src/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 "mesa/main/git_sha1.h"
  32 #include "util/strtod.h"
  33 #include "util/debug.h"
  34
  35 #include "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_errorf(VK_ERROR_INITIALIZATION_FAILED,
  66                        "failed to open %s: %m", path);
  67
  68    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
  69    device->instance = instance;
  70    device->path = path;
  71
  72    device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
  73    if (!device->chipset_id) {
  74       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
  75                          "failed to get chipset id: %m");
  76       goto fail;
  77    }
  78
  79    device->name = brw_get_device_name(device->chipset_id);
  80    device->info = brw_get_device_info(device->chipset_id);
  81    if (!device->info) {
  82       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
  83                          "failed to get device info");
  84       goto fail;
  85    }
  86
  87    if (device->info->is_haswell) {
  88       fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
  89    } else if (device->info->gen == 7 && !device->info->is_baytrail) {
  90       fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
  91    } else if (device->info->gen == 7 && device->info->is_baytrail) {
  92       fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
  93    } else if (device->info->gen >= 8) {
  94       /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
  95        * supported as anything */
  96    } else {
  97       result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
  98                          "Vulkan not yet supported on %s", device->name);
  99       goto fail;
 100    }
 101
 102    if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
 103       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 104                          "failed to get aperture size: %m");
 105       goto fail;
 106    }
 107
 108    if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
 109       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 110                          "kernel missing gem wait");
 111       goto fail;
 112    }
 113
 114    if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
 115       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 116                          "kernel missing execbuf2");
 117       goto fail;
 118    }
 119
 120    if (!device->info->has_llc &&
 121        anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
 122       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 123                          "kernel missing wc mmap");
 124       goto fail;
 125    }
 126
 127    bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
 128
 129    close(fd);
 130
 131    brw_process_intel_debug_variable();
 132
 133    device->compiler = brw_compiler_create(NULL, device->info);
 134    if (device->compiler == NULL) {
 135       result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 136       goto fail;
 137    }
 138    device->compiler->shader_debug_log = compiler_debug_log;
 139    device->compiler->shader_perf_log = compiler_perf_log;
 140
 141    /* XXX: Actually detect bit6 swizzling */
 142    isl_device_init(&device->isl_dev, device->info, swizzled);
 143
 144    return VK_SUCCESS;
 145
 146 fail:
 147    close(fd);
 148    return result;
 149 }
 150
 151 static void
 152 anv_physical_device_finish(struct anv_physical_device *device)
 153 {
 154    ralloc_free(device->compiler);
 155 }
 156
 157 static const VkExtensionProperties global_extensions[] = {
 158    {
 159       .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
 160       .specVersion = 25,
 161    },
 162    {
 163       .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
 164       .specVersion = 5,
 165    },
 166 #ifdef HAVE_WAYLAND_PLATFORM
 167    {
 168       .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
 169       .specVersion = 4,
 170    },
 171 #endif
 172 };
 173
 174 static const VkExtensionProperties device_extensions[] = {
 175    {
 176       .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
 177       .specVersion = 67,
 178    },
 179 };
 180
 181 static void *
 182 default_alloc_func(void *pUserData, size_t size, size_t align,
 183                    VkSystemAllocationScope allocationScope)
 184 {
 185    return malloc(size);
 186 }
 187
 188 static void *
 189 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
 190                      size_t align, VkSystemAllocationScope allocationScope)
 191 {
 192    return realloc(pOriginal, size);
 193 }
 194
 195 static void
 196 default_free_func(void *pUserData, void *pMemory)
 197 {
 198    free(pMemory);
 199 }
 200
 201 static const VkAllocationCallbacks default_alloc = {
 202    .pUserData = NULL,
 203    .pfnAllocation = default_alloc_func,
 204    .pfnReallocation = default_realloc_func,
 205    .pfnFree = default_free_func,
 206 };
 207
 208 VkResult anv_CreateInstance(
 209     const VkInstanceCreateInfo*                 pCreateInfo,
 210     const VkAllocationCallbacks*                pAllocator,
 211     VkInstance*                                 pInstance)
 212 {
 213    struct anv_instance *instance;
 214
 215    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 216
 217    uint32_t client_version = pCreateInfo->pApplicationInfo ?
 218                              pCreateInfo->pApplicationInfo->apiVersion :
 219                              VK_MAKE_VERSION(1, 0, 0);
 220    if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
 221        client_version > VK_MAKE_VERSION(1, 0, 3)) {
 222       return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 223                        "Client requested version %d.%d.%d",
 224                        VK_VERSION_MAJOR(client_version),
 225                        VK_VERSION_MINOR(client_version),
 226                        VK_VERSION_PATCH(client_version));
 227    }
 228
 229    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 230       bool found = false;
 231       for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
 232          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 233                     global_extensions[j].extensionName) == 0) {
 234             found = true;
 235             break;
 236          }
 237       }
 238       if (!found)
 239          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 240    }
 241
 242    instance = anv_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 243                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 244    if (!instance)
 245       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 246
 247    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 248
 249    if (pAllocator)
 250       instance->alloc = *pAllocator;
 251    else
 252       instance->alloc = default_alloc;
 253
 254    instance->apiVersion = client_version;
 255    instance->physicalDeviceCount = -1;
 256
 257    _mesa_locale_init();
 258
 259    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 260
 261    anv_init_wsi(instance);
 262
 263    *pInstance = anv_instance_to_handle(instance);
 264
 265    return VK_SUCCESS;
 266 }
 267
 268 void anv_DestroyInstance(
 269     VkInstance                                  _instance,
 270     const VkAllocationCallbacks*                pAllocator)
 271 {
 272    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 273
 274    if (instance->physicalDeviceCount > 0) {
 275       /* We support at most one physical device. */
 276       assert(instance->physicalDeviceCount == 1);
 277       anv_physical_device_finish(&instance->physicalDevice);
 278    }
 279
 280    anv_finish_wsi(instance);
 281
 282    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 283
 284    _mesa_locale_fini();
 285
 286    anv_free(&instance->alloc, instance);
 287 }
 288
 289 VkResult anv_EnumeratePhysicalDevices(
 290     VkInstance                                  _instance,
 291     uint32_t*                                   pPhysicalDeviceCount,
 292     VkPhysicalDevice*                           pPhysicalDevices)
 293 {
 294    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 295    VkResult result;
 296
 297    if (instance->physicalDeviceCount < 0) {
 298       result = anv_physical_device_init(&instance->physicalDevice,
 299                                         instance, "/dev/dri/renderD128");
 300       if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
 301          instance->physicalDeviceCount = 0;
 302       } else if (result == VK_SUCCESS) {
 303          instance->physicalDeviceCount = 1;
 304       } else {
 305          return result;
 306       }
 307    }
 308
 309    /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
 310     * otherwise it's an inout parameter.
 311     *
 312     * The Vulkan spec (git aaed022) says:
 313     *
 314     *    pPhysicalDeviceCount is a pointer to an unsigned integer variable
 315     *    that is initialized with the number of devices the application is
 316     *    prepared to receive handles to. pname:pPhysicalDevices is pointer to
 317     *    an array of at least this many VkPhysicalDevice handles [...].
 318     *
 319     *    Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
 320     *    overwrites the contents of the variable pointed to by
 321     *    pPhysicalDeviceCount with the number of physical devices in in the
 322     *    instance; otherwise, vkEnumeratePhysicalDevices overwrites
 323     *    pPhysicalDeviceCount with the number of physical handles written to
 324     *    pPhysicalDevices.
