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 <sys/sysinfo.h>
  29 #include <unistd.h>
  30 #include <fcntl.h>
  31 #include <xf86drm.h>
  32
  33 #include "anv_private.h"
  34 #include "util/strtod.h"
  35 #include "util/debug.h"
  36 #include "util/build_id.h"
  37 #include "util/mesa-sha1.h"
  38 #include "vk_util.h"
  39
  40 #include "genxml/gen7_pack.h"
  41
  42 static void
  43 compiler_debug_log(void *data, const char *fmt, ...)
  44 { }
  45
  46 static void
  47 compiler_perf_log(void *data, const char *fmt, ...)
  48 {
  49    va_list args;
  50    va_start(args, fmt);
  51
  52    if (unlikely(INTEL_DEBUG & DEBUG_PERF))
  53       vfprintf(stderr, fmt, args);
  54
  55    va_end(args);
  56 }
  57
  58 static VkResult
  59 anv_compute_heap_size(int fd, uint64_t *heap_size)
  60 {
  61    uint64_t gtt_size;
  62    if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
  63                                  &gtt_size) == -1) {
  64       /* If, for whatever reason, we can't actually get the GTT size from the
  65        * kernel (too old?) fall back to the aperture size.
  66        */
  67       anv_perf_warn("Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
  68
  69       if (anv_gem_get_aperture(fd, &gtt_size) == -1) {
  70          return vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
  71                           "failed to get aperture size: %m");
  72       }
  73    }
  74
  75    /* Query the total ram from the system */
  76    struct sysinfo info;
  77    sysinfo(&info);
  78
  79    uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
  80
  81    /* We don't want to burn too much ram with the GPU.  If the user has 4GiB
  82     * or less, we use at most half.  If they have more than 4GiB, we use 3/4.
  83     */
  84    uint64_t available_ram;
  85    if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
  86       available_ram = total_ram / 2;
  87    else
  88       available_ram = total_ram * 3 / 4;
  89
  90    /* We also want to leave some padding for things we allocate in the driver,
  91     * so don't go over 3/4 of the GTT either.
  92     */
  93    uint64_t available_gtt = gtt_size * 3 / 4;
  94
  95    *heap_size = MIN2(available_ram, available_gtt);
  96
  97    return VK_SUCCESS;
  98 }
  99
 100 static VkResult
 101 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
 102 {
 103    /* The kernel query only tells us whether or not the kernel supports the
 104     * EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and not whether or not the
 105     * hardware has actual 48bit address support.
 106     */
 107    device->supports_48bit_addresses =
 108       (device->info.gen >= 8) && anv_gem_supports_48b_addresses(fd);
 109
 110    uint64_t heap_size;
 111    VkResult result = anv_compute_heap_size(fd, &heap_size);
 112    if (result != VK_SUCCESS)
 113       return result;
 114
 115    if (heap_size <= 3ull * (1ull << 30)) {
 116       /* In this case, everything fits nicely into the 32-bit address space,
 117        * so there's no need for supporting 48bit addresses on client-allocated
 118        * memory objects.
 119        */
 120       device->memory.heap_count = 1;
 121       device->memory.heaps[0] = (struct anv_memory_heap) {
 122          .size = heap_size,
 123          .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 124          .supports_48bit_addresses = false,
 125       };
 126    } else {
 127       /* Not everything will fit nicely into a 32-bit address space.  In this
 128        * case we need a 64-bit heap.  Advertise a small 32-bit heap and a
 129        * larger 48-bit heap.  If we're in this case, then we have a total heap
 130        * size larger than 3GiB which most likely means they have 8 GiB of
 131        * video memory and so carving off 1 GiB for the 32-bit heap should be
 132        * reasonable.
 133        */
 134       const uint64_t heap_size_32bit = 1ull << 30;
 135       const uint64_t heap_size_48bit = heap_size - heap_size_32bit;
 136
 137       assert(device->supports_48bit_addresses);
 138
 139       device->memory.heap_count = 2;
 140       device->memory.heaps[0] = (struct anv_memory_heap) {
 141          .size = heap_size_48bit,
 142          .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 143          .supports_48bit_addresses = true,
 144       };
 145       device->memory.heaps[1] = (struct anv_memory_heap) {
 146          .size = heap_size_32bit,
 147          .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 148          .supports_48bit_addresses = false,
 149       };
 150    }
 151
 152    uint32_t type_count = 0;
 153    for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
 154       uint32_t valid_buffer_usage = ~0;
 155
 156       /* There appears to be a hardware issue in the VF cache where it only
 157        * considers the bottom 32 bits of memory addresses.  If you happen to
 158        * have two vertex buffers which get placed exactly 4 GiB apart and use
 159        * them in back-to-back draw calls, you can get collisions.  In order to
 160        * solve this problem, we require vertex and index buffers be bound to
 161        * memory allocated out of the 32-bit heap.
 162        */
 163       if (device->memory.heaps[heap].supports_48bit_addresses) {
 164          valid_buffer_usage &= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
 165                                  VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
 166       }
 167
 168       if (device->info.has_llc) {
 169          /* Big core GPUs share LLC with the CPU and thus one memory type can be
 170           * both cached and coherent at the same time.
 171           */
 172          device->memory.types[type_count++] = (struct anv_memory_type) {
 173             .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 174                              VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 175                              VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
 176                              VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 177             .heapIndex = heap,
 178             .valid_buffer_usage = valid_buffer_usage,
 179          };
 180       } else {
 181          /* The spec requires that we expose a host-visible, coherent memory
 182           * type, but Atom GPUs don't share LLC. Thus we offer two memory types
 183           * to give the application a choice between cached, but not coherent and
 184           * coherent but uncached (WC though).
 185           */
 186          device->memory.types[type_count++] = (struct anv_memory_type) {
 187             .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 188                              VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 189                              VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
 190             .heapIndex = heap,
 191             .valid_buffer_usage = valid_buffer_usage,
 192          };
 193          device->memory.types[type_count++] = (struct anv_memory_type) {
 194             .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 195                              VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 196                              VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 197             .heapIndex = heap,
 198             .valid_buffer_usage = valid_buffer_usage,
 199          };
 200       }
 201    }
 202    device->memory.type_count = type_count;
 203
 204    return VK_SUCCESS;
 205 }
 206
 207 static VkResult
 208 anv_physical_device_init_uuids(struct anv_physical_device *device)
 209 {
 210    const struct build_id_note *note = build_id_find_nhdr("libvulkan_intel.so");
 211    if (!note) {
 212       return vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 213                        "Failed to find build-id");
 214    }
 215
 216    unsigned build_id_len = build_id_length(note);
 217    if (build_id_len < 20) {
 218       return vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 219                        "build-id too short.  It needs to be a SHA");
 220    }
 221
 222    struct mesa_sha1 sha1_ctx;
 223    uint8_t sha1[20];
 224    STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
 225
 226    /* The pipeline cache UUID is used for determining when a pipeline cache is
 227     * invalid.  It needs both a driver build and the PCI ID of the device.
 228     */
 229    _mesa_sha1_init(&sha1_ctx);
 230    _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
 231    _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
 232                      sizeof(device->chipset_id));
 233    _mesa_sha1_final(&sha1_ctx, sha1);
 234    memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
 235
 236    /* The driver UUID is used for determining sharability of images and memory
 237     * between two Vulkan instances in separate processes.  People who want to
 238     * share memory need to also check the device UUID (below) so all this
 239     * needs to be is the build-id.
 240     */
 241    memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
 242
 243    /* The device UUID uniquely identifies the given device within the machine.
 244     * Since we never have more than one device, this doesn't need to be a real
 245     * UUID.  However, on the off-chance that someone tries to use this to
 246     * cache pre-tiled images or something of the like, we use the PCI ID and
 247     * some bits of ISL info to ensure that this is safe.
