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_EXTERNAL_MEMORY_CAPABILITIES_EXTENSION_NAME,
 412       .specVersion = 1,
 413    },
 414    {
 415       .extensionName = VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
 416       .specVersion = 1,
 417    },
 418    {
 419       .extensionName = VK_KHR_GET_SURFACE_CAPABILITIES_2_EXTENSION_NAME,
 420       .specVersion = 1,
 421    },
 422    {
 423       .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
 424       .specVersion = 25,
 425    },
 426 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
 427    {
 428       .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
 429       .specVersion = 6,
 430    },
 431 #endif
 432 #ifdef VK_USE_PLATFORM_XCB_KHR
 433    {
 434       .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
 435       .specVersion = 6,
 436    },
 437 #endif
 438 #ifdef VK_USE_PLATFORM_XLIB_KHR
 439    {
 440       .extensionName = VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
 441       .specVersion = 6,
 442    },
 443 #endif
 444 };
 445
 446 static const VkExtensionProperties device_extensions[] = {
 447    {
 448       .extensionName = VK_KHR_DEDICATED_ALLOCATION_EXTENSION_NAME,
 449       .specVersion = 1,
 450    },
 451    {
 452       .extensionName = VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME,
 453       .specVersion = 1,
 454    },
 455    {
 456       .extensionName = VK_KHR_EXTERNAL_MEMORY_EXTENSION_NAME,
 457       .specVersion = 1,
 458    },
 459    {
 460       .extensionName = VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME,
 461       .specVersion = 1,
 462    },
 463    {
 464       .extensionName = VK_KHR_GET_MEMORY_REQUIREMENTS_2_EXTENSION_NAME,
 465       .specVersion = 1,
 466    },
 467    {
 468       .extensionName = VK_KHR_INCREMENTAL_PRESENT_EXTENSION_NAME,
 469       .specVersion = 1,
 470    },
 471    {
 472       .extensionName = VK_KHR_MAINTENANCE1_EXTENSION_NAME,
 473       .specVersion = 1,
 474    },
 475    {
 476       .extensionName = VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME,
 477       .specVersion = 1,
 478    },
 479    {
 480       .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
 481       .specVersion = 1,
 482    },
 483    {
 484       .extensionName = VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
 485       .specVersion = 1,
 486    },
 487    {
 488       .extensionName = VK_KHR_STORAGE_BUFFER_STORAGE_CLASS_EXTENSION_NAME,
 489       .specVersion = 1,
 490    },
 491    {
 492       .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
 493       .specVersion = 68,
 494    },
 495    {
 496       .extensionName = VK_KHR_VARIABLE_POINTERS_EXTENSION_NAME,
 497       .specVersion = 1,
 498    },
 499    {
 500       .extensionName = VK_KHX_MULTIVIEW_EXTENSION_NAME,
 501       .specVersion = 1,
 502    },
 503 };
 504
 505 static void *
 506 default_alloc_func(void *pUserData, size_t size, size_t align,
 507                    VkSystemAllocationScope allocationScope)
 508 {
 509    return malloc(size);
 510 }
 511
 512 static void *
 513 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
 514                      size_t align, VkSystemAllocationScope allocationScope)
 515 {
 516    return realloc(pOriginal, size);
 517 }
 518
 519 static void
 520 default_free_func(void *pUserData, void *pMemory)
 521 {
 522    free(pMemory);
 523 }
 524
 525 static const VkAllocationCallbacks default_alloc = {
 526    .pUserData = NULL,
 527    .pfnAllocation = default_alloc_func,
 528    .pfnReallocation = default_realloc_func,
 529    .pfnFree = default_free_func,
 530 };
 531
 532 VkResult anv_CreateInstance(
 533     const VkInstanceCreateInfo*                 pCreateInfo,
 534     const VkAllocationCallbacks*                pAllocator,
 535     VkInstance*                                 pInstance)
 536 {
 537    struct anv_instance *instance;
 538
 539    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 540
 541    uint32_t client_version;
 542    if (pCreateInfo->pApplicationInfo &&
 543        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 544       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 545    } else {
 546       client_version = VK_MAKE_VERSION(1, 0, 0);
 547    }
 548
 549    if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
 550        client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
 551       return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 552                        "Client requested version %d.%d.%d",
 553                        VK_VERSION_MAJOR(client_version),
 554                        VK_VERSION_MINOR(client_version),
 555                        VK_VERSION_PATCH(client_version));
 556    }
 557
 558    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 559       bool found = false;
 560       for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
 561          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 562                     global_extensions[j].extensionName) == 0) {
 563             found = true;
 564             break;
 565          }
 566       }
 567       if (!found)
 568          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 569    }
 570
 571    instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 572                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 573    if (!instance)
 574       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 575
 576    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 577
 578    if (pAllocator)
 579       instance->alloc = *pAllocator;
 580    else
 581       instance->alloc = default_alloc;
 582
 583    instance->apiVersion = client_version;
 584    instance->physicalDeviceCount = -1;
 585
 586    _mesa_locale_init();
 587
 588    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 589
 590    *pInstance = anv_instance_to_handle(instance);
 591
 592    return VK_SUCCESS;
 593 }
 594
 595 void anv_DestroyInstance(
 596     VkInstance                                  _instance,
 597     const VkAllocationCallbacks*                pAllocator)
 598 {
 599    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 600
 601    if (!instance)
 602       return;
 603
 604    if (instance->physicalDeviceCount > 0) {
 605       /* We support at most one physical device. */
 606       assert(instance->physicalDeviceCount == 1);
 607       anv_physical_device_finish(&instance->physicalDevice);
 608    }
 609
 610    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 611
 612    _mesa_locale_fini();
 613
 614    vk_free(&instance->alloc, instance);
 615 }
 616
 617 static VkResult
 618 anv_enumerate_devices(struct anv_instance *instance)
 619 {
 620    /* TODO: Check for more devices ? */
 621    drmDevicePtr devices[8];
 622    VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
 623    int max_devices;
 624
 625    instance->physicalDeviceCount = 0;
 626
 627    max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
 628    if (max_devices < 1)
 629       return VK_ERROR_INCOMPATIBLE_DRIVER;
 630
 631    for (unsigned i = 0; i < (unsigned)max_devices; i++) {
 632       if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
 633           devices[i]->bustype == DRM_BUS_PCI &&
 634           devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
 635
 636          result = anv_physical_device_init(&instance->physicalDevice,
 637                         instance,
 638                         devices[i]->nodes[DRM_NODE_RENDER]);
 639          if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 640             break;
 641       }
 642    }
 643    drmFreeDevices(devices, max_devices);
 644
 645    if (result == VK_SUCCESS)
 646       instance->physicalDeviceCount = 1;
 647
 648    return result;
 649 }
 650
 651
 652 VkResult anv_EnumeratePhysicalDevices(
 653     VkInstance                                  _instance,
 654     uint32_t*                                   pPhysicalDeviceCount,
 655     VkPhysicalDevice*                           pPhysicalDevices)
 656 {
 657    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 658    VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
 659    VkResult result;
 660
 661    if (instance->physicalDeviceCount < 0) {
 662       result = anv_enumerate_devices(instance);
 663       if (result != VK_SUCCESS &&
 664           result != VK_ERROR_INCOMPATIBLE_DRIVER)
 665          return result;
 666    }
 667
 668    if (instance->physicalDeviceCount > 0) {
 669       assert(instance->physicalDeviceCount == 1);
 670       vk_outarray_append(&out, i) {
 671          *i = anv_physical_device_to_handle(&instance->physicalDevice);
 672       }
 673    }
 674
 675    return vk_outarray_status(&out);
 676 }
 677
 678 void anv_GetPhysicalDeviceFeatures(
 679     VkPhysicalDevice                            physicalDevice,
 680     VkPhysicalDeviceFeatures*                   pFeatures)
 681 {
 682    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 683
 684    *pFeatures = (VkPhysicalDeviceFeatures) {
 685       .robustBufferAccess                       = true,
 686       .fullDrawIndexUint32                      = true,
 687       .imageCubeArray                           = true,
 688       .independentBlend                         = true,
