projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
anv: use safer snprintf() to ensure NULL string-terminator
[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 #include <drm_fourcc.h>
33
34 #include "anv_private.h"
35 #include "util/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/disk_cache.h"
39 #include "util/mesa-sha1.h"
40 #include "vk_util.h"
41 #include "common/gen_defines.h"
42
43 #include "genxml/gen7_pack.h"
44
45 static void
46 compiler_debug_log(void *data, const char *fmt, ...)
47 { }
48
49 static void
50 compiler_perf_log(void *data, const char *fmt, ...)
51 {
52 va_list args;
53 va_start(args, fmt);
54
55 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
56 intel_logd_v(fmt, args);
57
58 va_end(args);
59 }
60
61 static VkResult
62 anv_compute_heap_size(int fd, uint64_t gtt_size, uint64_t *heap_size)
63 {
64 /* Query the total ram from the system */
65 struct sysinfo info;
66 sysinfo(&info);
67
68 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
69
70 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
71 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
72 */
73 uint64_t available_ram;
74 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
75 available_ram = total_ram / 2;
76 else
77 available_ram = total_ram * 3 / 4;
78
79 /* We also want to leave some padding for things we allocate in the driver,
80 * so don't go over 3/4 of the GTT either.
81 */
82 uint64_t available_gtt = gtt_size * 3 / 4;
83
84 *heap_size = MIN2(available_ram, available_gtt);
85
86 return VK_SUCCESS;
87 }
88
89 static VkResult
90 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
91 {
92 uint64_t gtt_size;
93 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
94 &gtt_size) == -1) {
95 /* If, for whatever reason, we can't actually get the GTT size from the
96 * kernel (too old?) fall back to the aperture size.
97 */
98 anv_perf_warn(NULL, NULL,
99 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
100
101 if (anv_gem_get_aperture(fd, &gtt_size) == -1) {
102 return vk_errorf(NULL, NULL, VK_ERROR_INITIALIZATION_FAILED,
103 "failed to get aperture size: %m");
104 }
105 }
106
107 device->supports_48bit_addresses = (device->info.gen >= 8) &&
108 gtt_size > (4ULL << 30 /* GiB */);
109
110 uint64_t heap_size = 0;
111 VkResult result = anv_compute_heap_size(fd, gtt_size, &heap_size);
112 if (result != VK_SUCCESS)
113 return result;
114
115 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
116 /* When running with an overridden PCI ID, we may get a GTT size from
117 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
118 * address support can still fail. Just clamp the address space size to
119 * 2 GiB if we don't have 48-bit support.
120 */
121 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
122 "not support for 48-bit addresses",
123 __FILE__, __LINE__);
124 heap_size = 2ull << 30;
125 }
126
127 if (heap_size <= 3ull * (1ull << 30)) {
128 /* In this case, everything fits nicely into the 32-bit address space,
129 * so there's no need for supporting 48bit addresses on client-allocated
130 * memory objects.
131 */
132 device->memory.heap_count = 1;
133 device->memory.heaps[0] = (struct anv_memory_heap) {
134 .size = heap_size,
135 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
136 .supports_48bit_addresses = false,
137 };
138 } else {
139 /* Not everything will fit nicely into a 32-bit address space. In this
140 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
141 * larger 48-bit heap. If we're in this case, then we have a total heap
142 * size larger than 3GiB which most likely means they have 8 GiB of
143 * video memory and so carving off 1 GiB for the 32-bit heap should be
144 * reasonable.
145 */
146 const uint64_t heap_size_32bit = 1ull << 30;
147 const uint64_t heap_size_48bit = heap_size - heap_size_32bit;
148
149 assert(device->supports_48bit_addresses);
150
151 device->memory.heap_count = 2;
152 device->memory.heaps[0] = (struct anv_memory_heap) {
153 .size = heap_size_48bit,
154 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
155 .supports_48bit_addresses = true,
156 };
157 device->memory.heaps[1] = (struct anv_memory_heap) {
158 .size = heap_size_32bit,
159 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
160 .supports_48bit_addresses = false,
161 };
162 }
163
164 uint32_t type_count = 0;
165 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
166 uint32_t valid_buffer_usage = ~0;
167
168 /* There appears to be a hardware issue in the VF cache where it only
169 * considers the bottom 32 bits of memory addresses. If you happen to
170 * have two vertex buffers which get placed exactly 4 GiB apart and use
171 * them in back-to-back draw calls, you can get collisions. In order to
172 * solve this problem, we require vertex and index buffers be bound to
173 * memory allocated out of the 32-bit heap.
174 */
175 if (device->memory.heaps[heap].supports_48bit_addresses) {
176 valid_buffer_usage &= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
177 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
178 }
179
180 if (device->info.has_llc) {
181 /* Big core GPUs share LLC with the CPU and thus one memory type can be
182 * both cached and coherent at the same time.
183 */
184 device->memory.types[type_count++] = (struct anv_memory_type) {
185 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
186 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
187 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
188 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
189 .heapIndex = heap,
190 .valid_buffer_usage = valid_buffer_usage,
191 };
192 } else {
193 /* The spec requires that we expose a host-visible, coherent memory
194 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
195 * to give the application a choice between cached, but not coherent and
196 * coherent but uncached (WC though).
197 */
198 device->memory.types[type_count++] = (struct anv_memory_type) {
199 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
200 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
201 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
202 .heapIndex = heap,
203 .valid_buffer_usage = valid_buffer_usage,
204 };
205 device->memory.types[type_count++] = (struct anv_memory_type) {
206 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
207 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
208 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
209 .heapIndex = heap,
210 .valid_buffer_usage = valid_buffer_usage,
211 };
212 }
213 }
214 device->memory.type_count = type_count;
215
216 return VK_SUCCESS;
217 }
218
219 static VkResult
220 anv_physical_device_init_uuids(struct anv_physical_device *device)
221 {
222 const struct build_id_note *note =
223 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
224 if (!note) {
225 return vk_errorf(device->instance, device,
226 VK_ERROR_INITIALIZATION_FAILED,
227 "Failed to find build-id");
228 }
229
230 unsigned build_id_len = build_id_length(note);
231 if (build_id_len < 20) {
232 return vk_errorf(device->instance, device,
233 VK_ERROR_INITIALIZATION_FAILED,
234 "build-id too short. It needs to be a SHA");
235 }
236
237 memcpy(device->driver_build_sha1, build_id_data(note), 20);
238
239 struct mesa_sha1 sha1_ctx;
240 uint8_t sha1[20];
241 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
242
243 /* The pipeline cache UUID is used for determining when a pipeline cache is
244 * invalid. It needs both a driver build and the PCI ID of the device.
245 */
246 _mesa_sha1_init(&sha1_ctx);
247 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
248 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
249 sizeof(device->chipset_id));
250 _mesa_sha1_final(&sha1_ctx, sha1);
251 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
252
253 /* The driver UUID is used for determining sharability of images and memory
254 * between two Vulkan instances in separate processes. People who want to
255 * share memory need to also check the device UUID (below) so all this
256 * needs to be is the build-id.
257 */
258 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
259
260 /* The device UUID uniquely identifies the given device within the machine.
261 * Since we never have more than one device, this doesn't need to be a real
262 * UUID. However, on the off-chance that someone tries to use this to
263 * cache pre-tiled images or something of the like, we use the PCI ID and
264 * some bits of ISL info to ensure that this is safe.
265 */
266 _mesa_sha1_init(&sha1_ctx);
267 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
268 sizeof(device->chipset_id));
269 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
270 sizeof(device->isl_dev.has_bit6_swizzling));
271 _mesa_sha1_final(&sha1_ctx, sha1);
272 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
273
274 return VK_SUCCESS;
275 }
276
277 static void
278 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
279 {
280 #ifdef ENABLE_SHADER_CACHE
281 char renderer[10];
282 MAYBE_UNUSED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
283 device->chipset_id);
284 assert(len == sizeof(renderer) - 2);
285
286 char timestamp[41];
287 _mesa_sha1_format(timestamp, device->driver_build_sha1);
288
289 const uint64_t driver_flags =
290 brw_get_compiler_config_value(device->compiler);
291 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
292 #else
293 device->disk_cache = NULL;
294 #endif
295 }
296
297 static void
298 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
299 {
300 #ifdef ENABLE_SHADER_CACHE
301 if (device->disk_cache)
302 disk_cache_destroy(device->disk_cache);
303 #else
304 assert(device->disk_cache == NULL);
305 #endif
306 }
307
308 static VkResult
309 anv_physical_device_init(struct anv_physical_device *device,
310 struct anv_instance *instance,
311 const char *primary_path,
312 const char *path)
313 {
314 VkResult result;
315 int fd;
316 int master_fd = -1;
317
318 brw_process_intel_debug_variable();
319
320 fd = open(path, O_RDWR | O_CLOEXEC);
321 if (fd < 0)
322 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
323
324 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
325 device->instance = instance;
326
327 assert(strlen(path) < ARRAY_SIZE(device->path));
328 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
329
330 device->no_hw = getenv("INTEL_NO_HW") != NULL;
331
332 const int pci_id_override = gen_get_pci_device_id_override();
333 if (pci_id_override < 0) {
334 device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
335 if (!device->chipset_id) {
336 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
337 goto fail;
338 }
339 } else {
340 device->chipset_id = pci_id_override;
341 device->no_hw = true;
342 }
343
344 device->name = gen_get_device_name(device->chipset_id);
345 if (!gen_get_device_info(device->chipset_id, &device->info)) {
346 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
347 goto fail;
348 }
349
350 if (device->info.is_haswell) {
351 intel_logw("Haswell Vulkan support is incomplete");
352 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
353 intel_logw("Ivy Bridge Vulkan support is incomplete");
354 } else if (device->info.gen == 7 && device->info.is_baytrail) {
355 intel_logw("Bay Trail Vulkan support is incomplete");
356 } else if (device->info.gen >= 8 && device->info.gen <= 10) {
357 /* Gen8-10 fully supported */
358 } else if (device->info.gen == 11) {
359 intel_logw("Vulkan is not yet fully supported on gen11.");
360 } else {
361 result = vk_errorf(device->instance, device,
362 VK_ERROR_INCOMPATIBLE_DRIVER,
363 "Vulkan not yet supported on %s", device->name);
364 goto fail;
365 }
366
367 device->cmd_parser_version = -1;
368 if (device->info.gen == 7) {
369 device->cmd_parser_version =
370 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
371 if (device->cmd_parser_version == -1) {
372 result = vk_errorf(device->instance, device,
373 VK_ERROR_INITIALIZATION_FAILED,
374 "failed to get command parser version");
375 goto fail;
376 }
377 }
378
379 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
380 result = vk_errorf(device->instance, device,
381 VK_ERROR_INITIALIZATION_FAILED,
382 "kernel missing gem wait");
383 goto fail;
384 }
385
386 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
387 result = vk_errorf(device->instance, device,
388 VK_ERROR_INITIALIZATION_FAILED,
389 "kernel missing execbuf2");
390 goto fail;
391 }
392
393 if (!device->info.has_llc &&
394 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
395 result = vk_errorf(device->instance, device,
396 VK_ERROR_INITIALIZATION_FAILED,
397 "kernel missing wc mmap");
398 goto fail;
399 }
400
401 result = anv_physical_device_init_heaps(device, fd);
402 if (result != VK_SUCCESS)
403 goto fail;
404
405 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
406 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
407 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
408 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
409 device->has_syncobj_wait = device->has_syncobj &&
410 anv_gem_supports_syncobj_wait(fd);
411 device->has_context_priority = anv_gem_has_context_priority(fd);
412
413 device->use_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN)
414 && device->supports_48bit_addresses;
415
416 device->has_context_isolation =
417 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
418
419 bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
420
421 /* Starting with Gen10, the timestamp frequency of the command streamer may
422 * vary from one part to another. We can query the value from the kernel.