 248     */
 249    _mesa_sha1_init(&sha1_ctx);
 250    _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
 251                      sizeof(device->chipset_id));
 252    _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
 253                      sizeof(device->isl_dev.has_bit6_swizzling));
 254    _mesa_sha1_final(&sha1_ctx, sha1);
 255    memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
 256
 257    return VK_SUCCESS;
 258 }
 259
 260 static VkResult
 261 anv_physical_device_init(struct anv_physical_device *device,
 262                          struct anv_instance *instance,
 263                          const char *path)
 264 {
 265    VkResult result;
 266    int fd;
 267
 268    fd = open(path, O_RDWR | O_CLOEXEC);
 269    if (fd < 0)
 270       return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
 271
 272    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 273    device->instance = instance;
 274
 275    assert(strlen(path) < ARRAY_SIZE(device->path));
 276    strncpy(device->path, path, ARRAY_SIZE(device->path));
 277
 278    device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
 279    if (!device->chipset_id) {
 280       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
 281       goto fail;
 282    }
 283
 284    device->name = gen_get_device_name(device->chipset_id);
 285    if (!gen_get_device_info(device->chipset_id, &device->info)) {
 286       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
 287       goto fail;
 288    }
 289
 290    if (device->info.is_haswell) {
 291       fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
 292    } else if (device->info.gen == 7 && !device->info.is_baytrail) {
 293       fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
 294    } else if (device->info.gen == 7 && device->info.is_baytrail) {
 295       fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
 296    } else if (device->info.gen >= 8 && device->info.gen <= 9) {
 297       /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
 298        * supported as anything */
 299    } else {
 300       result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 301                          "Vulkan not yet supported on %s", device->name);
 302       goto fail;
 303    }
 304
 305    device->cmd_parser_version = -1;
 306    if (device->info.gen == 7) {
 307       device->cmd_parser_version =
 308          anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
 309       if (device->cmd_parser_version == -1) {
 310          result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 311                             "failed to get command parser version");
 312          goto fail;
 313       }
 314    }
 315
 316    if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
 317       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 318                          "kernel missing gem wait");
 319       goto fail;
 320    }
 321
 322    if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
 323       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 324                          "kernel missing execbuf2");
 325       goto fail;
 326    }
 327
 328    if (!device->info.has_llc &&
 329        anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
 330       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 331                          "kernel missing wc mmap");
 332       goto fail;
 333    }
 334
 335    result = anv_physical_device_init_heaps(device, fd);
 336    if (result != VK_SUCCESS)
 337       goto fail;
 338
 339    device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
 340
 341    bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
 342
 343    /* GENs prior to 8 do not support EU/Subslice info */
 344    if (device->info.gen >= 8) {
 345       device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
 346       device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
 347
 348       /* Without this information, we cannot get the right Braswell
 349        * brandstrings, and we have to use conservative numbers for GPGPU on
 350        * many platforms, but otherwise, things will just work.
 351        */
 352       if (device->subslice_total < 1 || device->eu_total < 1) {
 353          fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
 354                          " query GPU properties.\n");
 355       }
 356    } else if (device->info.gen == 7) {
 357       device->subslice_total = 1 << (device->info.gt - 1);
 358    }
 359
 360    if (device->info.is_cherryview &&
 361        device->subslice_total > 0 && device->eu_total > 0) {
 362       /* Logical CS threads = EUs per subslice * num threads per EU */
 363       uint32_t max_cs_threads =
 364          device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
 365
 366       /* Fuse configurations may give more threads than expected, never less. */
 367       if (max_cs_threads > device->info.max_cs_threads)
 368          device->info.max_cs_threads = max_cs_threads;
 369    }
 370
 371    brw_process_intel_debug_variable();
 372
 373    device->compiler = brw_compiler_create(NULL, &device->info);
 374    if (device->compiler == NULL) {
 375       result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 376       goto fail;
 377    }
 378    device->compiler->shader_debug_log = compiler_debug_log;
 379    device->compiler->shader_perf_log = compiler_perf_log;
 380
 381    isl_device_init(&device->isl_dev, &device->info, swizzled);
 382
 383    result = anv_physical_device_init_uuids(device);
 384    if (result != VK_SUCCESS)
 385       goto fail;
 386
 387    result = anv_init_wsi(device);
 388    if (result != VK_SUCCESS) {
 389       ralloc_free(device->compiler);
 390       goto fail;
 391    }
 392
 393    device->local_fd = fd;
 394    return VK_SUCCESS;
 395
 396 fail:
 397    close(fd);
 398    return result;
 399 }
 400
 401 static void
 402 anv_physical_device_finish(struct anv_physical_device *device)
 403 {
 404    anv_finish_wsi(device);
 405    ralloc_free(device->compiler);
 406    close(device->local_fd);
 407 }
 408
 409 static const VkExtensionProperties global_extensions[] = {
 410    {
 411       .extensionName = VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
 412       .specVersion = 1,
 413    },
 414    {
 415       .extensionName = VK_KHR_GET_SURFACE_CAPABILITIES_2_EXTENSION_NAME,
 416       .specVersion = 1,
 417    },
 418    {
 419       .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
 420       .specVersion = 25,
 421    },
 422 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
 423    {
 424       .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
 425       .specVersion = 5,
 426    },
 427 #endif
 428 #ifdef VK_USE_PLATFORM_XCB_KHR
 429    {
 430       .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
 431       .specVersion = 6,
 432    },
 433 #endif
 434 #ifdef VK_USE_PLATFORM_XLIB_KHR
 435    {
 436       .extensionName = VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
 437       .specVersion = 6,
 438    },
 439 #endif
 440    {
 441       .extensionName = VK_KHX_EXTERNAL_MEMORY_CAPABILITIES_EXTENSION_NAME,
 442       .specVersion = 1,
 443    },
 444    {
 445       .extensionName = VK_KHX_EXTERNAL_SEMAPHORE_CAPABILITIES_EXTENSION_NAME,
 446       .specVersion = 1,
 447    },
 448 };
 449
 450 static const VkExtensionProperties device_extensions[] = {
 451    {
 452       .extensionName = VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME,
 453       .specVersion = 1,
 454    },
 455    {
 456       .extensionName = VK_KHR_INCREMENTAL_PRESENT_EXTENSION_NAME,
 457       .specVersion = 1,
 458    },
 459    {
 460       .extensionName = VK_KHR_MAINTENANCE1_EXTENSION_NAME,
 461       .specVersion = 1,
 462    },
 463    {
 464       .extensionName = VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME,
 465       .specVersion = 1,
 466    },
 467    {
 468       .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
 469       .specVersion = 1,
 470    },
 471    {
 472       .extensionName = VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
 473       .specVersion = 1,
 474    },
 475    {
 476       .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
 477       .specVersion = 68,
 478    },
 479    {
 480       .extensionName = VK_KHX_EXTERNAL_MEMORY_EXTENSION_NAME,
 481       .specVersion = 1,
 482    },
 483    {
 484       .extensionName = VK_KHX_EXTERNAL_MEMORY_FD_EXTENSION_NAME,
 485       .specVersion = 1,
 486    },
 487    {
 488       .extensionName = VK_KHX_EXTERNAL_SEMAPHORE_EXTENSION_NAME,
 489       .specVersion = 1,
 490    },
 491    {
 492       .extensionName = VK_KHX_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME,
 493       .specVersion = 1,
 494    },
 495    {
 496       .extensionName = VK_KHX_MULTIVIEW_EXTENSION_NAME,
 497       .specVersion = 1,
 498    },
 499 };
 500
 501 static void *
 502 default_alloc_func(void *pUserData, size_t size, size_t align,
 503                    VkSystemAllocationScope allocationScope)
 504 {
 505    return malloc(size);
 506 }
 507
 508 static void *
 509 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
 510                      size_t align, VkSystemAllocationScope allocationScope)
 511 {
 512    return realloc(pOriginal, size);
 513 }
 514
 515 static void
 516 default_free_func(void *pUserData, void *pMemory)
 517 {
 518    free(pMemory);
 519 }
 520
 521 static const VkAllocationCallbacks default_alloc = {
 522    .pUserData = NULL,
 523    .pfnAllocation = default_alloc_func,
 524    .pfnReallocation = default_realloc_func,
 525    .pfnFree = default_free_func,
 526 };
 527
 528 VkResult anv_CreateInstance(
 529     const VkInstanceCreateInfo*                 pCreateInfo,
 530     const VkAllocationCallbacks*                pAllocator,
