,
 531       .maxFragmentCombinedOutputResources       = 8,
 532       .maxComputeSharedMemorySize               = 32768,
 533       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 534       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 535       .maxComputeWorkGroupSize = {
 536          16 * devinfo->max_cs_threads,
 537          16 * devinfo->max_cs_threads,
 538          16 * devinfo->max_cs_threads,
 539       },
 540       .subPixelPrecisionBits                    = 4 /* FIXME */,
 541       .subTexelPrecisionBits                    = 4 /* FIXME */,
 542       .mipmapPrecisionBits                      = 4 /* FIXME */,
 543       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 544       .maxDrawIndirectCount                     = UINT32_MAX,
 545       .maxSamplerLodBias                        = 16,
 546       .maxSamplerAnisotropy                     = 16,
 547       .maxViewports                             = MAX_VIEWPORTS,
 548       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 549       .viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
 550       .viewportSubPixelBits                     = 13, /* We take a float? */
 551       .minMemoryMapAlignment                    = 4096, /* A page */
 552       .minTexelBufferOffsetAlignment            = 1,
 553       .minUniformBufferOffsetAlignment          = 1,
 554       .minStorageBufferOffsetAlignment          = 1,
 555       .minTexelOffset                           = -8,
 556       .maxTexelOffset                           = 7,
 557       .minTexelGatherOffset                     = -8,
 558       .maxTexelGatherOffset                     = 7,
 559       .minInterpolationOffset                   = -0.5,
 560       .maxInterpolationOffset                   = 0.4375,
 561       .subPixelInterpolationOffsetBits          = 4,
 562       .maxFramebufferWidth                      = (1 << 14),
 563       .maxFramebufferHeight                     = (1 << 14),
 564       .maxFramebufferLayers                     = (1 << 10),
 565       .framebufferColorSampleCounts             = sample_counts,
 566       .framebufferDepthSampleCounts             = sample_counts,
 567       .framebufferStencilSampleCounts           = sample_counts,
 568       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 569       .maxColorAttachments                      = MAX_RTS,
 570       .sampledImageColorSampleCounts            = sample_counts,
 571       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 572       .sampledImageDepthSampleCounts            = sample_counts,
 573       .sampledImageStencilSampleCounts          = sample_counts,
 574       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 575       .maxSampleMaskWords                       = 1,
 576       .timestampComputeAndGraphics              = false,
 577       .timestampPeriod                          = time_stamp_base,
 578       .maxClipDistances                         = 0 /* FIXME */,
 579       .maxCullDistances                         = 0 /* FIXME */,
 580       .maxCombinedClipAndCullDistances          = 0 /* FIXME */,
 581       .discreteQueuePriorities                  = 1,
 582       .pointSizeRange                           = { 0.125, 255.875 },
 583       .lineWidthRange                           = { 0.0, 7.9921875 },
 584       .pointSizeGranularity                     = (1.0 / 8.0),
 585       .lineWidthGranularity                     = (1.0 / 128.0),
 586       .strictLines                              = false, /* FINISHME */
 587       .standardSampleLocations                  = true,
 588       .optimalBufferCopyOffsetAlignment         = 128,
 589       .optimalBufferCopyRowPitchAlignment       = 128,
 590       .nonCoherentAtomSize                      = 64,
 591    };
 592
 593    *pProperties = (VkPhysicalDeviceProperties) {
 594       .apiVersion = VK_MAKE_VERSION(1, 0, 5),
 595       .driverVersion = 1,
 596       .vendorID = 0x8086,
 597       .deviceID = pdevice->chipset_id,
 598       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 599       .limits = limits,
 600       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 601    };
 602
 603    strcpy(pProperties->deviceName, pdevice->name);
 604    anv_device_get_cache_uuid(pProperties->pipelineCacheUUID);
 605 }
 606
 607 void anv_GetPhysicalDeviceQueueFamilyProperties(
 608     VkPhysicalDevice                            physicalDevice,
 609     uint32_t*                                   pCount,
 610     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 611 {
 612    if (pQueueFamilyProperties == NULL) {
 613       *pCount = 1;
 614       return;
 615    }
 616
 617    assert(*pCount >= 1);
 618
 619    *pQueueFamilyProperties = (VkQueueFamilyProperties) {
 620       .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 621                     VK_QUEUE_COMPUTE_BIT |
 622                     VK_QUEUE_TRANSFER_BIT,
 623       .queueCount = 1,
 624       .timestampValidBits = 36, /* XXX: Real value here */
 625       .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
 626    };
 627 }
 628
 629 void anv_GetPhysicalDeviceMemoryProperties(
 630     VkPhysicalDevice                            physicalDevice,
 631     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
 632 {
 633    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 634    VkDeviceSize heap_size;
 635
 636    /* Reserve some wiggle room for the driver by exposing only 75% of the
 637     * aperture to the heap.
 638     */
 639    heap_size = 3 * physical_device->aperture_size / 4;
 640
 641    if (physical_device->info.has_llc) {
 642       /* Big core GPUs share LLC with the CPU and thus one memory type can be
 643        * both cached and coherent at the same time.
 644        */
 645       pMemoryProperties->memoryTypeCount = 1;
 646       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 647          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 648                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 649                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
 650                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 651          .heapIndex = 0,
 652       };
 653    } else {
 654       /* The spec requires that we expose a host-visible, coherent memory
 655        * type, but Atom GPUs don't share LLC. Thus we offer two memory types
 656        * to give the application a choice between cached, but not coherent and
 657        * coherent but uncached (WC though).