423 */
424 if (device->info.gen >= 10) {
425 int timestamp_frequency =
426 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
427
428 if (timestamp_frequency < 0)
429 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
430 else
431 device->info.timestamp_frequency = timestamp_frequency;
432 }
433
434 /* GENs prior to 8 do not support EU/Subslice info */
435 if (device->info.gen >= 8) {
436 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
437 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
438
439 /* Without this information, we cannot get the right Braswell
440 * brandstrings, and we have to use conservative numbers for GPGPU on
441 * many platforms, but otherwise, things will just work.
442 */
443 if (device->subslice_total < 1 || device->eu_total < 1) {
444 intel_logw("Kernel 4.1 required to properly query GPU properties");
445 }
446 } else if (device->info.gen == 7) {
447 device->subslice_total = 1 << (device->info.gt - 1);
448 }
449
450 if (device->info.is_cherryview &&
451 device->subslice_total > 0 && device->eu_total > 0) {
452 /* Logical CS threads = EUs per subslice * num threads per EU */
453 uint32_t max_cs_threads =
454 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
455
456 /* Fuse configurations may give more threads than expected, never less. */
457 if (max_cs_threads > device->info.max_cs_threads)
458 device->info.max_cs_threads = max_cs_threads;
459 }
460
461 device->compiler = brw_compiler_create(NULL, &device->info);
462 if (device->compiler == NULL) {
463 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
464 goto fail;
465 }
466 device->compiler->shader_debug_log = compiler_debug_log;
467 device->compiler->shader_perf_log = compiler_perf_log;
468 device->compiler->supports_pull_constants = false;
469 device->compiler->constant_buffer_0_is_relative =
470 device->info.gen < 8 || !device->has_context_isolation;
471 device->compiler->supports_shader_constants = true;
472
473 isl_device_init(&device->isl_dev, &device->info, swizzled);
474
475 result = anv_physical_device_init_uuids(device);
476 if (result != VK_SUCCESS)
477 goto fail;
478
479 anv_physical_device_init_disk_cache(device);
480
481 if (instance->enabled_extensions.KHR_display) {
482 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
483 if (master_fd >= 0) {
484 /* prod the device with a GETPARAM call which will fail if
485 * we don't have permission to even render on this device
486 */
487 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
488 close(master_fd);
489 master_fd = -1;
490 }
491 }
492 }
493 device->master_fd = master_fd;
494
495 result = anv_init_wsi(device);
496 if (result != VK_SUCCESS) {
497 ralloc_free(device->compiler);
498 anv_physical_device_free_disk_cache(device);
499 goto fail;
500 }
501
502 anv_physical_device_get_supported_extensions(device,
503 &device->supported_extensions);
504
505
506 device->local_fd = fd;
507
508 return VK_SUCCESS;
509
510 fail:
511 close(fd);
512 if (master_fd != -1)
513 close(master_fd);
514 return result;
515 }
516
517 static void
518 anv_physical_device_finish(struct anv_physical_device *device)
519 {
520 anv_finish_wsi(device);
521 anv_physical_device_free_disk_cache(device);
522 ralloc_free(device->compiler);
523 close(device->local_fd);
524 if (device->master_fd >= 0)
525 close(device->master_fd);
526 }
527
528 static void *
529 default_alloc_func(void *pUserData, size_t size, size_t align,
530 VkSystemAllocationScope allocationScope)
531 {
532 return malloc(size);
533 }
534
535 static void *
536 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
537 size_t align, VkSystemAllocationScope allocationScope)
538 {
539 return realloc(pOriginal, size);
540 }
541
542 static void
543 default_free_func(void *pUserData, void *pMemory)
544 {
545 free(pMemory);
546 }
547
548 static const VkAllocationCallbacks default_alloc = {
549 .pUserData = NULL,
550 .pfnAllocation = default_alloc_func,
551 .pfnReallocation = default_realloc_func,
552 .pfnFree = default_free_func,
553 };
554
555 VkResult anv_EnumerateInstanceExtensionProperties(
556 const char* pLayerName,
557 uint32_t* pPropertyCount,
558 VkExtensionProperties* pProperties)
559 {
560 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
561
562 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
563 if (anv_instance_extensions_supported.extensions[i]) {
564 vk_outarray_append(&out, prop) {
565 *prop = anv_instance_extensions[i];
566 }
567 }
568 }
569
570 return vk_outarray_status(&out);
571 }
572
573 VkResult anv_CreateInstance(
574 const VkInstanceCreateInfo* pCreateInfo,
575 const VkAllocationCallbacks* pAllocator,
576 VkInstance* pInstance)
577 {
578 struct anv_instance *instance;
579 VkResult result;
580
581 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
582
583 struct anv_instance_extension_table enabled_extensions = {};
584 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
585 int idx;
586 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
587 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
588 anv_instance_extensions[idx].extensionName) == 0)
589 break;
590 }
591
592 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
593 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
594
595 if (!anv_instance_extensions_supported.extensions[idx])
596 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
597
598 enabled_extensions.extensions[idx] = true;
599 }
600
601 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
602 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
603 if (!instance)
604 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
605
606 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
607
608 if (pAllocator)
609 instance->alloc = *pAllocator;
610 else
611 instance->alloc = default_alloc;
612
613 if (pCreateInfo->pApplicationInfo &&
614 pCreateInfo->pApplicationInfo->apiVersion != 0) {
615 instance->apiVersion = pCreateInfo->pApplicationInfo->apiVersion;
616 } else {
617 anv_EnumerateInstanceVersion(&instance->apiVersion);
618 }
619
620 instance->enabled_extensions = enabled_extensions;
621
622 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
623 /* Vulkan requires that entrypoints for extensions which have not been
624 * enabled must not be advertised.