 531     VkInstance*                                 pInstance)
 532 {
 533    struct anv_instance *instance;
 534
 535    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 536
 537    uint32_t client_version;
 538    if (pCreateInfo->pApplicationInfo &&
 539        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 540       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 541    } else {
 542       client_version = VK_MAKE_VERSION(1, 0, 0);
 543    }
 544
 545    if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
 546        client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
 547       return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 548                        "Client requested version %d.%d.%d",
 549                        VK_VERSION_MAJOR(client_version),
 550                        VK_VERSION_MINOR(client_version),
 551                        VK_VERSION_PATCH(client_version));
 552    }
 553
 554    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 555       bool found = false;
 556       for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
 557          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 558                     global_extensions[j].extensionName) == 0) {
 559             found = true;
 560             break;
 561          }
 562       }
 563       if (!found)
 564          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 565    }
 566
 567    instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 568                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 569    if (!instance)
 570       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 571
 572    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 573
 574    if (pAllocator)
 575       instance->alloc = *pAllocator;
 576    else
 577       instance->alloc = default_alloc;
 578
 579    instance->apiVersion = client_version;
 580    instance->physicalDeviceCount = -1;
 581
 582    _mesa_locale_init();
 583
 584    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 585
 586    *pInstance = anv_instance_to_handle(instance);
 587
 588    return VK_SUCCESS;
 589 }
 590
 591 void anv_DestroyInstance(
 592     VkInstance                                  _instance,
 593     const VkAllocationCallbacks*                pAllocator)
 594 {
 595    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 596
 597    if (!instance)
 598       return;
 599
 600    if (instance->physicalDeviceCount > 0) {
 601       /* We support at most one physical device. */
 602       assert(instance->physicalDeviceCount == 1);
 603       anv_physical_device_finish(&instance->physicalDevice);
 604    }
 605
 606    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 607
 608    _mesa_locale_fini();
 609
 610    vk_free(&instance->alloc, instance);
 611 }
 612
 613 static VkResult
 614 anv_enumerate_devices(struct anv_instance *instance)
 615 {
 616    /* TODO: Check for more devices ? */
 617    drmDevicePtr devices[8];
 618    VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
 619    int max_devices;
 620
 621    instance->physicalDeviceCount = 0;
 622
 623    max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
 624    if (max_devices < 1)
 625       return VK_ERROR_INCOMPATIBLE_DRIVER;
 626
 627    for (unsigned i = 0; i < (unsigned)max_devices; i++) {
 628       if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
 629           devices[i]->bustype == DRM_BUS_PCI &&
 630           devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
 631
 632          result = anv_physical_device_init(&instance->physicalDevice,
 633                         instance,
 634                         devices[i]->nodes[DRM_NODE_RENDER]);
 635          if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 636             break;
 637       }
 638    }
 639    drmFreeDevices(devices, max_devices);
 640
 641    if (result == VK_SUCCESS)
 642       instance->physicalDeviceCount = 1;
 643
 644    return result;
 645 }
 646
 647
 648 VkResult anv_EnumeratePhysicalDevices(
 649     VkInstance                                  _instance,
 650     uint32_t*                                   pPhysicalDeviceCount,
 651     VkPhysicalDevice*                           pPhysicalDevices)
 652 {
 653    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 654    VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
 655    VkResult result;
 656
 657    if (instance->physicalDeviceCount < 0) {
 658       result = anv_enumerate_devices(instance);
 659       if (result != VK_SUCCESS &&
 660           result != VK_ERROR_INCOMPATIBLE_DRIVER)
 661          return result;
 662    }
 663
 664    if (instance->physicalDeviceCount > 0) {
 665       assert(instance->physicalDeviceCount == 1);
 666       vk_outarray_append(&out, i) {
 667          *i = anv_physical_device_to_handle(&instance->physicalDevice);
 668       }
 669    }
 670
 671    return vk_outarray_status(&out);
 672 }
 673
 674 void anv_GetPhysicalDeviceFeatures(
 675     VkPhysicalDevice                            physicalDevice,
 676     VkPhysicalDeviceFeatures*                   pFeatures)
 677 {
 678    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 679
 680    *pFeatures = (VkPhysicalDeviceFeatures) {
 681       .robustBufferAccess                       = true,
 682       .fullDrawIndexUint32                      = true,
 683       .imageCubeArray                           = true,
 684       .independentBlend                         = true,
 685       .geometryShader                           = true,
 686       .tessellationShader                       = true,
 687       .sampleRateShading                        = true,
 688       .dualSrcBlend                             = true,
 689       .logicOp                                  = true,
 690       .multiDrawIndirect                        = true,
 691       .drawIndirectFirstInstance                = true,
 692       .depthClamp                               = true,
 693       .depthBiasClamp                           = true,
 694       .fillModeNonSolid                         = true,
 695       .depthBounds                              = false,
 696       .wideLines                                = true,
 697       .largePoints                              = true,
 698       .alphaToOne                               = true,
 699       .multiViewport                            = true,
 700       .samplerAnisotropy                        = true,
 701       .textureCompressionETC2                   = pdevice->info.gen >= 8 ||
 702                                                   pdevice->info.is_baytrail,
 703       .textureCompressionASTC_LDR               = pdevice->info.gen >= 9, /* FINISHME CHV */
 704       .textureCompressionBC                     = true,
 705       .occlusionQueryPrecise                    = true,
 706       .pipelineStatisticsQuery                  = true,
 707       .fragmentStoresAndAtomics                 = true,
 708       .shaderTessellationAndGeometryPointSize   = true,
 709       .shaderImageGatherExtended                = true,
 710       .shaderStorageImageExtendedFormats        = true,
 711       .shaderStorageImageMultisample            = false,
 712       .shaderStorageImageReadWithoutFormat      = false,
 713       .shaderStorageImageWriteWithoutFormat     = true,
 714       .shaderUniformBufferArrayDynamicIndexing  = true,
 715       .shaderSampledImageArrayDynamicIndexing   = true,
 716       .shaderStorageBufferArrayDynamicIndexing  = true,
 717       .shaderStorageImageArrayDynamicIndexing   = true,
 718       .shaderClipDistance                       = true,
 719       .shaderCullDistance                       = true,
 720       .shaderFloat64                            = pdevice->info.gen >= 8,
 721       .shaderInt64                              = pdevice->info.gen >= 8,
 722       .shaderInt16                              = false,
 723       .shaderResourceMinLod                     = false,
 724       .variableMultisampleRate                  = false,
 725       .inheritedQueries                         = true,
 726    };
 727
 728    /* We can't do image stores in vec4 shaders */
 729    pFeatures->vertexPipelineStoresAndAtomics =
 730       pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
 731       pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
 732 }
 733
 734 void anv_GetPhysicalDeviceFeatures2KHR(
 735     VkPhysicalDevice                            physicalDevice,
 736     VkPhysicalDeviceFeatures2KHR*               pFeatures)
 737 {
 738    anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 739
 740    vk_foreach_struct(ext, pFeatures->pNext) {
 741       switch (ext->sType) {
 742       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX: {
 743          VkPhysicalDeviceMultiviewFeaturesKHX *features =
 744             (VkPhysicalDeviceMultiviewFeaturesKHX *)ext;
 745          features->multiview = true;
 746          features->multiviewGeometryShader = true;
 747          features->multiviewTessellationShader = true;
 748          break;
 749       }
 750
 751       default:
 752          anv_debug_ignored_stype(ext->sType);
 753          break;
 754       }
 755    }
 756 }
 757
 758 void anv_GetPhysicalDeviceProperties(
 759     VkPhysicalDevice                            physicalDevice,
 760     VkPhysicalDeviceProperties*                 pProperties)
 761 {
 762    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 763    const struct gen_device_info *devinfo = &pdevice->info;
 764
 765    /* See assertions made when programming the buffer surface state. */
 766    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
 767                                       (1ul << 30) : (1ul << 27);
 768
 769    const uint32_t max_samplers = (devinfo->gen >= 8 || devinfo->is_haswell) ?