 325     */
 326    if (!pPhysicalDevices) {
 327       *pPhysicalDeviceCount = instance->physicalDeviceCount;
 328    } else if (*pPhysicalDeviceCount >= 1) {
 329       pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
 330       *pPhysicalDeviceCount = 1;
 331    } else {
 332       *pPhysicalDeviceCount = 0;
 333    }
 334
 335    return VK_SUCCESS;
 336 }
 337
 338 void anv_GetPhysicalDeviceFeatures(
 339     VkPhysicalDevice                            physicalDevice,
 340     VkPhysicalDeviceFeatures*                   pFeatures)
 341 {
 342    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 343
 344    *pFeatures = (VkPhysicalDeviceFeatures) {
 345       .robustBufferAccess                       = true,
 346       .fullDrawIndexUint32                      = true,
 347       .imageCubeArray                           = false,
 348       .independentBlend                         = pdevice->info->gen >= 8,
 349       .geometryShader                           = true,
 350       .tessellationShader                       = false,
 351       .sampleRateShading                        = false,
 352       .dualSrcBlend                             = true,
 353       .logicOp                                  = true,
 354       .multiDrawIndirect                        = false,
 355       .drawIndirectFirstInstance                = false,
 356       .depthClamp                               = false,
 357       .depthBiasClamp                           = false,
 358       .fillModeNonSolid                         = true,
 359       .depthBounds                              = false,
 360       .wideLines                                = true,
 361       .largePoints                              = true,
 362       .alphaToOne                               = true,
 363       .multiViewport                            = true,
 364       .samplerAnisotropy                        = false, /* FINISHME */
 365       .textureCompressionETC2                   = true,
 366       .textureCompressionASTC_LDR               = true,
 367       .textureCompressionBC                     = true,
 368       .occlusionQueryPrecise                    = true,
 369       .pipelineStatisticsQuery                  = true,
 370       .vertexPipelineStoresAndAtomics           = pdevice->info->gen >= 8,
 371       .fragmentStoresAndAtomics                 = true,
 372       .shaderTessellationAndGeometryPointSize   = true,
 373       .shaderImageGatherExtended                = true,
 374       .shaderStorageImageExtendedFormats        = false,
 375       .shaderStorageImageMultisample            = false,
 376       .shaderUniformBufferArrayDynamicIndexing  = true,
 377       .shaderSampledImageArrayDynamicIndexing   = true,
 378       .shaderStorageBufferArrayDynamicIndexing  = true,
 379       .shaderStorageImageArrayDynamicIndexing   = true,
 380       .shaderStorageImageReadWithoutFormat      = false,
 381       .shaderStorageImageWriteWithoutFormat     = true,
 382       .shaderClipDistance                       = false,
 383       .shaderCullDistance                       = false,
 384       .shaderFloat64                            = false,
 385       .shaderInt64                              = false,
 386       .shaderInt16                              = false,
 387       .alphaToOne                               = true,
 388       .variableMultisampleRate                  = false,
 389       .inheritedQueries                         = false,
 390    };
 391 }
 392
 393 void
 394 anv_device_get_cache_uuid(void *uuid)
 395 {
 396    memset(uuid, 0, VK_UUID_SIZE);
 397    snprintf(uuid, VK_UUID_SIZE, "anv-%s", MESA_GIT_SHA1 + 4);
 398 }
 399
 400 void anv_GetPhysicalDeviceProperties(
 401     VkPhysicalDevice                            physicalDevice,
 402     VkPhysicalDeviceProperties*                 pProperties)
 403 {
 404    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 405    const struct brw_device_info *devinfo = pdevice->info;
 406
 407    anv_finishme("Get correct values for VkPhysicalDeviceLimits");
 408
 409    const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
 410
 411    VkSampleCountFlags sample_counts =
 412       isl_device_get_sample_counts(&pdevice->isl_dev);
 413
 414    VkPhysicalDeviceLimits limits = {
 415       .maxImageDimension1D                      = (1 << 14),
 416       .maxImageDimension2D                      = (1 << 14),
 417       .maxImageDimension3D                      = (1 << 10),
 418       .maxImageDimensionCube                    = (1 << 14),
 419       .maxImageArrayLayers                      = (1 << 10),
 420       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 421       .maxUniformBufferRange                    = UINT32_MAX,
 422       .maxStorageBufferRange                    = UINT32_MAX,
 423       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 424       .maxMemoryAllocationCount                 = UINT32_MAX,
 425       .maxSamplerAllocationCount                = 64 * 1024,
 426       .bufferImageGranularity                   = 64, /* A cache line */
 427       .sparseAddressSpaceSize                   = 0,
 428       .maxBoundDescriptorSets                   = MAX_SETS,
 429       .maxPerStageDescriptorSamplers            = 64,
 430       .maxPerStageDescriptorUniformBuffers      = 64,
 431       .maxPerStageDescriptorStorageBuffers      = 64,
 432       .maxPerStageDescriptorSampledImages       = 64,
 433       .maxPerStageDescriptorStorageImages       = 64,
 434       .maxPerStageDescriptorInputAttachments    = 64,
 435       .maxPerStageResources                     = 128,
 436       .maxDescriptorSetSamplers                 = 256,
 437       .maxDescriptorSetUniformBuffers           = 256,
 438       .maxDescriptorSetUniformBuffersDynamic    = 256,
 439       .maxDescriptorSetStorageBuffers           = 256,
 440       .maxDescriptorSetStorageBuffersDynamic    = 256,
 441       .maxDescriptorSetSampledImages            = 256,
 442       .maxDescriptorSetStorageImages            = 256,
 443       .maxDescriptorSetInputAttachments         = 256,
 444       .maxVertexInputAttributes                 = 32,
 445       .maxVertexInputBindings                   = 32,
 446       .maxVertexInputAttributeOffset            = 2047,
 447       .maxVertexInputBindingStride              = 2048,
 448       .maxVertexOutputComponents                = 128,
 449       .maxTessellationGenerationLevel           = 0,
 450       .maxTessellationPatchSize                 = 0,
 451       .maxTessellationControlPerVertexInputComponents = 0,
 452       .maxTessellationControlPerVertexOutputComponents = 0,
 453       .maxTessellationControlPerPatchOutputComponents = 0,
 454       .maxTessellationControlTotalOutputComponents = 0,
 455       .maxTessellationEvaluationInputComponents = 0,
 456       .maxTessellationEvaluationOutputComponents = 0,
 457       .maxGeometryShaderInvocations             = 32,
 458       .maxGeometryInputComponents               = 64,
 459       .maxGeometryOutputComponents              = 128,
 460       .maxGeometryOutputVertices                = 256,
 461       .maxGeometryTotalOutputComponents         = 1024,
 462       .maxFragmentInputComponents               = 128,
 463       .maxFragmentOutputAttachments             = 8,
 464       .maxFragmentDualSrcAttachments            = 2,
 465       .maxFragmentCombinedOutputResources       = 8,
 466       .maxComputeSharedMemorySize               = 32768,
 467       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 468       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 469       .maxComputeWorkGroupSize = {