 658        */
 659       pMemoryProperties->memoryTypeCount = 2;
 660       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 661          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 662                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 663                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
 664          .heapIndex = 0,
 665       };
 666       pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
 667          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 668                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 669                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 670          .heapIndex = 0,
 671       };
 672    }
 673
 674    pMemoryProperties->memoryHeapCount = 1;
 675    pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
 676       .size = heap_size,
 677       .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 678    };
 679 }
 680
 681 PFN_vkVoidFunction anv_GetInstanceProcAddr(
 682     VkInstance                                  instance,
 683     const char*                                 pName)
 684 {
 685    return anv_lookup_entrypoint(NULL, pName);
 686 }
 687
 688 /* With version 1+ of the loader interface the ICD should expose
 689  * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
 690  */
 691 PUBLIC
 692 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 693     VkInstance                                  instance,
 694     const char*                                 pName);
 695
 696 PUBLIC
 697 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 698     VkInstance                                  instance,
 699     const char*                                 pName)
 700 {
 701    return anv_GetInstanceProcAddr(instance, pName);
 702 }
 703
 704 PFN_vkVoidFunction anv_GetDeviceProcAddr(
 705     VkDevice                                    _device,
 706     const char*                                 pName)
 707 {
 708    ANV_FROM_HANDLE(anv_device, device, _device);
 709    return anv_lookup_entrypoint(&device->info, pName);
 710 }
 711
 712 static VkResult
 713 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
 714 {
 715    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 716    queue->device = device;
 717    queue->pool = &device->surface_state_pool;
 718
 719    return VK_SUCCESS;
 720 }
 721
 722 static void
 723 anv_queue_finish(struct anv_queue *queue)
 724 {
 725 }
 726
 727 static struct anv_state
 728 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
 729 {
 730    struct anv_state state;
 731
 732    state = anv_state_pool_alloc(pool, size, align);
 733    memcpy(state.map, p, size);
 734
 735    if (!pool->block_pool->device->info.has_llc)
 736       anv_state_clflush(state);
 737
 738    return state;
 739 }
 740
 741 struct gen8_border_color {
 742    union {
 743       float float32[4];
 744       uint32_t uint32[4];
 745    };
 746    /* Pad out to 64 bytes */
 747    uint32_t _pad[12];
 748 };
 749
 750 static void
 751 anv_device_init_border_colors(struct anv_device *device)
 752 {
 753    static const struct gen8_border_color border_colors[] = {
 754       [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] =  { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
 755       [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] =       { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
 756       [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] =       { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
 757       [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] =    { .uint32 = { 0, 0, 0, 0 } },
 758       [VK_BORDER_COLOR_INT_OPAQUE_BLACK] =         { .uint32 = { 0, 0, 0, 1 } },
 759       [VK_BORDER_COLOR_INT_OPAQUE_WHITE] =         { .uint32 = { 1, 1, 1, 1 } },
 760    };
 761
 762    device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
 763                                                     sizeof(border_colors), 64,
 764                                                     border_colors);
 765 }
 766
 767 VkResult
 768 anv_device_submit_simple_batch(struct anv_device *device,
 769                                struct anv_batch *batch)
 770 {
 771    struct drm_i915_gem_execbuffer2 execbuf;
 772    struct drm_i915_gem_exec_object2 exec2_objects[1];
 773    struct anv_bo bo, *exec_bos[1];
 774    VkResult result = VK_SUCCESS;
 775    uint32_t size;
 776    int64_t timeout;
 777    int ret;
 778
 779    /* Kernel driver requires 8 byte aligned batch length */
 780    size = align_u32(batch->next - batch->start, 8);
 781    result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
 782    if (result != VK_SUCCESS)
 783       return result;
 784
 785    memcpy(bo.map, batch->start, size);
 786    if (!device->info.has_llc)
 787       anv_clflush_range(bo.map, size);
 788
 789    exec_bos[0] = &bo;
 790    exec2_objects[0].handle = bo.gem_handle;
 791    exec2_objects[0].relocation_count = 0;
 792    exec2_objects[0].relocs_ptr = 0;
 793    exec2_objects[0].alignment = 0;
 794    exec2_objects[0].offset = bo.offset;
 795    exec2_objects[0].flags = 0;
 796    exec2_objects[0].rsvd1 = 0;
 797    exec2_objects[0].rsvd2 = 0;
 798
 799    execbuf.buffers_ptr = (uintptr_t) exec2_objects;
 800    execbuf.buffer_count = 1;
 801    execbuf.batch_start_offset = 0;
 802    execbuf.batch_len = size;
 803    execbuf.cliprects_ptr = 0;
 804    execbuf.num_cliprects = 0;
 805    execbuf.DR1 = 0;
 806    execbuf.DR4 = 0;
 807
 808    execbuf.flags =
 809       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
 810    execbuf.rsvd1 = device->context_id;
 811    execbuf.rsvd2 = 0;
 812
 813    result = anv_device_execbuf(device, &execbuf, exec_bos);
 814    if (result != VK_SUCCESS)
 815       goto fail;
 816
 817    timeout = INT64_MAX;
 818    ret = anv_gem_wait(device, bo.gem_handle, &timeout);
 819    if (ret != 0) {
 820       /* We don't know the real error. */
 821       result = vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
 822       goto fail;
 823    }
 824
 825  fail:
 826    anv_bo_pool_free(&device->batch_bo_pool, &bo);
 827
 828    return result;
 829 }
 830
 831 VkResult anv_CreateDevice(
 832     VkPhysicalDevice                            physicalDevice,
 833     const VkDeviceCreateInfo*                   pCreateInfo,
 834     const VkAllocationCallbacks*                pAllocator,
 835     VkDevice*                                   pDevice)
 836 {
 837    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 838    VkResult result;
 839    struct anv_device *device;
 840
 841    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
 842
 843    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 844       bool found = false;
 845       for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
 846          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 847                     device_extensions[j].extensionName) == 0) {
 848             found = true;
 849             break;
 850          }
 851       }
 852       if (!found)
 853          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 854    }
 855
 856    device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
 857                        sizeof(*device), 8,
 858                        VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
 859    if (!device)
 860       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 861
 862    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 863    device->instance = physical_device->instance;
 864    device->chipset_id = physical_device->chipset_id;
 865
 866    if (pAllocator)
 867       device->alloc = *pAllocator;
 868    else
 869       device->alloc = physical_device->instance->alloc;
 870
 871    /* XXX(chadv): Can we dup() physicalDevice->fd here? */
 872    device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
 873    if (device->fd == -1) {
 874       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
 875       goto fail_device;
 876    }
 877
 878    device->context_id = anv_gem_create_context(device);
 879    if (device->context_id == -1) {
 880       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
 881       goto fail_fd;
 882    }
 883
 884    device->info = physical_device->info;
 885    device->isl_dev = physical_device->isl_dev;
 886
 887    /* On Broadwell and later, we can use batch chaining to more efficiently
 888     * implement growing command buffers.  Prior to Haswell, the kernel
 889     * command parser gets in the way and we have to fall back to growing
 890     * the batch.