 770                                  128 : 16;
 771
 772    VkSampleCountFlags sample_counts =
 773       isl_device_get_sample_counts(&pdevice->isl_dev);
 774
 775    VkPhysicalDeviceLimits limits = {
 776       .maxImageDimension1D                      = (1 << 14),
 777       .maxImageDimension2D                      = (1 << 14),
 778       .maxImageDimension3D                      = (1 << 11),
 779       .maxImageDimensionCube                    = (1 << 14),
 780       .maxImageArrayLayers                      = (1 << 11),
 781       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 782       .maxUniformBufferRange                    = (1ul << 27),
 783       .maxStorageBufferRange                    = max_raw_buffer_sz,
 784       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 785       .maxMemoryAllocationCount                 = UINT32_MAX,
 786       .maxSamplerAllocationCount                = 64 * 1024,
 787       .bufferImageGranularity                   = 64, /* A cache line */
 788       .sparseAddressSpaceSize                   = 0,
 789       .maxBoundDescriptorSets                   = MAX_SETS,
 790       .maxPerStageDescriptorSamplers            = max_samplers,
 791       .maxPerStageDescriptorUniformBuffers      = 64,
 792       .maxPerStageDescriptorStorageBuffers      = 64,
 793       .maxPerStageDescriptorSampledImages       = max_samplers,
 794       .maxPerStageDescriptorStorageImages       = 64,
 795       .maxPerStageDescriptorInputAttachments    = 64,
 796       .maxPerStageResources                     = 250,
 797       .maxDescriptorSetSamplers                 = 256,
 798       .maxDescriptorSetUniformBuffers           = 256,
 799       .maxDescriptorSetUniformBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
 800       .maxDescriptorSetStorageBuffers           = 256,
 801       .maxDescriptorSetStorageBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
 802       .maxDescriptorSetSampledImages            = 256,
 803       .maxDescriptorSetStorageImages            = 256,
 804       .maxDescriptorSetInputAttachments         = 256,
 805       .maxVertexInputAttributes                 = MAX_VBS,
 806       .maxVertexInputBindings                   = MAX_VBS,
 807       .maxVertexInputAttributeOffset            = 2047,
 808       .maxVertexInputBindingStride              = 2048,
 809       .maxVertexOutputComponents                = 128,
 810       .maxTessellationGenerationLevel           = 64,
 811       .maxTessellationPatchSize                 = 32,
 812       .maxTessellationControlPerVertexInputComponents = 128,
 813       .maxTessellationControlPerVertexOutputComponents = 128,
 814       .maxTessellationControlPerPatchOutputComponents = 128,
 815       .maxTessellationControlTotalOutputComponents = 2048,
 816       .maxTessellationEvaluationInputComponents = 128,
 817       .maxTessellationEvaluationOutputComponents = 128,
 818       .maxGeometryShaderInvocations             = 32,
 819       .maxGeometryInputComponents               = 64,
 820       .maxGeometryOutputComponents              = 128,
 821       .maxGeometryOutputVertices                = 256,
 822       .maxGeometryTotalOutputComponents         = 1024,
 823       .maxFragmentInputComponents               = 128,
 824       .maxFragmentOutputAttachments             = 8,
 825       .maxFragmentDualSrcAttachments            = 1,
 826       .maxFragmentCombinedOutputResources       = 8,
 827       .maxComputeSharedMemorySize               = 32768,
 828       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 829       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 830       .maxComputeWorkGroupSize = {
 831          16 * devinfo->max_cs_threads,
 832          16 * devinfo->max_cs_threads,
 833          16 * devinfo->max_cs_threads,
 834       },
 835       .subPixelPrecisionBits                    = 4 /* FIXME */,
 836       .subTexelPrecisionBits                    = 4 /* FIXME */,
 837       .mipmapPrecisionBits                      = 4 /* FIXME */,
 838       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 839       .maxDrawIndirectCount                     = UINT32_MAX,
 840       .maxSamplerLodBias                        = 16,
 841       .maxSamplerAnisotropy                     = 16,
 842       .maxViewports                             = MAX_VIEWPORTS,
 843       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 844       .viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
 845       .viewportSubPixelBits                     = 13, /* We take a float? */
 846       .minMemoryMapAlignment                    = 4096, /* A page */
 847       .minTexelBufferOffsetAlignment            = 1,
 848       .minUniformBufferOffsetAlignment          = 16,
 849       .minStorageBufferOffsetAlignment          = 4,
 850       .minTexelOffset                           = -8,
 851       .maxTexelOffset                           = 7,
 852       .minTexelGatherOffset                     = -32,
 853       .maxTexelGatherOffset                     = 31,
 854       .minInterpolationOffset                   = -0.5,
 855       .maxInterpolationOffset                   = 0.4375,
 856       .subPixelInterpolationOffsetBits          = 4,
 857       .maxFramebufferWidth                      = (1 << 14),
 858       .maxFramebufferHeight                     = (1 << 14),
 859       .maxFramebufferLayers                     = (1 << 11),
 860       .framebufferColorSampleCounts             = sample_counts,
 861       .framebufferDepthSampleCounts             = sample_counts,
 862       .framebufferStencilSampleCounts           = sample_counts,
 863       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 864       .maxColorAttachments                      = MAX_RTS,
 865       .sampledImageColorSampleCounts            = sample_counts,
 866       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 867       .sampledImageDepthSampleCounts            = sample_counts,
 868       .sampledImageStencilSampleCounts          = sample_counts,
 869       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 870       .maxSampleMaskWords                       = 1,
 871       .timestampComputeAndGraphics              = false,
 872       .timestampPeriod                          = 1000000000.0 / devinfo->timestamp_frequency,
 873       .maxClipDistances                         = 8,
 874       .maxCullDistances                         = 8,
 875       .maxCombinedClipAndCullDistances          = 8,
 876       .discreteQueuePriorities                  = 1,
 877       .pointSizeRange                           = { 0.125, 255.875 },
 878       .lineWidthRange                           = { 0.0, 7.9921875 },
 879       .pointSizeGranularity                     = (1.0 / 8.0),
 880       .lineWidthGranularity                     = (1.0 / 128.0),
 881       .strictLines                              = false, /* FINISHME */
 882       .standardSampleLocations                  = true,
 883       .optimalBufferCopyOffsetAlignment         = 128,
 884       .optimalBufferCopyRowPitchAlignment       = 128,
 885       .nonCoherentAtomSize                      = 64,
 886    };
 887
 888    *pProperties = (VkPhysicalDeviceProperties) {
 889       .apiVersion = VK_MAKE_VERSION(1, 0, 42),
 890       .driverVersion = vk_get_driver_version(),
 891       .vendorID = 0x8086,
 892       .deviceID = pdevice->chipset_id,
 893       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 894       .limits = limits,
 895       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 896    };
 897
 898    strcpy(pProperties->deviceName, pdevice->name);
 899    memcpy(pProperties->pipelineCacheUUID,
 900           pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
 901 }
 902
 903 void anv_GetPhysicalDeviceProperties2KHR(
 904     VkPhysicalDevice                            physicalDevice,
 905     VkPhysicalDeviceProperties2KHR*             pProperties)
 906 {
 907    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 908
 909    anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 910
 911    vk_foreach_struct(ext, pProperties->pNext) {
 912       switch (ext->sType) {
 913       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 914          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 915             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 916
 917          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 918          break;
 919       }
 920
 921       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHX: {
 922          VkPhysicalDeviceIDPropertiesKHX *id_props =
 923             (VkPhysicalDeviceIDPropertiesKHX *)ext;
 924          memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
 925          memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
 926          /* The LUID is for Windows. */
 927          id_props->deviceLUIDValid = false;
 928          break;
 929       }
 930
 931       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX: {
 932          VkPhysicalDeviceMultiviewPropertiesKHX *properties =
 933             (VkPhysicalDeviceMultiviewPropertiesKHX *)ext;
 934          properties->maxMultiviewViewCount = 16;
 935          properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
 936          break;
 937       }
 938
 939       default:
 940          anv_debug_ignored_stype(ext->sType);
 941          break;
 942       }
 943    }
 944 }
 945
 946 /* We support exactly one queue family. */
 947 static const VkQueueFamilyProperties
 948 anv_queue_family_properties = {
 949    .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 950                  VK_QUEUE_COMPUTE_BIT |
 951                  VK_QUEUE_TRANSFER_BIT,
 952    .queueCount = 1,
 953    .timestampValidBits = 36, /* XXX: Real value here */
 954    .minImageTransferGranularity = { 1, 1, 1 },
 955 };
 956
 957 void anv_GetPhysicalDeviceQueueFamilyProperties(
 958     VkPhysicalDevice                            physicalDevice,
 959     uint32_t*                                   pCount,
 960     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 961 {
 962    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
 963
 964    vk_outarray_append(&out, p) {
 965       *p = anv_queue_family_properties;
 966    }
 967 }
 968
 969 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
 970     VkPhysicalDevice                            physicalDevice,
 971     uint32_t*                                   pQueueFamilyPropertyCount,
 972     VkQueueFamilyProperties2KHR*                pQueueFamilyProperties)
 973 {
 974
 975    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 976
 977    vk_outarray_append(&out, p) {
 978       p->queueFamilyProperties = anv_queue_family_properties;
 979
 980       vk_foreach_struct(s, p->pNext) {
 981          anv_debug_ignored_stype(s->sType);
 982       }
 983    }
 984 }
 985
 986 void anv_GetPhysicalDeviceMemoryProperties(
 987     VkPhysicalDevice                            physicalDevice,
 988     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
 989 {
 990    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 991
 992    pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
 993    for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
 994       pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
 995          .propertyFlags = physical_device->memory.types[i].propertyFlags,
 996          .heapIndex     = physical_device->memory.types[i].heapIndex,
 997       };
 998    }
 999
