src/intel/vulkan/anv_device.c

   1 /*
   2  * Copyright © 2015 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <assert.h>
  25 #include <stdbool.h>
  26 #include <string.h>
  27 #include <sys/mman.h>
  28 #include <unistd.h>
  29 #include <fcntl.h>
  30 #include <xf86drm.h>
  31
  32 #include "anv_private.h"
  33 #include "util/strtod.h"
  34 #include "util/debug.h"
  35 #include "util/build_id.h"
  36 #include "util/vk_util.h"
  37
  38 #include "genxml/gen7_pack.h"
  39
  40 static void
  41 compiler_debug_log(void *data, const char *fmt, ...)
  42 { }
  43
  44 static void
  45 compiler_perf_log(void *data, const char *fmt, ...)
  46 {
  47    va_list args;
  48    va_start(args, fmt);
  49
  50    if (unlikely(INTEL_DEBUG & DEBUG_PERF))
  51       vfprintf(stderr, fmt, args);
  52
  53    va_end(args);
  54 }
  55
  56 static bool
  57 anv_device_get_cache_uuid(void *uuid)
  58 {
  59    const struct build_id_note *note = build_id_find_nhdr("libvulkan_intel.so");
  60    if (!note)
  61       return false;
  62
  63    unsigned len = build_id_length(note);
  64    if (len < VK_UUID_SIZE)
  65       return false;
  66
  67    memcpy(uuid, build_id_data(note), VK_UUID_SIZE);
  68    return true;
  69 }
  70
  71 static VkResult
  72 anv_physical_device_init(struct anv_physical_device *device,
  73                          struct anv_instance *instance,
  74                          const char *path)
  75 {
  76    VkResult result;
  77    int fd;
  78
  79    fd = open(path, O_RDWR | O_CLOEXEC);
  80    if (fd < 0)
  81       return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  82
  83    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
  84    device->instance = instance;
  85
  86    assert(strlen(path) < ARRAY_SIZE(device->path));
  87    strncpy(device->path, path, ARRAY_SIZE(device->path));
  88
  89    device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
  90    if (!device->chipset_id) {
  91       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  92       goto fail;
  93    }
  94
  95    device->name = gen_get_device_name(device->chipset_id);
  96    if (!gen_get_device_info(device->chipset_id, &device->info)) {
  97       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  98       goto fail;
  99    }
 100
 101    if (device->info.is_haswell) {
 102       fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
 103    } else if (device->info.gen == 7 && !device->info.is_baytrail) {
 104       fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
 105    } else if (device->info.gen == 7 && device->info.is_baytrail) {
 106       fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
 107    } else if (device->info.gen >= 8) {
 108       /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
 109        * supported as anything */
 110    } else {
 111       result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 112                          "Vulkan not yet supported on %s", device->name);
 113       goto fail;
 114    }
 115
 116    device->cmd_parser_version = -1;
 117    if (device->info.gen == 7) {
 118       device->cmd_parser_version =
 119          anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
 120       if (device->cmd_parser_version == -1) {
 121          result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 122                             "failed to get command parser version");
 123          goto fail;
 124       }
 125    }
 126
 127    if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
 128       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 129                          "failed to get aperture size: %m");
 130       goto fail;
 131    }
 132
 133    if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
 134       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 135                          "kernel missing gem wait");
 136       goto fail;
 137    }
 138
 139    if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
 140       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 141                          "kernel missing execbuf2");
 142       goto fail;
 143    }
 144
 145    if (!device->info.has_llc &&
 146        anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
 147       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 148                          "kernel missing wc mmap");
 149       goto fail;
 150    }
 151
 152    if (!anv_device_get_cache_uuid(device->uuid)) {
 153       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 154                          "cannot generate UUID");
 155       goto fail;
 156    }
 157    bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
 158
 159    /* GENs prior to 8 do not support EU/Subslice info */
 160    if (device->info.gen >= 8) {
 161       device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
 162       device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
 163
 164       /* Without this information, we cannot get the right Braswell
 165        * brandstrings, and we have to use conservative numbers for GPGPU on
 166        * many platforms, but otherwise, things will just work.
 167        */
 168       if (device->subslice_total < 1 || device->eu_total < 1) {
 169          fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
 170                          " query GPU properties.\n");
 171       }
 172    } else if (device->info.gen == 7) {
 173       device->subslice_total = 1 << (device->info.gt - 1);
 174    }
 175
 176    if (device->info.is_cherryview &&
 177        device->subslice_total > 0 && device->eu_total > 0) {
 178       /* Logical CS threads = EUs per subslice * 7 threads per EU */
 179       uint32_t max_cs_threads = device->eu_total / device->subslice_total * 7;
 180
 181       /* Fuse configurations may give more threads than expected, never less. */
 182       if (max_cs_threads > device->info.max_cs_threads)
 183          device->info.max_cs_threads = max_cs_threads;
 184    }
 185
 186    brw_process_intel_debug_variable();
 187
 188    device->compiler = brw_compiler_create(NULL, &device->info);
 189    if (device->compiler == NULL) {
 190       result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 191       goto fail;
 192    }
 193    device->compiler->shader_debug_log = compiler_debug_log;
 194    device->compiler->shader_perf_log = compiler_perf_log;
 195
 196    result = anv_init_wsi(device);
 197    if (result != VK_SUCCESS) {
 198       ralloc_free(device->compiler);
 199       goto fail;
 200    }
 201
 202    isl_device_init(&device->isl_dev, &device->info, swizzled);
 203
 204    device->local_fd = fd;
 205    return VK_SUCCESS;
 206
 207 fail:
 208    close(fd);
 209    return result;
 210 }
 211
 212 static void
 213 anv_physical_device_finish(struct anv_physical_device *device)
 214 {
 215    anv_finish_wsi(device);
 216    ralloc_free(device->compiler);
 217    close(device->local_fd);
 218 }
 219
 220 static const VkExtensionProperties global_extensions[] = {
 221    {
 222       .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
 223       .specVersion = 25,
 224    },
 225 #ifdef VK_USE_PLATFORM_XCB_KHR
 226    {
 227       .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
 228       .specVersion = 6,
 229    },
 230 #endif
 231 #ifdef VK_USE_PLATFORM_XLIB_KHR
 232    {
 233       .extensionName = VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
 234       .specVersion = 6,
 235    },
 236 #endif
 237 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
 238    {
 239       .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
 240       .specVersion = 5,
 241    },
 242 #endif
 243    {
 244       .extensionName = VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
 245       .specVersion = 1,
 246    },
 247 };
 248
 249 static const VkExtensionProperties device_extensions[] = {
 250    {
 251       .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
 252       .specVersion = 68,
 253    },
 254    {
 255       .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
 256       .specVersion = 1,
 257    },
 258    {
 259       .extensionName = VK_KHR_MAINTENANCE1_EXTENSION_NAME,
 260       .specVersion = 1,
 261    },
 262    {
 263       .extensionName = VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
 264       .specVersion = 1,
 265    },
 266    {
 267       .extensionName = VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME,
 268       .specVersion = 1,
 269    },
 270    {
 271       .extensionName = VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME,
 272       .specVersion = 1,
 273    }
 274 };
 275
 276 static void *
 277 default_alloc_func(void *pUserData, size_t size, size_t align,
 278                    VkSystemAllocationScope allocationScope)
 279 {
 280    return malloc(size);
 281 }
 282
 283 static void *
 284 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
 285                      size_t align, VkSystemAllocationScope allocationScope)
 286 {
 287    return realloc(pOriginal, size);
 288 }
 289
 290 static void
 291 default_free_func(void *pUserData, void *pMemory)
 292 {
 293    free(pMemory);
 294 }
 295
 296 static const VkAllocationCallbacks default_alloc = {
 297    .pUserData = NULL,
 298    .pfnAllocation = default_alloc_func,
 299    .pfnReallocation = default_realloc_func,
 300    .pfnFree = default_free_func,
 301 };
 302
 303 VkResult anv_CreateInstance(
 304     const VkInstanceCreateInfo*                 pCreateInfo,
 305     const VkAllocationCallbacks*                pAllocator,
 306     VkInstance*                                 pInstance)
 307 {
 308    struct anv_instance *instance;
 309
 310    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 311
 312    uint32_t client_version;
 313    if (pCreateInfo->pApplicationInfo &&
 314        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 315       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 316    } else {
 317       client_version = VK_MAKE_VERSION(1, 0, 0);
 318    }
 319
 320    if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