 891     */
 892    device->can_chain_batches = device->info.gen >= 8;
 893
 894    device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
 895       pCreateInfo->pEnabledFeatures->robustBufferAccess;
 896
 897    pthread_mutex_init(&device->mutex, NULL);
 898
 899    anv_bo_pool_init(&device->batch_bo_pool, device);
 900
 901    anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
 902
 903    anv_state_pool_init(&device->dynamic_state_pool,
 904                        &device->dynamic_state_block_pool);
 905
 906    anv_block_pool_init(&device->instruction_block_pool, device, 128 * 1024);
 907    anv_state_pool_init(&device->instruction_state_pool,
 908                        &device->instruction_block_pool);
 909
 910    anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
 911
 912    anv_state_pool_init(&device->surface_state_pool,
 913                        &device->surface_state_block_pool);
 914
 915    anv_bo_init_new(&device->workaround_bo, device, 1024);
 916
 917    anv_scratch_pool_init(device, &device->scratch_pool);
 918
 919    anv_queue_init(device, &device->queue);
 920
 921    switch (device->info.gen) {
 922    case 7:
 923       if (!device->info.is_haswell)
 924          result = gen7_init_device_state(device);
 925       else
 926          result = gen75_init_device_state(device);
 927       break;
 928    case 8:
 929       result = gen8_init_device_state(device);
 930       break;
 931    case 9:
 932       result = gen9_init_device_state(device);
 933       break;
 934    default:
 935       /* Shouldn't get here as we don't create physical devices for any other
 936        * gens. */
 937       unreachable("unhandled gen");
 938    }
 939    if (result != VK_SUCCESS)
 940       goto fail_fd;
 941
 942    anv_device_init_blorp(device);
 943
 944    anv_device_init_border_colors(device);
 945
 946    *pDevice = anv_device_to_handle(device);
 947
 948    return VK_SUCCESS;
 949
 950  fail_fd:
 951    close(device->fd);
 952  fail_device:
 953    vk_free(&device->alloc, device);
 954
 955    return result;
 956 }
 957
 958 void anv_DestroyDevice(
 959     VkDevice                                    _device,
 960     const VkAllocationCallbacks*                pAllocator)
 961 {
 962    ANV_FROM_HANDLE(anv_device, device, _device);
 963
 964    anv_queue_finish(&device->queue);
 965
 966    anv_device_finish_blorp(device);
 967
 968 #ifdef HAVE_VALGRIND
 969    /* We only need to free these to prevent valgrind errors.  The backing
 970     * BO will go away in a couple of lines so we don't actually leak.