1000    pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1001    for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1002       pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1003          .size    = physical_device->memory.heaps[i].size,
1004          .flags   = physical_device->memory.heaps[i].flags,
1005       };
1006    }
1007 }
1008
1009 void anv_GetPhysicalDeviceMemoryProperties2KHR(
1010     VkPhysicalDevice                            physicalDevice,
1011     VkPhysicalDeviceMemoryProperties2KHR*       pMemoryProperties)
1012 {
1013    anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1014                                          &pMemoryProperties->memoryProperties);
1015
1016    vk_foreach_struct(ext, pMemoryProperties->pNext) {
1017       switch (ext->sType) {
1018       default:
1019          anv_debug_ignored_stype(ext->sType);
1020          break;
1021       }
1022    }
1023 }
1024
1025 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1026     VkInstance                                  instance,
1027     const char*                                 pName)
1028 {
1029    return anv_lookup_entrypoint(NULL, pName);
1030 }
1031
1032 /* With version 1+ of the loader interface the ICD should expose
1033  * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1034  */
1035 PUBLIC
1036 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1037     VkInstance                                  instance,
1038     const char*                                 pName);
1039
1040 PUBLIC
1041 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1042     VkInstance                                  instance,
1043     const char*                                 pName)
1044 {
1045    return anv_GetInstanceProcAddr(instance, pName);
1046 }
1047
1048 PFN_vkVoidFunction anv_GetDeviceProcAddr(
1049     VkDevice                                    _device,
1050     const char*                                 pName)
1051 {
1052    ANV_FROM_HANDLE(anv_device, device, _device);
1053    return anv_lookup_entrypoint(&device->info, pName);
1054 }
1055
1056 static void
1057 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
1058 {
1059    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1060    queue->device = device;
1061    queue->pool = &device->surface_state_pool;
1062 }
1063
1064 static void
1065 anv_queue_finish(struct anv_queue *queue)
1066 {
1067 }
1068
1069 static struct anv_state
1070 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
1071 {
1072    struct anv_state state;
1073
1074    state = anv_state_pool_alloc(pool, size, align);
1075    memcpy(state.map, p, size);
1076
1077    anv_state_flush(pool->block_pool.device, state);
1078
1079    return state;
1080 }
1081
1082 struct gen8_border_color {
1083    union {
1084       float float32[4];
1085       uint32_t uint32[4];
1086    };
1087    /* Pad out to 64 bytes */
1088    uint32_t _pad[12];
1089 };
1090
1091 static void
1092 anv_device_init_border_colors(struct anv_device *device)
1093 {
1094    static const struct gen8_border_color border_colors[] = {
1095       [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] =  { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
1096       [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] =       { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
1097       [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] =       { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
1098       [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] =    { .uint32 = { 0, 0, 0, 0 } },
1099       [VK_BORDER_COLOR_INT_OPAQUE_BLACK] =         { .uint32 = { 0, 0, 0, 1 } },
1100       [VK_BORDER_COLOR_INT_OPAQUE_WHITE] =         { .uint32 = { 1, 1, 1, 1 } },
1101    };
1102
1103    device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
1104                                                     sizeof(border_colors), 64,
1105                                                     border_colors);
1106 }
1107
1108 VkResult anv_CreateDevice(
1109     VkPhysicalDevice                            physicalDevice,
1110     const VkDeviceCreateInfo*                   pCreateInfo,
1111     const VkAllocationCallbacks*                pAllocator,
1112     VkDevice*                                   pDevice)
1113 {
1114    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1115    VkResult result;
1116    struct anv_device *device;
1117
1118    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
1119
1120    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1121       bool found = false;
1122       for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
1123          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
1124                     device_extensions[j].extensionName) == 0) {
1125             found = true;
1126             break;
1127          }
1128       }
1129       if (!found)
1130          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1131    }
1132
1133    /* Check enabled features */
1134    if (pCreateInfo->pEnabledFeatures) {
1135       VkPhysicalDeviceFeatures supported_features;
1136       anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1137       VkBool32 *supported_feature = (VkBool32 *)&supported_features;
1138       VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
1139       unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1140       for (uint32_t i = 0; i < num_features; i++) {
1141          if (enabled_feature[i] && !supported_feature[i])
1142             return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1143       }
1144    }
1145
1146    device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
1147                        sizeof(*device), 8,
1148                        VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1149    if (!device)
1150       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1151
1152    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1153    device->instance = physical_device->instance;
1154    device->chipset_id = physical_device->chipset_id;
1155    device->lost = false;
1156
1157    if (pAllocator)
1158       device->alloc = *pAllocator;
1159    else
1160       device->alloc = physical_device->instance->alloc;
1161
1162    /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1163    device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1164    if (device->fd == -1) {
1165       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1166       goto fail_device;
1167    }
1168
1169    device->context_id = anv_gem_create_context(device);
1170    if (device->context_id == -1) {
1171       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1172       goto fail_fd;
1173    }
1174
1175    device->info = physical_device->info;
1176    device->isl_dev = physical_device->isl_dev;
1177
1178    /* On Broadwell and later, we can use batch chaining to more efficiently
1179     * implement growing command buffers.  Prior to Haswell, the kernel
1180     * command parser gets in the way and we have to fall back to growing
1181     * the batch.
1182     */
1183    device->can_chain_batches = device->info.gen >= 8;
1184
1185    device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1186       pCreateInfo->pEnabledFeatures->robustBufferAccess;
1187
1188    if (pthread_mutex_init(&device->mutex, NULL) != 0) {
1189       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1190       goto fail_context_id;
1191    }
1192
1193    pthread_condattr_t condattr;
1194    if (pthread_condattr_init(&condattr) != 0) {
1195       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1196       goto fail_mutex;
1197    }
1198    if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
1199       pthread_condattr_destroy(&condattr);
1200       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1201       goto fail_mutex;
1202    }
1203    if (pthread_cond_init(&device->queue_submit, NULL) != 0) {
1204       pthread_condattr_destroy(&condattr);
1205       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1206       goto fail_mutex;
1207    }
1208    pthread_condattr_destroy(&condattr);
1209
1210    anv_bo_pool_init(&device->batch_bo_pool, device);
1211
1212    result = anv_bo_cache_init(&device->bo_cache);
1213    if (result != VK_SUCCESS)
1214       goto fail_batch_bo_pool;
1215
1216    result = anv_state_pool_init(&device->dynamic_state_pool, device, 16384);
1217    if (result != VK_SUCCESS)
1218       goto fail_bo_cache;
1219
1220    result = anv_state_pool_init(&device->instruction_state_pool, device, 16384);
1221    if (result != VK_SUCCESS)
1222       goto fail_dynamic_state_pool;
1223
1224    result = anv_state_pool_init(&device->surface_state_pool, device, 4096);
1225    if (result != VK_SUCCESS)
1226       goto fail_instruction_state_pool;
1227
1228    result = anv_bo_init_new(&device->workaround_bo, device, 1024);
1229    if (result != VK_SUCCESS)
1230       goto fail_surface_state_pool;
1231
1232    anv_scratch_pool_init(device, &device->scratch_pool);
1233
1234    anv_queue_init(device, &device->queue);
1235
1236    switch (device->info.gen) {
1237    case 7:
1238       if (!device->info.is_haswell)
1239          result = gen7_init_device_state(device);
1240       else
1241          result = gen75_init_device_state(device);
1242       break;
1243    case 8:
1244       result = gen8_init_device_state(device);
1245       break;
1246    case 9:
1247       result = gen9_init_device_state(device);
1248       break;
1249    case 10:
1250       result = gen10_init_device_state(device);
1251       break;
1252    default:
1253       /* Shouldn't get here as we don't create physical devices for any other
1254        * gens. */
1255       unreachable("unhandled gen");
1256    }
1257    if (result != VK_SUCCESS)
1258       goto fail_workaround_bo;
1259
1260    anv_device_init_blorp(device);
1261
1262    anv_device_init_border_colors(device);
1263
1264    *pDevice = anv_device_to_handle(device);
1265
1266    return VK_SUCCESS;
1267
1268  fail_workaround_bo:
1269    anv_queue_finish(&device->queue);
1270    anv_scratch_pool_finish(device, &device->scratch_pool);
1271    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1272    anv_gem_close(device, device->workaround_bo.gem_handle);
1273  fail_surface_state_pool:
1274    anv_state_pool_finish(&device->surface_state_pool);
1275  fail_instruction_state_pool:
1276    anv_state_pool_finish(&device->instruction_state_pool);
1277  fail_dynamic_state_pool:
1278    anv_state_pool_finish(&device->dynamic_state_pool);
1279  fail_bo_cache:
1280    anv_bo_cache_finish(&device->bo_cache);
1281  fail_batch_bo_pool:
1282    anv_bo_pool_finish(&device->batch_bo_pool);
1283    pthread_cond_destroy(&device->queue_submit);
1284  fail_mutex:
1285    pthread_mutex_destroy(&device->mutex);
1286  fail_context_id:
1287    anv_gem_destroy_context(device, device->context_id);
1288  fail_fd:
1289    close(device->fd);
1290  fail_device:
1291    vk_free(&device->alloc, device);
1292
1293    return result;
1294 }
1295
1296 void anv_DestroyDevice(
1297     VkDevice                                    _device,
1298     const VkAllocationCallbacks*                pAllocator)
1299 {
1300    ANV_FROM_HANDLE(anv_device, device, _device);
1301
1302    if (!device)
1303       return;
1304
1305    anv_device_finish_blorp(device);
1306
1307    anv_queue_finish(&device->queue);
1308
1309 #ifdef HAVE_VALGRIND
1310    /* We only need to free these to prevent valgrind errors.  The backing
1311     * BO will go away in a couple of lines so we don't actually leak.