 321        client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
 322       return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 323                        "Client requested version %d.%d.%d",
 324                        VK_VERSION_MAJOR(client_version),
 325                        VK_VERSION_MINOR(client_version),
 326                        VK_VERSION_PATCH(client_version));
 327    }
 328
 329    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 330       bool found = false;
 331       for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
 332          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 333                     global_extensions[j].extensionName) == 0) {
 334             found = true;
 335             break;
 336          }
 337       }
 338       if (!found)
 339          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 340    }
 341
 342    instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 343                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 344    if (!instance)
 345       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 346
 347    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 348
 349    if (pAllocator)
 350       instance->alloc = *pAllocator;
 351    else
 352       instance->alloc = default_alloc;
 353
 354    instance->apiVersion = client_version;
 355    instance->physicalDeviceCount = -1;
 356
 357    _mesa_locale_init();
 358
 359    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 360
 361    *pInstance = anv_instance_to_handle(instance);
 362
 363    return VK_SUCCESS;
 364 }
 365
 366 void anv_DestroyInstance(
 367     VkInstance                                  _instance,
 368     const VkAllocationCallbacks*                pAllocator)
 369 {
 370    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 371
 372    if (!instance)
 373       return;
 374
 375    if (instance->physicalDeviceCount > 0) {
 376       /* We support at most one physical device. */
 377       assert(instance->physicalDeviceCount == 1);
 378       anv_physical_device_finish(&instance->physicalDevice);
 379    }
 380
 381    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 382
 383    _mesa_locale_fini();
 384
 385    vk_free(&instance->alloc, instance);
 386 }
 387
 388 static VkResult
 389 anv_enumerate_devices(struct anv_instance *instance)
 390 {
 391    /* TODO: Check for more devices ? */
 392    drmDevicePtr devices[8];
 393    VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
 394    int max_devices;
 395
 396    instance->physicalDeviceCount = 0;
 397
 398    max_devices = drmGetDevices2(0, devices, sizeof(devices));
 399    if (max_devices < 1)
 400       return VK_ERROR_INCOMPATIBLE_DRIVER;
 401
 402    for (unsigned i = 0; i < (unsigned)max_devices; i++) {
 403       if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
 404           devices[i]->bustype == DRM_BUS_PCI &&
 405           devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
 406
 407          result = anv_physical_device_init(&instance->physicalDevice,
 408                         instance,
 409                         devices[i]->nodes[DRM_NODE_RENDER]);
 410          if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 411             break;
 412       }
 413    }
 414
 415    if (result == VK_SUCCESS)
 416       instance->physicalDeviceCount = 1;
 417
 418    return result;
 419 }
 420
 421
 422 VkResult anv_EnumeratePhysicalDevices(
 423     VkInstance                                  _instance,
 424     uint32_t*                                   pPhysicalDeviceCount,
 425     VkPhysicalDevice*                           pPhysicalDevices)
 426 {
 427    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 428    VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
 429    VkResult result;
 430
 431    if (instance->physicalDeviceCount < 0) {
 432       result = anv_enumerate_devices(instance);
 433       if (result != VK_SUCCESS &&
 434           result != VK_ERROR_INCOMPATIBLE_DRIVER)
 435          return result;
 436    }
 437
 438    if (instance->physicalDeviceCount > 0) {
 439       assert(instance->physicalDeviceCount == 1);
 440       vk_outarray_append(&out, i) {
 441          *i = anv_physical_device_to_handle(&instance->physicalDevice);
 442       }
 443    }
 444
 445    return vk_outarray_status(&out);
 446 }
 447
 448 void anv_GetPhysicalDeviceFeatures(
 449     VkPhysicalDevice                            physicalDevice,
 450     VkPhysicalDeviceFeatures*                   pFeatures)
 451 {
 452    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 453
 454    *pFeatures = (VkPhysicalDeviceFeatures) {
 455       .robustBufferAccess                       = true,
 456       .fullDrawIndexUint32                      = true,
 457       .imageCubeArray                           = true,
 458       .independentBlend                         = true,
 459       .geometryShader                           = true,
 460       .tessellationShader                       = true,
 461       .sampleRateShading                        = true,
 462       .dualSrcBlend                             = true,
 463       .logicOp                                  = true,
 464       .multiDrawIndirect                        = false,
 465       .drawIndirectFirstInstance                = true,
 466       .depthClamp                               = true,
 467       .depthBiasClamp                           = true,
 468       .fillModeNonSolid                         = true,
 469       .depthBounds                              = false,
 470       .wideLines                                = true,
 471       .largePoints                              = true,
 472       .alphaToOne                               = true,
 473       .multiViewport                            = true,
 474       .samplerAnisotropy                        = true,
 475       .textureCompressionETC2                   = pdevice->info.gen >= 8 ||
 476                                                   pdevice->info.is_baytrail,
 477       .textureCompressionASTC_LDR               = pdevice->info.gen >= 9, /* FINISHME CHV */
 478       .textureCompressionBC                     = true,
 479       .occlusionQueryPrecise                    = true,
 480       .pipelineStatisticsQuery                  = false,
 481       .fragmentStoresAndAtomics                 = true,
 482       .shaderTessellationAndGeometryPointSize   = true,
 483       .shaderImageGatherExtended                = true,
 484       .shaderStorageImageExtendedFormats        = true,
 485       .shaderStorageImageMultisample            = false,
 486       .shaderStorageImageReadWithoutFormat      = false,
 487       .shaderStorageImageWriteWithoutFormat     = true,
 488       .shaderUniformBufferArrayDynamicIndexing  = true,
 489       .shaderSampledImageArrayDynamicIndexing   = true,
 490       .shaderStorageBufferArrayDynamicIndexing  = true,
 491       .shaderStorageImageArrayDynamicIndexing   = true,
 492       .shaderClipDistance                       = true,
 493       .shaderCullDistance                       = true,
 494       .shaderFloat64                            = pdevice->info.gen >= 8,
 495       .shaderInt64                              = pdevice->info.gen >= 8,
 496       .shaderInt16                              = false,
 497       .shaderResourceMinLod                     = false,
 498       .variableMultisampleRate                  = false,
 499       .inheritedQueries                         = false,
 500    };
 501
 502    /* We can't do image stores in vec4 shaders */
 503    pFeatures->vertexPipelineStoresAndAtomics =
 504       pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
 505       pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
 506 }
 507
 508 void anv_GetPhysicalDeviceFeatures2KHR(
 509     VkPhysicalDevice                            physicalDevice,
 510     VkPhysicalDeviceFeatures2KHR*               pFeatures)
 511 {
 512    anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 513
 514    vk_foreach_struct(ext, pFeatures->pNext) {
 515       switch (ext->sType) {
 516       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 517          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 518             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 519
 520          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 521          break;
 522       }
 523
 524       default:
 525          anv_debug_ignored_stype(ext->sType);
 526          break;
 527       }
 528    }
 529 }
 530
 531 void anv_GetPhysicalDeviceProperties(
 532     VkPhysicalDevice                            physicalDevice,
 533     VkPhysicalDeviceProperties*                 pProperties)
 534 {
 535    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 536    const struct gen_device_info *devinfo = &pdevice->info;
 537
 538    /* See assertions made when programming the buffer surface state. */
 539    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
 540                                       (1ul << 30) : (1ul << 27);
 541
 542    VkSampleCountFlags sample_counts =
 543       isl_device_get_sample_counts(&pdevice->isl_dev);
 544
 545    VkPhysicalDeviceLimits limits = {
 546       .maxImageDimension1D                      = (1 << 14),
 547       .maxImageDimension2D                      = (1 << 14),
 548       .maxImageDimension3D                      = (1 << 11),
 549       .maxImageDimensionCube                    = (1 << 14),
 550       .maxImageArrayLayers                      = (1 << 11),
 551       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 552       .maxUniformBufferRange                    = (1ul << 27),
 553       .maxStorageBufferRange                    = max_raw_buffer_sz,
 554       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 555       .maxMemoryAllocationCount                 = UINT32_MAX,
 556       .maxSamplerAllocationCount                = 64 * 1024,