 470          16 * devinfo->max_cs_threads,
 471          16 * devinfo->max_cs_threads,
 472          16 * devinfo->max_cs_threads,
 473       },
 474       .subPixelPrecisionBits                    = 4 /* FIXME */,
 475       .subTexelPrecisionBits                    = 4 /* FIXME */,
 476       .mipmapPrecisionBits                      = 4 /* FIXME */,
 477       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 478       .maxDrawIndirectCount                     = UINT32_MAX,
 479       .maxSamplerLodBias                        = 16,
 480       .maxSamplerAnisotropy                     = 16,
 481       .maxViewports                             = MAX_VIEWPORTS,
 482       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 483       .viewportBoundsRange                      = { -16384.0, 16384.0 },
 484       .viewportSubPixelBits                     = 13, /* We take a float? */
 485       .minMemoryMapAlignment                    = 4096, /* A page */
 486       .minTexelBufferOffsetAlignment            = 1,
 487       .minUniformBufferOffsetAlignment          = 1,
 488       .minStorageBufferOffsetAlignment          = 1,
 489       .minTexelOffset                           = -8,
 490       .maxTexelOffset                           = 7,
 491       .minTexelGatherOffset                     = -8,
 492       .maxTexelGatherOffset                     = 7,
 493       .minInterpolationOffset                   = 0, /* FIXME */
 494       .maxInterpolationOffset                   = 0, /* FIXME */
 495       .subPixelInterpolationOffsetBits          = 0, /* FIXME */
 496       .maxFramebufferWidth                      = (1 << 14),
 497       .maxFramebufferHeight                     = (1 << 14),
 498       .maxFramebufferLayers                     = (1 << 10),
 499       .framebufferColorSampleCounts             = sample_counts,
 500       .framebufferDepthSampleCounts             = sample_counts,
 501       .framebufferStencilSampleCounts           = sample_counts,
 502       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 503       .maxColorAttachments                      = MAX_RTS,
 504       .sampledImageColorSampleCounts            = sample_counts,
 505       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 506       .sampledImageDepthSampleCounts            = sample_counts,
 507       .sampledImageStencilSampleCounts          = sample_counts,
 508       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 509       .maxSampleMaskWords                       = 1,
 510       .timestampComputeAndGraphics              = false,
 511       .timestampPeriod                          = time_stamp_base / (1000 * 1000 * 1000),
 512       .maxClipDistances                         = 0 /* FIXME */,
 513       .maxCullDistances                         = 0 /* FIXME */,
 514       .maxCombinedClipAndCullDistances          = 0 /* FIXME */,
 515       .discreteQueuePriorities                  = 1,
 516       .pointSizeRange                           = { 0.125, 255.875 },
 517       .lineWidthRange                           = { 0.0, 7.9921875 },
 518       .pointSizeGranularity                     = (1.0 / 8.0),
 519       .lineWidthGranularity                     = (1.0 / 128.0),
 520       .strictLines                              = false, /* FINISHME */
 521       .standardSampleLocations                  = true,
 522       .optimalBufferCopyOffsetAlignment         = 128,
 523       .optimalBufferCopyRowPitchAlignment       = 128,
 524       .nonCoherentAtomSize                      = 64,
 525    };
 526
 527    *pProperties = (VkPhysicalDeviceProperties) {
 528       .apiVersion = VK_MAKE_VERSION(1, 0, 2),
 529       .driverVersion = 1,
 530       .vendorID = 0x8086,
 531       .deviceID = pdevice->chipset_id,
 532       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 533       .limits = limits,
 534       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 535    };
 536
 537    strcpy(pProperties->deviceName, pdevice->name);
 538    anv_device_get_cache_uuid(pProperties->pipelineCacheUUID);
 539 }
 540
 541 void anv_GetPhysicalDeviceQueueFamilyProperties(
 542     VkPhysicalDevice                            physicalDevice,
 543     uint32_t*                                   pCount,
 544     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 545 {
 546    if (pQueueFamilyProperties == NULL) {
 547       *pCount = 1;
 548       return;
 549    }
 550
 551    assert(*pCount >= 1);
 552
 553    *pQueueFamilyProperties = (VkQueueFamilyProperties) {
 554       .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 555                     VK_QUEUE_COMPUTE_BIT |
 556                     VK_QUEUE_TRANSFER_BIT,
 557       .queueCount = 1,
 558       .timestampValidBits = 36, /* XXX: Real value here */
 559       .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
 560    };
 561 }
 562
 563 void anv_GetPhysicalDeviceMemoryProperties(
 564     VkPhysicalDevice                            physicalDevice,
 565     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
 566 {
 567    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 568    VkDeviceSize heap_size;
 569
 570    /* Reserve some wiggle room for the driver by exposing only 75% of the
 571     * aperture to the heap.
 572     */
 573    heap_size = 3 * physical_device->aperture_size / 4;
 574
 575    if (physical_device->info->has_llc) {
 576       /* Big core GPUs share LLC with the CPU and thus one memory type can be
 577        * both cached and coherent at the same time.
 578        */
 579       pMemoryProperties->memoryTypeCount = 1;
 580       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 581          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 582                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 583                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
 584                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 585          .heapIndex = 0,
 586       };
 587    } else {
 588       /* The spec requires that we expose a host-visible, coherent memory
 589        * type, but Atom GPUs don't share LLC. Thus we offer two memory types
 590        * to give the application a choice between cached, but not coherent and
 591        * coherent but uncached (WC though).
 592        */
 593       pMemoryProperties->memoryTypeCount = 2;
 594       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 595          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 596                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 597                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
 598          .heapIndex = 0,
 599       };
 600       pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
 601          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 602                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 603                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 604          .heapIndex = 0,
 605       };
 606    }
 607
 608    pMemoryProperties->memoryHeapCount = 1;
 609    pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
 610       .size = heap_size,
 611       .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 612    };
 613 }
 614
 615 PFN_vkVoidFunction anv_GetInstanceProcAddr(
 616     VkInstance                                  instance,
 617     const char*                                 pName)
 618 {
 619    return anv_lookup_entrypoint(pName);
 620 }
 621
 622 /* The loader wants us to expose a second GetInstanceProcAddr function
 623  * to work around certain LD_PRELOAD issues seen in apps.