625 */
626 if (!anv_entrypoint_is_enabled(i, instance->apiVersion,
627 &instance->enabled_extensions, NULL)) {
628 instance->dispatch.entrypoints[i] = NULL;
629 } else if (anv_dispatch_table.entrypoints[i] != NULL) {
630 instance->dispatch.entrypoints[i] = anv_dispatch_table.entrypoints[i];
631 } else {
632 instance->dispatch.entrypoints[i] =
633 anv_tramp_dispatch_table.entrypoints[i];
634 }
635 }
636
637 instance->physicalDeviceCount = -1;
638
639 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
640 if (result != VK_SUCCESS) {
641 vk_free2(&default_alloc, pAllocator, instance);
642 return vk_error(result);
643 }
644
645 instance->pipeline_cache_enabled =
646 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
647
648 _mesa_locale_init();
649
650 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
651
652 *pInstance = anv_instance_to_handle(instance);
653
654 return VK_SUCCESS;
655 }
656
657 void anv_DestroyInstance(
658 VkInstance _instance,
659 const VkAllocationCallbacks* pAllocator)
660 {
661 ANV_FROM_HANDLE(anv_instance, instance, _instance);
662
663 if (!instance)
664 return;
665
666 if (instance->physicalDeviceCount > 0) {
667 /* We support at most one physical device. */
668 assert(instance->physicalDeviceCount == 1);
669 anv_physical_device_finish(&instance->physicalDevice);
670 }
671
672 VG(VALGRIND_DESTROY_MEMPOOL(instance));
673
674 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
675
676 _mesa_locale_fini();
677
678 vk_free(&instance->alloc, instance);
679 }
680
681 static VkResult
682 anv_enumerate_devices(struct anv_instance *instance)
683 {
684 /* TODO: Check for more devices ? */
685 drmDevicePtr devices[8];
686 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
687 int max_devices;
688
689 instance->physicalDeviceCount = 0;
690
691 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
692 if (max_devices < 1)
693 return VK_ERROR_INCOMPATIBLE_DRIVER;
694
695 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
696 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
697 devices[i]->bustype == DRM_BUS_PCI &&
698 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
699
700 result = anv_physical_device_init(&instance->physicalDevice,
701 instance,
702 devices[i]->nodes[DRM_NODE_PRIMARY],
703 devices[i]->nodes[DRM_NODE_RENDER]);
704 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
705 break;
706 }
707 }
708 drmFreeDevices(devices, max_devices);
709
710 if (result == VK_SUCCESS)
711 instance->physicalDeviceCount = 1;
712
713 return result;
714 }
715
716 static VkResult
717 anv_instance_ensure_physical_device(struct anv_instance *instance)
718 {
719 if (instance->physicalDeviceCount < 0) {
720 VkResult result = anv_enumerate_devices(instance);
721 if (result != VK_SUCCESS &&
722 result != VK_ERROR_INCOMPATIBLE_DRIVER)
723 return result;
724 }
725
726 return VK_SUCCESS;
727 }
728
729 VkResult anv_EnumeratePhysicalDevices(
730 VkInstance _instance,
731 uint32_t* pPhysicalDeviceCount,
732 VkPhysicalDevice* pPhysicalDevices)
733 {
734 ANV_FROM_HANDLE(anv_instance, instance, _instance);
735 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
736
737 VkResult result = anv_instance_ensure_physical_device(instance);
738 if (result != VK_SUCCESS)
739 return result;
740
741 if (instance->physicalDeviceCount == 0)
742 return VK_SUCCESS;
743
744 assert(instance->physicalDeviceCount == 1);
745 vk_outarray_append(&out, i) {
746 *i = anv_physical_device_to_handle(&instance->physicalDevice);
747 }
748
749 return vk_outarray_status(&out);
750 }
751
752 VkResult anv_EnumeratePhysicalDeviceGroups(
753 VkInstance _instance,
754 uint32_t* pPhysicalDeviceGroupCount,
755 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
756 {
757 ANV_FROM_HANDLE(anv_instance, instance, _instance);
758 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
759 pPhysicalDeviceGroupCount);
760
761 VkResult result = anv_instance_ensure_physical_device(instance);
762 if (result != VK_SUCCESS)
763 return result;
764
765 if (instance->physicalDeviceCount == 0)
766 return VK_SUCCESS;
767
768 assert(instance->physicalDeviceCount == 1);
769
770 vk_outarray_append(&out, p) {
771 p->physicalDeviceCount = 1;
772 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
773 p->physicalDevices[0] =
774 anv_physical_device_to_handle(&instance->physicalDevice);
775 p->subsetAllocation = VK_FALSE;
776
777 vk_foreach_struct(ext, p->pNext)
778 anv_debug_ignored_stype(ext->sType);
779 }
780
781 return vk_outarray_status(&out);
782 }
783
784 void anv_GetPhysicalDeviceFeatures(
785 VkPhysicalDevice physicalDevice,
786 VkPhysicalDeviceFeatures* pFeatures)
787 {
788 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
789
790 *pFeatures = (VkPhysicalDeviceFeatures) {
791 .robustBufferAccess = true,
792 .fullDrawIndexUint32 = true,
793 .imageCubeArray = true,
794 .independentBlend = true,
795 .geometryShader = true,
796 .tessellationShader = true,
797 .sampleRateShading = true,
798 .dualSrcBlend = true,
799 .logicOp = true,
800 .multiDrawIndirect = true,
801 .drawIndirectFirstInstance = true,
802 .depthClamp = true,
803 .depthBiasClamp = true,
804 .fillModeNonSolid = true,
805 .depthBounds = false,
806 .wideLines = true,
807 .largePoints = true,
808 .alphaToOne = true,
809 .multiViewport = true,
810 .samplerAnisotropy = true,
811 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
812 pdevice->info.is_baytrail,
813 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
814 .textureCompressionBC = true,
815 .occlusionQueryPrecise = true,
816 .pipelineStatisticsQuery = true,
817 .fragmentStoresAndAtomics = true,
818 .shaderTessellationAndGeometryPointSize = true,
819 .shaderImageGatherExtended = true,
820 .shaderStorageImageExtendedFormats = true,
821 .shaderStorageImageMultisample = false,
822 .shaderStorageImageReadWithoutFormat = false,
823 .shaderStorageImageWriteWithoutFormat = true,
824 .shaderUniformBufferArrayDynamicIndexing = true,
825 .shaderSampledImageArrayDynamicIndexing = true,
826 .shaderStorageBufferArrayDynamicIndexing = true,
827 .shaderStorageImageArrayDynamicIndexing = true,
828 .shaderClipDistance = true,
829 .shaderCullDistance = true,
830 .shaderFloat64 = pdevice->info.gen >= 8 &&
831 pdevice->info.has_64bit_types,
832 .shaderInt64 = pdevice->info.gen >= 8 &&
833 pdevice->info.has_64bit_types,
834 .shaderInt16 = pdevice->info.gen >= 8,
835 .shaderResourceMinLod = false,
836 .variableMultisampleRate = true,
837 .inheritedQueries = true,
838 };
839
840 /* We can't do image stores in vec4 shaders */
841 pFeatures->vertexPipelineStoresAndAtomics =
842 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
843 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
844 }
845
846 void anv_GetPhysicalDeviceFeatures2(
847 VkPhysicalDevice physicalDevice,
848 VkPhysicalDeviceFeatures2* pFeatures)
849 {
850 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
851
852 vk_foreach_struct(ext, pFeatures->pNext) {
853 switch (ext->sType) {
854 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
855 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
856 features->protectedMemory = VK_FALSE;
857 break;
858 }
859
860 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
861 VkPhysicalDeviceMultiviewFeatures *features =
862 (VkPhysicalDeviceMultiviewFeatures *)ext;
863 features->multiview = true;
864 features->multiviewGeometryShader = true;
865 features->multiviewTessellationShader = true;
866 break;
867 }
868
869 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES: {
870 VkPhysicalDeviceVariablePointerFeatures *features = (void *)ext;
871 features->variablePointersStorageBuffer = true;
872 features->variablePointers = true;
873 break;
874 }
875
876 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
877 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
878 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
879 features->samplerYcbcrConversion = true;
880 break;
881 }
882
883 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES: {
884 VkPhysicalDeviceShaderDrawParameterFeatures *features = (void *)ext;
885 features->shaderDrawParameters = true;
886 break;
887 }
888
889 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR: {
890 VkPhysicalDevice16BitStorageFeaturesKHR *features =
891 (VkPhysicalDevice16BitStorageFeaturesKHR *)ext;
892 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
893
894 features->storageBuffer16BitAccess = pdevice->info.gen >= 8;
895 features->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
896 features->storagePushConstant16 = pdevice->info.gen >= 8;
897 features->storageInputOutput16 = false;
898 break;
899 }
900
901 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
902 VkPhysicalDevice8BitStorageFeaturesKHR *features =
903 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
904 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
905
906 features->storageBuffer8BitAccess = pdevice->info.gen >= 8;
907 features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
908 features->storagePushConstant8 = pdevice->info.gen >= 8;
909 break;
910 }
911
912 default:
913 anv_debug_ignored_stype(ext->sType);
914 break;
915 }
916 }
917 }
918
919 void anv_GetPhysicalDeviceProperties(
920 VkPhysicalDevice physicalDevice,
921 VkPhysicalDeviceProperties* pProperties)
922 {
923 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
924 const struct gen_device_info *devinfo = &pdevice->info;
925
926 /* See assertions made when programming the buffer surface state. */
927 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
928 (1ul << 30) : (1ul << 27);
929
930 const uint32_t max_samplers = (devinfo->gen >= 8 || devinfo->is_haswell) ?
931 128 : 16;
932
933 VkSampleCountFlags sample_counts =
934 isl_device_get_sample_counts(&pdevice->isl_dev);
935
936 VkPhysicalDeviceLimits limits = {
937 .maxImageDimension1D = (1 << 14),
938 .maxImageDimension2D = (1 << 14),
939 .maxImageDimension3D = (1 << 11),
940 .maxImageDimensionCube = (1 << 14),
941 .maxImageArrayLayers = (1 << 11),
942 .maxTexelBufferElements = 128 * 1024 * 1024,
943 .maxUniformBufferRange = (1ul << 27),
944 .maxStorageBufferRange = max_raw_buffer_sz,
945 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
946 .maxMemoryAllocationCount = UINT32_MAX,
947 .maxSamplerAllocationCount = 64 * 1024,
948 .bufferImageGranularity = 64, /* A cache line */
949 .sparseAddressSpaceSize = 0,
950 .maxBoundDescriptorSets = MAX_SETS,
951 .maxPerStageDescriptorSamplers = max_samplers,
952 .maxPerStageDescriptorUniformBuffers = 64,
953 .maxPerStageDescriptorStorageBuffers = 64,
954 .maxPerStageDescriptorSampledImages = max_samplers,
955 .maxPerStageDescriptorStorageImages = 64,
956 .maxPerStageDescriptorInputAttachments = 64,
957 .maxPerStageResources = 250,
958 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
959 .maxDescriptorSetUniformBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
960 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
961 .maxDescriptorSetStorageBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
962 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