 557       .bufferImageGranularity                   = 64, /* A cache line */
 558       .sparseAddressSpaceSize                   = 0,
 559       .maxBoundDescriptorSets                   = MAX_SETS,
 560       .maxPerStageDescriptorSamplers            = 64,
 561       .maxPerStageDescriptorUniformBuffers      = 64,
 562       .maxPerStageDescriptorStorageBuffers      = 64,
 563       .maxPerStageDescriptorSampledImages       = 64,
 564       .maxPerStageDescriptorStorageImages       = 64,
 565       .maxPerStageDescriptorInputAttachments    = 64,
 566       .maxPerStageResources                     = 128,
 567       .maxDescriptorSetSamplers                 = 256,
 568       .maxDescriptorSetUniformBuffers           = 256,
 569       .maxDescriptorSetUniformBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
 570       .maxDescriptorSetStorageBuffers           = 256,
 571       .maxDescriptorSetStorageBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
 572       .maxDescriptorSetSampledImages            = 256,
 573       .maxDescriptorSetStorageImages            = 256,
 574       .maxDescriptorSetInputAttachments         = 256,
 575       .maxVertexInputAttributes                 = MAX_VBS,
 576       .maxVertexInputBindings                   = MAX_VBS,
 577       .maxVertexInputAttributeOffset            = 2047,
 578       .maxVertexInputBindingStride              = 2048,
 579       .maxVertexOutputComponents                = 128,
 580       .maxTessellationGenerationLevel           = 64,
 581       .maxTessellationPatchSize                 = 32,
 582       .maxTessellationControlPerVertexInputComponents = 128,
 583       .maxTessellationControlPerVertexOutputComponents = 128,
 584       .maxTessellationControlPerPatchOutputComponents = 128,
 585       .maxTessellationControlTotalOutputComponents = 2048,
 586       .maxTessellationEvaluationInputComponents = 128,
 587       .maxTessellationEvaluationOutputComponents = 128,
 588       .maxGeometryShaderInvocations             = 32,
 589       .maxGeometryInputComponents               = 64,
 590       .maxGeometryOutputComponents              = 128,
 591       .maxGeometryOutputVertices                = 256,
 592       .maxGeometryTotalOutputComponents         = 1024,
 593       .maxFragmentInputComponents               = 128,
 594       .maxFragmentOutputAttachments             = 8,
 595       .maxFragmentDualSrcAttachments            = 1,
 596       .maxFragmentCombinedOutputResources       = 8,
 597       .maxComputeSharedMemorySize               = 32768,
 598       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 599       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 600       .maxComputeWorkGroupSize = {
 601          16 * devinfo->max_cs_threads,
 602          16 * devinfo->max_cs_threads,
 603          16 * devinfo->max_cs_threads,
 604       },
 605       .subPixelPrecisionBits                    = 4 /* FIXME */,
 606       .subTexelPrecisionBits                    = 4 /* FIXME */,
 607       .mipmapPrecisionBits                      = 4 /* FIXME */,
 608       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 609       .maxDrawIndirectCount                     = UINT32_MAX,
 610       .maxSamplerLodBias                        = 16,
 611       .maxSamplerAnisotropy                     = 16,
 612       .maxViewports                             = MAX_VIEWPORTS,
 613       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 614       .viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
 615       .viewportSubPixelBits                     = 13, /* We take a float? */
 616       .minMemoryMapAlignment                    = 4096, /* A page */
 617       .minTexelBufferOffsetAlignment            = 1,
 618       .minUniformBufferOffsetAlignment          = 16,
 619       .minStorageBufferOffsetAlignment          = 4,
 620       .minTexelOffset                           = -8,
 621       .maxTexelOffset                           = 7,
 622       .minTexelGatherOffset                     = -32,
 623       .maxTexelGatherOffset                     = 31,
 624       .minInterpolationOffset                   = -0.5,
 625       .maxInterpolationOffset                   = 0.4375,
 626       .subPixelInterpolationOffsetBits          = 4,
 627       .maxFramebufferWidth                      = (1 << 14),
 628       .maxFramebufferHeight                     = (1 << 14),
 629       .maxFramebufferLayers                     = (1 << 11),
 630       .framebufferColorSampleCounts             = sample_counts,
 631       .framebufferDepthSampleCounts             = sample_counts,
 632       .framebufferStencilSampleCounts           = sample_counts,
 633       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 634       .maxColorAttachments                      = MAX_RTS,
 635       .sampledImageColorSampleCounts            = sample_counts,
 636       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 637       .sampledImageDepthSampleCounts            = sample_counts,
 638       .sampledImageStencilSampleCounts          = sample_counts,
 639       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 640       .maxSampleMaskWords                       = 1,
 641       .timestampComputeAndGraphics              = false,
 642       .timestampPeriod                          = devinfo->timebase_scale,
 643       .maxClipDistances                         = 8,
 644       .maxCullDistances                         = 8,
 645       .maxCombinedClipAndCullDistances          = 8,
 646       .discreteQueuePriorities                  = 1,
 647       .pointSizeRange                           = { 0.125, 255.875 },
 648       .lineWidthRange                           = { 0.0, 7.9921875 },
 649       .pointSizeGranularity                     = (1.0 / 8.0),
 650       .lineWidthGranularity                     = (1.0 / 128.0),
 651       .strictLines                              = false, /* FINISHME */
 652       .standardSampleLocations                  = true,
 653       .optimalBufferCopyOffsetAlignment         = 128,
 654       .optimalBufferCopyRowPitchAlignment       = 128,
 655       .nonCoherentAtomSize                      = 64,
 656    };
 657
 658    *pProperties = (VkPhysicalDeviceProperties) {
 659       .apiVersion = VK_MAKE_VERSION(1, 0, 42),
 660       .driverVersion = 1,
 661       .vendorID = 0x8086,
 662       .deviceID = pdevice->chipset_id,
 663       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 664       .limits = limits,
 665       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 666    };
 667
 668    strcpy(pProperties->deviceName, pdevice->name);
 669    memcpy(pProperties->pipelineCacheUUID, pdevice->uuid, VK_UUID_SIZE);
 670 }
 671
 672 void anv_GetPhysicalDeviceProperties2KHR(
 673     VkPhysicalDevice                            physicalDevice,
 674     VkPhysicalDeviceProperties2KHR*             pProperties)
 675 {
 676    anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 677
 678    vk_foreach_struct(ext, pProperties->pNext) {
 679       switch (ext->sType) {
 680       default:
 681          anv_debug_ignored_stype(ext->sType);
 682          break;
 683       }
 684    }
 685 }
 686
 687 /* We support exactly one queue family. */
 688 static const VkQueueFamilyProperties
 689 anv_queue_family_properties = {
 690    .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 691                  VK_QUEUE_COMPUTE_BIT |
 692                  VK_QUEUE_TRANSFER_BIT,
 693    .queueCount = 1,
 694    .timestampValidBits = 36, /* XXX: Real value here */
 695    .minImageTransferGranularity = { 1, 1, 1 },
 696 };
 697
 698 void anv_GetPhysicalDeviceQueueFamilyProperties(
 699     VkPhysicalDevice                            physicalDevice,
 700     uint32_t*                                   pCount,
 701     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 702 {
 703    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
 704
 705    vk_outarray_append(&out, p) {
 706       *p = anv_queue_family_properties;
 707    }
 708 }
 709
 710 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
 711     VkPhysicalDevice                            physicalDevice,
 712     uint32_t*                                   pQueueFamilyPropertyCount,
 713     VkQueueFamilyProperties2KHR*                pQueueFamilyProperties)
 714 {
 715
 716    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 717
 718    vk_outarray_append(&out, p) {
 719       p->queueFamilyProperties = anv_queue_family_properties;
 720
 721       vk_foreach_struct(s, p->pNext) {
 722          anv_debug_ignored_stype(s->sType);
 723       }
 724    }
 725 }
 726
 727 void anv_GetPhysicalDeviceMemoryProperties(
 728     VkPhysicalDevice                            physicalDevice,
 729     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
 730 {
 731    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 732    VkDeviceSize heap_size;
 733
 734    /* Reserve some wiggle room for the driver by exposing only 75% of the
 735     * aperture to the heap.
 736     */
 737    heap_size = 3 * physical_device->aperture_size / 4;
 738
 739    if (physical_device->info.has_llc) {
 740       /* Big core GPUs share LLC with the CPU and thus one memory type can be
 741        * both cached and coherent at the same time.
 742        */
 743       pMemoryProperties->memoryTypeCount = 1;
 744       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 745          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 746                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 747                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
 748                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 749          .heapIndex = 0,
 750       };
 751    } else {
 752       /* The spec requires that we expose a host-visible, coherent memory
 753        * type, but Atom GPUs don't share LLC. Thus we offer two memory types
 754        * to give the application a choice between cached, but not coherent and
 755        * coherent but uncached (WC though).