 689       .geometryShader                           = true,
 690       .tessellationShader                       = true,
 691       .sampleRateShading                        = true,
 692       .dualSrcBlend                             = true,
 693       .logicOp                                  = true,
 694       .multiDrawIndirect                        = true,
 695       .drawIndirectFirstInstance                = true,
 696       .depthClamp                               = true,
 697       .depthBiasClamp                           = true,
 698       .fillModeNonSolid                         = true,
 699       .depthBounds                              = false,
 700       .wideLines                                = true,
 701       .largePoints                              = true,
 702       .alphaToOne                               = true,
 703       .multiViewport                            = true,
 704       .samplerAnisotropy                        = true,
 705       .textureCompressionETC2                   = pdevice->info.gen >= 8 ||
 706                                                   pdevice->info.is_baytrail,
 707       .textureCompressionASTC_LDR               = pdevice->info.gen >= 9, /* FINISHME CHV */
 708       .textureCompressionBC                     = true,
 709       .occlusionQueryPrecise                    = true,
 710       .pipelineStatisticsQuery                  = true,
 711       .fragmentStoresAndAtomics                 = true,
 712       .shaderTessellationAndGeometryPointSize   = true,
 713       .shaderImageGatherExtended                = true,
 714       .shaderStorageImageExtendedFormats        = true,
 715       .shaderStorageImageMultisample            = false,
 716       .shaderStorageImageReadWithoutFormat      = false,
 717       .shaderStorageImageWriteWithoutFormat     = true,
 718       .shaderUniformBufferArrayDynamicIndexing  = true,
 719       .shaderSampledImageArrayDynamicIndexing   = true,
 720       .shaderStorageBufferArrayDynamicIndexing  = true,
 721       .shaderStorageImageArrayDynamicIndexing   = true,
 722       .shaderClipDistance                       = true,
 723       .shaderCullDistance                       = true,
 724       .shaderFloat64                            = pdevice->info.gen >= 8,
 725       .shaderInt64                              = pdevice->info.gen >= 8,
 726       .shaderInt16                              = false,
 727       .shaderResourceMinLod                     = false,
 728       .variableMultisampleRate                  = false,
 729       .inheritedQueries                         = true,
 730    };
 731
 732    /* We can't do image stores in vec4 shaders */
 733    pFeatures->vertexPipelineStoresAndAtomics =
 734       pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
 735       pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
 736 }
 737
 738 void anv_GetPhysicalDeviceFeatures2KHR(
 739     VkPhysicalDevice                            physicalDevice,
 740     VkPhysicalDeviceFeatures2KHR*               pFeatures)
 741 {
 742    anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 743
 744    vk_foreach_struct(ext, pFeatures->pNext) {
 745       switch (ext->sType) {
 746       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX: {
 747          VkPhysicalDeviceMultiviewFeaturesKHX *features =
 748             (VkPhysicalDeviceMultiviewFeaturesKHX *)ext;
 749          features->multiview = true;
 750          features->multiviewGeometryShader = true;
 751          features->multiviewTessellationShader = true;
 752          break;
 753       }
 754
 755       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR: {
 756          VkPhysicalDeviceVariablePointerFeaturesKHR *features = (void *)ext;
 757          features->variablePointersStorageBuffer = true;
 758          features->variablePointers = false;
 759          break;
 760       }
 761
 762       default:
 763          anv_debug_ignored_stype(ext->sType);
 764          break;
 765       }
 766    }
 767 }
 768
 769 void anv_GetPhysicalDeviceProperties(
 770     VkPhysicalDevice                            physicalDevice,
 771     VkPhysicalDeviceProperties*                 pProperties)
 772 {
 773    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 774    const struct gen_device_info *devinfo = &pdevice->info;
 775
 776    /* See assertions made when programming the buffer surface state. */
 777    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
 778                                       (1ul << 30) : (1ul << 27);
 779
 780    const uint32_t max_samplers = (devinfo->gen >= 8 || devinfo->is_haswell) ?
 781                                  128 : 16;
 782
 783    VkSampleCountFlags sample_counts =
 784       isl_device_get_sample_counts(&pdevice->isl_dev);
 785
 786    VkPhysicalDeviceLimits limits = {
 787       .maxImageDimension1D                      = (1 << 14),
 788       .maxImageDimension2D                      = (1 << 14),
 789       .maxImageDimension3D                      = (1 << 11),
 790       .maxImageDimensionCube                    = (1 << 14),
 791       .maxImageArrayLayers                      = (1 << 11),
 792       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 793       .maxUniformBufferRange                    = (1ul << 27),
 794       .maxStorageBufferRange                    = max_raw_buffer_sz,
 795       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 796       .maxMemoryAllocationCount                 = UINT32_MAX,
 797       .maxSamplerAllocationCount                = 64 * 1024,
 798       .bufferImageGranularity                   = 64, /* A cache line */
 799       .sparseAddressSpaceSize                   = 0,
 800       .maxBoundDescriptorSets                   = MAX_SETS,
 801       .maxPerStageDescriptorSamplers            = max_samplers,
 802       .maxPerStageDescriptorUniformBuffers      = 64,
 803       .maxPerStageDescriptorStorageBuffers      = 64,
 804       .maxPerStageDescriptorSampledImages       = max_samplers,
 805       .maxPerStageDescriptorStorageImages       = 64,
 806       .maxPerStageDescriptorInputAttachments    = 64,
 807       .maxPerStageResources                     = 250,
 808       .maxDescriptorSetSamplers                 = 256,
 809       .maxDescriptorSetUniformBuffers           = 256,
 810       .maxDescriptorSetUniformBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
 811       .maxDescriptorSetStorageBuffers           = 256,
 812       .maxDescriptorSetStorageBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
 813       .maxDescriptorSetSampledImages            = 256,
 814       .maxDescriptorSetStorageImages            = 256,
 815       .maxDescriptorSetInputAttachments         = 256,
 816       .maxVertexInputAttributes                 = MAX_VBS,
 817       .maxVertexInputBindings                   = MAX_VBS,
 818       .maxVertexInputAttributeOffset            = 2047,
 819       .maxVertexInputBindingStride              = 2048,
 820       .maxVertexOutputComponents                = 128,
 821       .maxTessellationGenerationLevel           = 64,
 822       .maxTessellationPatchSize                 = 32,
 823       .maxTessellationControlPerVertexInputComponents = 128,
 824       .maxTessellationControlPerVertexOutputComponents = 128,
 825       .maxTessellationControlPerPatchOutputComponents = 128,
 826       .maxTessellationControlTotalOutputComponents = 2048,
 827       .maxTessellationEvaluationInputComponents = 128,
 828       .maxTessellationEvaluationOutputComponents = 128,
 829       .maxGeometryShaderInvocations             = 32,
 830       .maxGeometryInputComponents               = 64,
 831       .maxGeometryOutputComponents              = 128,
 832       .maxGeometryOutputVertices                = 256,
 833       .maxGeometryTotalOutputComponents         = 1024,
 834       .maxFragmentInputComponents               = 128,
 835       .maxFragmentOutputAttachments             = 8,
 836       .maxFragmentDualSrcAttachments            = 1,
 837       .maxFragmentCombinedOutputResources       = 8,
 838       .maxComputeSharedMemorySize               = 32768,
 839       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 840       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 841       .maxComputeWorkGroupSize = {
 842          16 * devinfo->max_cs_threads,
 843          16 * devinfo->max_cs_threads,
 844          16 * devinfo->max_cs_threads,
 845       },
 846       .subPixelPrecisionBits                    = 4 /* FIXME */,
 847       .subTexelPrecisionBits                    = 4 /* FIXME */,
 848       .mipmapPrecisionBits                      = 4 /* FIXME */,
 849       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 850       .maxDrawIndirectCount                     = UINT32_MAX,
 851       .maxSamplerLodBias                        = 16,
 852       .maxSamplerAnisotropy                     = 16,
 853       .maxViewports                             = MAX_VIEWPORTS,
 854       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 855       .viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
 856       .viewportSubPixelBits                     = 13, /* We take a float? */