963 .maxDescriptorSetSampledImages = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSampledImages */
964 .maxDescriptorSetStorageImages = 6 * 64, /* number of stages * maxPerStageDescriptorStorageImages */
965 .maxDescriptorSetInputAttachments = 256,
966 .maxVertexInputAttributes = MAX_VBS,
967 .maxVertexInputBindings = MAX_VBS,
968 .maxVertexInputAttributeOffset = 2047,
969 .maxVertexInputBindingStride = 2048,
970 .maxVertexOutputComponents = 128,
971 .maxTessellationGenerationLevel = 64,
972 .maxTessellationPatchSize = 32,
973 .maxTessellationControlPerVertexInputComponents = 128,
974 .maxTessellationControlPerVertexOutputComponents = 128,
975 .maxTessellationControlPerPatchOutputComponents = 128,
976 .maxTessellationControlTotalOutputComponents = 2048,
977 .maxTessellationEvaluationInputComponents = 128,
978 .maxTessellationEvaluationOutputComponents = 128,
979 .maxGeometryShaderInvocations = 32,
980 .maxGeometryInputComponents = 64,
981 .maxGeometryOutputComponents = 128,
982 .maxGeometryOutputVertices = 256,
983 .maxGeometryTotalOutputComponents = 1024,
984 .maxFragmentInputComponents = 112, /* 128 components - (POS, PSIZ, CLIP_DIST0, CLIP_DIST1) */
985 .maxFragmentOutputAttachments = 8,
986 .maxFragmentDualSrcAttachments = 1,
987 .maxFragmentCombinedOutputResources = 8,
988 .maxComputeSharedMemorySize = 32768,
989 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
990 .maxComputeWorkGroupInvocations = 16 * devinfo->max_cs_threads,
991 .maxComputeWorkGroupSize = {
992 16 * devinfo->max_cs_threads,
993 16 * devinfo->max_cs_threads,
994 16 * devinfo->max_cs_threads,
995 },
996 .subPixelPrecisionBits = 4 /* FIXME */,
997 .subTexelPrecisionBits = 4 /* FIXME */,
998 .mipmapPrecisionBits = 4 /* FIXME */,
999 .maxDrawIndexedIndexValue = UINT32_MAX,
1000 .maxDrawIndirectCount = UINT32_MAX,
1001 .maxSamplerLodBias = 16,
1002 .maxSamplerAnisotropy = 16,
1003 .maxViewports = MAX_VIEWPORTS,
1004 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1005 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1006 .viewportSubPixelBits = 13, /* We take a float? */
1007 .minMemoryMapAlignment = 4096, /* A page */
1008 .minTexelBufferOffsetAlignment = 1,
1009 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1010 .minUniformBufferOffsetAlignment = 32,
1011 .minStorageBufferOffsetAlignment = 4,
1012 .minTexelOffset = -8,
1013 .maxTexelOffset = 7,
1014 .minTexelGatherOffset = -32,
1015 .maxTexelGatherOffset = 31,
1016 .minInterpolationOffset = -0.5,
1017 .maxInterpolationOffset = 0.4375,
1018 .subPixelInterpolationOffsetBits = 4,
1019 .maxFramebufferWidth = (1 << 14),
1020 .maxFramebufferHeight = (1 << 14),
1021 .maxFramebufferLayers = (1 << 11),
1022 .framebufferColorSampleCounts = sample_counts,
1023 .framebufferDepthSampleCounts = sample_counts,
1024 .framebufferStencilSampleCounts = sample_counts,
1025 .framebufferNoAttachmentsSampleCounts = sample_counts,
1026 .maxColorAttachments = MAX_RTS,
1027 .sampledImageColorSampleCounts = sample_counts,
1028 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1029 .sampledImageDepthSampleCounts = sample_counts,
1030 .sampledImageStencilSampleCounts = sample_counts,
1031 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1032 .maxSampleMaskWords = 1,
1033 .timestampComputeAndGraphics = false,
1034 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1035 .maxClipDistances = 8,
1036 .maxCullDistances = 8,
1037 .maxCombinedClipAndCullDistances = 8,
1038 .discreteQueuePriorities = 1,
1039 .pointSizeRange = { 0.125, 255.875 },
1040 .lineWidthRange = { 0.0, 7.9921875 },
1041 .pointSizeGranularity = (1.0 / 8.0),
1042 .lineWidthGranularity = (1.0 / 128.0),
1043 .strictLines = false, /* FINISHME */
1044 .standardSampleLocations = true,
1045 .optimalBufferCopyOffsetAlignment = 128,
1046 .optimalBufferCopyRowPitchAlignment = 128,
1047 .nonCoherentAtomSize = 64,
1048 };
1049
1050 *pProperties = (VkPhysicalDeviceProperties) {
1051 .apiVersion = anv_physical_device_api_version(pdevice),
1052 .driverVersion = vk_get_driver_version(),
1053 .vendorID = 0x8086,
1054 .deviceID = pdevice->chipset_id,
1055 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1056 .limits = limits,
1057 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1058 };
1059
1060 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1061 "%s", pdevice->name);
1062 memcpy(pProperties->pipelineCacheUUID,
1063 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1064 }
1065
1066 void anv_GetPhysicalDeviceProperties2(
1067 VkPhysicalDevice physicalDevice,
1068 VkPhysicalDeviceProperties2* pProperties)
1069 {
1070 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1071
1072 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1073
1074 vk_foreach_struct(ext, pProperties->pNext) {
1075 switch (ext->sType) {
1076 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1077 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1078 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1079
1080 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1081 break;
1082 }
1083
1084 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1085 VkPhysicalDeviceIDProperties *id_props =
1086 (VkPhysicalDeviceIDProperties *)ext;
1087 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1088 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1089 /* The LUID is for Windows. */
1090 id_props->deviceLUIDValid = false;
1091 break;
1092 }
1093
1094 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1095 VkPhysicalDeviceMaintenance3Properties *props =
1096 (VkPhysicalDeviceMaintenance3Properties *)ext;
1097 /* This value doesn't matter for us today as our per-stage
1098 * descriptors are the real limit.
1099 */
1100 props->maxPerSetDescriptors = 1024;
1101 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1102 break;
1103 }
1104
1105 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1106 VkPhysicalDeviceMultiviewProperties *properties =
1107 (VkPhysicalDeviceMultiviewProperties *)ext;
1108 properties->maxMultiviewViewCount = 16;
1109 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1110 break;
1111 }
1112
1113 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1114 VkPhysicalDevicePointClippingProperties *properties =
1115 (VkPhysicalDevicePointClippingProperties *) ext;
1116 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
1117 anv_finishme("Implement pop-free point clipping");
1118 break;
1119 }
1120
1121 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1122 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1123
1124 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1125
1126 VkShaderStageFlags scalar_stages = 0;
1127 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1128 if (pdevice->compiler->scalar_stage[stage])
1129 scalar_stages |= mesa_to_vk_shader_stage(stage);
1130 }
1131 properties->supportedStages = scalar_stages;
1132
1133 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1134 VK_SUBGROUP_FEATURE_VOTE_BIT |
1135 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1136 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1137 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1138 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1139 VK_SUBGROUP_FEATURE_CLUSTERED_BIT |
1140 VK_SUBGROUP_FEATURE_QUAD_BIT;
1141 properties->quadOperationsInAllStages = VK_TRUE;
1142 break;
1143 }
1144
1145 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1146 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1147 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1148 /* We have to restrict this a bit for multiview */
1149 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1150 break;
1151 }
1152
1153 default:
1154 anv_debug_ignored_stype(ext->sType);
1155 break;
1156 }
1157 }
1158 }
1159
1160 /* We support exactly one queue family. */
1161 static const VkQueueFamilyProperties
1162 anv_queue_family_properties = {
1163 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1164 VK_QUEUE_COMPUTE_BIT |
1165 VK_QUEUE_TRANSFER_BIT,
1166 .queueCount = 1,
1167 .timestampValidBits = 36, /* XXX: Real value here */
1168 .minImageTransferGranularity = { 1, 1, 1 },
1169 };
1170
1171 void anv_GetPhysicalDeviceQueueFamilyProperties(
1172 VkPhysicalDevice physicalDevice,
1173 uint32_t* pCount,
1174 VkQueueFamilyProperties* pQueueFamilyProperties)
1175 {
1176 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1177
1178 vk_outarray_append(&out, p) {
1179 *p = anv_queue_family_properties;
1180 }
1181 }
1182
1183 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1184 VkPhysicalDevice physicalDevice,
1185 uint32_t* pQueueFamilyPropertyCount,
1186 VkQueueFamilyProperties2* pQueueFamilyProperties)
1187 {
1188
1189 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1190
1191 vk_outarray_append(&out, p) {
1192 p->queueFamilyProperties = anv_queue_family_properties;
1193
1194 vk_foreach_struct(s, p->pNext) {
1195 anv_debug_ignored_stype(s->sType);
1196 }
1197 }
1198 }
1199
1200 void anv_GetPhysicalDeviceMemoryProperties(
1201 VkPhysicalDevice physicalDevice,
1202 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1203 {
1204 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1205
1206 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1207 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1208 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1209 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1210 .heapIndex = physical_device->memory.types[i].heapIndex,
1211 };
1212 }
1213
1214 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1215 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1216 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1217 .size = physical_device->memory.heaps[i].size,
1218 .flags = physical_device->memory.heaps[i].flags,
1219 };
1220 }
1221 }
1222
1223 void anv_GetPhysicalDeviceMemoryProperties2(
1224 VkPhysicalDevice physicalDevice,
1225 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1226 {
1227 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1228 &pMemoryProperties->memoryProperties);
1229
1230 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1231 switch (ext->sType) {
1232 default:
1233 anv_debug_ignored_stype(ext->sType);
1234 break;
1235 }
1236 }
1237 }
1238
1239 void
1240 anv_GetDeviceGroupPeerMemoryFeatures(
1241 VkDevice device,
1242 uint32_t heapIndex,
1243 uint32_t localDeviceIndex,
1244 uint32_t remoteDeviceIndex,
1245 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1246 {
1247 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1248 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1249 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1250 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1251 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1252 }
1253
1254 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1255 VkInstance _instance,
1256 const char* pName)
1257 {
1258 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1259
1260 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1261 * when we have to return valid function pointers, NULL, or it's left
1262 * undefined. See the table for exact details.