 756        */
 757       pMemoryProperties->memoryTypeCount = 2;
 758       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 759          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 760                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 761                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
 762          .heapIndex = 0,
 763       };
 764       pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
 765          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 766                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 767                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 768          .heapIndex = 0,
 769       };
 770    }
 771
 772    pMemoryProperties->memoryHeapCount = 1;
 773    pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
 774       .size = heap_size,
 775       .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 776    };
 777 }
 778
 779 void anv_GetPhysicalDeviceMemoryProperties2KHR(
 780     VkPhysicalDevice                            physicalDevice,
 781     VkPhysicalDeviceMemoryProperties2KHR*       pMemoryProperties)
 782 {
 783    anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
 784                                          &pMemoryProperties->memoryProperties);
 785
 786    vk_foreach_struct(ext, pMemoryProperties->pNext) {
 787       switch (ext->sType) {
 788       default:
 789          anv_debug_ignored_stype(ext->sType);
 790          break;
 791       }
 792    }
 793 }
 794
 795 PFN_vkVoidFunction anv_GetInstanceProcAddr(
 796     VkInstance                                  instance,
 797     const char*                                 pName)
 798 {
 799    return anv_lookup_entrypoint(NULL, pName);
 800 }
 801
 802 /* With version 1+ of the loader interface the ICD should expose
 803  * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
 804  */
 805 PUBLIC
 806 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 807     VkInstance                                  instance,
 808     const char*                                 pName);
 809
 810 PUBLIC
 811 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 812     VkInstance                                  instance,
 813     const char*                                 pName)
 814 {
 815    return anv_GetInstanceProcAddr(instance, pName);
 816 }
 817
 818 PFN_vkVoidFunction anv_GetDeviceProcAddr(
 819     VkDevice                                    _device,
 820     const char*                                 pName)
 821 {
 822    ANV_FROM_HANDLE(anv_device, device, _device);
 823    return anv_lookup_entrypoint(&device->info, pName);
 824 }
 825
 826 static void
 827 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
 828 {
 829    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 830    queue->device = device;
 831    queue->pool = &device->surface_state_pool;
 832 }
 833
 834 static void
 835 anv_queue_finish(struct anv_queue *queue)
 836 {
 837 }
 838
 839 static struct anv_state
 840 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
 841 {
 842    struct anv_state state;
 843
 844    state = anv_state_pool_alloc(pool, size, align);
 845    memcpy(state.map, p, size);
 846
 847    anv_state_flush(pool->block_pool->device, state);
 848
 849    return state;
 850 }
 851
 852 struct gen8_border_color {
 853    union {
 854       float float32[4];
 855       uint32_t uint32[4];
 856    };
 857    /* Pad out to 64 bytes */
 858    uint32_t _pad[12];
 859 };
 860
 861 static void
 862 anv_device_init_border_colors(struct anv_device *device)
 863 {
 864    static const struct gen8_border_color border_colors[] = {
 865       [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] =  { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
 866       [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] =       { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
 867       [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] =       { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
 868       [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] =    { .uint32 = { 0, 0, 0, 0 } },
 869       [VK_BORDER_COLOR_INT_OPAQUE_BLACK] =         { .uint32 = { 0, 0, 0, 1 } },
 870       [VK_BORDER_COLOR_INT_OPAQUE_WHITE] =         { .uint32 = { 1, 1, 1, 1 } },
 871    };
 872
 873    device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
 874                                                     sizeof(border_colors), 64,
 875                                                     border_colors);
 876 }
 877
 878 VkResult
 879 anv_device_submit_simple_batch(struct anv_device *device,
 880                                struct anv_batch *batch)
 881 {
 882    struct drm_i915_gem_execbuffer2 execbuf;
 883    struct drm_i915_gem_exec_object2 exec2_objects[1];
 884    struct anv_bo bo, *exec_bos[1];
 885    VkResult result = VK_SUCCESS;
 886    uint32_t size;
 887    int64_t timeout;
 888    int ret;
 889
 890    /* Kernel driver requires 8 byte aligned batch length */
 891    size = align_u32(batch->next - batch->start, 8);
 892    result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
 893    if (result != VK_SUCCESS)
 894       return result;
 895
 896    memcpy(bo.map, batch->start, size);
 897    if (!device->info.has_llc)
 898       anv_flush_range(bo.map, size);
 899
 900    exec_bos[0] = &bo;
 901    exec2_objects[0].handle = bo.gem_handle;
 902    exec2_objects[0].relocation_count = 0;
 903    exec2_objects[0].relocs_ptr = 0;
 904    exec2_objects[0].alignment = 0;
 905    exec2_objects[0].offset = bo.offset;
 906    exec2_objects[0].flags = 0;
 907    exec2_objects[0].rsvd1 = 0;
 908    exec2_objects[0].rsvd2 = 0;
 909
 910    execbuf.buffers_ptr = (uintptr_t) exec2_objects;
 911    execbuf.buffer_count = 1;
 912    execbuf.batch_start_offset = 0;
 913    execbuf.batch_len = size;
 914    execbuf.cliprects_ptr = 0;
 915    execbuf.num_cliprects = 0;
 916    execbuf.DR1 = 0;
 917    execbuf.DR4 = 0;
 918
 919    execbuf.flags =
 920       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
 921    execbuf.rsvd1 = device->context_id;
 922    execbuf.rsvd2 = 0;
 923
 924    result = anv_device_execbuf(device, &execbuf, exec_bos);
 925    if (result != VK_SUCCESS)
 926       goto fail;
 927
 928    timeout = INT64_MAX;
 929    ret = anv_gem_wait(device, bo.gem_handle, &timeout);
 930    if (ret != 0) {
 931       /* We don't know the real error. */
 932       result = vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
 933       goto fail;
 934    }
 935
 936  fail:
 937    anv_bo_pool_free(&device->batch_bo_pool, &bo);
 938
 939    return result;
 940 }
 941
 942 VkResult anv_CreateDevice(
 943     VkPhysicalDevice                            physicalDevice,
 944     const VkDeviceCreateInfo*                   pCreateInfo,
 945     const VkAllocationCallbacks*                pAllocator,
 946     VkDevice*                                   pDevice)
 947 {
 948    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 949    VkResult result;
 950    struct anv_device *device;
 951
 952    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
 953
 954    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 955       bool found = false;
 956       for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
 957          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 958                     device_extensions[j].extensionName) == 0) {
 959             found = true;
 960             break;
 961          }
 962       }
 963       if (!found)
 964          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 965    }
 966
 967    device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
 968                        sizeof(*device), 8,
 969                        VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
 970    if (!device)
 971       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 972
 973    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 974    device->instance = physical_device->instance;
 975    device->chipset_id = physical_device->chipset_id;
 976
 977    if (pAllocator)
 978       device->alloc = *pAllocator;
 979    else
 980       device->alloc = physical_device->instance->alloc;
 981
 982    /* XXX(chadv): Can we dup() physicalDevice->fd here? */
 983    device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
 984    if (device->fd == -1) {
 985       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
 986       goto fail_device;
 987    }
 988
 989    device->context_id = anv_gem_create_context(device);
 990    if (device->context_id == -1) {
 991       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
 992       goto fail_fd;
 993    }
 994
 995    device->info = physical_device->info;
 996    device->isl_dev = physical_device->isl_dev;
 997
 998    /* On Broadwell and later, we can use batch chaining to more efficiently
 999     * implement growing command buffers.  Prior to Haswell, the kernel
1000     * command parser gets in the way and we have to fall back to growing
1001     * the batch.
1002     */
1003    device->can_chain_batches = device->info.gen >= 8;
1004
1005    device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1006       pCreateInfo->pEnabledFeatures->robustBufferAccess;
1007
1008    if (pthread_mutex_init(&device->mutex, NULL) != 0) {
1009       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1010       goto fail_context_id;
1011    }
1012
1013    pthread_condattr_t condattr;
1014    if (pthread_condattr_init(&condattr) != 0) {
1015       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1016       goto fail_mutex;
1017    }
1018    if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
1019       pthread_condattr_destroy(&condattr);
1020       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1021       goto fail_mutex;
1022    }
1023    if (pthread_cond_init(&device->queue_submit, NULL) != 0) {
1024       pthread_condattr_destroy(&condattr);
1025       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1026       goto fail_mutex;
1027    }
1028    pthread_condattr_destroy(&condattr);
1029
1030    anv_bo_pool_init(&device->batch_bo_pool, device);
1031
1032    result = anv_block_pool_init(&device->dynamic_state_block_pool, device,
1033                                 16384);
1034    if (result != VK_SUCCESS)
1035       goto fail_batch_bo_pool;
1036
1037    anv_state_pool_init(&device->dynamic_state_pool,
1038                        &device->dynamic_state_block_pool);
1039
1040    result = anv_block_pool_init(&device->instruction_block_pool, device,
1041                                 1024 * 1024);
1042    if (result != VK_SUCCESS)
1043       goto fail_dynamic_state_pool;
1044
1045    anv_state_pool_init(&device->instruction_state_pool,
1046                        &device->instruction_block_pool);
1047
1048    result = anv_block_pool_init(&device->surface_state_block_pool, device,
1049                                 4096);
1050    if (result != VK_SUCCESS)
1051       goto fail_instruction_state_pool;
1052
1053    anv_state_pool_init(&device->surface_state_pool,
1054                        &device->surface_state_block_pool);
1055
1056    result = anv_bo_init_new(&device->workaround_bo, device, 1024);
1057    if (result != VK_SUCCESS)
1058       goto fail_surface_state_pool;
1059
1060    anv_scratch_pool_init(device, &device->scratch_pool);
1061
1062    anv_queue_init(device, &device->queue);
1063
1064    switch (device->info.gen) {
1065    case 7:
1066       if (!device->info.is_haswell)
1067          result = gen7_init_device_state(device);
1068       else
1069          result = gen75_init_device_state(device);
1070       break;
1071    case 8:
1072       result = gen8_init_device_state(device);
1073       break;
1074    case 9:
1075       result = gen9_init_device_state(device);
1076       break;
1077    default:
1078       /* Shouldn't get here as we don't create physical devices for any other
1079        * gens. */
1080       unreachable("unhandled gen");
1081    }
1082    if (result != VK_SUCCESS)
1083       goto fail_workaround_bo;
1084
1085    anv_device_init_blorp(device);
1086
1087    anv_device_init_border_colors(device);
1088
1089    *pDevice = anv_device_to_handle(device);
1090
1091    return VK_SUCCESS;
1092
1093  fail_workaround_bo:
1094    anv_queue_finish(&device->queue);
1095    anv_scratch_pool_finish(device, &device->scratch_pool);
1096    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1097    anv_gem_close(device, device->workaround_bo.gem_handle);
1098  fail_surface_state_pool:
1099    anv_state_pool_finish(&device->surface_state_pool);
1100    anv_block_pool_finish(&device->surface_state_block_pool);
1101  fail_instruction_state_pool:
1102    anv_state_pool_finish(&device->instruction_state_pool);
1103    anv_block_pool_finish(&device->instruction_block_pool);
1104  fail_dynamic_state_pool:
1105    anv_state_pool_finish(&device->dynamic_state_pool);
1106    anv_block_pool_finish(&device->dynamic_state_block_pool);
1107  fail_batch_bo_pool:
1108    anv_bo_pool_finish(&device->batch_bo_pool);
1109    pthread_cond_destroy(&device->queue_submit);
1110  fail_mutex:
1111    pthread_mutex_destroy(&device->mutex);
1112  fail_context_id:
1113    anv_gem_destroy_context(device, device->context_id);
1114  fail_fd:
1115    close(device->fd);
1116  fail_device:
1117    vk_free(&device->alloc, device);
1118
1119    return result;
1120 }
1121
1122 void anv_DestroyDevice(
1123     VkDevice                                    _device,
1124     const VkAllocationCallbacks*                pAllocator)
1125 {
1126    ANV_FROM_HANDLE(anv_device, device, _device);
1127
1128    if (!device)
1129       return;
1130
1131    anv_device_finish_blorp(device);
1132
1133    anv_queue_finish(&device->queue);
1134
1135 #ifdef HAVE_VALGRIND
1136    /* We only need to free these to prevent valgrind errors.  The backing
1137     * BO will go away in a couple of lines so we don't actually leak.