 624  */
 625 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 626     VkInstance                                  instance,
 627     const char*                                 pName);
 628
 629 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 630     VkInstance                                  instance,
 631     const char*                                 pName)
 632 {
 633    return anv_GetInstanceProcAddr(instance, pName);
 634 }
 635
 636 PFN_vkVoidFunction anv_GetDeviceProcAddr(
 637     VkDevice                                    device,
 638     const char*                                 pName)
 639 {
 640    return anv_lookup_entrypoint(pName);
 641 }
 642
 643 static VkResult
 644 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
 645 {
 646    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 647    queue->device = device;
 648    queue->pool = &device->surface_state_pool;
 649
 650    return VK_SUCCESS;
 651 }
 652
 653 static void
 654 anv_queue_finish(struct anv_queue *queue)
 655 {
 656 }
 657
 658 static struct anv_state
 659 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
 660 {
 661    struct anv_state state;
 662
 663    state = anv_state_pool_alloc(pool, size, align);
 664    memcpy(state.map, p, size);
 665
 666    if (!pool->block_pool->device->info.has_llc)
 667       anv_state_clflush(state);
 668
 669    return state;
 670 }
 671
 672 struct gen8_border_color {
 673    union {
 674       float float32[4];
 675       uint32_t uint32[4];
 676    };
 677    /* Pad out to 64 bytes */
 678    uint32_t _pad[12];
 679 };
 680
 681 static void
 682 anv_device_init_border_colors(struct anv_device *device)
 683 {
 684    static const struct gen8_border_color border_colors[] = {
 685       [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] =  { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
 686       [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] =       { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
 687       [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] =       { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
 688       [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] =    { .uint32 = { 0, 0, 0, 0 } },
 689       [VK_BORDER_COLOR_INT_OPAQUE_BLACK] =         { .uint32 = { 0, 0, 0, 1 } },
 690       [VK_BORDER_COLOR_INT_OPAQUE_WHITE] =         { .uint32 = { 1, 1, 1, 1 } },
 691    };
 692
 693    device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
 694                                                     sizeof(border_colors), 64,
 695                                                     border_colors);
 696 }
 697
 698 VkResult
 699 anv_device_submit_simple_batch(struct anv_device *device,
 700                                struct anv_batch *batch)
 701 {
 702    struct drm_i915_gem_execbuffer2 execbuf;
 703    struct drm_i915_gem_exec_object2 exec2_objects[1];
 704    struct anv_bo bo;
 705    VkResult result = VK_SUCCESS;
 706    uint32_t size;
 707    int64_t timeout;
 708    int ret;
 709
 710    /* Kernel driver requires 8 byte aligned batch length */
 711    size = align_u32(batch->next - batch->start, 8);
 712    assert(size < device->batch_bo_pool.bo_size);
 713    result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo);
 714    if (result != VK_SUCCESS)
 715       return result;
 716
 717    memcpy(bo.map, batch->start, size);
 718    if (!device->info.has_llc)
 719       anv_clflush_range(bo.map, size);
 720
 721    exec2_objects[0].handle = bo.gem_handle;
 722    exec2_objects[0].relocation_count = 0;
 723    exec2_objects[0].relocs_ptr = 0;
 724    exec2_objects[0].alignment = 0;
 725    exec2_objects[0].offset = bo.offset;
 726    exec2_objects[0].flags = 0;
 727    exec2_objects[0].rsvd1 = 0;
 728    exec2_objects[0].rsvd2 = 0;
 729
 730    execbuf.buffers_ptr = (uintptr_t) exec2_objects;
 731    execbuf.buffer_count = 1;
 732    execbuf.batch_start_offset = 0;
 733    execbuf.batch_len = size;
 734    execbuf.cliprects_ptr = 0;
 735    execbuf.num_cliprects = 0;
 736    execbuf.DR1 = 0;
 737    execbuf.DR4 = 0;
 738
 739    execbuf.flags =
 740       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
 741    execbuf.rsvd1 = device->context_id;
 742    execbuf.rsvd2 = 0;
 743
 744    ret = anv_gem_execbuffer(device, &execbuf);
 745    if (ret != 0) {
 746       /* We don't know the real error. */
 747       result = vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY, "execbuf2 failed: %m");
 748       goto fail;
 749    }
 750
 751    timeout = INT64_MAX;
 752    ret = anv_gem_wait(device, bo.gem_handle, &timeout);
 753    if (ret != 0) {
 754       /* We don't know the real error. */
 755       result = vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY, "execbuf2 failed: %m");
 756       goto fail;
 757    }
 758
 759  fail:
 760    anv_bo_pool_free(&device->batch_bo_pool, &bo);
 761
 762    return result;
 763 }
 764
 765 VkResult anv_CreateDevice(
 766     VkPhysicalDevice                            physicalDevice,
 767     const VkDeviceCreateInfo*                   pCreateInfo,
 768     const VkAllocationCallbacks*                pAllocator,
 769     VkDevice*                                   pDevice)
 770 {
 771    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 772    VkResult result;
 773    struct anv_device *device;
 774
 775    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
 776
 777    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 778       bool found = false;
 779       for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
 780          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 781                     device_extensions[j].extensionName) == 0) {
 782             found = true;
 783             break;
 784          }
 785       }
 786       if (!found)
 787          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 788    }
 789
 790    anv_set_dispatch_devinfo(physical_device->info);
 791
 792    device = anv_alloc2(&physical_device->instance->alloc, pAllocator,
 793                        sizeof(*device), 8,
 794                        VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
 795    if (!device)
 796       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 797
 798    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 799    device->instance = physical_device->instance;
 800    device->chipset_id = physical_device->chipset_id;
 801
 802    if (pAllocator)
 803       device->alloc = *pAllocator;
 804    else
 805       device->alloc = physical_device->instance->alloc;
 806
 807    /* XXX(chadv): Can we dup() physicalDevice->fd here? */
 808    device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
 809    if (device->fd == -1) {
 810       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
 811       goto fail_device;
 812    }
 813
 814    device->context_id = anv_gem_create_context(device);
 815    if (device->context_id == -1) {
 816       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
 817       goto fail_fd;
 818    }
 819
 820    device->info = *physical_device->info;
 821    device->isl_dev = physical_device->isl_dev;
 822
 823    pthread_mutex_init(&device->mutex, NULL);
 824
 825    anv_bo_pool_init(&device->batch_bo_pool, device, ANV_CMD_BUFFER_BATCH_SIZE);
 826
 827    anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
 828
 829    anv_state_pool_init(&device->dynamic_state_pool,
 830                        &device->dynamic_state_block_pool);
 831
 832    anv_block_pool_init(&device->instruction_block_pool, device, 128 * 1024);
 833    anv_pipeline_cache_init(&device->default_pipeline_cache, device);
 834
 835    anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
 836
 837    anv_state_pool_init(&device->surface_state_pool,
 838                        &device->surface_state_block_pool);
 839
 840    anv_bo_init_new(&device->workaround_bo, device, 1024);
 841
 842    anv_block_pool_init(&device->scratch_block_pool, device, 0x10000);
 843
 844    anv_queue_init(device, &device->queue);
 845
 846    switch (device->info.gen) {
 847    case 7:
 848       if (!device->info.is_haswell)
 849          result = gen7_init_device_state(device);
 850       else
 851          result = gen75_init_device_state(device);
 852       break;
 853    case 8:
 854       result = gen8_init_device_state(device);
 855       break;
 856    case 9:
 857       result = gen9_init_device_state(device);
 858       break;
 859    default:
 860       /* Shouldn't get here as we don't create physical devices for any other
 861        * gens. */
 862       unreachable("unhandled gen");
 863    }
 864    if (result != VK_SUCCESS)
 865       goto fail_fd;
 866
 867    result = anv_device_init_meta(device);
 868    if (result != VK_SUCCESS)
 869       goto fail_fd;
 870
 871    anv_device_init_border_colors(device);
 872
 873    *pDevice = anv_device_to_handle(device);
 874
 875    return VK_SUCCESS;
 876
 877  fail_fd:
 878    close(device->fd);
 879  fail_device:
 880    anv_free(&device->alloc, device);
 881
 882    return result;
 883 }
 884
 885 void anv_DestroyDevice(
 886     VkDevice                                    _device,
 887     const VkAllocationCallbacks*                pAllocator)
 888 {
 889    ANV_FROM_HANDLE(anv_device, device, _device);
 890
 891    anv_queue_finish(&device->queue);
 892
 893    anv_device_finish_meta(device);
 894
 895 #ifdef HAVE_VALGRIND
 896    /* We only need to free these to prevent valgrind errors.  The backing
 897     * BO will go away in a couple of lines so we don't actually leak.