 857       .minMemoryMapAlignment                    = 4096, /* A page */
 858       .minTexelBufferOffsetAlignment            = 1,
 859       .minUniformBufferOffsetAlignment          = 16,
 860       .minStorageBufferOffsetAlignment          = 4,
 861       .minTexelOffset                           = -8,
 862       .maxTexelOffset                           = 7,
 863       .minTexelGatherOffset                     = -32,
 864       .maxTexelGatherOffset                     = 31,
 865       .minInterpolationOffset                   = -0.5,
 866       .maxInterpolationOffset                   = 0.4375,
 867       .subPixelInterpolationOffsetBits          = 4,
 868       .maxFramebufferWidth                      = (1 << 14),
 869       .maxFramebufferHeight                     = (1 << 14),
 870       .maxFramebufferLayers                     = (1 << 11),
 871       .framebufferColorSampleCounts             = sample_counts,
 872       .framebufferDepthSampleCounts             = sample_counts,
 873       .framebufferStencilSampleCounts           = sample_counts,
 874       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 875       .maxColorAttachments                      = MAX_RTS,
 876       .sampledImageColorSampleCounts            = sample_counts,
 877       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 878       .sampledImageDepthSampleCounts            = sample_counts,
 879       .sampledImageStencilSampleCounts          = sample_counts,
 880       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 881       .maxSampleMaskWords                       = 1,
 882       .timestampComputeAndGraphics              = false,
 883       .timestampPeriod                          = 1000000000.0 / devinfo->timestamp_frequency,
 884       .maxClipDistances                         = 8,
 885       .maxCullDistances                         = 8,
 886       .maxCombinedClipAndCullDistances          = 8,
 887       .discreteQueuePriorities                  = 1,
 888       .pointSizeRange                           = { 0.125, 255.875 },
 889       .lineWidthRange                           = { 0.0, 7.9921875 },
 890       .pointSizeGranularity                     = (1.0 / 8.0),
 891       .lineWidthGranularity                     = (1.0 / 128.0),
 892       .strictLines                              = false, /* FINISHME */
 893       .standardSampleLocations                  = true,
 894       .optimalBufferCopyOffsetAlignment         = 128,
 895       .optimalBufferCopyRowPitchAlignment       = 128,
 896       .nonCoherentAtomSize                      = 64,
 897    };
 898
 899    *pProperties = (VkPhysicalDeviceProperties) {
 900       .apiVersion = VK_MAKE_VERSION(1, 0, 54),
 901       .driverVersion = vk_get_driver_version(),
 902       .vendorID = 0x8086,
 903       .deviceID = pdevice->chipset_id,
 904       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 905       .limits = limits,
 906       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 907    };
 908
 909    snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
 910             "%s", pdevice->name);
 911    memcpy(pProperties->pipelineCacheUUID,
 912           pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
 913 }
 914
 915 void anv_GetPhysicalDeviceProperties2KHR(
 916     VkPhysicalDevice                            physicalDevice,
 917     VkPhysicalDeviceProperties2KHR*             pProperties)
 918 {
 919    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 920
 921    anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 922
 923    vk_foreach_struct(ext, pProperties->pNext) {
 924       switch (ext->sType) {
 925       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 926          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 927             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 928
 929          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 930          break;
 931       }
 932
 933       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR: {
 934          VkPhysicalDeviceIDPropertiesKHR *id_props =
 935             (VkPhysicalDeviceIDPropertiesKHR *)ext;
 936          memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
 937          memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
 938          /* The LUID is for Windows. */
 939          id_props->deviceLUIDValid = false;
 940          break;
 941       }
 942
 943       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX: {
 944          VkPhysicalDeviceMultiviewPropertiesKHX *properties =
 945             (VkPhysicalDeviceMultiviewPropertiesKHX *)ext;
 946          properties->maxMultiviewViewCount = 16;
 947          properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
 948          break;
 949       }
 950
 951       default:
 952          anv_debug_ignored_stype(ext->sType);
 953          break;
 954       }
 955    }
 956 }
 957
 958 /* We support exactly one queue family. */
 959 static const VkQueueFamilyProperties
 960 anv_queue_family_properties = {
 961    .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 962                  VK_QUEUE_COMPUTE_BIT |
 963                  VK_QUEUE_TRANSFER_BIT,
 964    .queueCount = 1,
 965    .timestampValidBits = 36, /* XXX: Real value here */
 966    .minImageTransferGranularity = { 1, 1, 1 },
 967 };
 968
 969 void anv_GetPhysicalDeviceQueueFamilyProperties(
 970     VkPhysicalDevice                            physicalDevice,
 971     uint32_t*                                   pCount,
 972     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 973 {
 974    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
 975
 976    vk_outarray_append(&out, p) {
 977       *p = anv_queue_family_properties;
 978    }
 979 }
 980
 981 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
 982     VkPhysicalDevice                            physicalDevice,
 983     uint32_t*                                   pQueueFamilyPropertyCount,
 984     VkQueueFamilyProperties2KHR*                pQueueFamilyProperties)
 985 {
 986
 987    VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 988
 989    vk_outarray_append(&out, p) {
 990       p->queueFamilyProperties = anv_queue_family_properties;
 991
 992       vk_foreach_struct(s, p->pNext) {
 993          anv_debug_ignored_stype(s->sType);
 994       }
 995    }
 996 }
 997
 998 void anv_GetPhysicalDeviceMemoryProperties(
 999     VkPhysicalDevice                            physicalDevice,
1000     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
1001 {
1002    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1003
1004    pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1005    for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1006       pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1007          .propertyFlags = physical_device->memory.types[i].propertyFlags,
1008          .heapIndex     = physical_device->memory.types[i].heapIndex,
1009       };
1010    }
1011
1012    pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1013    for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1014       pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1015          .size    = physical_device->memory.heaps[i].size,
1016          .flags   = physical_device->memory.heaps[i].flags,
1017       };
1018    }
1019 }
1020
1021 void anv_GetPhysicalDeviceMemoryProperties2KHR(
1022     VkPhysicalDevice                            physicalDevice,
1023     VkPhysicalDeviceMemoryProperties2KHR*       pMemoryProperties)
1024 {
1025    anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1026                                          &pMemoryProperties->memoryProperties);
1027
1028    vk_foreach_struct(ext, pMemoryProperties->pNext) {
1029       switch (ext->sType) {
1030       default:
1031          anv_debug_ignored_stype(ext->sType);
1032          break;
1033       }
1034    }
1035 }
1036
1037 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1038     VkInstance                                  instance,
1039     const char*                                 pName)
1040 {
1041    return anv_lookup_entrypoint(NULL, pName);
1042 }
1043
1044 /* With version 1+ of the loader interface the ICD should expose
1045  * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1046  */
1047 PUBLIC
1048 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1049     VkInstance                                  instance,
1050     const char*                                 pName);
1051
1052 PUBLIC
1053 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1054     VkInstance                                  instance,
1055     const char*                                 pName)
1056 {
1057    return anv_GetInstanceProcAddr(instance, pName);
1058 }
1059
1060 PFN_vkVoidFunction anv_GetDeviceProcAddr(
1061     VkDevice                                    _device,
1062     const char*                                 pName)
1063 {
1064    ANV_FROM_HANDLE(anv_device, device, _device);
1065    return anv_lookup_entrypoint(&device->info, pName);
1066 }
1067
1068 static void
1069 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
1070 {
1071    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1072    queue->device = device;
1073    queue->pool = &device->surface_state_pool;