 971     */
 972    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
 973 #endif
 974
 975    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
 976    anv_gem_close(device, device->workaround_bo.gem_handle);
 977
 978    anv_bo_pool_finish(&device->batch_bo_pool);
 979    anv_state_pool_finish(&device->dynamic_state_pool);
 980    anv_block_pool_finish(&device->dynamic_state_block_pool);
 981    anv_state_pool_finish(&device->instruction_state_pool);
 982    anv_block_pool_finish(&device->instruction_block_pool);
 983    anv_state_pool_finish(&device->surface_state_pool);
 984    anv_block_pool_finish(&device->surface_state_block_pool);
 985    anv_scratch_pool_finish(device, &device->scratch_pool);
 986
 987    close(device->fd);
 988
 989    pthread_mutex_destroy(&device->mutex);
 990
 991    vk_free(&device->alloc, device);
 992 }
 993
 994 VkResult anv_EnumerateInstanceExtensionProperties(
 995     const char*                                 pLayerName,
 996     uint32_t*                                   pPropertyCount,
 997     VkExtensionProperties*                      pProperties)
 998 {
 999    if (pProperties == NULL) {
1000       *pPropertyCount = ARRAY_SIZE(global_extensions);
1001       return VK_SUCCESS;
1002    }
1003
1004    assert(*pPropertyCount >= ARRAY_SIZE(global_extensions));
1005
1006    *pPropertyCount = ARRAY_SIZE(global_extensions);
1007    memcpy(pProperties, global_extensions, sizeof(global_extensions));
1008
1009    return VK_SUCCESS;
1010 }
1011
1012 VkResult anv_EnumerateDeviceExtensionProperties(
1013     VkPhysicalDevice                            physicalDevice,
1014     const char*                                 pLayerName,
1015     uint32_t*                                   pPropertyCount,
1016     VkExtensionProperties*                      pProperties)
1017 {
1018    if (pProperties == NULL) {
1019       *pPropertyCount = ARRAY_SIZE(device_extensions);
1020       return VK_SUCCESS;
1021    }
1022
1023    assert(*pPropertyCount >= ARRAY_SIZE(device_extensions));
1024
1025    *pPropertyCount = ARRAY_SIZE(device_extensions);
1026    memcpy(pProperties, device_extensions, sizeof(device_extensions));
1027
1028    return VK_SUCCESS;
1029 }
1030
1031 VkResult anv_EnumerateInstanceLayerProperties(
1032     uint32_t*                                   pPropertyCount,
1033     VkLayerProperties*                          pProperties)
1034 {
1035    if (pProperties == NULL) {
1036       *pPropertyCount = 0;
1037       return VK_SUCCESS;
1038    }
1039
1040    /* None supported at this time */
1041    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1042 }
1043
1044 VkResult anv_EnumerateDeviceLayerProperties(
1045     VkPhysicalDevice                            physicalDevice,
1046     uint32_t*                                   pPropertyCount,
1047     VkLayerProperties*                          pProperties)
1048 {
1049    if (pProperties == NULL) {
1050       *pPropertyCount = 0;
1051       return VK_SUCCESS;
1052    }
1053
1054    /* None supported at this time */
1055    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1056 }
1057
1058 void anv_GetDeviceQueue(
1059     VkDevice                                    _device,
1060     uint32_t                                    queueNodeIndex,
1061     uint32_t                                    queueIndex,
1062     VkQueue*                                    pQueue)
1063 {
1064    ANV_FROM_HANDLE(anv_device, device, _device);
1065
1066    assert(queueIndex == 0);
1067
1068    *pQueue = anv_queue_to_handle(&device->queue);
1069 }
1070
1071 VkResult
1072 anv_device_execbuf(struct anv_device *device,
1073                    struct drm_i915_gem_execbuffer2 *execbuf,
1074                    struct anv_bo **execbuf_bos)
1075 {
1076    int ret = anv_gem_execbuffer(device, execbuf);
1077    if (ret != 0) {
1078       /* We don't know the real error. */
1079       return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
1080    }
1081
1082    struct drm_i915_gem_exec_object2 *objects = (void *)execbuf->buffers_ptr;
1083    for (uint32_t k = 0; k < execbuf->buffer_count; k++)
1084       execbuf_bos[k]->offset = objects[k].offset;
1085
1086    return VK_SUCCESS;
1087 }
1088
1089 VkResult anv_QueueSubmit(
1090     VkQueue                                     _queue,
1091     uint32_t                                    submitCount,
1092     const VkSubmitInfo*                         pSubmits,
1093     VkFence                                     _fence)
1094 {
1095    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1096    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1097    struct anv_device *device = queue->device;
1098    VkResult result = VK_SUCCESS;
1099
1100    /* We lock around QueueSubmit for three main reasons:
1101     *
1102     *  1) When a block pool is resized, we create a new gem handle with a
1103     *     different size and, in the case of surface states, possibly a
1104     *     different center offset but we re-use the same anv_bo struct when
1105     *     we do so.  If this happens in the middle of setting up an execbuf,
1106     *     we could end up with our list of BOs out of sync with our list of
1107     *     gem handles.
1108     *
1109     *  2) The algorithm we use for building the list of unique buffers isn't
1110     *     thread-safe.  While the client is supposed to syncronize around
1111     *     QueueSubmit, this would be extremely difficult to debug if it ever
1112     *     came up in the wild due to a broken app.  It's better to play it
1113     *     safe and just lock around QueueSubmit.
1114     *
1115     *  3)  The anv_cmd_buffer_execbuf function may perform relocations in
1116     *      userspace.  Due to the fact that the surface state buffer is shared
1117     *      between batches, we can't afford to have that happen from multiple
1118     *      threads at the same time.  Even though the user is supposed to
1119     *      ensure this doesn't happen, we play it safe as in (2) above.
1120     *
1121     * Since the only other things that ever take the device lock such as block
1122     * pool resize only rarely happen, this will almost never be contended so
1123     * taking a lock isn't really an expensive operation in this case.