 898     */
 899    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
 900 #endif
 901
 902    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
 903    anv_gem_close(device, device->workaround_bo.gem_handle);
 904
 905    anv_bo_pool_finish(&device->batch_bo_pool);
 906    anv_state_pool_finish(&device->dynamic_state_pool);
 907    anv_block_pool_finish(&device->dynamic_state_block_pool);
 908    anv_block_pool_finish(&device->instruction_block_pool);
 909    anv_state_pool_finish(&device->surface_state_pool);
 910    anv_block_pool_finish(&device->surface_state_block_pool);
 911    anv_block_pool_finish(&device->scratch_block_pool);
 912
 913    close(device->fd);
 914
 915    pthread_mutex_destroy(&device->mutex);
 916
 917    anv_free(&device->alloc, device);
 918 }
 919
 920 VkResult anv_EnumerateInstanceExtensionProperties(
 921     const char*                                 pLayerName,
 922     uint32_t*                                   pPropertyCount,
 923     VkExtensionProperties*                      pProperties)
 924 {
 925    if (pProperties == NULL) {
 926       *pPropertyCount = ARRAY_SIZE(global_extensions);
 927       return VK_SUCCESS;
 928    }
 929
 930    assert(*pPropertyCount >= ARRAY_SIZE(global_extensions));
 931
 932    *pPropertyCount = ARRAY_SIZE(global_extensions);
 933    memcpy(pProperties, global_extensions, sizeof(global_extensions));
 934
 935    return VK_SUCCESS;
 936 }
 937
 938 VkResult anv_EnumerateDeviceExtensionProperties(
 939     VkPhysicalDevice                            physicalDevice,
 940     const char*                                 pLayerName,
 941     uint32_t*                                   pPropertyCount,
 942     VkExtensionProperties*                      pProperties)
 943 {
 944    if (pProperties == NULL) {
 945       *pPropertyCount = ARRAY_SIZE(device_extensions);
 946       return VK_SUCCESS;
 947    }
 948
 949    assert(*pPropertyCount >= ARRAY_SIZE(device_extensions));
 950
 951    *pPropertyCount = ARRAY_SIZE(device_extensions);
 952    memcpy(pProperties, device_extensions, sizeof(device_extensions));
 953
 954    return VK_SUCCESS;
 955 }
 956
 957 VkResult anv_EnumerateInstanceLayerProperties(
 958     uint32_t*                                   pPropertyCount,
 959     VkLayerProperties*                          pProperties)
 960 {
 961    if (pProperties == NULL) {
 962       *pPropertyCount = 0;
 963       return VK_SUCCESS;
 964    }
 965
 966    /* None supported at this time */
 967    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
 968 }
 969
 970 VkResult anv_EnumerateDeviceLayerProperties(
 971     VkPhysicalDevice                            physicalDevice,
 972     uint32_t*                                   pPropertyCount,
 973     VkLayerProperties*                          pProperties)
 974 {
 975    if (pProperties == NULL) {
 976       *pPropertyCount = 0;
 977       return VK_SUCCESS;
 978    }
 979
 980    /* None supported at this time */
 981    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
 982 }
 983
 984 void anv_GetDeviceQueue(
 985     VkDevice                                    _device,
 986     uint32_t                                    queueNodeIndex,
 987     uint32_t                                    queueIndex,
 988     VkQueue*                                    pQueue)
 989 {
 990    ANV_FROM_HANDLE(anv_device, device, _device);
 991
 992    assert(queueIndex == 0);
 993
 994    *pQueue = anv_queue_to_handle(&device->queue);
 995 }
 996
 997 VkResult anv_QueueSubmit(
 998     VkQueue                                     _queue,
 999     uint32_t                                    submitCount,
1000     const VkSubmitInfo*                         pSubmits,
1001     VkFence                                     _fence)
1002 {
1003    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1004    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1005    struct anv_device *device = queue->device;
1006    int ret;
1007
1008    for (uint32_t i = 0; i < submitCount; i++) {
1009       for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1010          ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1011                          pSubmits[i].pCommandBuffers[j]);
1012          assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1013
1014          ret = anv_gem_execbuffer(device, &cmd_buffer->execbuf2.execbuf);
1015          if (ret != 0) {
1016             /* We don't know the real error. */
1017             return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1018                              "execbuf2 failed: %m");
1019          }
1020
1021          for (uint32_t k = 0; k < cmd_buffer->execbuf2.bo_count; k++)
1022             cmd_buffer->execbuf2.bos[k]->offset = cmd_buffer->execbuf2.objects[k].offset;
1023       }
1024    }
1025
1026    if (fence) {
1027       ret = anv_gem_execbuffer(device, &fence->execbuf);
1028       if (ret != 0) {
1029          /* We don't know the real error. */
1030          return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1031                           "execbuf2 failed: %m");
1032       }
1033    }
1034
1035    return VK_SUCCESS;
1036 }
1037
1038 VkResult anv_QueueWaitIdle(
1039     VkQueue                                     _queue)
1040 {
1041    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1042
1043    return ANV_CALL(DeviceWaitIdle)(anv_device_to_handle(queue->device));
1044 }
1045
1046 VkResult anv_DeviceWaitIdle(
1047     VkDevice                                    _device)
1048 {
1049    ANV_FROM_HANDLE(anv_device, device, _device);
1050    struct anv_batch batch;
1051
1052    uint32_t cmds[8];
1053    batch.start = batch.next = cmds;
1054    batch.end = (void *) cmds + sizeof(cmds);
1055
1056    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END);
1057    anv_batch_emit(&batch, GEN7_MI_NOOP);
1058
1059    return anv_device_submit_simple_batch(device, &batch);
1060 }
1061
1062 VkResult
1063 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1064 {
1065    bo->gem_handle = anv_gem_create(device, size);
1066    if (!bo->gem_handle)
1067       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1068
1069    bo->map = NULL;
1070    bo->index = 0;
1071    bo->offset = 0;
1072    bo->size = size;
1073    bo->is_winsys_bo = false;
1074
1075    return VK_SUCCESS;
1076 }
1077
1078 VkResult anv_AllocateMemory(
1079     VkDevice                                    _device,
1080     const VkMemoryAllocateInfo*                 pAllocateInfo,
1081     const VkAllocationCallbacks*                pAllocator,
1082     VkDeviceMemory*                             pMem)
1083 {
1084    ANV_FROM_HANDLE(anv_device, device, _device);
1085    struct anv_device_memory *mem;
1086    VkResult result;
1087
1088    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1089
1090    if (pAllocateInfo->allocationSize == 0) {
1091       /* Apparently, this is allowed */
1092       *pMem = VK_NULL_HANDLE;
1093       return VK_SUCCESS;
1094    }
1095
1096    /* We support exactly one memory heap. */