1138     */
1139    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1140 #endif
1141
1142    anv_scratch_pool_finish(device, &device->scratch_pool);
1143
1144    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1145    anv_gem_close(device, device->workaround_bo.gem_handle);
1146
1147    anv_state_pool_finish(&device->surface_state_pool);
1148    anv_block_pool_finish(&device->surface_state_block_pool);
1149    anv_state_pool_finish(&device->instruction_state_pool);
1150    anv_block_pool_finish(&device->instruction_block_pool);
1151    anv_state_pool_finish(&device->dynamic_state_pool);
1152    anv_block_pool_finish(&device->dynamic_state_block_pool);
1153
1154    anv_bo_pool_finish(&device->batch_bo_pool);
1155
1156    pthread_cond_destroy(&device->queue_submit);
1157    pthread_mutex_destroy(&device->mutex);
1158
1159    anv_gem_destroy_context(device, device->context_id);
1160
1161    close(device->fd);
1162
1163    vk_free(&device->alloc, device);
1164 }
1165
1166 VkResult anv_EnumerateInstanceExtensionProperties(
1167     const char*                                 pLayerName,
1168     uint32_t*                                   pPropertyCount,
1169     VkExtensionProperties*                      pProperties)
1170 {
1171    if (pProperties == NULL) {
1172       *pPropertyCount = ARRAY_SIZE(global_extensions);
1173       return VK_SUCCESS;
1174    }
1175
1176    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(global_extensions));
1177    typed_memcpy(pProperties, global_extensions, *pPropertyCount);
1178
1179    if (*pPropertyCount < ARRAY_SIZE(global_extensions))
1180       return VK_INCOMPLETE;
1181
1182    return VK_SUCCESS;
1183 }
1184
1185 VkResult anv_EnumerateDeviceExtensionProperties(
1186     VkPhysicalDevice                            physicalDevice,
1187     const char*                                 pLayerName,
1188     uint32_t*                                   pPropertyCount,
1189     VkExtensionProperties*                      pProperties)
1190 {
1191    if (pProperties == NULL) {
1192       *pPropertyCount = ARRAY_SIZE(device_extensions);
1193       return VK_SUCCESS;
1194    }
1195
1196    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(device_extensions));
1197    typed_memcpy(pProperties, device_extensions, *pPropertyCount);
1198
1199    if (*pPropertyCount < ARRAY_SIZE(device_extensions))
1200       return VK_INCOMPLETE;
1201
1202    return VK_SUCCESS;
1203 }
1204
1205 VkResult anv_EnumerateInstanceLayerProperties(
1206     uint32_t*                                   pPropertyCount,
1207     VkLayerProperties*                          pProperties)
1208 {
1209    if (pProperties == NULL) {
1210       *pPropertyCount = 0;
1211       return VK_SUCCESS;
1212    }
1213
1214    /* None supported at this time */
1215    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1216 }
1217
1218 VkResult anv_EnumerateDeviceLayerProperties(
1219     VkPhysicalDevice                            physicalDevice,
1220     uint32_t*                                   pPropertyCount,
1221     VkLayerProperties*                          pProperties)
1222 {
1223    if (pProperties == NULL) {
1224       *pPropertyCount = 0;
1225       return VK_SUCCESS;
1226    }
1227
1228    /* None supported at this time */
1229    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1230 }
1231
1232 void anv_GetDeviceQueue(
1233     VkDevice                                    _device,
1234     uint32_t                                    queueNodeIndex,
1235     uint32_t                                    queueIndex,
1236     VkQueue*                                    pQueue)
1237 {
1238    ANV_FROM_HANDLE(anv_device, device, _device);
1239
1240    assert(queueIndex == 0);
1241
1242    *pQueue = anv_queue_to_handle(&device->queue);
1243 }
1244
1245 VkResult
1246 anv_device_execbuf(struct anv_device *device,
1247                    struct drm_i915_gem_execbuffer2 *execbuf,
1248                    struct anv_bo **execbuf_bos)
1249 {
1250    int ret = anv_gem_execbuffer(device, execbuf);
1251    if (ret != 0) {
1252       /* We don't know the real error. */
1253       return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
1254    }
1255
1256    struct drm_i915_gem_exec_object2 *objects =
1257       (void *)(uintptr_t)execbuf->buffers_ptr;
1258    for (uint32_t k = 0; k < execbuf->buffer_count; k++)
1259       execbuf_bos[k]->offset = objects[k].offset;
1260
1261    return VK_SUCCESS;
1262 }
1263
1264 VkResult anv_QueueSubmit(
1265     VkQueue                                     _queue,
1266     uint32_t                                    submitCount,
1267     const VkSubmitInfo*                         pSubmits,
1268     VkFence                                     _fence)
1269 {
1270    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1271    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1272    struct anv_device *device = queue->device;
1273    VkResult result = VK_SUCCESS;
1274
1275    /* We lock around QueueSubmit for three main reasons:
1276     *
1277     *  1) When a block pool is resized, we create a new gem handle with a
1278     *     different size and, in the case of surface states, possibly a
1279     *     different center offset but we re-use the same anv_bo struct when
1280     *     we do so.  If this happens in the middle of setting up an execbuf,
1281     *     we could end up with our list of BOs out of sync with our list of
1282     *     gem handles.
1283     *
1284     *  2) The algorithm we use for building the list of unique buffers isn't
1285     *     thread-safe.  While the client is supposed to syncronize around
1286     *     QueueSubmit, this would be extremely difficult to debug if it ever
1287     *     came up in the wild due to a broken app.  It's better to play it
1288     *     safe and just lock around QueueSubmit.
1289     *
1290     *  3)  The anv_cmd_buffer_execbuf function may perform relocations in
1291     *      userspace.  Due to the fact that the surface state buffer is shared
1292     *      between batches, we can't afford to have that happen from multiple
1293     *      threads at the same time.  Even though the user is supposed to
1294     *      ensure this doesn't happen, we play it safe as in (2) above.
1295     *
1296     * Since the only other things that ever take the device lock such as block
1297     * pool resize only rarely happen, this will almost never be contended so
1298     * taking a lock isn't really an expensive operation in this case.