1074 }
1075
1076 static void
1077 anv_queue_finish(struct anv_queue *queue)
1078 {
1079 }
1080
1081 static struct anv_state
1082 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
1083 {
1084    struct anv_state state;
1085
1086    state = anv_state_pool_alloc(pool, size, align);
1087    memcpy(state.map, p, size);
1088
1089    anv_state_flush(pool->block_pool.device, state);
1090
1091    return state;
1092 }
1093
1094 struct gen8_border_color {
1095    union {
1096       float float32[4];
1097       uint32_t uint32[4];
1098    };
1099    /* Pad out to 64 bytes */
1100    uint32_t _pad[12];
1101 };
1102
1103 static void
1104 anv_device_init_border_colors(struct anv_device *device)
1105 {
1106    static const struct gen8_border_color border_colors[] = {
1107       [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] =  { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
1108       [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] =       { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
1109       [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] =       { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
1110       [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] =    { .uint32 = { 0, 0, 0, 0 } },
1111       [VK_BORDER_COLOR_INT_OPAQUE_BLACK] =         { .uint32 = { 0, 0, 0, 1 } },
1112       [VK_BORDER_COLOR_INT_OPAQUE_WHITE] =         { .uint32 = { 1, 1, 1, 1 } },
1113    };
1114
1115    device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
1116                                                     sizeof(border_colors), 64,
1117                                                     border_colors);
1118 }
1119
1120 VkResult anv_CreateDevice(
1121     VkPhysicalDevice                            physicalDevice,
1122     const VkDeviceCreateInfo*                   pCreateInfo,
1123     const VkAllocationCallbacks*                pAllocator,
1124     VkDevice*                                   pDevice)
1125 {
1126    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1127    VkResult result;
1128    struct anv_device *device;
1129
1130    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
1131
1132    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1133       bool found = false;
1134       for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
1135          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
1136                     device_extensions[j].extensionName) == 0) {
1137             found = true;
1138             break;
1139          }
1140       }
1141       if (!found)
1142          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1143    }
1144
1145    /* Check enabled features */
1146    if (pCreateInfo->pEnabledFeatures) {
1147       VkPhysicalDeviceFeatures supported_features;
1148       anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1149       VkBool32 *supported_feature = (VkBool32 *)&supported_features;
1150       VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
1151       unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1152       for (uint32_t i = 0; i < num_features; i++) {
1153          if (enabled_feature[i] && !supported_feature[i])
1154             return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1155       }
1156    }
1157
1158    device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
1159                        sizeof(*device), 8,
1160                        VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1161    if (!device)
1162       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1163
1164    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1165    device->instance = physical_device->instance;
1166    device->chipset_id = physical_device->chipset_id;
1167    device->lost = false;
1168
1169    if (pAllocator)
1170       device->alloc = *pAllocator;
1171    else
1172       device->alloc = physical_device->instance->alloc;
1173
1174    /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1175    device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1176    if (device->fd == -1) {
1177       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1178       goto fail_device;
1179    }
1180
1181    device->context_id = anv_gem_create_context(device);
1182    if (device->context_id == -1) {
1183       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1184       goto fail_fd;
1185    }
1186
1187    device->info = physical_device->info;
1188    device->isl_dev = physical_device->isl_dev;
1189
1190    /* On Broadwell and later, we can use batch chaining to more efficiently
1191     * implement growing command buffers.  Prior to Haswell, the kernel
1192     * command parser gets in the way and we have to fall back to growing
1193     * the batch.
1194     */
1195    device->can_chain_batches = device->info.gen >= 8;
1196
1197    device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1198       pCreateInfo->pEnabledFeatures->robustBufferAccess;
1199
1200    if (pthread_mutex_init(&device->mutex, NULL) != 0) {
1201       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1202       goto fail_context_id;
1203    }
1204
1205    pthread_condattr_t condattr;
1206    if (pthread_condattr_init(&condattr) != 0) {
1207       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1208       goto fail_mutex;
1209    }
1210    if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
1211       pthread_condattr_destroy(&condattr);
1212       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1213       goto fail_mutex;
1214    }
1215    if (pthread_cond_init(&device->queue_submit, NULL) != 0) {
1216       pthread_condattr_destroy(&condattr);
1217       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1218       goto fail_mutex;
1219    }
1220    pthread_condattr_destroy(&condattr);
1221
1222    anv_bo_pool_init(&device->batch_bo_pool, device);
1223
1224    result = anv_bo_cache_init(&device->bo_cache);
1225    if (result != VK_SUCCESS)
1226       goto fail_batch_bo_pool;
1227
1228    result = anv_state_pool_init(&device->dynamic_state_pool, device, 16384);
1229    if (result != VK_SUCCESS)
1230       goto fail_bo_cache;
1231
1232    result = anv_state_pool_init(&device->instruction_state_pool, device, 16384);
1233    if (result != VK_SUCCESS)
1234       goto fail_dynamic_state_pool;
1235
1236    result = anv_state_pool_init(&device->surface_state_pool, device, 4096);
1237    if (result != VK_SUCCESS)
1238       goto fail_instruction_state_pool;
1239
1240    result = anv_bo_init_new(&device->workaround_bo, device, 1024);
1241    if (result != VK_SUCCESS)
1242       goto fail_surface_state_pool;
1243
1244    anv_scratch_pool_init(device, &device->scratch_pool);
1245
1246    anv_queue_init(device, &device->queue);
1247
1248    switch (device->info.gen) {
1249    case 7:
1250       if (!device->info.is_haswell)
1251          result = gen7_init_device_state(device);
1252       else
1253          result = gen75_init_device_state(device);
1254       break;
1255    case 8:
1256       result = gen8_init_device_state(device);
1257       break;
1258    case 9:
1259       result = gen9_init_device_state(device);
1260       break;
1261    case 10:
1262       result = gen10_init_device_state(device);
1263       break;
1264    default:
1265       /* Shouldn't get here as we don't create physical devices for any other
1266        * gens. */
1267       unreachable("unhandled gen");
1268    }
1269    if (result != VK_SUCCESS)
1270       goto fail_workaround_bo;
1271
1272    anv_device_init_blorp(device);
1273
1274    anv_device_init_border_colors(device);
1275
1276    *pDevice = anv_device_to_handle(device);
1277
1278    return VK_SUCCESS;
1279
1280  fail_workaround_bo:
1281    anv_queue_finish(&device->queue);
1282    anv_scratch_pool_finish(device, &device->scratch_pool);
1283    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1284    anv_gem_close(device, device->workaround_bo.gem_handle);
1285  fail_surface_state_pool:
1286    anv_state_pool_finish(&device->surface_state_pool);
1287  fail_instruction_state_pool:
1288    anv_state_pool_finish(&device->instruction_state_pool);
1289  fail_dynamic_state_pool:
1290    anv_state_pool_finish(&device->dynamic_state_pool);
1291  fail_bo_cache:
1292    anv_bo_cache_finish(&device->bo_cache);
1293  fail_batch_bo_pool:
1294    anv_bo_pool_finish(&device->batch_bo_pool);
1295    pthread_cond_destroy(&device->queue_submit);
1296  fail_mutex:
1297    pthread_mutex_destroy(&device->mutex);
1298  fail_context_id:
1299    anv_gem_destroy_context(device, device->context_id);
1300  fail_fd:
1301    close(device->fd);
1302  fail_device:
1303    vk_free(&device->alloc, device);
1304
1305    return result;
1306 }
1307
1308 void anv_DestroyDevice(
1309     VkDevice                                    _device,
1310     const VkAllocationCallbacks*                pAllocator)
1311 {
1312    ANV_FROM_HANDLE(anv_device, device, _device);
1313
1314    if (!device)
1315       return;
1316
1317    anv_device_finish_blorp(device);
1318
1319    anv_queue_finish(&device->queue);
1320
1321 #ifdef HAVE_VALGRIND
1322    /* We only need to free these to prevent valgrind errors.  The backing
1323     * BO will go away in a couple of lines so we don't actually leak.