1097    assert(pAllocateInfo->memoryTypeIndex == 0 ||
1098           (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1099
1100    /* FINISHME: Fail if allocation request exceeds heap size. */
1101
1102    mem = anv_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1103                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1104    if (mem == NULL)
1105       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1106
1107    /* The kernel is going to give us whole pages anyway */
1108    uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1109
1110    result = anv_bo_init_new(&mem->bo, device, alloc_size);
1111    if (result != VK_SUCCESS)
1112       goto fail;
1113
1114    mem->type_index = pAllocateInfo->memoryTypeIndex;
1115
1116    *pMem = anv_device_memory_to_handle(mem);
1117
1118    return VK_SUCCESS;
1119
1120  fail:
1121    anv_free2(&device->alloc, pAllocator, mem);
1122
1123    return result;
1124 }
1125
1126 void anv_FreeMemory(
1127     VkDevice                                    _device,
1128     VkDeviceMemory                              _mem,
1129     const VkAllocationCallbacks*                pAllocator)
1130 {
1131    ANV_FROM_HANDLE(anv_device, device, _device);
1132    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1133
1134    if (mem == NULL)
1135       return;
1136
1137    if (mem->bo.map)
1138       anv_gem_munmap(mem->bo.map, mem->bo.size);
1139
1140    if (mem->bo.gem_handle != 0)
1141       anv_gem_close(device, mem->bo.gem_handle);
1142
1143    anv_free2(&device->alloc, pAllocator, mem);
1144 }
1145
1146 VkResult anv_MapMemory(
1147     VkDevice                                    _device,
1148     VkDeviceMemory                              _memory,
1149     VkDeviceSize                                offset,
1150     VkDeviceSize                                size,
1151     VkMemoryMapFlags                            flags,
1152     void**                                      ppData)
1153 {
1154    ANV_FROM_HANDLE(anv_device, device, _device);
1155    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1156
1157    if (mem == NULL) {
1158       *ppData = NULL;
1159       return VK_SUCCESS;
1160    }
1161
1162    if (size == VK_WHOLE_SIZE)
1163       size = mem->bo.size - offset;
1164
1165    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1166     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1167     * at a time is valid. We could just mmap up front and return an offset
1168     * pointer here, but that may exhaust virtual memory on 32 bit
1169     * userspace. */
1170
1171    uint32_t gem_flags = 0;
1172    if (!device->info.has_llc && mem->type_index == 0)
1173       gem_flags |= I915_MMAP_WC;
1174
1175    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1176    uint64_t map_offset = offset & ~4095ull;
1177    assert(offset >= map_offset);
1178    uint64_t map_size = (offset + size) - map_offset;
1179
1180    /* Let's map whole pages */
1181    map_size = align_u64(map_size, 4096);
1182
1183    mem->map = anv_gem_mmap(device, mem->bo.gem_handle,
1184                            map_offset, map_size, gem_flags);
1185    mem->map_size = map_size;
1186
1187    *ppData = mem->map + (offset - map_offset);
1188
1189    return VK_SUCCESS;
1190 }
1191
1192 void anv_UnmapMemory(
1193     VkDevice                                    _device,
1194     VkDeviceMemory                              _memory)
1195 {
1196    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1197
1198    if (mem == NULL)
1199       return;
1200
1201    anv_gem_munmap(mem->map, mem->map_size);
1202 }
1203
1204 static void
1205 clflush_mapped_ranges(struct anv_device         *device,
1206                       uint32_t                   count,
1207                       const VkMappedMemoryRange *ranges)
1208 {
1209    for (uint32_t i = 0; i < count; i++) {
1210       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1211       void *p = mem->map + (ranges[i].offset & ~CACHELINE_MASK);
1212       void *end;
1213
1214       if (ranges[i].offset + ranges[i].size > mem->map_size)
1215          end = mem->map + mem->map_size;
1216       else
1217          end = mem->map + ranges[i].offset + ranges[i].size;
1218
1219       while (p < end) {
1220          __builtin_ia32_clflush(p);
1221          p += CACHELINE_SIZE;
1222       }
1223    }
1224 }
1225
1226 VkResult anv_FlushMappedMemoryRanges(
1227     VkDevice                                    _device,
1228     uint32_t                                    memoryRangeCount,
1229     const VkMappedMemoryRange*                  pMemoryRanges)
1230 {
1231    ANV_FROM_HANDLE(anv_device, device, _device);
1232
1233    if (device->info.has_llc)
1234       return VK_SUCCESS;
1235
1236    /* Make sure the writes we're flushing have landed. */
1237    __builtin_ia32_mfence();
1238
1239    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1240
1241    return VK_SUCCESS;
1242 }
1243
1244 VkResult anv_InvalidateMappedMemoryRanges(
1245     VkDevice                                    _device,
1246     uint32_t                                    memoryRangeCount,
1247     const VkMappedMemoryRange*                  pMemoryRanges)
1248 {
1249    ANV_FROM_HANDLE(anv_device, device, _device);
1250
1251    if (device->info.has_llc)
1252       return VK_SUCCESS;
1253
1254    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1255
1256    /* Make sure no reads get moved up above the invalidate. */
1257    __builtin_ia32_mfence();
1258
1259    return VK_SUCCESS;
1260 }
1261
1262 void anv_GetBufferMemoryRequirements(
1263     VkDevice                                    device,
1264     VkBuffer                                    _buffer,
1265     VkMemoryRequirements*                       pMemoryRequirements)
1266 {
1267    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1268
1269    /* The Vulkan spec (git aaed022) says:
1270     *
1271     *    memoryTypeBits is a bitfield and contains one bit set for every
1272     *    supported memory type for the resource. The bit `1<<i` is set if and
1273     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1274     *    structure for the physical device is supported.
1275     *
1276     * We support exactly one memory type.
1277     */
1278    pMemoryRequirements->memoryTypeBits = 1;
1279
1280    pMemoryRequirements->size = buffer->size;
1281    pMemoryRequirements->alignment = 16;
1282 }
1283
1284 void anv_GetImageMemoryRequirements(
1285     VkDevice                                    device,
1286     VkImage                                     _image,
1287     VkMemoryRequirements*                       pMemoryRequirements)
1288 {
1289    ANV_FROM_HANDLE(anv_image, image, _image);
1290
1291    /* The Vulkan spec (git aaed022) says:
1292     *
1293     *    memoryTypeBits is a bitfield and contains one bit set for every
1294     *    supported memory type for the resource. The bit `1<<i` is set if and
1295     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1296     *    structure for the physical device is supported.
1297     *
1298     * We support exactly one memory type.