1299     */
1300    pthread_mutex_lock(&device->mutex);
1301
1302    for (uint32_t i = 0; i < submitCount; i++) {
1303       for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1304          ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1305                          pSubmits[i].pCommandBuffers[j]);
1306          assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1307          assert(!anv_batch_has_error(&cmd_buffer->batch));
1308
1309          result = anv_cmd_buffer_execbuf(device, cmd_buffer);
1310          if (result != VK_SUCCESS)
1311             goto out;
1312       }
1313    }
1314
1315    if (fence) {
1316       struct anv_bo *fence_bo = &fence->bo;
1317       result = anv_device_execbuf(device, &fence->execbuf, &fence_bo);
1318       if (result != VK_SUCCESS)
1319          goto out;
1320
1321       /* Update the fence and wake up any waiters */
1322       assert(fence->state == ANV_FENCE_STATE_RESET);
1323       fence->state = ANV_FENCE_STATE_SUBMITTED;
1324       pthread_cond_broadcast(&device->queue_submit);
1325    }
1326
1327 out:
1328    pthread_mutex_unlock(&device->mutex);
1329
1330    return result;
1331 }
1332
1333 VkResult anv_QueueWaitIdle(
1334     VkQueue                                     _queue)
1335 {
1336    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1337
1338    return anv_DeviceWaitIdle(anv_device_to_handle(queue->device));
1339 }
1340
1341 VkResult anv_DeviceWaitIdle(
1342     VkDevice                                    _device)
1343 {
1344    ANV_FROM_HANDLE(anv_device, device, _device);
1345    struct anv_batch batch;
1346
1347    uint32_t cmds[8];
1348    batch.start = batch.next = cmds;
1349    batch.end = (void *) cmds + sizeof(cmds);
1350
1351    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1352    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1353
1354    return anv_device_submit_simple_batch(device, &batch);
1355 }
1356
1357 VkResult
1358 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1359 {
1360    uint32_t gem_handle = anv_gem_create(device, size);
1361    if (!gem_handle)
1362       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1363
1364    anv_bo_init(bo, gem_handle, size);
1365
1366    return VK_SUCCESS;
1367 }
1368
1369 VkResult anv_AllocateMemory(
1370     VkDevice                                    _device,
1371     const VkMemoryAllocateInfo*                 pAllocateInfo,
1372     const VkAllocationCallbacks*                pAllocator,
1373     VkDeviceMemory*                             pMem)
1374 {
1375    ANV_FROM_HANDLE(anv_device, device, _device);
1376    struct anv_device_memory *mem;
1377    VkResult result;
1378
1379    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1380
1381    /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1382    assert(pAllocateInfo->allocationSize > 0);
1383
1384    /* We support exactly one memory heap. */
1385    assert(pAllocateInfo->memoryTypeIndex == 0 ||
1386           (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1387
1388    /* FINISHME: Fail if allocation request exceeds heap size. */
1389
1390    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1391                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1392    if (mem == NULL)
1393       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1394
1395    /* The kernel is going to give us whole pages anyway */
1396    uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1397
1398    result = anv_bo_init_new(&mem->bo, device, alloc_size);
1399    if (result != VK_SUCCESS)
1400       goto fail;
1401
1402    mem->type_index = pAllocateInfo->memoryTypeIndex;
1403
1404    mem->map = NULL;
1405    mem->map_size = 0;
1406
1407    *pMem = anv_device_memory_to_handle(mem);
1408
1409    return VK_SUCCESS;
1410
1411  fail:
1412    vk_free2(&device->alloc, pAllocator, mem);
1413
1414    return result;
1415 }
1416
1417 void anv_FreeMemory(
1418     VkDevice                                    _device,
1419     VkDeviceMemory                              _mem,
1420     const VkAllocationCallbacks*                pAllocator)
1421 {
1422    ANV_FROM_HANDLE(anv_device, device, _device);
1423    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1424
1425    if (mem == NULL)
1426       return;
1427
1428    if (mem->map)
1429       anv_UnmapMemory(_device, _mem);
1430
1431    if (mem->bo.map)
1432       anv_gem_munmap(mem->bo.map, mem->bo.size);
1433
1434    if (mem->bo.gem_handle != 0)
1435       anv_gem_close(device, mem->bo.gem_handle);
1436
1437    vk_free2(&device->alloc, pAllocator, mem);
1438 }
1439
1440 VkResult anv_MapMemory(
1441     VkDevice                                    _device,
1442     VkDeviceMemory                              _memory,
1443     VkDeviceSize                                offset,
1444     VkDeviceSize                                size,
1445     VkMemoryMapFlags                            flags,
1446     void**                                      ppData)
1447 {
1448    ANV_FROM_HANDLE(anv_device, device, _device);
1449    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1450
1451    if (mem == NULL) {
1452       *ppData = NULL;
1453       return VK_SUCCESS;
1454    }
1455
1456    if (size == VK_WHOLE_SIZE)
1457       size = mem->bo.size - offset;
1458
1459    /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1460     *
1461     *  * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1462     *    assert(size != 0);
1463     *  * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1464     *    equal to the size of the memory minus offset
1465     */
1466    assert(size > 0);
1467    assert(offset + size <= mem->bo.size);
1468
1469    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1470     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1471     * at a time is valid. We could just mmap up front and return an offset
1472     * pointer here, but that may exhaust virtual memory on 32 bit
1473     * userspace. */
1474
1475    uint32_t gem_flags = 0;
1476    if (!device->info.has_llc && mem->type_index == 0)
1477       gem_flags |= I915_MMAP_WC;
1478
1479    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1480    uint64_t map_offset = offset & ~4095ull;
1481    assert(offset >= map_offset);
1482    uint64_t map_size = (offset + size) - map_offset;
1483
1484    /* Let's map whole pages */
1485    map_size = align_u64(map_size, 4096);
1486
1487    void *map = anv_gem_mmap(device, mem->bo.gem_handle,
1488                             map_offset, map_size, gem_flags);
1489    if (map == MAP_FAILED)
1490       return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
1491
1492    mem->map = map;
1493    mem->map_size = map_size;
1494
1495    *ppData = mem->map + (offset - map_offset);
1496
1497    return VK_SUCCESS;
1498 }
1499
1500 void anv_UnmapMemory(
1501     VkDevice                                    _device,
1502     VkDeviceMemory                              _memory)
1503 {
1504    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1505
1506    if (mem == NULL)
1507       return;
1508
1509    anv_gem_munmap(mem->map, mem->map_size);
1510
1511    mem->map = NULL;
1512    mem->map_size = 0;
1513 }
1514
1515 static void
1516 clflush_mapped_ranges(struct anv_device         *device,
1517                       uint32_t                   count,
1518                       const VkMappedMemoryRange *ranges)
1519 {
1520    for (uint32_t i = 0; i < count; i++) {
1521       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1522       if (ranges[i].offset >= mem->map_size)
1523          continue;
1524
1525       anv_clflush_range(mem->map + ranges[i].offset,
1526                         MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
1527    }
1528 }
1529
1530 VkResult anv_FlushMappedMemoryRanges(
1531     VkDevice                                    _device,
1532     uint32_t                                    memoryRangeCount,
1533     const VkMappedMemoryRange*                  pMemoryRanges)
1534 {
1535    ANV_FROM_HANDLE(anv_device, device, _device);
1536
1537    if (device->info.has_llc)
1538       return VK_SUCCESS;
1539
1540    /* Make sure the writes we're flushing have landed. */
1541    __builtin_ia32_mfence();
1542
1543    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1544
1545    return VK_SUCCESS;
1546 }
1547
1548 VkResult anv_InvalidateMappedMemoryRanges(
1549     VkDevice                                    _device,
1550     uint32_t                                    memoryRangeCount,
1551     const VkMappedMemoryRange*                  pMemoryRanges)
1552 {
1553    ANV_FROM_HANDLE(anv_device, device, _device);
1554
1555    if (device->info.has_llc)
1556       return VK_SUCCESS;
1557
1558    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1559
1560    /* Make sure no reads get moved up above the invalidate. */
1561    __builtin_ia32_mfence();
1562
1563    return VK_SUCCESS;
1564 }
1565
1566 void anv_GetBufferMemoryRequirements(
1567     VkDevice                                    _device,
1568     VkBuffer                                    _buffer,
1569     VkMemoryRequirements*                       pMemoryRequirements)
1570 {
1571    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1572    ANV_FROM_HANDLE(anv_device, device, _device);
1573
1574    /* The Vulkan spec (git aaed022) says:
1575     *
1576     *    memoryTypeBits is a bitfield and contains one bit set for every
1577     *    supported memory type for the resource. The bit `1<<i` is set if and
1578     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1579     *    structure for the physical device is supported.
1580     *
1581     * We support exactly one memory type on LLC, two on non-LLC.
1582     */
1583    pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1584
1585    pMemoryRequirements->size = buffer->size;
1586    pMemoryRequirements->alignment = 16;
1587 }
1588
1589 void anv_GetImageMemoryRequirements(
1590     VkDevice                                    _device,
1591     VkImage                                     _image,
1592     VkMemoryRequirements*                       pMemoryRequirements)
1593 {
1594    ANV_FROM_HANDLE(anv_image, image, _image);
1595    ANV_FROM_HANDLE(anv_device, device, _device);
1596
1597    /* The Vulkan spec (git aaed022) says:
1598     *
1599     *    memoryTypeBits is a bitfield and contains one bit set for every
1600     *    supported memory type for the resource. The bit `1<<i` is set if and
1601     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1602     *    structure for the physical device is supported.
1603     *
1604     * We support exactly one memory type on LLC, two on non-LLC.