1312     */
1313    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1314 #endif
1315
1316    anv_scratch_pool_finish(device, &device->scratch_pool);
1317
1318    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1319    anv_gem_close(device, device->workaround_bo.gem_handle);
1320
1321    anv_state_pool_finish(&device->surface_state_pool);
1322    anv_state_pool_finish(&device->instruction_state_pool);
1323    anv_state_pool_finish(&device->dynamic_state_pool);
1324
1325    anv_bo_cache_finish(&device->bo_cache);
1326
1327    anv_bo_pool_finish(&device->batch_bo_pool);
1328
1329    pthread_cond_destroy(&device->queue_submit);
1330    pthread_mutex_destroy(&device->mutex);
1331
1332    anv_gem_destroy_context(device, device->context_id);
1333
1334    close(device->fd);
1335
1336    vk_free(&device->alloc, device);
1337 }
1338
1339 VkResult anv_EnumerateInstanceExtensionProperties(
1340     const char*                                 pLayerName,
1341     uint32_t*                                   pPropertyCount,
1342     VkExtensionProperties*                      pProperties)
1343 {
1344    if (pProperties == NULL) {
1345       *pPropertyCount = ARRAY_SIZE(global_extensions);
1346       return VK_SUCCESS;
1347    }
1348
1349    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(global_extensions));
1350    typed_memcpy(pProperties, global_extensions, *pPropertyCount);
1351
1352    if (*pPropertyCount < ARRAY_SIZE(global_extensions))
1353       return VK_INCOMPLETE;
1354
1355    return VK_SUCCESS;
1356 }
1357
1358 VkResult anv_EnumerateDeviceExtensionProperties(
1359     VkPhysicalDevice                            physicalDevice,
1360     const char*                                 pLayerName,
1361     uint32_t*                                   pPropertyCount,
1362     VkExtensionProperties*                      pProperties)
1363 {
1364    if (pProperties == NULL) {
1365       *pPropertyCount = ARRAY_SIZE(device_extensions);
1366       return VK_SUCCESS;
1367    }
1368
1369    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(device_extensions));
1370    typed_memcpy(pProperties, device_extensions, *pPropertyCount);
1371
1372    if (*pPropertyCount < ARRAY_SIZE(device_extensions))
1373       return VK_INCOMPLETE;
1374
1375    return VK_SUCCESS;
1376 }
1377
1378 VkResult anv_EnumerateInstanceLayerProperties(
1379     uint32_t*                                   pPropertyCount,
1380     VkLayerProperties*                          pProperties)
1381 {
1382    if (pProperties == NULL) {
1383       *pPropertyCount = 0;
1384       return VK_SUCCESS;
1385    }
1386
1387    /* None supported at this time */
1388    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1389 }
1390
1391 VkResult anv_EnumerateDeviceLayerProperties(
1392     VkPhysicalDevice                            physicalDevice,
1393     uint32_t*                                   pPropertyCount,
1394     VkLayerProperties*                          pProperties)
1395 {
1396    if (pProperties == NULL) {
1397       *pPropertyCount = 0;
1398       return VK_SUCCESS;
1399    }
1400
1401    /* None supported at this time */
1402    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1403 }
1404
1405 void anv_GetDeviceQueue(
1406     VkDevice                                    _device,
1407     uint32_t                                    queueNodeIndex,
1408     uint32_t                                    queueIndex,
1409     VkQueue*                                    pQueue)
1410 {
1411    ANV_FROM_HANDLE(anv_device, device, _device);
1412
1413    assert(queueIndex == 0);
1414
1415    *pQueue = anv_queue_to_handle(&device->queue);
1416 }
1417
1418 VkResult
1419 anv_device_query_status(struct anv_device *device)
1420 {
1421    /* This isn't likely as most of the callers of this function already check
1422     * for it.  However, it doesn't hurt to check and it potentially lets us
1423     * avoid an ioctl.
1424     */
1425    if (unlikely(device->lost))
1426       return VK_ERROR_DEVICE_LOST;
1427
1428    uint32_t active, pending;
1429    int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
1430    if (ret == -1) {
1431       /* We don't know the real error. */
1432       device->lost = true;
1433       return vk_errorf(VK_ERROR_DEVICE_LOST, "get_reset_stats failed: %m");
1434    }
1435
1436    if (active) {
1437       device->lost = true;
1438       return vk_errorf(VK_ERROR_DEVICE_LOST,
1439                        "GPU hung on one of our command buffers");
1440    } else if (pending) {
1441       device->lost = true;
1442       return vk_errorf(VK_ERROR_DEVICE_LOST,
1443                        "GPU hung with commands in-flight");
1444    }
1445
1446    return VK_SUCCESS;
1447 }
1448
1449 VkResult
1450 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
1451 {
1452    /* Note:  This only returns whether or not the BO is in use by an i915 GPU.
1453     * Other usages of the BO (such as on different hardware) will not be
1454     * flagged as "busy" by this ioctl.  Use with care.
1455     */
1456    int ret = anv_gem_busy(device, bo->gem_handle);
1457    if (ret == 1) {
1458       return VK_NOT_READY;
1459    } else if (ret == -1) {
1460       /* We don't know the real error. */
1461       device->lost = true;
1462       return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1463    }
1464
1465    /* Query for device status after the busy call.  If the BO we're checking
1466     * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1467     * client because it clearly doesn't have valid data.  Yes, this most
1468     * likely means an ioctl, but we just did an ioctl to query the busy status
1469     * so it's no great loss.
1470     */
1471    return anv_device_query_status(device);
1472 }
1473
1474 VkResult
1475 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
1476                 int64_t timeout)
1477 {
1478    int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
1479    if (ret == -1 && errno == ETIME) {
1480       return VK_TIMEOUT;
1481    } else if (ret == -1) {
1482       /* We don't know the real error. */
1483       device->lost = true;
1484       return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1485    }
1486
1487    /* Query for device status after the wait.  If the BO we're waiting on got
1488     * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1489     * because it clearly doesn't have valid data.  Yes, this most likely means
1490     * an ioctl, but we just did an ioctl to wait so it's no great loss.
1491     */
1492    return anv_device_query_status(device);
1493 }
1494
1495 VkResult anv_DeviceWaitIdle(
1496     VkDevice                                    _device)
1497 {
1498    ANV_FROM_HANDLE(anv_device, device, _device);
1499    if (unlikely(device->lost))
1500       return VK_ERROR_DEVICE_LOST;
1501
1502    struct anv_batch batch;
1503
1504    uint32_t cmds[8];
1505    batch.start = batch.next = cmds;
1506    batch.end = (void *) cmds + sizeof(cmds);
1507
1508    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1509    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1510
1511    return anv_device_submit_simple_batch(device, &batch);
1512 }
1513
1514 VkResult
1515 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1516 {
1517    uint32_t gem_handle = anv_gem_create(device, size);
1518    if (!gem_handle)
1519       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1520
1521    anv_bo_init(bo, gem_handle, size);
1522
1523    return VK_SUCCESS;
1524 }
1525
1526 VkResult anv_AllocateMemory(
1527     VkDevice                                    _device,
1528     const VkMemoryAllocateInfo*                 pAllocateInfo,
1529     const VkAllocationCallbacks*                pAllocator,
1530     VkDeviceMemory*                             pMem)
1531 {
1532    ANV_FROM_HANDLE(anv_device, device, _device);
1533    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
1534    struct anv_device_memory *mem;
1535    VkResult result = VK_SUCCESS;
1536
1537    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1538
1539    /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1540    assert(pAllocateInfo->allocationSize > 0);
1541
1542    /* The kernel relocation API has a limitation of a 32-bit delta value
1543     * applied to the address before it is written which, in spite of it being
1544     * unsigned, is treated as signed .  Because of the way that this maps to
1545     * the Vulkan API, we cannot handle an offset into a buffer that does not
1546     * fit into a signed 32 bits.  The only mechanism we have for dealing with
1547     * this at the moment is to limit all VkDeviceMemory objects to a maximum
1548     * of 2GB each.  The Vulkan spec allows us to do this:
1549     *
1550     *    "Some platforms may have a limit on the maximum size of a single
1551     *    allocation. For example, certain systems may fail to create
1552     *    allocations with a size greater than or equal to 4GB. Such a limit is
1553     *    implementation-dependent, and if such a failure occurs then the error
1554     *    VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1555     *
1556     * We don't use vk_error here because it's not an error so much as an
1557     * indication to the application that the allocation is too large.
1558     */
1559    if (pAllocateInfo->allocationSize > (1ull << 31))
1560       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
1561
1562    /* FINISHME: Fail if allocation request exceeds heap size. */
1563
1564    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1565                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1566    if (mem == NULL)
1567       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1568
1569    assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
1570    mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
1571    mem->map = NULL;
1572    mem->map_size = 0;
1573
1574    const VkImportMemoryFdInfoKHX *fd_info =
1575       vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHX);
1576
1577    /* The Vulkan spec permits handleType to be 0, in which case the struct is
1578     * ignored.
1579     */
1580    if (fd_info && fd_info->handleType) {
1581       /* At the moment, we only support the OPAQUE_FD memory type which is
1582        * just a GEM buffer.