1299     */
1300    pMemoryRequirements->memoryTypeBits = 1;
1301
1302    pMemoryRequirements->size = image->size;
1303    pMemoryRequirements->alignment = image->alignment;
1304 }
1305
1306 void anv_GetImageSparseMemoryRequirements(
1307     VkDevice                                    device,
1308     VkImage                                     image,
1309     uint32_t*                                   pSparseMemoryRequirementCount,
1310     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1311 {
1312    stub();
1313 }
1314
1315 void anv_GetDeviceMemoryCommitment(
1316     VkDevice                                    device,
1317     VkDeviceMemory                              memory,
1318     VkDeviceSize*                               pCommittedMemoryInBytes)
1319 {
1320    *pCommittedMemoryInBytes = 0;
1321 }
1322
1323 VkResult anv_BindBufferMemory(
1324     VkDevice                                    device,
1325     VkBuffer                                    _buffer,
1326     VkDeviceMemory                              _memory,
1327     VkDeviceSize                                memoryOffset)
1328 {
1329    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1330    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1331
1332    if (mem) {
1333       buffer->bo = &mem->bo;
1334       buffer->offset = memoryOffset;
1335    } else {
1336       buffer->bo = NULL;
1337       buffer->offset = 0;
1338    }
1339
1340    return VK_SUCCESS;
1341 }
1342
1343 VkResult anv_BindImageMemory(
1344     VkDevice                                    device,
1345     VkImage                                     _image,
1346     VkDeviceMemory                              _memory,
1347     VkDeviceSize                                memoryOffset)
1348 {
1349    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1350    ANV_FROM_HANDLE(anv_image, image, _image);
1351
1352    if (mem) {
1353       image->bo = &mem->bo;
1354       image->offset = memoryOffset;
1355    } else {
1356       image->bo = NULL;
1357       image->offset = 0;
1358    }
1359
1360    return VK_SUCCESS;
1361 }
1362
1363 VkResult anv_QueueBindSparse(
1364     VkQueue                                     queue,
1365     uint32_t                                    bindInfoCount,
1366     const VkBindSparseInfo*                     pBindInfo,
1367     VkFence                                     fence)
1368 {
1369    stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1370 }
1371
1372 VkResult anv_CreateFence(
1373     VkDevice                                    _device,
1374     const VkFenceCreateInfo*                    pCreateInfo,
1375     const VkAllocationCallbacks*                pAllocator,
1376     VkFence*                                    pFence)
1377 {
1378    ANV_FROM_HANDLE(anv_device, device, _device);
1379    struct anv_fence *fence;
1380    struct anv_batch batch;
1381    VkResult result;
1382
1383    const uint32_t fence_size = 128;
1384
1385    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1386
1387    fence = anv_alloc2(&device->alloc, pAllocator, sizeof(*fence), 8,
1388                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1389    if (fence == NULL)
1390       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1391
1392    result = anv_bo_init_new(&fence->bo, device, fence_size);
1393    if (result != VK_SUCCESS)
1394       goto fail;
1395
1396    fence->bo.map =
1397       anv_gem_mmap(device, fence->bo.gem_handle, 0, fence->bo.size, 0);
1398    batch.next = batch.start = fence->bo.map;
1399    batch.end = fence->bo.map + fence->bo.size;
1400    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END);
1401    anv_batch_emit(&batch, GEN7_MI_NOOP);
1402
1403    if (!device->info.has_llc) {
1404       assert(((uintptr_t) fence->bo.map & CACHELINE_MASK) == 0);
1405       assert(batch.next - fence->bo.map <= CACHELINE_SIZE);
1406       __builtin_ia32_mfence();
1407       __builtin_ia32_clflush(fence->bo.map);
1408    }
1409
1410    fence->exec2_objects[0].handle = fence->bo.gem_handle;
1411    fence->exec2_objects[0].relocation_count = 0;
1412    fence->exec2_objects[0].relocs_ptr = 0;
1413    fence->exec2_objects[0].alignment = 0;
1414    fence->exec2_objects[0].offset = fence->bo.offset;
1415    fence->exec2_objects[0].flags = 0;
1416    fence->exec2_objects[0].rsvd1 = 0;
1417    fence->exec2_objects[0].rsvd2 = 0;
1418
1419    fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1420    fence->execbuf.buffer_count = 1;
1421    fence->execbuf.batch_start_offset = 0;
1422    fence->execbuf.batch_len = batch.next - fence->bo.map;
1423    fence->execbuf.cliprects_ptr = 0;
1424    fence->execbuf.num_cliprects = 0;
1425    fence->execbuf.DR1 = 0;
1426    fence->execbuf.DR4 = 0;
1427
1428    fence->execbuf.flags =
1429       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1430    fence->execbuf.rsvd1 = device->context_id;
1431    fence->execbuf.rsvd2 = 0;
1432
1433    fence->ready = false;
1434
1435    *pFence = anv_fence_to_handle(fence);
1436
1437    return VK_SUCCESS;
1438
1439  fail:
1440    anv_free2(&device->alloc, pAllocator, fence);
1441
1442    return result;
1443 }
1444
1445 void anv_DestroyFence(
1446     VkDevice                                    _device,
1447     VkFence                                     _fence,
1448     const VkAllocationCallbacks*                pAllocator)
1449 {
1450    ANV_FROM_HANDLE(anv_device, device, _device);
1451    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1452
1453    anv_gem_munmap(fence->bo.map, fence->bo.size);
1454    anv_gem_close(device, fence->bo.gem_handle);
1455    anv_free2(&device->alloc, pAllocator, fence);
1456 }
1457
1458 VkResult anv_ResetFences(
1459     VkDevice                                    _device,
1460     uint32_t                                    fenceCount,
1461     const VkFence*                              pFences)
1462 {
1463    for (uint32_t i = 0; i < fenceCount; i++) {
1464       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1465       fence->ready = false;
1466    }
1467
1468    return VK_SUCCESS;
1469 }
1470
1471 VkResult anv_GetFenceStatus(
1472     VkDevice                                    _device,
1473     VkFence                                     _fence)
1474 {
1475    ANV_FROM_HANDLE(anv_device, device, _device);
1476    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1477    int64_t t = 0;
1478    int ret;
1479
1480    if (fence->ready)
1481       return VK_SUCCESS;
1482
1483    ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1484    if (ret == 0) {
1485       fence->ready = true;
1486       return VK_SUCCESS;
1487    }
1488
1489    return VK_NOT_READY;
1490 }
1491
1492 VkResult anv_WaitForFences(
1493     VkDevice                                    _device,
1494     uint32_t                                    fenceCount,
1495     const VkFence*                              pFences,
1496     VkBool32                                    waitAll,
1497     uint64_t                                    timeout)
1498 {
1499    ANV_FROM_HANDLE(anv_device, device, _device);
1500
1501    /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1502     * to block indefinitely timeouts <= 0.  Unfortunately, this was broken
1503     * for a couple of kernel releases.  Since there's no way to know
1504     * whether or not the kernel we're using is one of the broken ones, the
1505     * best we can do is to clamp the timeout to INT64_MAX.  This limits the
1506     * maximum timeout from 584 years to 292 years - likely not a big deal.
1507     */
1508    if (timeout > INT64_MAX)
1509       timeout = INT64_MAX;
1510
1511    int64_t t = timeout;
1512
1513    /* FIXME: handle !waitAll */
1514
1515    for (uint32_t i = 0; i < fenceCount; i++) {
1516       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1517       int ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1518       if (ret == -1 && errno == ETIME) {
1519          return VK_TIMEOUT;
1520       } else if (ret == -1) {
1521          /* We don't know the real error. */
1522          return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1523                           "gem wait failed: %m");
1524       }
1525    }
1526
1527    return VK_SUCCESS;
1528 }
1529
1530 // Queue semaphore functions
1531
1532 VkResult anv_CreateSemaphore(
1533     VkDevice                                    device,
1534     const VkSemaphoreCreateInfo*                pCreateInfo,
1535     const VkAllocationCallbacks*                pAllocator,
1536     VkSemaphore*                                pSemaphore)
1537 {
1538    /* The DRM execbuffer ioctl always execute in-oder, even between different
1539     * rings. As such, there's nothing to do for the user space semaphore.