1124     */
1125    pthread_mutex_lock(&device->mutex);
1126
1127    for (uint32_t i = 0; i < submitCount; i++) {
1128       for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1129          ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1130                          pSubmits[i].pCommandBuffers[j]);
1131          assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1132
1133          result = anv_cmd_buffer_execbuf(device, cmd_buffer);
1134          if (result != VK_SUCCESS)
1135             goto out;
1136       }
1137    }
1138
1139    if (fence) {
1140       struct anv_bo *fence_bo = &fence->bo;
1141       result = anv_device_execbuf(device, &fence->execbuf, &fence_bo);
1142       if (result != VK_SUCCESS)
1143          goto out;
1144    }
1145
1146 out:
1147    pthread_mutex_unlock(&device->mutex);
1148
1149    return result;
1150 }
1151
1152 VkResult anv_QueueWaitIdle(
1153     VkQueue                                     _queue)
1154 {
1155    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1156
1157    return anv_DeviceWaitIdle(anv_device_to_handle(queue->device));
1158 }
1159
1160 VkResult anv_DeviceWaitIdle(
1161     VkDevice                                    _device)
1162 {
1163    ANV_FROM_HANDLE(anv_device, device, _device);
1164    struct anv_batch batch;
1165
1166    uint32_t cmds[8];
1167    batch.start = batch.next = cmds;
1168    batch.end = (void *) cmds + sizeof(cmds);
1169
1170    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1171    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1172
1173    return anv_device_submit_simple_batch(device, &batch);
1174 }
1175
1176 VkResult
1177 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1178 {
1179    uint32_t gem_handle = anv_gem_create(device, size);
1180    if (!gem_handle)
1181       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1182
1183    anv_bo_init(bo, gem_handle, size);
1184
1185    return VK_SUCCESS;
1186 }
1187
1188 VkResult anv_AllocateMemory(
1189     VkDevice                                    _device,
1190     const VkMemoryAllocateInfo*                 pAllocateInfo,
1191     const VkAllocationCallbacks*                pAllocator,
1192     VkDeviceMemory*                             pMem)
1193 {
1194    ANV_FROM_HANDLE(anv_device, device, _device);
1195    struct anv_device_memory *mem;
1196    VkResult result;
1197
1198    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1199
1200    if (pAllocateInfo->allocationSize == 0) {
1201       /* Apparently, this is allowed */
1202       *pMem = VK_NULL_HANDLE;
1203       return VK_SUCCESS;
1204    }
1205
1206    /* We support exactly one memory heap. */
1207    assert(pAllocateInfo->memoryTypeIndex == 0 ||
1208           (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1209
1210    /* FINISHME: Fail if allocation request exceeds heap size. */
1211
1212    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1213                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1214    if (mem == NULL)
1215       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1216
1217    /* The kernel is going to give us whole pages anyway */
1218    uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1219
1220    result = anv_bo_init_new(&mem->bo, device, alloc_size);
1221    if (result != VK_SUCCESS)
1222       goto fail;
1223
1224    mem->type_index = pAllocateInfo->memoryTypeIndex;
1225
1226    *pMem = anv_device_memory_to_handle(mem);
1227
1228    return VK_SUCCESS;
1229
1230  fail:
1231    vk_free2(&device->alloc, pAllocator, mem);
1232
1233    return result;
1234 }
1235
1236 void anv_FreeMemory(
1237     VkDevice                                    _device,
1238     VkDeviceMemory                              _mem,
1239     const VkAllocationCallbacks*                pAllocator)
1240 {
1241    ANV_FROM_HANDLE(anv_device, device, _device);
1242    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1243
1244    if (mem == NULL)
1245       return;
1246
1247    if (mem->bo.map)
1248       anv_gem_munmap(mem->bo.map, mem->bo.size);
1249
1250    if (mem->bo.gem_handle != 0)
1251       anv_gem_close(device, mem->bo.gem_handle);
1252
1253    vk_free2(&device->alloc, pAllocator, mem);
1254 }
1255
1256 VkResult anv_MapMemory(
1257     VkDevice                                    _device,
1258     VkDeviceMemory                              _memory,
1259     VkDeviceSize                                offset,
1260     VkDeviceSize                                size,
1261     VkMemoryMapFlags                            flags,
1262     void**                                      ppData)
1263 {
1264    ANV_FROM_HANDLE(anv_device, device, _device);
1265    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1266
1267    if (mem == NULL) {
1268       *ppData = NULL;
1269       return VK_SUCCESS;
1270    }
1271
1272    if (size == VK_WHOLE_SIZE)
1273       size = mem->bo.size - offset;
1274
1275    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1276     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1277     * at a time is valid. We could just mmap up front and return an offset
1278     * pointer here, but that may exhaust virtual memory on 32 bit
1279     * userspace. */
1280
1281    uint32_t gem_flags = 0;
1282    if (!device->info.has_llc && mem->type_index == 0)
1283       gem_flags |= I915_MMAP_WC;
1284
1285    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1286    uint64_t map_offset = offset & ~4095ull;
1287    assert(offset >= map_offset);
1288    uint64_t map_size = (offset + size) - map_offset;
1289
1290    /* Let's map whole pages */
1291    map_size = align_u64(map_size, 4096);
1292
1293    mem->map = anv_gem_mmap(device, mem->bo.gem_handle,
1294                            map_offset, map_size, gem_flags);
1295    mem->map_size = map_size;
1296
1297    *ppData = mem->map + (offset - map_offset);
1298
1299    return VK_SUCCESS;