1583        */
1584       assert(fd_info->handleType ==
1585              VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHX);
1586
1587       result = anv_bo_cache_import(device, &device->bo_cache,
1588                                    fd_info->fd, pAllocateInfo->allocationSize,
1589                                    &mem->bo);
1590       if (result != VK_SUCCESS)
1591          goto fail;
1592    } else {
1593       result = anv_bo_cache_alloc(device, &device->bo_cache,
1594                                   pAllocateInfo->allocationSize,
1595                                   &mem->bo);
1596       if (result != VK_SUCCESS)
1597          goto fail;
1598    }
1599
1600    assert(mem->type->heapIndex < pdevice->memory.heap_count);
1601    if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
1602       mem->bo->flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
1603
1604    if (pdevice->has_exec_async)
1605       mem->bo->flags |= EXEC_OBJECT_ASYNC;
1606
1607    *pMem = anv_device_memory_to_handle(mem);
1608
1609    return VK_SUCCESS;
1610
1611  fail:
1612    vk_free2(&device->alloc, pAllocator, mem);
1613
1614    return result;
1615 }
1616
1617 VkResult anv_GetMemoryFdKHX(
1618     VkDevice                                    device_h,
1619     VkDeviceMemory                              memory_h,
1620     VkExternalMemoryHandleTypeFlagBitsKHX       handleType,
1621     int*                                        pFd)
1622 {
1623    ANV_FROM_HANDLE(anv_device, dev, device_h);
1624    ANV_FROM_HANDLE(anv_device_memory, mem, memory_h);
1625
1626    /* We support only one handle type. */
1627    assert(handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHX);
1628
1629    return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
1630 }
1631
1632 VkResult anv_GetMemoryFdPropertiesKHX(
1633     VkDevice                                    device_h,
1634     VkExternalMemoryHandleTypeFlagBitsKHX       handleType,
1635     int                                         fd,
1636     VkMemoryFdPropertiesKHX*                    pMemoryFdProperties)
1637 {
1638    /* The valid usage section for this function says:
1639     *
1640     *    "handleType must not be one of the handle types defined as opaque."
1641     *
1642     * Since we only handle opaque handles for now, there are no FD properties.
1643     */
1644    return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHX;
1645 }
1646
1647 void anv_FreeMemory(
1648     VkDevice                                    _device,
1649     VkDeviceMemory                              _mem,
1650     const VkAllocationCallbacks*                pAllocator)
1651 {
1652    ANV_FROM_HANDLE(anv_device, device, _device);
1653    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1654
1655    if (mem == NULL)
1656       return;
1657
1658    if (mem->map)
1659       anv_UnmapMemory(_device, _mem);
1660
1661    anv_bo_cache_release(device, &device->bo_cache, mem->bo);
1662
1663    vk_free2(&device->alloc, pAllocator, mem);
1664 }
1665
1666 VkResult anv_MapMemory(
1667     VkDevice                                    _device,
1668     VkDeviceMemory                              _memory,
1669     VkDeviceSize                                offset,
1670     VkDeviceSize                                size,
1671     VkMemoryMapFlags                            flags,
1672     void**                                      ppData)
1673 {
1674    ANV_FROM_HANDLE(anv_device, device, _device);
1675    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1676
1677    if (mem == NULL) {
1678       *ppData = NULL;
1679       return VK_SUCCESS;
1680    }
1681
1682    if (size == VK_WHOLE_SIZE)
1683       size = mem->bo->size - offset;
1684
1685    /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1686     *
1687     *  * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1688     *    assert(size != 0);
1689     *  * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1690     *    equal to the size of the memory minus offset
1691     */
1692    assert(size > 0);
1693    assert(offset + size <= mem->bo->size);
1694
1695    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1696     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1697     * at a time is valid. We could just mmap up front and return an offset
1698     * pointer here, but that may exhaust virtual memory on 32 bit
1699     * userspace. */
1700
1701    uint32_t gem_flags = 0;
1702
1703    if (!device->info.has_llc &&
1704        (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
1705       gem_flags |= I915_MMAP_WC;
1706
1707    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1708    uint64_t map_offset = offset & ~4095ull;
1709    assert(offset >= map_offset);
1710    uint64_t map_size = (offset + size) - map_offset;
1711
1712    /* Let's map whole pages */
1713    map_size = align_u64(map_size, 4096);
1714
1715    void *map = anv_gem_mmap(device, mem->bo->gem_handle,
1716                             map_offset, map_size, gem_flags);
1717    if (map == MAP_FAILED)
1718       return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
1719
1720    mem->map = map;
1721    mem->map_size = map_size;
1722
1723    *ppData = mem->map + (offset - map_offset);
1724
1725    return VK_SUCCESS;
1726 }
1727
1728 void anv_UnmapMemory(
1729     VkDevice                                    _device,
1730     VkDeviceMemory                              _memory)
1731 {
1732    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1733
1734    if (mem == NULL)
1735       return;
1736
1737    anv_gem_munmap(mem->map, mem->map_size);
1738
1739    mem->map = NULL;
1740    mem->map_size = 0;
1741 }
1742
1743 static void
1744 clflush_mapped_ranges(struct anv_device         *device,
1745                       uint32_t                   count,
1746                       const VkMappedMemoryRange *ranges)
1747 {
1748    for (uint32_t i = 0; i < count; i++) {
1749       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1750       if (ranges[i].offset >= mem->map_size)
1751          continue;
1752
1753       anv_clflush_range(mem->map + ranges[i].offset,
1754                         MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
1755    }
1756 }
1757
1758 VkResult anv_FlushMappedMemoryRanges(
1759     VkDevice                                    _device,
1760     uint32_t                                    memoryRangeCount,
1761     const VkMappedMemoryRange*                  pMemoryRanges)
1762 {
1763    ANV_FROM_HANDLE(anv_device, device, _device);
1764
1765    if (device->info.has_llc)
1766       return VK_SUCCESS;
1767
1768    /* Make sure the writes we're flushing have landed. */
1769    __builtin_ia32_mfence();
1770
1771    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1772
1773    return VK_SUCCESS;
1774 }
1775
1776 VkResult anv_InvalidateMappedMemoryRanges(
1777     VkDevice                                    _device,
1778     uint32_t                                    memoryRangeCount,
1779     const VkMappedMemoryRange*                  pMemoryRanges)
1780 {
1781    ANV_FROM_HANDLE(anv_device, device, _device);
1782
1783    if (device->info.has_llc)
1784       return VK_SUCCESS;
1785
1786    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1787
1788    /* Make sure no reads get moved up above the invalidate. */
1789    __builtin_ia32_mfence();
1790
1791    return VK_SUCCESS;
1792 }
1793
1794 void anv_GetBufferMemoryRequirements(
1795     VkDevice                                    _device,
1796     VkBuffer                                    _buffer,
1797     VkMemoryRequirements*                       pMemoryRequirements)
1798 {
1799    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1800    ANV_FROM_HANDLE(anv_device, device, _device);
1801    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
1802
1803    /* The Vulkan spec (git aaed022) says:
1804     *
1805     *    memoryTypeBits is a bitfield and contains one bit set for every
1806     *    supported memory type for the resource. The bit `1<<i` is set if and
1807     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1808     *    structure for the physical device is supported.
1809     */
1810    uint32_t memory_types = 0;
1811    for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
1812       uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
1813       if ((valid_usage & buffer->usage) == buffer->usage)
1814          memory_types |= (1u << i);
1815    }
1816
1817    pMemoryRequirements->size = buffer->size;
1818    pMemoryRequirements->alignment = 16;
1819    pMemoryRequirements->memoryTypeBits = memory_types;
1820 }
1821
1822 void anv_GetImageMemoryRequirements(
1823     VkDevice                                    _device,
1824     VkImage                                     _image,
1825     VkMemoryRequirements*                       pMemoryRequirements)
1826 {
1827    ANV_FROM_HANDLE(anv_image, image, _image);
1828    ANV_FROM_HANDLE(anv_device, device, _device);
1829    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
1830
1831    /* The Vulkan spec (git aaed022) says:
1832     *
1833     *    memoryTypeBits is a bitfield and contains one bit set for every
1834     *    supported memory type for the resource. The bit `1<<i` is set if and
1835     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1836     *    structure for the physical device is supported.
1837     *
1838     * All types are currently supported for images.