1540     */
1541
1542    *pSemaphore = (VkSemaphore)1;
1543
1544    return VK_SUCCESS;
1545 }
1546
1547 void anv_DestroySemaphore(
1548     VkDevice                                    device,
1549     VkSemaphore                                 semaphore,
1550     const VkAllocationCallbacks*                pAllocator)
1551 {
1552 }
1553
1554 // Event functions
1555
1556 VkResult anv_CreateEvent(
1557     VkDevice                                    _device,
1558     const VkEventCreateInfo*                    pCreateInfo,
1559     const VkAllocationCallbacks*                pAllocator,
1560     VkEvent*                                    pEvent)
1561 {
1562    ANV_FROM_HANDLE(anv_device, device, _device);
1563    struct anv_state state;
1564    struct anv_event *event;
1565
1566    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1567
1568    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1569                                 sizeof(*event), 8);
1570    event = state.map;
1571    event->state = state;
1572    event->semaphore = VK_EVENT_RESET;
1573
1574    if (!device->info.has_llc) {
1575       /* Make sure the writes we're flushing have landed. */
1576       __builtin_ia32_mfence();
1577       __builtin_ia32_clflush(event);
1578    }
1579
1580    *pEvent = anv_event_to_handle(event);
1581
1582    return VK_SUCCESS;
1583 }
1584
1585 void anv_DestroyEvent(
1586     VkDevice                                    _device,
1587     VkEvent                                     _event,
1588     const VkAllocationCallbacks*                pAllocator)
1589 {
1590    ANV_FROM_HANDLE(anv_device, device, _device);
1591    ANV_FROM_HANDLE(anv_event, event, _event);
1592
1593    anv_state_pool_free(&device->dynamic_state_pool, event->state);
1594 }
1595
1596 VkResult anv_GetEventStatus(
1597     VkDevice                                    _device,
1598     VkEvent                                     _event)
1599 {
1600    ANV_FROM_HANDLE(anv_device, device, _device);
1601    ANV_FROM_HANDLE(anv_event, event, _event);
1602
1603    if (!device->info.has_llc) {
1604       /* Invalidate read cache before reading event written by GPU. */
1605       __builtin_ia32_clflush(event);
1606       __builtin_ia32_mfence();
1607
1608    }
1609
1610    return event->semaphore;
1611 }
1612
1613 VkResult anv_SetEvent(
1614     VkDevice                                    _device,
1615     VkEvent                                     _event)
1616 {
1617    ANV_FROM_HANDLE(anv_device, device, _device);
1618    ANV_FROM_HANDLE(anv_event, event, _event);
1619
1620    event->semaphore = VK_EVENT_SET;
1621
1622    if (!device->info.has_llc) {
1623       /* Make sure the writes we're flushing have landed. */
1624       __builtin_ia32_mfence();
1625       __builtin_ia32_clflush(event);
1626    }
1627
1628    return VK_SUCCESS;
1629 }
1630
1631 VkResult anv_ResetEvent(
1632     VkDevice                                    _device,
1633     VkEvent                                     _event)
1634 {
1635    ANV_FROM_HANDLE(anv_device, device, _device);
1636    ANV_FROM_HANDLE(anv_event, event, _event);
1637
1638    event->semaphore = VK_EVENT_RESET;
1639
1640    if (!device->info.has_llc) {
1641       /* Make sure the writes we're flushing have landed. */
1642       __builtin_ia32_mfence();
1643       __builtin_ia32_clflush(event);
1644    }
1645
1646    return VK_SUCCESS;
1647 }
1648
1649 // Buffer functions
1650
1651 VkResult anv_CreateBuffer(
1652     VkDevice                                    _device,
1653     const VkBufferCreateInfo*                   pCreateInfo,
1654     const VkAllocationCallbacks*                pAllocator,
1655     VkBuffer*                                   pBuffer)
1656 {
1657    ANV_FROM_HANDLE(anv_device, device, _device);
1658    struct anv_buffer *buffer;
1659
1660    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1661
1662    buffer = anv_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1663                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1664    if (buffer == NULL)
1665       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1666
1667    buffer->size = pCreateInfo->size;
1668    buffer->usage = pCreateInfo->usage;
1669    buffer->bo = NULL;
1670    buffer->offset = 0;
1671
1672    *pBuffer = anv_buffer_to_handle(buffer);
1673
1674    return VK_SUCCESS;
1675 }
1676
1677 void anv_DestroyBuffer(
1678     VkDevice                                    _device,
1679     VkBuffer                                    _buffer,
1680     const VkAllocationCallbacks*                pAllocator)
1681 {
1682    ANV_FROM_HANDLE(anv_device, device, _device);
1683    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1684
1685    anv_free2(&device->alloc, pAllocator, buffer);
1686 }
1687
1688 void
1689 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
1690                               enum isl_format format,
1691                               uint32_t offset, uint32_t range, uint32_t stride)
1692 {
1693    switch (device->info.gen) {
1694    case 7:
1695       if (device->info.is_haswell)
1696          gen75_fill_buffer_surface_state(state.map, format, offset, range,
1697                                          stride);
1698       else
1699          gen7_fill_buffer_surface_state(state.map, format, offset, range,
1700                                         stride);
1701       break;
1702    case 8:
1703       gen8_fill_buffer_surface_state(state.map, format, offset, range, stride);
1704       break;
1705    case 9:
1706       gen9_fill_buffer_surface_state(state.map, format, offset, range, stride);
1707       break;
1708    default:
1709       unreachable("unsupported gen\n");
1710    }
1711
1712    if (!device->info.has_llc)
1713       anv_state_clflush(state);
1714 }
1715
1716 void anv_DestroySampler(
1717     VkDevice                                    _device,
1718     VkSampler                                   _sampler,
1719     const VkAllocationCallbacks*                pAllocator)
1720 {
1721    ANV_FROM_HANDLE(anv_device, device, _device);
1722    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
1723
1724    anv_free2(&device->alloc, pAllocator, sampler);
1725 }
1726
1727 VkResult anv_CreateFramebuffer(
1728     VkDevice                                    _device,
1729     const VkFramebufferCreateInfo*              pCreateInfo,
1730     const VkAllocationCallbacks*                pAllocator,
1731     VkFramebuffer*                              pFramebuffer)
1732 {
1733    ANV_FROM_HANDLE(anv_device, device, _device);
1734    struct anv_framebuffer *framebuffer;
1735
1736    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
1737
1738    size_t size = sizeof(*framebuffer) +
1739                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
1740    framebuffer = anv_alloc2(&device->alloc, pAllocator, size, 8,
1741                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1742    if (framebuffer == NULL)
1743       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1744
1745    framebuffer->attachment_count = pCreateInfo->attachmentCount;
1746    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
1747       VkImageView _iview = pCreateInfo->pAttachments[i];
1748       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
1749    }
1750
1751    framebuffer->width = pCreateInfo->width;
1752    framebuffer->height = pCreateInfo->height;
1753    framebuffer->layers = pCreateInfo->layers;
1754
1755    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
1756
1757    return VK_SUCCESS;
1758 }
1759
1760 void anv_DestroyFramebuffer(
1761     VkDevice                                    _device,
1762     VkFramebuffer                               _fb,
1763     const VkAllocationCallbacks*                pAllocator)
1764 {
1765    ANV_FROM_HANDLE(anv_device, device, _device);
1766    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
1767
1768    anv_free2(&device->alloc, pAllocator, fb);
1769 }
1770
1771 void vkCmdDbgMarkerBegin(
1772     VkCommandBuffer                              commandBuffer,
1773     const char*                                 pMarker)
1774    __attribute__ ((visibility ("default")));
1775
1776 void vkCmdDbgMarkerEnd(
1777    VkCommandBuffer                              commandBuffer)
1778    __attribute__ ((visibility ("default")));
1779
1780 void vkCmdDbgMarkerBegin(
1781     VkCommandBuffer                              commandBuffer,
1782     const char*                                 pMarker)
1783 {
1784 }
1785
1786 void vkCmdDbgMarkerEnd(
1787     VkCommandBuffer                              commandBuffer)
1788 {
1789 }