1263 */
1264 if (pName == NULL)
1265 return NULL;
1266
1267 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1268 if (strcmp(pName, "vk" #entrypoint) == 0) \
1269 return (PFN_vkVoidFunction)anv_##entrypoint
1270
1271 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1272 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1273 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1274 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1275
1276 #undef LOOKUP_ANV_ENTRYPOINT
1277
1278 if (instance == NULL)
1279 return NULL;
1280
1281 int idx = anv_get_entrypoint_index(pName);
1282 if (idx < 0)
1283 return NULL;
1284
1285 return instance->dispatch.entrypoints[idx];
1286 }
1287
1288 /* With version 1+ of the loader interface the ICD should expose
1289 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1290 */
1291 PUBLIC
1292 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1293 VkInstance instance,
1294 const char* pName);
1295
1296 PUBLIC
1297 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1298 VkInstance instance,
1299 const char* pName)
1300 {
1301 return anv_GetInstanceProcAddr(instance, pName);
1302 }
1303
1304 PFN_vkVoidFunction anv_GetDeviceProcAddr(
1305 VkDevice _device,
1306 const char* pName)
1307 {
1308 ANV_FROM_HANDLE(anv_device, device, _device);
1309
1310 if (!device || !pName)
1311 return NULL;
1312
1313 int idx = anv_get_entrypoint_index(pName);
1314 if (idx < 0)
1315 return NULL;
1316
1317 return device->dispatch.entrypoints[idx];
1318 }
1319
1320 VkResult
1321 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
1322 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
1323 const VkAllocationCallbacks* pAllocator,
1324 VkDebugReportCallbackEXT* pCallback)
1325 {
1326 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1327 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
1328 pCreateInfo, pAllocator, &instance->alloc,
1329 pCallback);
1330 }
1331
1332 void
1333 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
1334 VkDebugReportCallbackEXT _callback,
1335 const VkAllocationCallbacks* pAllocator)
1336 {
1337 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1338 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
1339 _callback, pAllocator, &instance->alloc);
1340 }
1341
1342 void
1343 anv_DebugReportMessageEXT(VkInstance _instance,
1344 VkDebugReportFlagsEXT flags,
1345 VkDebugReportObjectTypeEXT objectType,
1346 uint64_t object,
1347 size_t location,
1348 int32_t messageCode,
1349 const char* pLayerPrefix,
1350 const char* pMessage)
1351 {
1352 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1353 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
1354 object, location, messageCode, pLayerPrefix, pMessage);
1355 }
1356
1357 static void
1358 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
1359 {
1360 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1361 queue->device = device;
1362 queue->flags = 0;
1363 }
1364
1365 static void
1366 anv_queue_finish(struct anv_queue *queue)
1367 {
1368 }
1369
1370 static struct anv_state
1371 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
1372 {
1373 struct anv_state state;
1374
1375 state = anv_state_pool_alloc(pool, size, align);
1376 memcpy(state.map, p, size);
1377
1378 anv_state_flush(pool->block_pool.device, state);
1379
1380 return state;
1381 }
1382
1383 struct gen8_border_color {
1384 union {
1385 float float32[4];
1386 uint32_t uint32[4];
1387 };
1388 /* Pad out to 64 bytes */
1389 uint32_t _pad[12];
1390 };
1391
1392 static void
1393 anv_device_init_border_colors(struct anv_device *device)
1394 {
1395 static const struct gen8_border_color border_colors[] = {
1396 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
1397 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
1398 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
1399 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
1400 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
1401 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
1402 };
1403
1404 device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
1405 sizeof(border_colors), 64,
1406 border_colors);
1407 }
1408
1409 static void
1410 anv_device_init_trivial_batch(struct anv_device *device)
1411 {
1412 anv_bo_init_new(&device->trivial_batch_bo, device, 4096);
1413
1414 if (device->instance->physicalDevice.has_exec_async)
1415 device->trivial_batch_bo.flags |= EXEC_OBJECT_ASYNC;
1416
1417 if (device->instance->physicalDevice.use_softpin)
1418 device->trivial_batch_bo.flags |= EXEC_OBJECT_PINNED;
1419
1420 anv_vma_alloc(device, &device->trivial_batch_bo);
1421
1422 void *map = anv_gem_mmap(device, device->trivial_batch_bo.gem_handle,
1423 0, 4096, 0);
1424
1425 struct anv_batch batch = {
1426 .start = map,
1427 .next = map,
1428 .end = map + 4096,
1429 };
1430
1431 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1432 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1433
1434 if (!device->info.has_llc)
1435 gen_clflush_range(map, batch.next - map);
1436
1437 anv_gem_munmap(map, device->trivial_batch_bo.size);
1438 }
1439
1440 VkResult anv_EnumerateDeviceExtensionProperties(
1441 VkPhysicalDevice physicalDevice,
1442 const char* pLayerName,
1443 uint32_t* pPropertyCount,
1444 VkExtensionProperties* pProperties)
1445 {
1446 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1447 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1448 (void)device;
1449
1450 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
1451 if (device->supported_extensions.extensions[i]) {
1452 vk_outarray_append(&out, prop) {
1453 *prop = anv_device_extensions[i];
1454 }
1455 }
1456 }
1457
1458 return vk_outarray_status(&out);
1459 }
1460
1461 static void
1462 anv_device_init_dispatch(struct anv_device *device)
1463 {
1464 const struct anv_dispatch_table *genX_table;
1465 switch (device->info.gen) {
1466 case 11:
1467 genX_table = &gen11_dispatch_table;
1468 break;
1469 case 10:
1470 genX_table = &gen10_dispatch_table;
1471 break;
1472 case 9:
1473 genX_table = &gen9_dispatch_table;
1474 break;
1475 case 8:
1476 genX_table = &gen8_dispatch_table;
1477 break;
1478 case 7:
1479 if (device->info.is_haswell)
1480 genX_table = &gen75_dispatch_table;
1481 else
1482 genX_table = &gen7_dispatch_table;
1483 break;
1484 default:
1485 unreachable("unsupported gen\n");
1486 }
1487
1488 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
1489 /* Vulkan requires that entrypoints for extensions which have not been
1490 * enabled must not be advertised.
1491 */
1492 if (!anv_entrypoint_is_enabled(i, device->instance->apiVersion,
1493 &device->instance->enabled_extensions,
1494 &device->enabled_extensions)) {
1495 device->dispatch.entrypoints[i] = NULL;
1496 } else if (genX_table->entrypoints[i]) {
1497 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
1498 } else {
1499 device->dispatch.entrypoints[i] = anv_dispatch_table.entrypoints[i];
1500 }
1501 }
1502 }
1503
1504 static int
1505 vk_priority_to_gen(int priority)
1506 {
1507 switch (priority) {
1508 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
1509 return GEN_CONTEXT_LOW_PRIORITY;
1510 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
1511 return GEN_CONTEXT_MEDIUM_PRIORITY;
1512 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
1513 return GEN_CONTEXT_HIGH_PRIORITY;
1514 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
1515 return GEN_CONTEXT_REALTIME_PRIORITY;
1516 default:
1517 unreachable("Invalid priority");
1518 }
1519 }
1520
1521 static void
1522 anv_device_init_hiz_clear_batch(struct anv_device *device)
1523 {
1524 anv_bo_init_new(&device->hiz_clear_bo, device, 4096);
1525 uint32_t *map = anv_gem_mmap(device, device->hiz_clear_bo.gem_handle,
1526 0, 4096, 0);
1527
1528 union isl_color_value hiz_clear = { .u32 = { 0, } };
1529 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
1530
1531 memcpy(map, hiz_clear.u32, sizeof(hiz_clear.u32));
1532 anv_gem_munmap(map, device->hiz_clear_bo.size);
1533 }
1534
1535 VkResult anv_CreateDevice(
1536 VkPhysicalDevice physicalDevice,
1537 const VkDeviceCreateInfo* pCreateInfo,
1538 const VkAllocationCallbacks* pAllocator,
1539 VkDevice* pDevice)
1540 {
1541 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1542 VkResult result;
1543 struct anv_device *device;
1544
1545 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
1546
1547 struct anv_device_extension_table enabled_extensions = { };
1548 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1549 int idx;
1550 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
1551 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
1552 anv_device_extensions[idx].extensionName) == 0)
1553 break;
1554 }
1555
1556 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
1557 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1558
1559 if (!physical_device->supported_extensions.extensions[idx])
1560 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1561
1562 enabled_extensions.extensions[idx] = true;
1563 }
1564
1565 /* Check enabled features */
1566 if (pCreateInfo->pEnabledFeatures) {
1567 VkPhysicalDeviceFeatures supported_features;
1568 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1569 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
1570 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
1571 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1572 for (uint32_t i = 0; i < num_features; i++) {
1573 if (enabled_feature[i] && !supported_feature[i])
1574 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1575 }
1576 }
1577
1578 /* Check requested queues and fail if we are requested to create any
1579 * queues with flags we don't support.
1580 */
1581 assert(pCreateInfo->queueCreateInfoCount > 0);
1582 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1583 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
1584 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
1585 }
1586
1587 /* Check if client specified queue priority. */
1588 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
1589 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
1590 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
1591
1592 VkQueueGlobalPriorityEXT priority =
1593 queue_priority ? queue_priority->globalPriority :
1594 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
1595
1596 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
1597 sizeof(*device), 8,
1598 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1599 if (!device)
1600 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1601
1602 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1603 device->instance = physical_device->instance;
1604 device->chipset_id = physical_device->chipset_id;
1605 device->no_hw = physical_device->no_hw;
1606 device->lost = false;
1607
1608 if (pAllocator)
1609 device->alloc = *pAllocator;
1610 else
1611 device->alloc = physical_device->instance->alloc;
1612
1613 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1614 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1615 if (device->fd == -1) {
1616 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1617 goto fail_device;
1618 }
1619
1620 device->context_id = anv_gem_create_context(device);
1621 if (device->context_id == -1) {
1622 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1623 goto fail_fd;
1624 }
1625
1626 if (physical_device->use_softpin) {
1627 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
1628 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1629 goto fail_fd;
1630 }
1631
1632 /* keep the page with address zero out of the allocator */
1633 util_vma_heap_init(&device->vma_lo, LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
1634 device->vma_lo_available =
1635 physical_device->memory.heaps[physical_device->memory.heap_count - 1].size;
1636
1637 /* Leave the last 4GiB out of the high vma range, so that no state base
1638 * address + size can overflow 48 bits. For more information see the
1639 * comment about Wa32bitGeneralStateOffset in anv_allocator.c
1640 */
1641 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
1642 HIGH_HEAP_SIZE);
1643 device->vma_hi_available = physical_device->memory.heap_count == 1 ? 0 :
1644 physical_device->memory.heaps[0].size;
1645 }
1646
1647 /* As per spec, the driver implementation may deny requests to acquire
1648 * a priority above the default priority (MEDIUM) if the caller does not
1649 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1650 * is returned.
1651 */
1652 if (physical_device->has_context_priority) {
1653 int err = anv_gem_set_context_param(device->fd, device->context_id,
1654 I915_CONTEXT_PARAM_PRIORITY,
1655 vk_priority_to_gen(priority));
1656 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
1657 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
1658 goto fail_fd;
1659 }
1660 }
1661
1662 device->info = physical_device->info;
1663 device->isl_dev = physical_device->isl_dev;
1664
1665 /* On Broadwell and later, we can use batch chaining to more efficiently
1666 * implement growing command buffers. Prior to Haswell, the kernel
1667 * command parser gets in the way and we have to fall back to growing
1668 * the batch.