1324     */
1325    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1326 #endif
1327
1328    anv_scratch_pool_finish(device, &device->scratch_pool);
1329
1330    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1331    anv_gem_close(device, device->workaround_bo.gem_handle);
1332
1333    anv_state_pool_finish(&device->surface_state_pool);
1334    anv_state_pool_finish(&device->instruction_state_pool);
1335    anv_state_pool_finish(&device->dynamic_state_pool);
1336
1337    anv_bo_cache_finish(&device->bo_cache);
1338
1339    anv_bo_pool_finish(&device->batch_bo_pool);
1340
1341    pthread_cond_destroy(&device->queue_submit);
1342    pthread_mutex_destroy(&device->mutex);
1343
1344    anv_gem_destroy_context(device, device->context_id);
1345
1346    close(device->fd);
1347
1348    vk_free(&device->alloc, device);
1349 }
1350
1351 VkResult anv_EnumerateInstanceExtensionProperties(
1352     const char*                                 pLayerName,
1353     uint32_t*                                   pPropertyCount,
1354     VkExtensionProperties*                      pProperties)
1355 {
1356    if (pProperties == NULL) {
1357       *pPropertyCount = ARRAY_SIZE(global_extensions);
1358       return VK_SUCCESS;
1359    }
1360
1361    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(global_extensions));
1362    typed_memcpy(pProperties, global_extensions, *pPropertyCount);
1363
1364    if (*pPropertyCount < ARRAY_SIZE(global_extensions))
1365       return VK_INCOMPLETE;
1366
1367    return VK_SUCCESS;
1368 }
1369
1370 VkResult anv_EnumerateDeviceExtensionProperties(
1371     VkPhysicalDevice                            physicalDevice,
1372     const char*                                 pLayerName,
1373     uint32_t*                                   pPropertyCount,
1374     VkExtensionProperties*                      pProperties)
1375 {
1376    if (pProperties == NULL) {
1377       *pPropertyCount = ARRAY_SIZE(device_extensions);
1378       return VK_SUCCESS;
1379    }
1380
1381    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(device_extensions));
1382    typed_memcpy(pProperties, device_extensions, *pPropertyCount);
1383
1384    if (*pPropertyCount < ARRAY_SIZE(device_extensions))
1385       return VK_INCOMPLETE;
1386
1387    return VK_SUCCESS;
1388 }
1389
1390 VkResult anv_EnumerateInstanceLayerProperties(
1391     uint32_t*                                   pPropertyCount,
1392     VkLayerProperties*                          pProperties)
1393 {
1394    if (pProperties == NULL) {
1395       *pPropertyCount = 0;
1396       return VK_SUCCESS;
1397    }
1398
1399    /* None supported at this time */
1400    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1401 }
1402
1403 VkResult anv_EnumerateDeviceLayerProperties(
1404     VkPhysicalDevice                            physicalDevice,
1405     uint32_t*                                   pPropertyCount,
1406     VkLayerProperties*                          pProperties)
1407 {
1408    if (pProperties == NULL) {
1409       *pPropertyCount = 0;
1410       return VK_SUCCESS;
1411    }
1412
1413    /* None supported at this time */
1414    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1415 }
1416
1417 void anv_GetDeviceQueue(
1418     VkDevice                                    _device,
1419     uint32_t                                    queueNodeIndex,
1420     uint32_t                                    queueIndex,
1421     VkQueue*                                    pQueue)
1422 {
1423    ANV_FROM_HANDLE(anv_device, device, _device);
1424
1425    assert(queueIndex == 0);
1426
1427    *pQueue = anv_queue_to_handle(&device->queue);
1428 }
1429
1430 VkResult
1431 anv_device_query_status(struct anv_device *device)
1432 {
1433    /* This isn't likely as most of the callers of this function already check
1434     * for it.  However, it doesn't hurt to check and it potentially lets us
1435     * avoid an ioctl.
1436     */
1437    if (unlikely(device->lost))
1438       return VK_ERROR_DEVICE_LOST;
1439
1440    uint32_t active, pending;
1441    int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
1442    if (ret == -1) {
1443       /* We don't know the real error. */
1444       device->lost = true;
1445       return vk_errorf(VK_ERROR_DEVICE_LOST, "get_reset_stats failed: %m");
1446    }
1447
1448    if (active) {
1449       device->lost = true;
1450       return vk_errorf(VK_ERROR_DEVICE_LOST,
1451                        "GPU hung on one of our command buffers");
1452    } else if (pending) {
1453       device->lost = true;
1454       return vk_errorf(VK_ERROR_DEVICE_LOST,
1455                        "GPU hung with commands in-flight");
1456    }
1457
1458    return VK_SUCCESS;
1459 }
1460
1461 VkResult
1462 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
1463 {
1464    /* Note:  This only returns whether or not the BO is in use by an i915 GPU.
1465     * Other usages of the BO (such as on different hardware) will not be
1466     * flagged as "busy" by this ioctl.  Use with care.
1467     */
1468    int ret = anv_gem_busy(device, bo->gem_handle);
1469    if (ret == 1) {
1470       return VK_NOT_READY;
1471    } else if (ret == -1) {
1472       /* We don't know the real error. */
1473       device->lost = true;
1474       return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1475    }
1476
1477    /* Query for device status after the busy call.  If the BO we're checking
1478     * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1479     * client because it clearly doesn't have valid data.  Yes, this most
1480     * likely means an ioctl, but we just did an ioctl to query the busy status
1481     * so it's no great loss.
1482     */
1483    return anv_device_query_status(device);
1484 }
1485
1486 VkResult
1487 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
1488                 int64_t timeout)
1489 {
1490    int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
1491    if (ret == -1 && errno == ETIME) {
1492       return VK_TIMEOUT;
1493    } else if (ret == -1) {
1494       /* We don't know the real error. */
1495       device->lost = true;
1496       return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1497    }
1498
1499    /* Query for device status after the wait.  If the BO we're waiting on got
1500     * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1501     * because it clearly doesn't have valid data.  Yes, this most likely means
1502     * an ioctl, but we just did an ioctl to wait so it's no great loss.