1300 }
1301
1302 void anv_UnmapMemory(
1303     VkDevice                                    _device,
1304     VkDeviceMemory                              _memory)
1305 {
1306    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1307
1308    if (mem == NULL)
1309       return;
1310
1311    anv_gem_munmap(mem->map, mem->map_size);
1312 }
1313
1314 static void
1315 clflush_mapped_ranges(struct anv_device         *device,
1316                       uint32_t                   count,
1317                       const VkMappedMemoryRange *ranges)
1318 {
1319    for (uint32_t i = 0; i < count; i++) {
1320       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1321       void *p = mem->map + (ranges[i].offset & ~CACHELINE_MASK);
1322       void *end;
1323
1324       if (ranges[i].offset + ranges[i].size > mem->map_size)
1325          end = mem->map + mem->map_size;
1326       else
1327          end = mem->map + ranges[i].offset + ranges[i].size;
1328
1329       while (p < end) {
1330          __builtin_ia32_clflush(p);
1331          p += CACHELINE_SIZE;
1332       }
1333    }
1334 }
1335
1336 VkResult anv_FlushMappedMemoryRanges(
1337     VkDevice                                    _device,
1338     uint32_t                                    memoryRangeCount,
1339     const VkMappedMemoryRange*                  pMemoryRanges)
1340 {
1341    ANV_FROM_HANDLE(anv_device, device, _device);
1342
1343    if (device->info.has_llc)
1344       return VK_SUCCESS;
1345
1346    /* Make sure the writes we're flushing have landed. */
1347    __builtin_ia32_mfence();
1348
1349    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1350
1351    return VK_SUCCESS;
1352 }
1353
1354 VkResult anv_InvalidateMappedMemoryRanges(
1355     VkDevice                                    _device,
1356     uint32_t                                    memoryRangeCount,
1357     const VkMappedMemoryRange*                  pMemoryRanges)
1358 {
1359    ANV_FROM_HANDLE(anv_device, device, _device);
1360
1361    if (device->info.has_llc)
1362       return VK_SUCCESS;
1363
1364    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1365
1366    /* Make sure no reads get moved up above the invalidate. */
1367    __builtin_ia32_mfence();
1368
1369    return VK_SUCCESS;
1370 }
1371
1372 void anv_GetBufferMemoryRequirements(
1373     VkDevice                                    device,
1374     VkBuffer                                    _buffer,
1375     VkMemoryRequirements*                       pMemoryRequirements)
1376 {
1377    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1378
1379    /* The Vulkan spec (git aaed022) says:
1380     *
1381     *    memoryTypeBits is a bitfield and contains one bit set for every
1382     *    supported memory type for the resource. The bit `1<<i` is set if and
1383     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1384     *    structure for the physical device is supported.
1385     *
1386     * We support exactly one memory type.
1387     */
1388    pMemoryRequirements->memoryTypeBits = 1;
1389
1390    pMemoryRequirements->size = buffer->size;
1391    pMemoryRequirements->alignment = 16;
1392 }
1393
1394 void anv_GetImageMemoryRequirements(
1395     VkDevice                                    device,
1396     VkImage                                     _image,
1397     VkMemoryRequirements*                       pMemoryRequirements)
1398 {
1399    ANV_FROM_HANDLE(anv_image, image, _image);
1400
1401    /* The Vulkan spec (git aaed022) says:
1402     *
1403     *    memoryTypeBits is a bitfield and contains one bit set for every
1404     *    supported memory type for the resource. The bit `1<<i` is set if and
1405     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1406     *    structure for the physical device is supported.
1407     *
1408     * We support exactly one memory type.
1409     */
1410    pMemoryRequirements->memoryTypeBits = 1;
1411
1412    pMemoryRequirements->size = image->size;
1413    pMemoryRequirements->alignment = image->alignment;
1414 }
1415
1416 void anv_GetImageSparseMemoryRequirements(
1417     VkDevice                                    device,
1418     VkImage                                     image,
1419     uint32_t*                                   pSparseMemoryRequirementCount,
1420     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1421 {
1422    stub();
1423 }
1424
1425 void anv_GetDeviceMemoryCommitment(
1426     VkDevice                                    device,
1427     VkDeviceMemory                              memory,
1428     VkDeviceSize*                               pCommittedMemoryInBytes)
1429 {
1430    *pCommittedMemoryInBytes = 0;
1431 }
1432
1433 VkResult anv_BindBufferMemory(
1434     VkDevice                                    device,
1435     VkBuffer                                    _buffer,
1436     VkDeviceMemory                              _memory,
1437     VkDeviceSize                                memoryOffset)
1438 {
1439    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1440    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1441
1442    if (mem) {
1443       buffer->bo = &mem->bo;
1444       buffer->offset = memoryOffset;
1445    } else {
1446       buffer->bo = NULL;
1447       buffer->offset = 0;
1448    }
1449
1450    return VK_SUCCESS;
1451 }
1452
1453 VkResult anv_QueueBindSparse(
1454     VkQueue                                     queue,
1455     uint32_t                                    bindInfoCount,
1456     const VkBindSparseInfo*                     pBindInfo,
1457     VkFence                                     fence)
1458 {
1459    stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1460 }
1461
1462 VkResult anv_CreateFence(
1463     VkDevice                                    _device,
1464     const VkFenceCreateInfo*                    pCreateInfo,