1669 */
1670 device->can_chain_batches = device->info.gen >= 8;
1671
1672 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1673 pCreateInfo->pEnabledFeatures->robustBufferAccess;
1674 device->enabled_extensions = enabled_extensions;
1675
1676 anv_device_init_dispatch(device);
1677
1678 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
1679 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1680 goto fail_context_id;
1681 }
1682
1683 pthread_condattr_t condattr;
1684 if (pthread_condattr_init(&condattr) != 0) {
1685 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1686 goto fail_mutex;
1687 }
1688 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
1689 pthread_condattr_destroy(&condattr);
1690 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1691 goto fail_mutex;
1692 }
1693 if (pthread_cond_init(&device->queue_submit, NULL) != 0) {
1694 pthread_condattr_destroy(&condattr);
1695 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1696 goto fail_mutex;
1697 }
1698 pthread_condattr_destroy(&condattr);
1699
1700 uint64_t bo_flags =
1701 (physical_device->supports_48bit_addresses ? EXEC_OBJECT_SUPPORTS_48B_ADDRESS : 0) |
1702 (physical_device->has_exec_async ? EXEC_OBJECT_ASYNC : 0) |
1703 (physical_device->has_exec_capture ? EXEC_OBJECT_CAPTURE : 0) |
1704 (physical_device->use_softpin ? EXEC_OBJECT_PINNED : 0);
1705
1706 anv_bo_pool_init(&device->batch_bo_pool, device, bo_flags);
1707
1708 result = anv_bo_cache_init(&device->bo_cache);
1709 if (result != VK_SUCCESS)
1710 goto fail_batch_bo_pool;
1711
1712 if (!physical_device->use_softpin)
1713 bo_flags &= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
1714
1715 result = anv_state_pool_init(&device->dynamic_state_pool, device,
1716 DYNAMIC_STATE_POOL_MIN_ADDRESS,
1717 16384,
1718 bo_flags);
1719 if (result != VK_SUCCESS)
1720 goto fail_bo_cache;
1721
1722 result = anv_state_pool_init(&device->instruction_state_pool, device,
1723 INSTRUCTION_STATE_POOL_MIN_ADDRESS,
1724 16384,
1725 bo_flags);
1726 if (result != VK_SUCCESS)
1727 goto fail_dynamic_state_pool;
1728
1729 result = anv_state_pool_init(&device->surface_state_pool, device,
1730 SURFACE_STATE_POOL_MIN_ADDRESS,
1731 4096,
1732 bo_flags);
1733 if (result != VK_SUCCESS)
1734 goto fail_instruction_state_pool;
1735
1736 if (physical_device->use_softpin) {
1737 result = anv_state_pool_init(&device->binding_table_pool, device,
1738 BINDING_TABLE_POOL_MIN_ADDRESS,
1739 4096,
1740 bo_flags);
1741 if (result != VK_SUCCESS)
1742 goto fail_surface_state_pool;
1743 }
1744
1745 result = anv_bo_init_new(&device->workaround_bo, device, 1024);
1746 if (result != VK_SUCCESS)
1747 goto fail_binding_table_pool;
1748
1749 if (physical_device->use_softpin)
1750 device->workaround_bo.flags |= EXEC_OBJECT_PINNED;
1751
1752 if (!anv_vma_alloc(device, &device->workaround_bo))
1753 goto fail_workaround_bo;
1754
1755 anv_device_init_trivial_batch(device);
1756
1757 if (device->info.gen >= 10)
1758 anv_device_init_hiz_clear_batch(device);
1759
1760 anv_scratch_pool_init(device, &device->scratch_pool);
1761
1762 anv_queue_init(device, &device->queue);
1763
1764 switch (device->info.gen) {
1765 case 7:
1766 if (!device->info.is_haswell)
1767 result = gen7_init_device_state(device);
1768 else
1769 result = gen75_init_device_state(device);
1770 break;
1771 case 8:
1772 result = gen8_init_device_state(device);
1773 break;
1774 case 9:
1775 result = gen9_init_device_state(device);
1776 break;
1777 case 10:
1778 result = gen10_init_device_state(device);
1779 break;
1780 case 11:
1781 result = gen11_init_device_state(device);
1782 break;
1783 default:
1784 /* Shouldn't get here as we don't create physical devices for any other
1785 * gens. */
1786 unreachable("unhandled gen");
1787 }
1788 if (result != VK_SUCCESS)
1789 goto fail_workaround_bo;
1790
1791 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
1792
1793 anv_device_init_blorp(device);
1794
1795 anv_device_init_border_colors(device);
1796
1797 *pDevice = anv_device_to_handle(device);
1798
1799 return VK_SUCCESS;
1800
1801 fail_workaround_bo:
1802 anv_queue_finish(&device->queue);
1803 anv_scratch_pool_finish(device, &device->scratch_pool);
1804 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1805 anv_gem_close(device, device->workaround_bo.gem_handle);
1806 fail_binding_table_pool:
1807 if (physical_device->use_softpin)
1808 anv_state_pool_finish(&device->binding_table_pool);
1809 fail_surface_state_pool:
1810 anv_state_pool_finish(&device->surface_state_pool);
1811 fail_instruction_state_pool:
1812 anv_state_pool_finish(&device->instruction_state_pool);
1813 fail_dynamic_state_pool:
1814 anv_state_pool_finish(&device->dynamic_state_pool);
1815 fail_bo_cache:
1816 anv_bo_cache_finish(&device->bo_cache);
1817 fail_batch_bo_pool:
1818 anv_bo_pool_finish(&device->batch_bo_pool);
1819 pthread_cond_destroy(&device->queue_submit);
1820 fail_mutex:
1821 pthread_mutex_destroy(&device->mutex);
1822 fail_context_id:
1823 anv_gem_destroy_context(device, device->context_id);
1824 fail_fd:
1825 close(device->fd);
1826 fail_device:
1827 vk_free(&device->alloc, device);
1828
1829 return result;
1830 }
1831
1832 void anv_DestroyDevice(
1833 VkDevice _device,
1834 const VkAllocationCallbacks* pAllocator)
1835 {
1836 ANV_FROM_HANDLE(anv_device, device, _device);
1837 struct anv_physical_device *physical_device;
1838
1839 if (!device)
1840 return;
1841
1842 physical_device = &device->instance->physicalDevice;
1843
1844 anv_device_finish_blorp(device);
1845
1846 anv_pipeline_cache_finish(&device->default_pipeline_cache);
1847
1848 anv_queue_finish(&device->queue);
1849
1850 #ifdef HAVE_VALGRIND
1851 /* We only need to free these to prevent valgrind errors. The backing
1852 * BO will go away in a couple of lines so we don't actually leak.
1853 */
1854 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1855 #endif
1856
1857 anv_scratch_pool_finish(device, &device->scratch_pool);
1858
1859 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1860 anv_vma_free(device, &device->workaround_bo);
1861 anv_gem_close(device, device->workaround_bo.gem_handle);
1862
1863 anv_vma_free(device, &device->trivial_batch_bo);
1864 anv_gem_close(device, device->trivial_batch_bo.gem_handle);
1865 if (device->info.gen >= 10)
1866 anv_gem_close(device, device->hiz_clear_bo.gem_handle);
1867
1868 if (physical_device->use_softpin)
1869 anv_state_pool_finish(&device->binding_table_pool);
1870 anv_state_pool_finish(&device->surface_state_pool);
1871 anv_state_pool_finish(&device->instruction_state_pool);
1872 anv_state_pool_finish(&device->dynamic_state_pool);
1873
1874 anv_bo_cache_finish(&device->bo_cache);
1875
1876 anv_bo_pool_finish(&device->batch_bo_pool);
1877
1878 pthread_cond_destroy(&device->queue_submit);
1879 pthread_mutex_destroy(&device->mutex);
1880
1881 anv_gem_destroy_context(device, device->context_id);
1882
1883 close(device->fd);
1884
1885 vk_free(&device->alloc, device);
1886 }
1887
1888 VkResult anv_EnumerateInstanceLayerProperties(
1889 uint32_t* pPropertyCount,
1890 VkLayerProperties* pProperties)
1891 {
1892 if (pProperties == NULL) {
1893 *pPropertyCount = 0;
1894 return VK_SUCCESS;
1895 }
1896
1897 /* None supported at this time */
1898 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1899 }
1900
1901 VkResult anv_EnumerateDeviceLayerProperties(
1902 VkPhysicalDevice physicalDevice,
1903 uint32_t* pPropertyCount,
1904 VkLayerProperties* pProperties)
1905 {
1906 if (pProperties == NULL) {
1907 *pPropertyCount = 0;
1908 return VK_SUCCESS;
1909 }
1910
1911 /* None supported at this time */
1912 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1913 }
1914
1915 void anv_GetDeviceQueue(
1916 VkDevice _device,
1917 uint32_t queueNodeIndex,
1918 uint32_t queueIndex,
1919 VkQueue* pQueue)
1920 {
1921 ANV_FROM_HANDLE(anv_device, device, _device);
1922
1923 assert(queueIndex == 0);
1924
1925 *pQueue = anv_queue_to_handle(&device->queue);
1926 }
1927
1928 void anv_GetDeviceQueue2(
1929 VkDevice _device,
1930 const VkDeviceQueueInfo2* pQueueInfo,
1931 VkQueue* pQueue)
1932 {
1933 ANV_FROM_HANDLE(anv_device, device, _device);
1934
1935 assert(pQueueInfo->queueIndex == 0);
1936
1937 if (pQueueInfo->flags == device->queue.flags)
1938 *pQueue = anv_queue_to_handle(&device->queue);
1939 else
1940 *pQueue = NULL;
1941 }
1942
1943 VkResult
1944 anv_device_query_status(struct anv_device *device)
1945 {
1946 /* This isn't likely as most of the callers of this function already check
1947 * for it. However, it doesn't hurt to check and it potentially lets us
1948 * avoid an ioctl.
1949 */
1950 if (unlikely(device->lost))
1951 return VK_ERROR_DEVICE_LOST;
1952
1953 uint32_t active, pending;
1954 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
1955 if (ret == -1) {
1956 /* We don't know the real error. */
1957 device->lost = true;
1958 return vk_errorf(device->instance, device, VK_ERROR_DEVICE_LOST,
1959 "get_reset_stats failed: %m");
1960 }
1961
1962 if (active) {
1963 device->lost = true;
1964 return vk_errorf(device->instance, device, VK_ERROR_DEVICE_LOST,
1965 "GPU hung on one of our command buffers");
1966 } else if (pending) {
1967 device->lost = true;
1968 return vk_errorf(device->instance, device, VK_ERROR_DEVICE_LOST,
1969 "GPU hung with commands in-flight");
1970 }
1971
1972 return VK_SUCCESS;
1973 }
1974
1975 VkResult
1976 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
1977 {
1978 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1979 * Other usages of the BO (such as on different hardware) will not be
1980 * flagged as "busy" by this ioctl. Use with care.