1503     */
1504    return anv_device_query_status(device);
1505 }
1506
1507 VkResult anv_DeviceWaitIdle(
1508     VkDevice                                    _device)
1509 {
1510    ANV_FROM_HANDLE(anv_device, device, _device);
1511    if (unlikely(device->lost))
1512       return VK_ERROR_DEVICE_LOST;
1513
1514    struct anv_batch batch;
1515
1516    uint32_t cmds[8];
1517    batch.start = batch.next = cmds;
1518    batch.end = (void *) cmds + sizeof(cmds);
1519
1520    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1521    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1522
1523    return anv_device_submit_simple_batch(device, &batch);
1524 }
1525
1526 VkResult
1527 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1528 {
1529    uint32_t gem_handle = anv_gem_create(device, size);
1530    if (!gem_handle)
1531       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1532
1533    anv_bo_init(bo, gem_handle, size);
1534
1535    return VK_SUCCESS;
1536 }
1537
1538 VkResult anv_AllocateMemory(
1539     VkDevice                                    _device,
1540     const VkMemoryAllocateInfo*                 pAllocateInfo,
1541     const VkAllocationCallbacks*                pAllocator,
1542     VkDeviceMemory*                             pMem)
1543 {
1544    ANV_FROM_HANDLE(anv_device, device, _device);
1545    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
1546    struct anv_device_memory *mem;
1547    VkResult result = VK_SUCCESS;
1548
1549    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1550
1551    /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1552    assert(pAllocateInfo->allocationSize > 0);
1553
1554    /* The kernel relocation API has a limitation of a 32-bit delta value
1555     * applied to the address before it is written which, in spite of it being
1556     * unsigned, is treated as signed .  Because of the way that this maps to
1557     * the Vulkan API, we cannot handle an offset into a buffer that does not
1558     * fit into a signed 32 bits.  The only mechanism we have for dealing with
1559     * this at the moment is to limit all VkDeviceMemory objects to a maximum
1560     * of 2GB each.  The Vulkan spec allows us to do this:
1561     *
1562     *    "Some platforms may have a limit on the maximum size of a single
1563     *    allocation. For example, certain systems may fail to create
1564     *    allocations with a size greater than or equal to 4GB. Such a limit is
1565     *    implementation-dependent, and if such a failure occurs then the error
1566     *    VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1567     *
1568     * We don't use vk_error here because it's not an error so much as an
1569     * indication to the application that the allocation is too large.
1570     */
1571    if (pAllocateInfo->allocationSize > (1ull << 31))
1572       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
1573
1574    /* FINISHME: Fail if allocation request exceeds heap size. */
1575
1576    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1577                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1578    if (mem == NULL)
1579       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1580
1581    assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
1582    mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
1583    mem->map = NULL;
1584    mem->map_size = 0;
1585
1586    const VkImportMemoryFdInfoKHR *fd_info =
1587       vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
1588
1589    /* The Vulkan spec permits handleType to be 0, in which case the struct is
1590     * ignored.
1591     */
1592    if (fd_info && fd_info->handleType) {
1593       /* At the moment, we only support the OPAQUE_FD memory type which is
1594        * just a GEM buffer.
1595        */
1596       assert(fd_info->handleType ==
1597              VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR);
1598
1599       result = anv_bo_cache_import(device, &device->bo_cache,
1600                                    fd_info->fd, pAllocateInfo->allocationSize,
1601                                    &mem->bo);
1602       if (result != VK_SUCCESS)
1603          goto fail;
1604    } else {
1605       result = anv_bo_cache_alloc(device, &device->bo_cache,
1606                                   pAllocateInfo->allocationSize,
1607                                   &mem->bo);
1608       if (result != VK_SUCCESS)
1609          goto fail;
1610    }
1611
1612    assert(mem->type->heapIndex < pdevice->memory.heap_count);
1613    if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
1614       mem->bo->flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
1615
1616    if (pdevice->has_exec_async)
1617       mem->bo->flags |= EXEC_OBJECT_ASYNC;
1618
1619    *pMem = anv_device_memory_to_handle(mem);
1620
1621    return VK_SUCCESS;
1622
1623  fail:
1624    vk_free2(&device->alloc, pAllocator, mem);
1625
1626    return result;
1627 }
1628
1629 VkResult anv_GetMemoryFdKHR(
1630     VkDevice                                    device_h,
1631     const VkMemoryGetFdInfoKHR*                 pGetFdInfo,
1632     int*                                        pFd)
1633 {
1634    ANV_FROM_HANDLE(anv_device, dev, device_h);
1635    ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
1636
1637    assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
1638
1639    /* We support only one handle type. */
1640    assert(pGetFdInfo->handleType ==
1641           VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR);
1642
1643    return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
1644 }
1645
1646 VkResult anv_GetMemoryFdPropertiesKHR(
1647     VkDevice                                    device_h,
1648     VkExternalMemoryHandleTypeFlagBitsKHR       handleType,
1649     int                                         fd,
1650     VkMemoryFdPropertiesKHR*                    pMemoryFdProperties)
1651 {
1652    /* The valid usage section for this function says:
1653     *
1654     *    "handleType must not be one of the handle types defined as opaque."
1655     *
1656     * Since we only handle opaque handles for now, there are no FD properties.
1657     */
1658    return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR;
1659 }
1660
1661 void anv_FreeMemory(
1662     VkDevice                                    _device,
1663     VkDeviceMemory                              _mem,
1664     const VkAllocationCallbacks*                pAllocator)
1665 {
1666    ANV_FROM_HANDLE(anv_device, device, _device);
1667    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1668
1669    if (mem == NULL)
1670       return;
1671
1672    if (mem->map)
1673       anv_UnmapMemory(_device, _mem);
1674
1675    anv_bo_cache_release(device, &device->bo_cache, mem->bo);
1676
1677    vk_free2(&device->alloc, pAllocator, mem);
1678 }
1679
1680 VkResult anv_MapMemory(
1681     VkDevice                                    _device,
1682     VkDeviceMemory                              _memory,
1683     VkDeviceSize                                offset,
1684     VkDeviceSize                                size,
1685     VkMemoryMapFlags                            flags,
1686     void**                                      ppData)
1687 {
1688    ANV_FROM_HANDLE(anv_device, device, _device);
1689    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1690
1691    if (mem == NULL) {
1692       *ppData = NULL;
1693       return VK_SUCCESS;
1694    }
1695
1696    if (size == VK_WHOLE_SIZE)
1697       size = mem->bo->size - offset;
1698
1699    /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1700     *
1701     *  * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1702     *    assert(size != 0);
1703     *  * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1704     *    equal to the size of the memory minus offset
1705     */
1706    assert(size > 0);
1707    assert(offset + size <= mem->bo->size);
1708
1709    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1710     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1711     * at a time is valid. We could just mmap up front and return an offset
1712     * pointer here, but that may exhaust virtual memory on 32 bit
1713     * userspace. */
1714
1715    uint32_t gem_flags = 0;
1716
1717    if (!device->info.has_llc &&
1718        (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
1719       gem_flags |= I915_MMAP_WC;
1720
1721    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1722    uint64_t map_offset = offset & ~4095ull;
1723    assert(offset >= map_offset);
1724    uint64_t map_size = (offset + size) - map_offset;
1725
1726    /* Let's map whole pages */
1727    map_size = align_u64(map_size, 4096);
1728
1729    void *map = anv_gem_mmap(device, mem->bo->gem_handle,
1730                             map_offset, map_size, gem_flags);
1731    if (map == MAP_FAILED)
1732       return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
1733
1734    mem->map = map;
1735    mem->map_size = map_size;
1736
1737    *ppData = mem->map + (offset - map_offset);
1738
1739    return VK_SUCCESS;
1740 }
1741
1742 void anv_UnmapMemory(
1743     VkDevice                                    _device,
1744     VkDeviceMemory                              _memory)
1745 {
1746    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1747
1748    if (mem == NULL)
1749       return;
1750
1751    anv_gem_munmap(mem->map, mem->map_size);
1752
1753    mem->map = NULL;
1754    mem->map_size = 0;
1755 }
1756
1757 static void
1758 clflush_mapped_ranges(struct anv_device         *device,
1759                       uint32_t                   count,
1760                       const VkMappedMemoryRange *ranges)
1761 {
1762    for (uint32_t i = 0; i < count; i++) {
1763       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1764       if (ranges[i].offset >= mem->map_size)
1765          continue;
1766
1767       gen_clflush_range(mem->map + ranges[i].offset,
1768                         MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
1769    }
1770 }
1771
1772 VkResult anv_FlushMappedMemoryRanges(
1773     VkDevice                                    _device,
1774     uint32_t                                    memoryRangeCount,
1775     const VkMappedMemoryRange*                  pMemoryRanges)
1776 {
1777    ANV_FROM_HANDLE(anv_device, device, _device);
1778
1779    if (device->info.has_llc)
1780       return VK_SUCCESS;
1781
1782    /* Make sure the writes we're flushing have landed. */
1783    __builtin_ia32_mfence();
1784
1785    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1786
1787    return VK_SUCCESS;
1788 }
1789
1790 VkResult anv_InvalidateMappedMemoryRanges(
1791     VkDevice                                    _device,
1792     uint32_t                                    memoryRangeCount,
1793     const VkMappedMemoryRange*                  pMemoryRanges)
1794 {
1795    ANV_FROM_HANDLE(anv_device, device, _device);
1796
1797    if (device->info.has_llc)
1798       return VK_SUCCESS;
1799
1800    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1801
1802    /* Make sure no reads get moved up above the invalidate. */
1803    __builtin_ia32_mfence();
1804
1805    return VK_SUCCESS;
1806 }
1807
1808 void anv_GetBufferMemoryRequirements(
1809     VkDevice                                    _device,
1810     VkBuffer                                    _buffer,
1811     VkMemoryRequirements*                       pMemoryRequirements)
1812 {
1813    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1814    ANV_FROM_HANDLE(anv_device, device, _device);
1815    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
1816
1817    /* The Vulkan spec (git aaed022) says:
1818     *
1819     *    memoryTypeBits is a bitfield and contains one bit set for every
1820     *    supported memory type for the resource. The bit `1<<i` is set if and
1821     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1822     *    structure for the physical device is supported.
