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