1981 */
1982 int ret = anv_gem_busy(device, bo->gem_handle);
1983 if (ret == 1) {
1984 return VK_NOT_READY;
1985 } else if (ret == -1) {
1986 /* We don't know the real error. */
1987 device->lost = true;
1988 return vk_errorf(device->instance, device, VK_ERROR_DEVICE_LOST,
1989 "gem wait failed: %m");
1990 }
1991
1992 /* Query for device status after the busy call. If the BO we're checking
1993 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1994 * client because it clearly doesn't have valid data. Yes, this most
1995 * likely means an ioctl, but we just did an ioctl to query the busy status
1996 * so it's no great loss.
1997 */
1998 return anv_device_query_status(device);
1999 }
2000
2001 VkResult
2002 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2003 int64_t timeout)
2004 {
2005 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2006 if (ret == -1 && errno == ETIME) {
2007 return VK_TIMEOUT;
2008 } else if (ret == -1) {
2009 /* We don't know the real error. */
2010 device->lost = true;
2011 return vk_errorf(device->instance, device, VK_ERROR_DEVICE_LOST,
2012 "gem wait failed: %m");
2013 }
2014
2015 /* Query for device status after the wait. If the BO we're waiting on got
2016 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2017 * because it clearly doesn't have valid data. Yes, this most likely means
2018 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2019 */
2020 return anv_device_query_status(device);
2021 }
2022
2023 VkResult anv_DeviceWaitIdle(
2024 VkDevice _device)
2025 {
2026 ANV_FROM_HANDLE(anv_device, device, _device);
2027 if (unlikely(device->lost))
2028 return VK_ERROR_DEVICE_LOST;
2029
2030 struct anv_batch batch;
2031
2032 uint32_t cmds[8];
2033 batch.start = batch.next = cmds;
2034 batch.end = (void *) cmds + sizeof(cmds);
2035
2036 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2037 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2038
2039 return anv_device_submit_simple_batch(device, &batch);
2040 }
2041
2042 bool
2043 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo)
2044 {
2045 if (!(bo->flags & EXEC_OBJECT_PINNED))
2046 return true;
2047
2048 pthread_mutex_lock(&device->vma_mutex);
2049
2050 bo->offset = 0;
2051
2052 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS &&
2053 device->vma_hi_available >= bo->size) {
2054 uint64_t addr = util_vma_heap_alloc(&device->vma_hi, bo->size, 4096);
2055 if (addr) {
2056 bo->offset = gen_canonical_address(addr);
2057 assert(addr == gen_48b_address(bo->offset));
2058 device->vma_hi_available -= bo->size;
2059 }
2060 }
2061
2062 if (bo->offset == 0 && device->vma_lo_available >= bo->size) {
2063 uint64_t addr = util_vma_heap_alloc(&device->vma_lo, bo->size, 4096);
2064 if (addr) {
2065 bo->offset = gen_canonical_address(addr);
2066 assert(addr == gen_48b_address(bo->offset));
2067 device->vma_lo_available -= bo->size;
2068 }
2069 }
2070
2071 pthread_mutex_unlock(&device->vma_mutex);
2072
2073 return bo->offset != 0;
2074 }
2075
2076 void
2077 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
2078 {
2079 if (!(bo->flags & EXEC_OBJECT_PINNED))
2080 return;
2081
2082 const uint64_t addr_48b = gen_48b_address(bo->offset);
2083
2084 pthread_mutex_lock(&device->vma_mutex);
2085
2086 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
2087 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
2088 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
2089 device->vma_lo_available += bo->size;
2090 } else {
2091 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS &&
2092 addr_48b <= HIGH_HEAP_MAX_ADDRESS);
2093 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
2094 device->vma_hi_available += bo->size;
2095 }
2096
2097 pthread_mutex_unlock(&device->vma_mutex);
2098
2099 bo->offset = 0;
2100 }
2101
2102 VkResult
2103 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
2104 {
2105 uint32_t gem_handle = anv_gem_create(device, size);
2106 if (!gem_handle)
2107 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
2108
2109 anv_bo_init(bo, gem_handle, size);
2110
2111 return VK_SUCCESS;
2112 }
2113
2114 VkResult anv_AllocateMemory(
2115 VkDevice _device,
2116 const VkMemoryAllocateInfo* pAllocateInfo,
2117 const VkAllocationCallbacks* pAllocator,
2118 VkDeviceMemory* pMem)
2119 {
2120 ANV_FROM_HANDLE(anv_device, device, _device);
2121 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2122 struct anv_device_memory *mem;
2123 VkResult result = VK_SUCCESS;
2124
2125 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
2126
2127 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2128 assert(pAllocateInfo->allocationSize > 0);
2129
2130 if (pAllocateInfo->allocationSize > MAX_MEMORY_ALLOCATION_SIZE)
2131 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
2132
2133 /* FINISHME: Fail if allocation request exceeds heap size. */
2134
2135 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
2136 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2137 if (mem == NULL)
2138 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2139
2140 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
2141 mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
2142 mem->map = NULL;
2143 mem->map_size = 0;
2144
2145 uint64_t bo_flags = 0;
2146
2147 assert(mem->type->heapIndex < pdevice->memory.heap_count);
2148 if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
2149 bo_flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2150
2151 const struct wsi_memory_allocate_info *wsi_info =
2152 vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
2153 if (wsi_info && wsi_info->implicit_sync) {
2154 /* We need to set the WRITE flag on window system buffers so that GEM
2155 * will know we're writing to them and synchronize uses on other rings
2156 * (eg if the display server uses the blitter ring).
2157 */
2158 bo_flags |= EXEC_OBJECT_WRITE;
2159 } else if (pdevice->has_exec_async) {
2160 bo_flags |= EXEC_OBJECT_ASYNC;
2161 }
2162
2163 if (pdevice->use_softpin)
2164 bo_flags |= EXEC_OBJECT_PINNED;
2165
2166 const VkImportMemoryFdInfoKHR *fd_info =
2167 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
2168
2169 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2170 * ignored.
2171 */
2172 if (fd_info && fd_info->handleType) {
2173 /* At the moment, we support only the below handle types. */
2174 assert(fd_info->handleType ==
2175 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2176 fd_info->handleType ==
2177 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2178
2179 result = anv_bo_cache_import(device, &device->bo_cache,
2180 fd_info->fd, bo_flags, &mem->bo);
2181 if (result != VK_SUCCESS)
2182 goto fail;
2183
2184 VkDeviceSize aligned_alloc_size =
2185 align_u64(pAllocateInfo->allocationSize, 4096);
2186
2187 /* For security purposes, we reject importing the bo if it's smaller
2188 * than the requested allocation size. This prevents a malicious client
2189 * from passing a buffer to a trusted client, lying about the size, and
2190 * telling the trusted client to try and texture from an image that goes
2191 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2192 * in the trusted client. The trusted client can protect itself against
2193 * this sort of attack but only if it can trust the buffer size.
2194 */
2195 if (mem->bo->size < aligned_alloc_size) {
2196 result = vk_errorf(device->instance, device,
2197 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR,
2198 "aligned allocationSize too large for "
2199 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
2200 "%"PRIu64"B > %"PRIu64"B",
2201 aligned_alloc_size, mem->bo->size);
2202 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2203 goto fail;
2204 }
2205
2206 /* From the Vulkan spec:
2207 *
2208 * "Importing memory from a file descriptor transfers ownership of
2209 * the file descriptor from the application to the Vulkan
2210 * implementation. The application must not perform any operations on
2211 * the file descriptor after a successful import."
2212 *
2213 * If the import fails, we leave the file descriptor open.
2214 */
2215 close(fd_info->fd);
2216 } else {
2217 result = anv_bo_cache_alloc(device, &device->bo_cache,
2218 pAllocateInfo->allocationSize, bo_flags,
2219 &mem->bo);
2220 if (result != VK_SUCCESS)
2221 goto fail;
2222
2223 const VkMemoryDedicatedAllocateInfoKHR *dedicated_info =
2224 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_DEDICATED_ALLOCATE_INFO_KHR);
2225 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
2226 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
2227
2228 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2229 * the BO. In this case, we have a dedicated allocation.
2230 */
2231 if (image->needs_set_tiling) {
2232 const uint32_t i915_tiling =
2233 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
2234 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
2235 image->planes[0].surface.isl.row_pitch,
2236 i915_tiling);
2237 if (ret) {
2238 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2239 return vk_errorf(device->instance, NULL,
2240 VK_ERROR_OUT_OF_DEVICE_MEMORY,
2241 "failed to set BO tiling: %m");
2242 }
2243 }
2244 }
2245 }
2246
2247 *pMem = anv_device_memory_to_handle(mem);
2248
2249 return VK_SUCCESS;
2250
2251 fail:
2252 vk_free2(&device->alloc, pAllocator, mem);
2253
2254 return result;
2255 }
2256
2257 VkResult anv_GetMemoryFdKHR(
2258 VkDevice device_h,
2259 const VkMemoryGetFdInfoKHR* pGetFdInfo,
2260 int* pFd)
2261 {
2262 ANV_FROM_HANDLE(anv_device, dev, device_h);
2263 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
2264
2265 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2266
2267 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2268 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2269
2270 return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
2271 }
2272
2273 VkResult anv_GetMemoryFdPropertiesKHR(
2274 VkDevice _device,
2275 VkExternalMemoryHandleTypeFlagBitsKHR handleType,
2276 int fd,
2277 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
2278 {
2279 ANV_FROM_HANDLE(anv_device, device, _device);
2280 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2281
2282 switch (handleType) {
2283 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
2284 /* dma-buf can be imported as any memory type */
2285 pMemoryFdProperties->memoryTypeBits =
2286 (1 << pdevice->memory.type_count) - 1;
2287 return VK_SUCCESS;
2288
2289 default:
2290 /* The valid usage section for this function says:
2291 *
2292 * "handleType must not be one of the handle types defined as
2293 * opaque."
2294 *
2295 * So opaque handle types fall into the default "unsupported" case.