1823     */
1824    uint32_t memory_types = 0;
1825    for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
1826       uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
1827       if ((valid_usage & buffer->usage) == buffer->usage)
1828          memory_types |= (1u << i);
1829    }
1830
1831    pMemoryRequirements->size = buffer->size;
1832    pMemoryRequirements->alignment = 16;
1833    pMemoryRequirements->memoryTypeBits = memory_types;
1834 }
1835
1836 void anv_GetBufferMemoryRequirements2KHR(
1837     VkDevice                                    _device,
1838     const VkBufferMemoryRequirementsInfo2KHR*   pInfo,
1839     VkMemoryRequirements2KHR*                   pMemoryRequirements)
1840 {
1841    anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
1842                                    &pMemoryRequirements->memoryRequirements);
1843
1844    vk_foreach_struct(ext, pMemoryRequirements->pNext) {
1845       switch (ext->sType) {
1846       case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR: {
1847          VkMemoryDedicatedRequirementsKHR *requirements = (void *)ext;
1848          requirements->prefersDedicatedAllocation = VK_FALSE;
1849          requirements->requiresDedicatedAllocation = VK_FALSE;
1850          break;
1851       }
1852
1853       default:
1854          anv_debug_ignored_stype(ext->sType);
1855          break;
1856       }
1857    }
1858 }
1859
1860 void anv_GetImageMemoryRequirements(
1861     VkDevice                                    _device,
1862     VkImage                                     _image,
1863     VkMemoryRequirements*                       pMemoryRequirements)
1864 {
1865    ANV_FROM_HANDLE(anv_image, image, _image);
1866    ANV_FROM_HANDLE(anv_device, device, _device);
1867    struct anv_physical_device *pdevice = &device->instance->physicalDevice;
1868
1869    /* The Vulkan spec (git aaed022) says:
1870     *
1871     *    memoryTypeBits is a bitfield and contains one bit set for every
1872     *    supported memory type for the resource. The bit `1<<i` is set if and
1873     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1874     *    structure for the physical device is supported.
1875     *
1876     * All types are currently supported for images.
1877     */
1878    uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
1879
1880    pMemoryRequirements->size = image->size;
1881    pMemoryRequirements->alignment = image->alignment;
1882    pMemoryRequirements->memoryTypeBits = memory_types;
1883 }
1884
1885 void anv_GetImageMemoryRequirements2KHR(
1886     VkDevice                                    _device,
1887     const VkImageMemoryRequirementsInfo2KHR*    pInfo,
1888     VkMemoryRequirements2KHR*                   pMemoryRequirements)
1889 {
1890    anv_GetImageMemoryRequirements(_device, pInfo->image,
1891                                   &pMemoryRequirements->memoryRequirements);
1892
1893    vk_foreach_struct(ext, pMemoryRequirements->pNext) {
1894       switch (ext->sType) {
1895       case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR: {
1896          VkMemoryDedicatedRequirementsKHR *requirements = (void *)ext;
1897          requirements->prefersDedicatedAllocation = VK_FALSE;
1898          requirements->requiresDedicatedAllocation = VK_FALSE;
1899          break;
1900       }
1901
1902       default:
1903          anv_debug_ignored_stype(ext->sType);
1904          break;
1905       }
1906    }
1907 }
1908
1909 void anv_GetImageSparseMemoryRequirements(
1910     VkDevice                                    device,
1911     VkImage                                     image,
1912     uint32_t*                                   pSparseMemoryRequirementCount,
1913     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1914 {
1915    *pSparseMemoryRequirementCount = 0;
1916 }
1917
1918 void anv_GetImageSparseMemoryRequirements2KHR(
1919     VkDevice                                    device,
1920     const VkImageSparseMemoryRequirementsInfo2KHR* pInfo,
1921     uint32_t*                                   pSparseMemoryRequirementCount,
1922     VkSparseImageMemoryRequirements2KHR*        pSparseMemoryRequirements)
1923 {
1924    *pSparseMemoryRequirementCount = 0;
1925 }
1926
1927 void anv_GetDeviceMemoryCommitment(
1928     VkDevice                                    device,
1929     VkDeviceMemory                              memory,
1930     VkDeviceSize*                               pCommittedMemoryInBytes)
1931 {
1932    *pCommittedMemoryInBytes = 0;
1933 }
1934
1935 VkResult anv_BindBufferMemory(
1936     VkDevice                                    device,
1937     VkBuffer                                    _buffer,
1938     VkDeviceMemory                              _memory,
1939     VkDeviceSize                                memoryOffset)
1940 {
1941    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1942    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1943
1944    if (mem) {
1945       assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
1946       buffer->bo = mem->bo;
1947       buffer->offset = memoryOffset;
1948    } else {
1949       buffer->bo = NULL;
1950       buffer->offset = 0;
1951    }
1952
1953    return VK_SUCCESS;
1954 }
1955
1956 VkResult anv_QueueBindSparse(
1957     VkQueue                                     _queue,
1958     uint32_t                                    bindInfoCount,
1959     const VkBindSparseInfo*                     pBindInfo,
1960     VkFence                                     fence)
1961 {
1962    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1963    if (unlikely(queue->device->lost))
1964       return VK_ERROR_DEVICE_LOST;
1965
1966    return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1967 }
1968
1969 // Event functions
1970
1971 VkResult anv_CreateEvent(
1972     VkDevice                                    _device,
1973     const VkEventCreateInfo*                    pCreateInfo,
1974     const VkAllocationCallbacks*                pAllocator,
1975     VkEvent*                                    pEvent)
1976 {
1977    ANV_FROM_HANDLE(anv_device, device, _device);
1978    struct anv_state state;
1979    struct anv_event *event;
1980
1981    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1982
1983    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1984                                 sizeof(*event), 8);
1985    event = state.map;
1986    event->state = state;
1987    event->semaphore = VK_EVENT_RESET;
1988
1989    if (!device->info.has_llc) {
1990       /* Make sure the writes we're flushing have landed. */
1991       __builtin_ia32_mfence();
1992       __builtin_ia32_clflush(event);
1993    }
1994
1995    *pEvent = anv_event_to_handle(event);
1996
1997    return VK_SUCCESS;
1998 }
1999
2000 void anv_DestroyEvent(
2001     VkDevice                                    _device,
2002     VkEvent                                     _event,
2003     const VkAllocationCallbacks*                pAllocator)
2004 {
2005    ANV_FROM_HANDLE(anv_device, device, _device);
2006    ANV_FROM_HANDLE(anv_event, event, _event);
2007
2008    if (!event)
2009       return;
2010
2011    anv_state_pool_free(&device->dynamic_state_pool, event->state);
2012 }
2013
2014 VkResult anv_GetEventStatus(
2015     VkDevice                                    _device,
2016     VkEvent                                     _event)
2017 {
2018    ANV_FROM_HANDLE(anv_device, device, _device);
2019    ANV_FROM_HANDLE(anv_event, event, _event);
2020
2021    if (unlikely(device->lost))
2022       return VK_ERROR_DEVICE_LOST;
2023
2024    if (!device->info.has_llc) {