1839     */
1840    uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
1841
1842    pMemoryRequirements->size = image->size;
1843    pMemoryRequirements->alignment = image->alignment;
1844    pMemoryRequirements->memoryTypeBits = memory_types;
1845 }
1846
1847 void anv_GetImageSparseMemoryRequirements(
1848     VkDevice                                    device,
1849     VkImage                                     image,
1850     uint32_t*                                   pSparseMemoryRequirementCount,
1851     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1852 {
1853    *pSparseMemoryRequirementCount = 0;
1854 }
1855
1856 void anv_GetDeviceMemoryCommitment(
1857     VkDevice                                    device,
1858     VkDeviceMemory                              memory,
1859     VkDeviceSize*                               pCommittedMemoryInBytes)
1860 {
1861    *pCommittedMemoryInBytes = 0;
1862 }
1863
1864 VkResult anv_BindBufferMemory(
1865     VkDevice                                    device,
1866     VkBuffer                                    _buffer,
1867     VkDeviceMemory                              _memory,
1868     VkDeviceSize                                memoryOffset)
1869 {
1870    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1871    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1872
1873    if (mem) {
1874       assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
1875       buffer->bo = mem->bo;
1876       buffer->offset = memoryOffset;
1877    } else {
1878       buffer->bo = NULL;
1879       buffer->offset = 0;
1880    }
1881
1882    return VK_SUCCESS;
1883 }
1884
1885 VkResult anv_QueueBindSparse(
1886     VkQueue                                     _queue,
1887     uint32_t                                    bindInfoCount,
1888     const VkBindSparseInfo*                     pBindInfo,
1889     VkFence                                     fence)
1890 {
1891    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1892    if (unlikely(queue->device->lost))
1893       return VK_ERROR_DEVICE_LOST;
1894
1895    return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1896 }
1897
1898 // Event functions
1899
1900 VkResult anv_CreateEvent(
1901     VkDevice                                    _device,
1902     const VkEventCreateInfo*                    pCreateInfo,
1903     const VkAllocationCallbacks*                pAllocator,
1904     VkEvent*                                    pEvent)
1905 {
1906    ANV_FROM_HANDLE(anv_device, device, _device);
1907    struct anv_state state;
1908    struct anv_event *event;
1909
1910    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1911
1912    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1913                                 sizeof(*event), 8);
1914    event = state.map;
1915    event->state = state;
1916    event->semaphore = VK_EVENT_RESET;
1917
1918    if (!device->info.has_llc) {
1919       /* Make sure the writes we're flushing have landed. */
1920       __builtin_ia32_mfence();
1921       __builtin_ia32_clflush(event);
1922    }
1923
1924    *pEvent = anv_event_to_handle(event);
1925
1926    return VK_SUCCESS;
1927 }
1928
1929 void anv_DestroyEvent(
1930     VkDevice                                    _device,
1931     VkEvent                                     _event,
1932     const VkAllocationCallbacks*                pAllocator)
1933 {
1934    ANV_FROM_HANDLE(anv_device, device, _device);
1935    ANV_FROM_HANDLE(anv_event, event, _event);
1936
1937    if (!event)
1938       return;
1939
1940    anv_state_pool_free(&device->dynamic_state_pool, event->state);
1941 }
1942
1943 VkResult anv_GetEventStatus(
1944     VkDevice                                    _device,
1945     VkEvent                                     _event)
1946 {
1947    ANV_FROM_HANDLE(anv_device, device, _device);
1948    ANV_FROM_HANDLE(anv_event, event, _event);
1949
1950    if (unlikely(device->lost))
1951       return VK_ERROR_DEVICE_LOST;
1952
1953    if (!device->info.has_llc) {
1954       /* Invalidate read cache before reading event written by GPU. */
1955       __builtin_ia32_clflush(event);
1956       __builtin_ia32_mfence();
1957
1958    }
1959
1960    return event->semaphore;
1961 }
1962
1963 VkResult anv_SetEvent(
1964     VkDevice                                    _device,
1965     VkEvent                                     _event)
1966 {
1967    ANV_FROM_HANDLE(anv_device, device, _device);
1968    ANV_FROM_HANDLE(anv_event, event, _event);
1969
1970    event->semaphore = VK_EVENT_SET;
1971
1972    if (!device->info.has_llc) {
1973       /* Make sure the writes we're flushing have landed. */
1974       __builtin_ia32_mfence();
1975       __builtin_ia32_clflush(event);
1976    }
1977
1978    return VK_SUCCESS;
1979 }
1980
1981 VkResult anv_ResetEvent(
1982     VkDevice                                    _device,
1983     VkEvent                                     _event)
1984 {
1985    ANV_FROM_HANDLE(anv_device, device, _device);
1986    ANV_FROM_HANDLE(anv_event, event, _event);
1987
1988    event->semaphore = VK_EVENT_RESET;
1989
1990    if (!device->info.has_llc) {
1991       /* Make sure the writes we're flushing have landed. */
1992       __builtin_ia32_mfence();
1993       __builtin_ia32_clflush(event);
1994    }
1995
1996    return VK_SUCCESS;
1997 }
1998
1999 // Buffer functions
2000
2001 VkResult anv_CreateBuffer(
2002     VkDevice                                    _device,
2003     const VkBufferCreateInfo*                   pCreateInfo,
2004     const VkAllocationCallbacks*                pAllocator,
2005     VkBuffer*                                   pBuffer)
2006 {
2007    ANV_FROM_HANDLE(anv_device, device, _device);
2008    struct anv_buffer *buffer;
2009
2010    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2011
2012    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2013                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2014    if (buffer == NULL)
2015       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2016
2017    buffer->size = pCreateInfo->size;
2018    buffer->usage = pCreateInfo->usage;
2019    buffer->bo = NULL;
2020    buffer->offset = 0;
2021
2022    *pBuffer = anv_buffer_to_handle(buffer);
2023
2024    return VK_SUCCESS;
2025 }
2026
2027 void anv_DestroyBuffer(
2028     VkDevice                                    _device,
2029     VkBuffer                                    _buffer,
2030     const VkAllocationCallbacks*                pAllocator)
2031 {
2032    ANV_FROM_HANDLE(anv_device, device, _device);
2033    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2034
2035    if (!buffer)
2036       return;
2037
2038    vk_free2(&device->alloc, pAllocator, buffer);
2039 }
2040
2041 void
2042 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
2043                               enum isl_format format,
2044                               uint32_t offset, uint32_t range, uint32_t stride)
2045 {
2046    isl_buffer_fill_state(&device->isl_dev, state.map,
2047                          .address = offset,
2048                          .mocs = device->default_mocs,
2049                          .size = range,
2050                          .format = format,
2051                          .stride = stride);
2052
2053    anv_state_flush(device, state);
2054 }
2055
2056 void anv_DestroySampler(
2057     VkDevice                                    _device,
2058     VkSampler                                   _sampler,
2059     const VkAllocationCallbacks*                pAllocator)
2060 {
2061    ANV_FROM_HANDLE(anv_device, device, _device);
2062    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
2063
2064    if (!sampler)
2065       return;
2066
2067    vk_free2(&device->alloc, pAllocator, sampler);
2068 }
2069
2070 VkResult anv_CreateFramebuffer(
2071     VkDevice                                    _device,
2072     const VkFramebufferCreateInfo*              pCreateInfo,
2073     const VkAllocationCallbacks*                pAllocator,
2074     VkFramebuffer*                              pFramebuffer)
2075 {
2076    ANV_FROM_HANDLE(anv_device, device, _device);
2077    struct anv_framebuffer *framebuffer;
2078
2079    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2080
2081    size_t size = sizeof(*framebuffer) +
2082                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2083    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2084                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2085    if (framebuffer == NULL)
2086       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2087
2088    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2089    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2090       VkImageView _iview = pCreateInfo->pAttachments[i];
2091       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2092    }
2093
2094    framebuffer->width = pCreateInfo->width;
2095    framebuffer->height = pCreateInfo->height;
2096    framebuffer->layers = pCreateInfo->layers;
2097
2098    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2099
2100    return VK_SUCCESS;
2101 }
2102
2103 void anv_DestroyFramebuffer(
2104     VkDevice                                    _device,
2105     VkFramebuffer                               _fb,
2106     const VkAllocationCallbacks*                pAllocator)
2107 {
2108    ANV_FROM_HANDLE(anv_device, device, _device);
2109    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2110
2111    if (!fb)
2112       return;
2113
2114    vk_free2(&device->alloc, pAllocator, fb);
2115 }
2116
2117 /* vk_icd.h does not declare this function, so we declare it here to
2118  * suppress Wmissing-prototypes.
2119  */
2120 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2121 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2122
2123 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2124 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2125 {
2126    /* For the full details on loader interface versioning, see
2127     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2128     * What follows is a condensed summary, to help you navigate the large and
2129     * confusing official doc.
2130     *
2131     *   - Loader interface v0 is incompatible with later versions. We don't
2132     *     support it.
2133     *
2134     *   - In loader interface v1:
2135     *       - The first ICD entrypoint called by the loader is
2136     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2137     *         entrypoint.
2138     *       - The ICD must statically expose no other Vulkan symbol unless it is
2139     *         linked with -Bsymbolic.
2140     *       - Each dispatchable Vulkan handle created by the ICD must be
2141     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2142     *         ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2143     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2144     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2145     *         such loader-managed surfaces.
2146     *
2147     *    - Loader interface v2 differs from v1 in:
2148     *       - The first ICD entrypoint called by the loader is
2149     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2150     *         statically expose this entrypoint.
2151     *
2152     *    - Loader interface v3 differs from v2 in:
2153     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2154     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2155     *          because the loader no longer does so.
2156     */
2157    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2158    return VK_SUCCESS;
2159 }