1465     const VkAllocationCallbacks*                pAllocator,
1466     VkFence*                                    pFence)
1467 {
1468    ANV_FROM_HANDLE(anv_device, device, _device);
1469    struct anv_bo fence_bo;
1470    struct anv_fence *fence;
1471    struct anv_batch batch;
1472    VkResult result;
1473
1474    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1475
1476    result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1477    if (result != VK_SUCCESS)
1478       return result;
1479
1480    /* Fences are small.  Just store the CPU data structure in the BO. */
1481    fence = fence_bo.map;
1482    fence->bo = fence_bo;
1483
1484    /* Place the batch after the CPU data but on its own cache line. */
1485    const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1486    batch.next = batch.start = fence->bo.map + batch_offset;
1487    batch.end = fence->bo.map + fence->bo.size;
1488    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1489    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1490
1491    if (!device->info.has_llc) {
1492       assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1493       assert(batch.next - batch.start <= CACHELINE_SIZE);
1494       __builtin_ia32_mfence();
1495       __builtin_ia32_clflush(batch.start);
1496    }
1497
1498    fence->exec2_objects[0].handle = fence->bo.gem_handle;
1499    fence->exec2_objects[0].relocation_count = 0;
1500    fence->exec2_objects[0].relocs_ptr = 0;
1501    fence->exec2_objects[0].alignment = 0;
1502    fence->exec2_objects[0].offset = fence->bo.offset;
1503    fence->exec2_objects[0].flags = 0;
1504    fence->exec2_objects[0].rsvd1 = 0;
1505    fence->exec2_objects[0].rsvd2 = 0;
1506
1507    fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1508    fence->execbuf.buffer_count = 1;
1509    fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1510    fence->execbuf.batch_len = batch.next - batch.start;
1511    fence->execbuf.cliprects_ptr = 0;
1512    fence->execbuf.num_cliprects = 0;
1513    fence->execbuf.DR1 = 0;
1514    fence->execbuf.DR4 = 0;
1515
1516    fence->execbuf.flags =
1517       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1518    fence->execbuf.rsvd1 = device->context_id;
1519    fence->execbuf.rsvd2 = 0;
1520
1521    fence->ready = false;
1522
1523    *pFence = anv_fence_to_handle(fence);
1524
1525    return VK_SUCCESS;
1526 }
1527
1528 void anv_DestroyFence(
1529     VkDevice                                    _device,
1530     VkFence                                     _fence,
1531     const VkAllocationCallbacks*                pAllocator)
1532 {
1533    ANV_FROM_HANDLE(anv_device, device, _device);
1534    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1535
1536    assert(fence->bo.map == fence);
1537    anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1538 }
1539
1540 VkResult anv_ResetFences(
1541     VkDevice                                    _device,
1542     uint32_t                                    fenceCount,
1543     const VkFence*                              pFences)
1544 {
1545    for (uint32_t i = 0; i < fenceCount; i++) {
1546       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1547       fence->ready = false;
1548    }
1549
1550    return VK_SUCCESS;
1551 }
1552
1553 VkResult anv_GetFenceStatus(
1554     VkDevice                                    _device,
1555     VkFence                                     _fence)
1556 {
1557    ANV_FROM_HANDLE(anv_device, device, _device);
1558    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1559    int64_t t = 0;
1560    int ret;
1561
1562    if (fence->ready)
1563       return VK_SUCCESS;
1564
1565    ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1566    if (ret == 0) {
1567       fence->ready = true;
1568       return VK_SUCCESS;
1569    }
1570
1571    return VK_NOT_READY;
1572 }
1573
1574 VkResult anv_WaitForFences(
1575     VkDevice                                    _device,
1576     uint32_t                                    fenceCount,
1577     const VkFence*                              pFences,
1578     VkBool32                                    waitAll,
1579     uint64_t                                    timeout)
1580 {
1581    ANV_FROM_HANDLE(anv_device, device, _device);
1582
1583    /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1584     * to block indefinitely timeouts <= 0.  Unfortunately, this was broken
1585     * for a couple of kernel releases.  Since there's no way to know
1586     * whether or not the kernel we're using is one of the broken ones, the
1587     * best we can do is to clamp the timeout to INT64_MAX.  This limits the
1588     * maximum timeout from 584 years to 292 years - likely not a big deal.
1589     */
1590    if (timeout > INT64_MAX)
1591       timeout = INT64_MAX;
1592
1593    int64_t t = timeout;
1594
1595    /* FIXME: handle !waitAll */
1596
1597    for (uint32_t i = 0; i < fenceCount; i++) {
1598       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1599       int ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1600       if (ret == -1 && errno == ETIME) {
1601          return VK_TIMEOUT;
1602       } else if (ret == -1) {
1603          /* We don't know the real error. */
1604          return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1605                           "gem wait failed: %m");
1606       }
1607    }
1608
1609    return VK_SUCCESS;
1610 }
1611
1612 // Queue semaphore functions
1613
1614 VkResult anv_CreateSemaphore(
1615     VkDevice                                    device,
1616     const VkSemaphoreCreateInfo*                pCreateInfo,
1617     const VkAllocationCallbacks*                pAllocator,
1618     VkSemaphore*                                pSemaphore)
1619 {
1620    /* The DRM execbuffer ioctl always execute in-oder, even between different
1621     * rings. As such, there's nothing to do for the user space semaphore.