2025       /* Invalidate read cache before reading event written by GPU. */
2026       __builtin_ia32_clflush(event);
2027       __builtin_ia32_mfence();
2028
2029    }
2030
2031    return event->semaphore;
2032 }
2033
2034 VkResult anv_SetEvent(
2035     VkDevice                                    _device,
2036     VkEvent                                     _event)
2037 {
2038    ANV_FROM_HANDLE(anv_device, device, _device);
2039    ANV_FROM_HANDLE(anv_event, event, _event);
2040
2041    event->semaphore = VK_EVENT_SET;
2042
2043    if (!device->info.has_llc) {
2044       /* Make sure the writes we're flushing have landed. */
2045       __builtin_ia32_mfence();
2046       __builtin_ia32_clflush(event);
2047    }
2048
2049    return VK_SUCCESS;
2050 }
2051
2052 VkResult anv_ResetEvent(
2053     VkDevice                                    _device,
2054     VkEvent                                     _event)
2055 {
2056    ANV_FROM_HANDLE(anv_device, device, _device);
2057    ANV_FROM_HANDLE(anv_event, event, _event);
2058
2059    event->semaphore = VK_EVENT_RESET;
2060
2061    if (!device->info.has_llc) {
2062       /* Make sure the writes we're flushing have landed. */
2063       __builtin_ia32_mfence();
2064       __builtin_ia32_clflush(event);
2065    }
2066
2067    return VK_SUCCESS;
2068 }
2069
2070 // Buffer functions
2071
2072 VkResult anv_CreateBuffer(
2073     VkDevice                                    _device,
2074     const VkBufferCreateInfo*                   pCreateInfo,
2075     const VkAllocationCallbacks*                pAllocator,
2076     VkBuffer*                                   pBuffer)
2077 {
2078    ANV_FROM_HANDLE(anv_device, device, _device);
2079    struct anv_buffer *buffer;
2080
2081    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2082
2083    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2084                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2085    if (buffer == NULL)
2086       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2087
2088    buffer->size = pCreateInfo->size;
2089    buffer->usage = pCreateInfo->usage;
2090    buffer->bo = NULL;
2091    buffer->offset = 0;
2092
2093    *pBuffer = anv_buffer_to_handle(buffer);
2094
2095    return VK_SUCCESS;
2096 }
2097
2098 void anv_DestroyBuffer(
2099     VkDevice                                    _device,
2100     VkBuffer                                    _buffer,
2101     const VkAllocationCallbacks*                pAllocator)
2102 {
2103    ANV_FROM_HANDLE(anv_device, device, _device);
2104    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2105
2106    if (!buffer)
2107       return;
2108
2109    vk_free2(&device->alloc, pAllocator, buffer);
2110 }
2111
2112 void
2113 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
2114                               enum isl_format format,
2115                               uint32_t offset, uint32_t range, uint32_t stride)
2116 {
2117    isl_buffer_fill_state(&device->isl_dev, state.map,
2118                          .address = offset,
2119                          .mocs = device->default_mocs,
2120                          .size = range,
2121                          .format = format,
2122                          .stride = stride);
2123
2124    anv_state_flush(device, state);
2125 }
2126
2127 void anv_DestroySampler(
2128     VkDevice                                    _device,
2129     VkSampler                                   _sampler,
2130     const VkAllocationCallbacks*                pAllocator)
2131 {
2132    ANV_FROM_HANDLE(anv_device, device, _device);
2133    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
2134
2135    if (!sampler)
2136       return;
2137
2138    vk_free2(&device->alloc, pAllocator, sampler);
2139 }
2140
2141 VkResult anv_CreateFramebuffer(
2142     VkDevice                                    _device,
2143     const VkFramebufferCreateInfo*              pCreateInfo,
2144     const VkAllocationCallbacks*                pAllocator,
2145     VkFramebuffer*                              pFramebuffer)
2146 {
2147    ANV_FROM_HANDLE(anv_device, device, _device);
2148    struct anv_framebuffer *framebuffer;
2149
2150    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2151
2152    size_t size = sizeof(*framebuffer) +
2153                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2154    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2155                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2156    if (framebuffer == NULL)
2157       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2158
2159    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2160    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2161       VkImageView _iview = pCreateInfo->pAttachments[i];
2162       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2163    }
2164
2165    framebuffer->width = pCreateInfo->width;
2166    framebuffer->height = pCreateInfo->height;
2167    framebuffer->layers = pCreateInfo->layers;
2168
2169    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2170
2171    return VK_SUCCESS;
2172 }
2173
2174 void anv_DestroyFramebuffer(
2175     VkDevice                                    _device,
2176     VkFramebuffer                               _fb,
2177     const VkAllocationCallbacks*                pAllocator)
2178 {
2179    ANV_FROM_HANDLE(anv_device, device, _device);
2180    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2181
2182    if (!fb)
2183       return;
2184
2185    vk_free2(&device->alloc, pAllocator, fb);
2186 }
2187
2188 /* vk_icd.h does not declare this function, so we declare it here to
2189  * suppress Wmissing-prototypes.
2190  */
2191 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2192 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2193
2194 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2195 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2196 {
2197    /* For the full details on loader interface versioning, see
2198     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2199     * What follows is a condensed summary, to help you navigate the large and
2200     * confusing official doc.
2201     *
2202     *   - Loader interface v0 is incompatible with later versions. We don't
2203     *     support it.
2204     *
2205     *   - In loader interface v1:
2206     *       - The first ICD entrypoint called by the loader is
2207     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2208     *         entrypoint.
2209     *       - The ICD must statically expose no other Vulkan symbol unless it is
2210     *         linked with -Bsymbolic.
2211     *       - Each dispatchable Vulkan handle created by the ICD must be
2212     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2213     *         ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2214     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2215     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2216     *         such loader-managed surfaces.
2217     *
2218     *    - Loader interface v2 differs from v1 in:
2219     *       - The first ICD entrypoint called by the loader is
2220     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2221     *         statically expose this entrypoint.
2222     *
2223     *    - Loader interface v3 differs from v2 in:
2224     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2225     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2226     *          because the loader no longer does so.
2227     */
2228    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2229    return VK_SUCCESS;
2230 }