1605     */
1606    pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1607
1608    pMemoryRequirements->size = image->size;
1609    pMemoryRequirements->alignment = image->alignment;
1610 }
1611
1612 void anv_GetImageSparseMemoryRequirements(
1613     VkDevice                                    device,
1614     VkImage                                     image,
1615     uint32_t*                                   pSparseMemoryRequirementCount,
1616     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1617 {
1618    *pSparseMemoryRequirementCount = 0;
1619 }
1620
1621 void anv_GetDeviceMemoryCommitment(
1622     VkDevice                                    device,
1623     VkDeviceMemory                              memory,
1624     VkDeviceSize*                               pCommittedMemoryInBytes)
1625 {
1626    *pCommittedMemoryInBytes = 0;
1627 }
1628
1629 VkResult anv_BindBufferMemory(
1630     VkDevice                                    device,
1631     VkBuffer                                    _buffer,
1632     VkDeviceMemory                              _memory,
1633     VkDeviceSize                                memoryOffset)
1634 {
1635    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1636    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1637
1638    if (mem) {
1639       buffer->bo = &mem->bo;
1640       buffer->offset = memoryOffset;
1641    } else {
1642       buffer->bo = NULL;
1643       buffer->offset = 0;
1644    }
1645
1646    return VK_SUCCESS;
1647 }
1648
1649 VkResult anv_QueueBindSparse(
1650     VkQueue                                     queue,
1651     uint32_t                                    bindInfoCount,
1652     const VkBindSparseInfo*                     pBindInfo,
1653     VkFence                                     fence)
1654 {
1655    return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1656 }
1657
1658 VkResult anv_CreateFence(
1659     VkDevice                                    _device,
1660     const VkFenceCreateInfo*                    pCreateInfo,
1661     const VkAllocationCallbacks*                pAllocator,
1662     VkFence*                                    pFence)
1663 {
1664    ANV_FROM_HANDLE(anv_device, device, _device);
1665    struct anv_bo fence_bo;
1666    struct anv_fence *fence;
1667    struct anv_batch batch;
1668    VkResult result;
1669
1670    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1671
1672    result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1673    if (result != VK_SUCCESS)
1674       return result;
1675
1676    /* Fences are small.  Just store the CPU data structure in the BO. */
1677    fence = fence_bo.map;
1678    fence->bo = fence_bo;
1679
1680    /* Place the batch after the CPU data but on its own cache line. */
1681    const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1682    batch.next = batch.start = fence->bo.map + batch_offset;
1683    batch.end = fence->bo.map + fence->bo.size;
1684    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1685    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1686
1687    if (!device->info.has_llc) {
1688       assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1689       assert(batch.next - batch.start <= CACHELINE_SIZE);
1690       __builtin_ia32_mfence();
1691       __builtin_ia32_clflush(batch.start);
1692    }
1693
1694    fence->exec2_objects[0].handle = fence->bo.gem_handle;
1695    fence->exec2_objects[0].relocation_count = 0;
1696    fence->exec2_objects[0].relocs_ptr = 0;
1697    fence->exec2_objects[0].alignment = 0;
1698    fence->exec2_objects[0].offset = fence->bo.offset;
1699    fence->exec2_objects[0].flags = 0;
1700    fence->exec2_objects[0].rsvd1 = 0;
1701    fence->exec2_objects[0].rsvd2 = 0;
1702
1703    fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1704    fence->execbuf.buffer_count = 1;
1705    fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1706    fence->execbuf.batch_len = batch.next - batch.start;
1707    fence->execbuf.cliprects_ptr = 0;
1708    fence->execbuf.num_cliprects = 0;
1709    fence->execbuf.DR1 = 0;
1710    fence->execbuf.DR4 = 0;
1711
1712    fence->execbuf.flags =
1713       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1714    fence->execbuf.rsvd1 = device->context_id;
1715    fence->execbuf.rsvd2 = 0;
1716
1717    if (pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT) {
1718       fence->state = ANV_FENCE_STATE_SIGNALED;
1719    } else {
1720       fence->state = ANV_FENCE_STATE_RESET;
1721    }
1722
1723    *pFence = anv_fence_to_handle(fence);
1724
1725    return VK_SUCCESS;
1726 }
1727
1728 void anv_DestroyFence(
1729     VkDevice                                    _device,
1730     VkFence                                     _fence,
1731     const VkAllocationCallbacks*                pAllocator)
1732 {
1733    ANV_FROM_HANDLE(anv_device, device, _device);
1734    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1735
1736    if (!fence)
1737       return;
1738
1739    assert(fence->bo.map == fence);
1740    anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1741 }
1742
1743 VkResult anv_ResetFences(
1744     VkDevice                                    _device,
1745     uint32_t                                    fenceCount,
1746     const VkFence*                              pFences)
1747 {
1748    for (uint32_t i = 0; i < fenceCount; i++) {
1749       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1750       fence->state = ANV_FENCE_STATE_RESET;
1751    }
1752
1753    return VK_SUCCESS;
1754 }
1755
1756 VkResult anv_GetFenceStatus(
1757     VkDevice                                    _device,
1758     VkFence                                     _fence)
1759 {
1760    ANV_FROM_HANDLE(anv_device, device, _device);
1761    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1762    int64_t t = 0;
1763    int ret;
1764
1765    switch (fence->state) {
1766    case ANV_FENCE_STATE_RESET:
1767       /* If it hasn't even been sent off to the GPU yet, it's not ready */
1768       return VK_NOT_READY;
1769
1770    case ANV_FENCE_STATE_SIGNALED:
1771       /* It's been signaled, return success */
1772       return VK_SUCCESS;
1773
1774    case ANV_FENCE_STATE_SUBMITTED:
1775       /* It's been submitted to the GPU but we don't know if it's done yet. */
1776       ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1777       if (ret == 0) {
1778          fence->state = ANV_FENCE_STATE_SIGNALED;
1779          return VK_SUCCESS;
1780       } else {
1781          return VK_NOT_READY;
1782       }
1783    default:
1784       unreachable("Invalid fence status");
1785    }
1786 }
1787
1788 #define NSEC_PER_SEC 1000000000
1789 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
1790
1791 VkResult anv_WaitForFences(
1792     VkDevice                                    _device,
1793     uint32_t                                    fenceCount,
1794     const VkFence*                              pFences,
1795     VkBool32                                    waitAll,
1796     uint64_t                                    _timeout)
1797 {
1798    ANV_FROM_HANDLE(anv_device, device, _device);
1799    int ret;
1800
1801    /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1802     * to block indefinitely timeouts <= 0.  Unfortunately, this was broken
1803     * for a couple of kernel releases.  Since there's no way to know
1804     * whether or not the kernel we're using is one of the broken ones, the
1805     * best we can do is to clamp the timeout to INT64_MAX.  This limits the
1806     * maximum timeout from 584 years to 292 years - likely not a big deal.
1807     */
1808    int64_t timeout = MIN2(_timeout, INT64_MAX);
1809
1810    uint32_t pending_fences = fenceCount;
1811    while (pending_fences) {
1812       pending_fences = 0;
1813       bool signaled_fences = false;
1814       for (uint32_t i = 0; i < fenceCount; i++) {
1815          ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1816          switch (fence->state) {
1817          case ANV_FENCE_STATE_RESET:
1818             /* This fence hasn't been submitted yet, we'll catch it the next
1819              * time around.  Yes, this may mean we dead-loop but, short of
1820              * lots of locking and a condition variable, there's not much that
1821              * we can do about that.
1822              */
1823             pending_fences++;
1824             continue;
1825
1826          case ANV_FENCE_STATE_SIGNALED:
1827             /* This fence is not pending.  If waitAll isn't set, we can return
1828              * early.  Otherwise, we have to keep going.
1829              */
1830             if (!waitAll)
1831                return VK_SUCCESS;
1832             continue;
1833
1834          case ANV_FENCE_STATE_SUBMITTED:
1835             /* These are the fences we really care about.  Go ahead and wait
1836              * on it until we hit a timeout.
1837              */
1838             ret = anv_gem_wait(device, fence->bo.gem_handle, &timeout);
1839             if (ret == -1 && errno == ETIME) {
1840                return VK_TIMEOUT;
1841             } else if (ret == -1) {
1842                /* We don't know the real error. */
1843                return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1844             } else {
1845                fence->state = ANV_FENCE_STATE_SIGNALED;
1846                signaled_fences = true;
1847                if (!waitAll)
1848                   return VK_SUCCESS;
1849                continue;
1850             }
1851          }
1852       }
1853
1854       if (pending_fences && !signaled_fences) {
1855          /* If we've hit this then someone decided to vkWaitForFences before
1856           * they've actually submitted any of them to a queue.  This is a
1857           * fairly pessimal case, so it's ok to lock here and use a standard
1858           * pthreads condition variable.
1859           */
1860          pthread_mutex_lock(&device->mutex);
1861
1862          /* It's possible that some of the fences have changed state since the
1863           * last time we checked.  Now that we have the lock, check for
1864           * pending fences again and don't wait if it's changed.
1865           */
1866          uint32_t now_pending_fences = 0;
1867          for (uint32_t i = 0; i < fenceCount; i++) {
1868             ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1869             if (fence->state == ANV_FENCE_STATE_RESET)
1870                now_pending_fences++;
1871          }
1872          assert(now_pending_fences <= pending_fences);
1873
1874          if (now_pending_fences == pending_fences) {
1875             struct timespec before;
1876             clock_gettime(CLOCK_MONOTONIC, &before);
1877
1878             uint32_t abs_nsec = before.tv_nsec + timeout % NSEC_PER_SEC;
1879             uint64_t abs_sec = before.tv_sec + (abs_nsec / NSEC_PER_SEC) +
1880                                (timeout / NSEC_PER_SEC);
1881             abs_nsec %= NSEC_PER_SEC;
1882
1883             /* Avoid roll-over in tv_sec on 32-bit systems if the user
1884              * provided timeout is UINT64_MAX
1885              */
1886             struct timespec abstime;
1887             abstime.tv_nsec = abs_nsec;
1888             abstime.tv_sec = MIN2(abs_sec, INT_TYPE_MAX(abstime.tv_sec));
1889
1890             ret = pthread_cond_timedwait(&device->queue_submit,
1891                                          &device->mutex, &abstime);
1892             assert(ret != EINVAL);
1893
1894             struct timespec after;
1895             clock_gettime(CLOCK_MONOTONIC, &after);
1896             uint64_t time_elapsed =
1897                ((uint64_t)after.tv_sec * NSEC_PER_SEC + after.tv_nsec) -
1898                ((uint64_t)before.tv_sec * NSEC_PER_SEC + before.tv_nsec);
1899
1900             if (time_elapsed >= timeout) {
1901                pthread_mutex_unlock(&device->mutex);
1902                return VK_TIMEOUT;
1903             }
1904
1905             timeout -= time_elapsed;
1906          }
1907
1908          pthread_mutex_unlock(&device->mutex);
1909       }
1910    }
1911
1912    return VK_SUCCESS;
1913 }
1914
1915 // Queue semaphore functions
1916
1917 VkResult anv_CreateSemaphore(
1918     VkDevice                                    device,
1919     const VkSemaphoreCreateInfo*                pCreateInfo,
1920     const VkAllocationCallbacks*                pAllocator,
1921     VkSemaphore*                                pSemaphore)
1922 {
1923    /* The DRM execbuffer ioctl always execute in-oder, even between different
1924     * rings. As such, there's nothing to do for the user space semaphore.