1622     */
1623
1624    *pSemaphore = (VkSemaphore)1;
1625
1626    return VK_SUCCESS;
1627 }
1628
1629 void anv_DestroySemaphore(
1630     VkDevice                                    device,
1631     VkSemaphore                                 semaphore,
1632     const VkAllocationCallbacks*                pAllocator)
1633 {
1634 }
1635
1636 // Event functions
1637
1638 VkResult anv_CreateEvent(
1639     VkDevice                                    _device,
1640     const VkEventCreateInfo*                    pCreateInfo,
1641     const VkAllocationCallbacks*                pAllocator,
1642     VkEvent*                                    pEvent)
1643 {
1644    ANV_FROM_HANDLE(anv_device, device, _device);
1645    struct anv_state state;
1646    struct anv_event *event;
1647
1648    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1649
1650    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1651                                 sizeof(*event), 8);
1652    event = state.map;
1653    event->state = state;
1654    event->semaphore = VK_EVENT_RESET;
1655
1656    if (!device->info.has_llc) {
1657       /* Make sure the writes we're flushing have landed. */
1658       __builtin_ia32_mfence();
1659       __builtin_ia32_clflush(event);
1660    }
1661
1662    *pEvent = anv_event_to_handle(event);
1663
1664    return VK_SUCCESS;
1665 }
1666
1667 void anv_DestroyEvent(
1668     VkDevice                                    _device,
1669     VkEvent                                     _event,
1670     const VkAllocationCallbacks*                pAllocator)
1671 {
1672    ANV_FROM_HANDLE(anv_device, device, _device);
1673    ANV_FROM_HANDLE(anv_event, event, _event);
1674
1675    anv_state_pool_free(&device->dynamic_state_pool, event->state);
1676 }
1677
1678 VkResult anv_GetEventStatus(
1679     VkDevice                                    _device,
1680     VkEvent                                     _event)
1681 {
1682    ANV_FROM_HANDLE(anv_device, device, _device);
1683    ANV_FROM_HANDLE(anv_event, event, _event);
1684
1685    if (!device->info.has_llc) {
1686       /* Invalidate read cache before reading event written by GPU. */
1687       __builtin_ia32_clflush(event);
1688       __builtin_ia32_mfence();
1689
1690    }
1691
1692    return event->semaphore;
1693 }
1694
1695 VkResult anv_SetEvent(
1696     VkDevice                                    _device,
1697     VkEvent                                     _event)
1698 {
1699    ANV_FROM_HANDLE(anv_device, device, _device);
1700    ANV_FROM_HANDLE(anv_event, event, _event);
1701
1702    event->semaphore = VK_EVENT_SET;
1703
1704    if (!device->info.has_llc) {
1705       /* Make sure the writes we're flushing have landed. */
1706       __builtin_ia32_mfence();
1707       __builtin_ia32_clflush(event);
1708    }
1709
1710    return VK_SUCCESS;
1711 }
1712
1713 VkResult anv_ResetEvent(
1714     VkDevice                                    _device,
1715     VkEvent                                     _event)
1716 {
1717    ANV_FROM_HANDLE(anv_device, device, _device);
1718    ANV_FROM_HANDLE(anv_event, event, _event);
1719
1720    event->semaphore = VK_EVENT_RESET;
1721
1722    if (!device->info.has_llc) {
1723       /* Make sure the writes we're flushing have landed. */
1724       __builtin_ia32_mfence();
1725       __builtin_ia32_clflush(event);
1726    }
1727
1728    return VK_SUCCESS;
1729 }
1730
1731 // Buffer functions
1732
1733 VkResult anv_CreateBuffer(
1734     VkDevice                                    _device,
1735     const VkBufferCreateInfo*                   pCreateInfo,
1736     const VkAllocationCallbacks*                pAllocator,
1737     VkBuffer*                                   pBuffer)
1738 {
1739    ANV_FROM_HANDLE(anv_device, device, _device);
1740    struct anv_buffer *buffer;
1741
1742    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1743
1744    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1745                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1746    if (buffer == NULL)
1747       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1748
1749    buffer->size = pCreateInfo->size;
1750    buffer->usage = pCreateInfo->usage;
1751    buffer->bo = NULL;
1752    buffer->offset = 0;
1753
1754    *pBuffer = anv_buffer_to_handle(buffer);
1755
1756    return VK_SUCCESS;
1757 }
1758
1759 void anv_DestroyBuffer(
1760     VkDevice                                    _device,
1761     VkBuffer                                    _buffer,
1762     const VkAllocationCallbacks*                pAllocator)
1763 {
1764    ANV_FROM_HANDLE(anv_device, device, _device);
1765    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1766
1767    vk_free2(&device->alloc, pAllocator, buffer);
1768 }
1769
1770 void
1771 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
1772                               enum isl_format format,
1773                               uint32_t offset, uint32_t range, uint32_t stride)
1774 {
1775    isl_buffer_fill_state(&device->isl_dev, state.map,
1776                          .address = offset,
1777                          .mocs = device->default_mocs,
1778                          .size = range,
1779                          .format = format,
1780                          .stride = stride);
1781
1782    if (!device->info.has_llc)
1783       anv_state_clflush(state);
1784 }
1785
1786 void anv_DestroySampler(
1787     VkDevice                                    _device,
1788     VkSampler                                   _sampler,
1789     const VkAllocationCallbacks*                pAllocator)
1790 {
1791    ANV_FROM_HANDLE(anv_device, device, _device);
1792    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
1793
1794    vk_free2(&device->alloc, pAllocator, sampler);
1795 }
1796
1797 VkResult anv_CreateFramebuffer(
1798     VkDevice                                    _device,
1799     const VkFramebufferCreateInfo*              pCreateInfo,
1800     const VkAllocationCallbacks*                pAllocator,
1801     VkFramebuffer*                              pFramebuffer)
1802 {
1803    ANV_FROM_HANDLE(anv_device, device, _device);
1804    struct anv_framebuffer *framebuffer;
1805
1806    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
1807
1808    size_t size = sizeof(*framebuffer) +
1809                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
1810    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
1811                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1812    if (framebuffer == NULL)
1813       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1814
1815    framebuffer->attachment_count = pCreateInfo->attachmentCount;
1816    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
1817       VkImageView _iview = pCreateInfo->pAttachments[i];
1818       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
1819    }
1820
1821    framebuffer->width = pCreateInfo->width;
1822    framebuffer->height = pCreateInfo->height;
1823    framebuffer->layers = pCreateInfo->layers;
1824
1825    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
1826
1827    return VK_SUCCESS;
1828 }
1829
1830 void anv_DestroyFramebuffer(
1831     VkDevice                                    _device,
1832     VkFramebuffer                               _fb,
1833     const VkAllocationCallbacks*                pAllocator)
1834 {
1835    ANV_FROM_HANDLE(anv_device, device, _device);
1836    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
1837
1838    vk_free2(&device->alloc, pAllocator, fb);
1839 }