1925     */
1926
1927    *pSemaphore = (VkSemaphore)1;
1928
1929    return VK_SUCCESS;
1930 }
1931
1932 void anv_DestroySemaphore(
1933     VkDevice                                    device,
1934     VkSemaphore                                 semaphore,
1935     const VkAllocationCallbacks*                pAllocator)
1936 {
1937 }
1938
1939 // Event functions
1940
1941 VkResult anv_CreateEvent(
1942     VkDevice                                    _device,
1943     const VkEventCreateInfo*                    pCreateInfo,
1944     const VkAllocationCallbacks*                pAllocator,
1945     VkEvent*                                    pEvent)
1946 {
1947    ANV_FROM_HANDLE(anv_device, device, _device);
1948    struct anv_state state;
1949    struct anv_event *event;
1950
1951    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1952
1953    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1954                                 sizeof(*event), 8);
1955    event = state.map;
1956    event->state = state;
1957    event->semaphore = VK_EVENT_RESET;
1958
1959    if (!device->info.has_llc) {
1960       /* Make sure the writes we're flushing have landed. */
1961       __builtin_ia32_mfence();
1962       __builtin_ia32_clflush(event);
1963    }
1964
1965    *pEvent = anv_event_to_handle(event);
1966
1967    return VK_SUCCESS;
1968 }
1969
1970 void anv_DestroyEvent(
1971     VkDevice                                    _device,
1972     VkEvent                                     _event,
1973     const VkAllocationCallbacks*                pAllocator)
1974 {
1975    ANV_FROM_HANDLE(anv_device, device, _device);
1976    ANV_FROM_HANDLE(anv_event, event, _event);
1977
1978    if (!event)
1979       return;
1980
1981    anv_state_pool_free(&device->dynamic_state_pool, event->state);
1982 }
1983
1984 VkResult anv_GetEventStatus(
1985     VkDevice                                    _device,
1986     VkEvent                                     _event)
1987 {
1988    ANV_FROM_HANDLE(anv_device, device, _device);
1989    ANV_FROM_HANDLE(anv_event, event, _event);
1990
1991    if (!device->info.has_llc) {
1992       /* Invalidate read cache before reading event written by GPU. */
1993       __builtin_ia32_clflush(event);
1994       __builtin_ia32_mfence();
1995
1996    }
1997
1998    return event->semaphore;
1999 }
2000
2001 VkResult anv_SetEvent(
2002     VkDevice                                    _device,
2003     VkEvent                                     _event)
2004 {
2005    ANV_FROM_HANDLE(anv_device, device, _device);
2006    ANV_FROM_HANDLE(anv_event, event, _event);
2007
2008    event->semaphore = VK_EVENT_SET;
2009
2010    if (!device->info.has_llc) {
2011       /* Make sure the writes we're flushing have landed. */
2012       __builtin_ia32_mfence();
2013       __builtin_ia32_clflush(event);
2014    }
2015
2016    return VK_SUCCESS;
2017 }
2018
2019 VkResult anv_ResetEvent(
2020     VkDevice                                    _device,
2021     VkEvent                                     _event)
2022 {
2023    ANV_FROM_HANDLE(anv_device, device, _device);
2024    ANV_FROM_HANDLE(anv_event, event, _event);
2025
2026    event->semaphore = VK_EVENT_RESET;
2027
2028    if (!device->info.has_llc) {
2029       /* Make sure the writes we're flushing have landed. */
2030       __builtin_ia32_mfence();
2031       __builtin_ia32_clflush(event);
2032    }
2033
2034    return VK_SUCCESS;
2035 }
2036
2037 // Buffer functions
2038
2039 VkResult anv_CreateBuffer(
2040     VkDevice                                    _device,
2041     const VkBufferCreateInfo*                   pCreateInfo,
2042     const VkAllocationCallbacks*                pAllocator,
2043     VkBuffer*                                   pBuffer)
2044 {
2045    ANV_FROM_HANDLE(anv_device, device, _device);
2046    struct anv_buffer *buffer;
2047
2048    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2049
2050    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2051                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2052    if (buffer == NULL)
2053       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2054
2055    buffer->size = pCreateInfo->size;
2056    buffer->usage = pCreateInfo->usage;
2057    buffer->bo = NULL;
2058    buffer->offset = 0;
2059
2060    *pBuffer = anv_buffer_to_handle(buffer);
2061
2062    return VK_SUCCESS;
2063 }
2064
2065 void anv_DestroyBuffer(
2066     VkDevice                                    _device,
2067     VkBuffer                                    _buffer,
2068     const VkAllocationCallbacks*                pAllocator)
2069 {
2070    ANV_FROM_HANDLE(anv_device, device, _device);
2071    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2072
2073    if (!buffer)
2074       return;
2075
2076    vk_free2(&device->alloc, pAllocator, buffer);
2077 }
2078
2079 void
2080 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
2081                               enum isl_format format,
2082                               uint32_t offset, uint32_t range, uint32_t stride)
2083 {
2084    isl_buffer_fill_state(&device->isl_dev, state.map,
2085                          .address = offset,
2086                          .mocs = device->default_mocs,
2087                          .size = range,
2088                          .format = format,
2089                          .stride = stride);
2090
2091    anv_state_flush(device, state);
2092 }
2093
2094 void anv_DestroySampler(
2095     VkDevice                                    _device,
2096     VkSampler                                   _sampler,
2097     const VkAllocationCallbacks*                pAllocator)
2098 {
2099    ANV_FROM_HANDLE(anv_device, device, _device);
2100    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
2101
2102    if (!sampler)
2103       return;
2104
2105    vk_free2(&device->alloc, pAllocator, sampler);
2106 }
2107
2108 VkResult anv_CreateFramebuffer(
2109     VkDevice                                    _device,
2110     const VkFramebufferCreateInfo*              pCreateInfo,
2111     const VkAllocationCallbacks*                pAllocator,
2112     VkFramebuffer*                              pFramebuffer)
2113 {
2114    ANV_FROM_HANDLE(anv_device, device, _device);
2115    struct anv_framebuffer *framebuffer;
2116
2117    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2118
2119    size_t size = sizeof(*framebuffer) +
2120                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2121    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2122                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2123    if (framebuffer == NULL)
2124       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2125
2126    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2127    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2128       VkImageView _iview = pCreateInfo->pAttachments[i];
2129       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2130    }
2131
2132    framebuffer->width = pCreateInfo->width;
2133    framebuffer->height = pCreateInfo->height;
2134    framebuffer->layers = pCreateInfo->layers;
2135
2136    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2137
2138    return VK_SUCCESS;
2139 }
2140
2141 void anv_DestroyFramebuffer(
2142     VkDevice                                    _device,
2143     VkFramebuffer                               _fb,
2144     const VkAllocationCallbacks*                pAllocator)
2145 {
2146    ANV_FROM_HANDLE(anv_device, device, _device);
2147    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2148
2149    if (!fb)
2150       return;
2151
2152    vk_free2(&device->alloc, pAllocator, fb);
2153 }
2154
2155 /* vk_icd.h does not declare this function, so we declare it here to
2156  * suppress Wmissing-prototypes.
2157  */
2158 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2159 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2160
2161 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2162 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2163 {
2164    /* For the full details on loader interface versioning, see
2165     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2166     * What follows is a condensed summary, to help you navigate the large and
2167     * confusing official doc.
2168     *
2169     *   - Loader interface v0 is incompatible with later versions. We don't
2170     *     support it.
2171     *
2172     *   - In loader interface v1:
2173     *       - The first ICD entrypoint called by the loader is
2174     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2175     *         entrypoint.
2176     *       - The ICD must statically expose no other Vulkan symbol unless it is
2177     *         linked with -Bsymbolic.
2178     *       - Each dispatchable Vulkan handle created by the ICD must be
2179     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2180     *         ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2181     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2182     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2183     *         such loader-managed surfaces.
2184     *
2185     *    - Loader interface v2 differs from v1 in:
2186     *       - The first ICD entrypoint called by the loader is
2187     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2188     *         statically expose this entrypoint.
2189     *
2190     *    - Loader interface v3 differs from v2 in:
2191     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2192     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2193     *          because the loader no longer does so.
2194     */
2195    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2196    return VK_SUCCESS;
2197 }