2296 */
2297 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
2298 }
2299 }
2300
2301 void anv_FreeMemory(
2302 VkDevice _device,
2303 VkDeviceMemory _mem,
2304 const VkAllocationCallbacks* pAllocator)
2305 {
2306 ANV_FROM_HANDLE(anv_device, device, _device);
2307 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
2308
2309 if (mem == NULL)
2310 return;
2311
2312 if (mem->map)
2313 anv_UnmapMemory(_device, _mem);
2314
2315 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2316
2317 vk_free2(&device->alloc, pAllocator, mem);
2318 }
2319
2320 VkResult anv_MapMemory(
2321 VkDevice _device,
2322 VkDeviceMemory _memory,
2323 VkDeviceSize offset,
2324 VkDeviceSize size,
2325 VkMemoryMapFlags flags,
2326 void** ppData)
2327 {
2328 ANV_FROM_HANDLE(anv_device, device, _device);
2329 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
2330
2331 if (mem == NULL) {
2332 *ppData = NULL;
2333 return VK_SUCCESS;
2334 }
2335
2336 if (size == VK_WHOLE_SIZE)
2337 size = mem->bo->size - offset;
2338
2339 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2340 *
2341 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2342 * assert(size != 0);
2343 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2344 * equal to the size of the memory minus offset
2345 */
2346 assert(size > 0);
2347 assert(offset + size <= mem->bo->size);
2348
2349 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2350 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2351 * at a time is valid. We could just mmap up front and return an offset
2352 * pointer here, but that may exhaust virtual memory on 32 bit
2353 * userspace. */
2354
2355 uint32_t gem_flags = 0;
2356
2357 if (!device->info.has_llc &&
2358 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
2359 gem_flags |= I915_MMAP_WC;
2360
2361 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2362 uint64_t map_offset = offset & ~4095ull;
2363 assert(offset >= map_offset);
2364 uint64_t map_size = (offset + size) - map_offset;
2365
2366 /* Let's map whole pages */
2367 map_size = align_u64(map_size, 4096);
2368
2369 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
2370 map_offset, map_size, gem_flags);
2371 if (map == MAP_FAILED)
2372 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
2373
2374 mem->map = map;
2375 mem->map_size = map_size;
2376
2377 *ppData = mem->map + (offset - map_offset);
2378
2379 return VK_SUCCESS;
2380 }
2381
2382 void anv_UnmapMemory(
2383 VkDevice _device,
2384 VkDeviceMemory _memory)
2385 {
2386 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
2387
2388 if (mem == NULL)
2389 return;
2390
2391 anv_gem_munmap(mem->map, mem->map_size);
2392
2393 mem->map = NULL;
2394 mem->map_size = 0;
2395 }
2396
2397 static void
2398 clflush_mapped_ranges(struct anv_device *device,
2399 uint32_t count,
2400 const VkMappedMemoryRange *ranges)
2401 {
2402 for (uint32_t i = 0; i < count; i++) {
2403 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
2404 if (ranges[i].offset >= mem->map_size)
2405 continue;
2406
2407 gen_clflush_range(mem->map + ranges[i].offset,
2408 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
2409 }
2410 }
2411
2412 VkResult anv_FlushMappedMemoryRanges(
2413 VkDevice _device,
2414 uint32_t memoryRangeCount,
2415 const VkMappedMemoryRange* pMemoryRanges)
2416 {
2417 ANV_FROM_HANDLE(anv_device, device, _device);
2418
2419 if (device->info.has_llc)
2420 return VK_SUCCESS;
2421
2422 /* Make sure the writes we're flushing have landed. */
2423 __builtin_ia32_mfence();
2424
2425 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
2426
2427 return VK_SUCCESS;
2428 }
2429
2430 VkResult anv_InvalidateMappedMemoryRanges(
2431 VkDevice _device,
2432 uint32_t memoryRangeCount,
2433 const VkMappedMemoryRange* pMemoryRanges)
2434 {
2435 ANV_FROM_HANDLE(anv_device, device, _device);
2436
2437 if (device->info.has_llc)
2438 return VK_SUCCESS;
2439
2440 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
2441
2442 /* Make sure no reads get moved up above the invalidate. */
2443 __builtin_ia32_mfence();
2444
2445 return VK_SUCCESS;
2446 }
2447
2448 void anv_GetBufferMemoryRequirements(
2449 VkDevice _device,
2450 VkBuffer _buffer,
2451 VkMemoryRequirements* pMemoryRequirements)
2452 {
2453 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2454 ANV_FROM_HANDLE(anv_device, device, _device);
2455 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2456
2457 /* The Vulkan spec (git aaed022) says:
2458 *
2459 * memoryTypeBits is a bitfield and contains one bit set for every
2460 * supported memory type for the resource. The bit `1<<i` is set if and
2461 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2462 * structure for the physical device is supported.
2463 */
2464 uint32_t memory_types = 0;
2465 for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
2466 uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
2467 if ((valid_usage & buffer->usage) == buffer->usage)
2468 memory_types |= (1u << i);
2469 }
2470
2471 /* Base alignment requirement of a cache line */
2472 uint32_t alignment = 16;
2473
2474 /* We need an alignment of 32 for pushing UBOs */
2475 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
2476 alignment = MAX2(alignment, 32);
2477
2478 pMemoryRequirements->size = buffer->size;
2479 pMemoryRequirements->alignment = alignment;
2480
2481 /* Storage and Uniform buffers should have their size aligned to
2482 * 32-bits to avoid boundary checks when last DWord is not complete.
2483 * This would ensure that not internal padding would be needed for
2484 * 16-bit types.
2485 */
2486 if (device->robust_buffer_access &&
2487 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
2488 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
2489 pMemoryRequirements->size = align_u64(buffer->size, 4);
2490
2491 pMemoryRequirements->memoryTypeBits = memory_types;
2492 }
2493
2494 void anv_GetBufferMemoryRequirements2(
2495 VkDevice _device,
2496 const VkBufferMemoryRequirementsInfo2* pInfo,
2497 VkMemoryRequirements2* pMemoryRequirements)
2498 {
2499 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
2500 &pMemoryRequirements->memoryRequirements);
2501
2502 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
2503 switch (ext->sType) {
2504 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
2505 VkMemoryDedicatedRequirements *requirements = (void *)ext;
2506 requirements->prefersDedicatedAllocation = VK_FALSE;
2507 requirements->requiresDedicatedAllocation = VK_FALSE;
2508 break;
2509 }
2510
2511 default:
2512 anv_debug_ignored_stype(ext->sType);
2513 break;
2514 }
2515 }
2516 }
2517
2518 void anv_GetImageMemoryRequirements(
2519 VkDevice _device,
2520 VkImage _image,
2521 VkMemoryRequirements* pMemoryRequirements)
2522 {
2523 ANV_FROM_HANDLE(anv_image, image, _image);
2524 ANV_FROM_HANDLE(anv_device, device, _device);
2525 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2526
2527 /* The Vulkan spec (git aaed022) says:
2528 *
2529 * memoryTypeBits is a bitfield and contains one bit set for every
2530 * supported memory type for the resource. The bit `1<<i` is set if and
2531 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2532 * structure for the physical device is supported.
2533 *
2534 * All types are currently supported for images.
2535 */
2536 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
2537
2538 pMemoryRequirements->size = image->size;
2539 pMemoryRequirements->alignment = image->alignment;
2540 pMemoryRequirements->memoryTypeBits = memory_types;
2541 }
2542
2543 void anv_GetImageMemoryRequirements2(
2544 VkDevice _device,
2545 const VkImageMemoryRequirementsInfo2* pInfo,
2546 VkMemoryRequirements2* pMemoryRequirements)
2547 {
2548 ANV_FROM_HANDLE(anv_device, device, _device);
2549 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
2550
2551 anv_GetImageMemoryRequirements(_device, pInfo->image,
2552 &pMemoryRequirements->memoryRequirements);
2553
2554 vk_foreach_struct_const(ext, pInfo->pNext) {
2555 switch (ext->sType) {
2556 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
2557 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2558 const VkImagePlaneMemoryRequirementsInfoKHR *plane_reqs =
2559 (const VkImagePlaneMemoryRequirementsInfoKHR *) ext;
2560 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
2561 plane_reqs->planeAspect);
2562
2563 assert(image->planes[plane].offset == 0);
2564
2565 /* The Vulkan spec (git aaed022) says:
2566 *
2567 * memoryTypeBits is a bitfield and contains one bit set for every
2568 * supported memory type for the resource. The bit `1<<i` is set
2569 * if and only if the memory type `i` in the
2570 * VkPhysicalDeviceMemoryProperties structure for the physical
2571 * device is supported.
2572 *
2573 * All types are currently supported for images.
2574 */
2575 pMemoryRequirements->memoryRequirements.memoryTypeBits =
2576 (1ull << pdevice->memory.type_count) - 1;
2577
2578 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
2579 pMemoryRequirements->memoryRequirements.alignment =
2580 image->planes[plane].alignment;
2581 break;
2582 }
2583
2584 default:
2585 anv_debug_ignored_stype(ext->sType);
2586 break;
2587 }
2588 }
2589
2590 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
2591 switch (ext->sType) {
2592 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
2593 VkMemoryDedicatedRequirements *requirements = (void *)ext;
2594 if (image->needs_set_tiling) {
2595 /* If we need to set the tiling for external consumers, we need a
2596 * dedicated allocation.
2597 *
2598 * See also anv_AllocateMemory.
2599 */
2600 requirements->prefersDedicatedAllocation = VK_TRUE;
2601 requirements->requiresDedicatedAllocation = VK_TRUE;
2602 } else {
2603 requirements->prefersDedicatedAllocation = VK_FALSE;
2604 requirements->requiresDedicatedAllocation = VK_FALSE;
2605 }
2606 break;
2607 }
2608
2609 default:
2610 anv_debug_ignored_stype(ext->sType);
2611 break;
2612 }
2613 }
2614 }
2615
2616 void anv_GetImageSparseMemoryRequirements(
2617 VkDevice device,
2618 VkImage image,
2619 uint32_t* pSparseMemoryRequirementCount,
2620 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
2621 {
2622 *pSparseMemoryRequirementCount = 0;
2623 }
2624
2625 void anv_GetImageSparseMemoryRequirements2(
2626 VkDevice device,
2627 const VkImageSparseMemoryRequirementsInfo2* pInfo,
2628 uint32_t* pSparseMemoryRequirementCount,
2629 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
2630 {
2631 *pSparseMemoryRequirementCount = 0;
2632 }
2633
2634 void anv_GetDeviceMemoryCommitment(
2635 VkDevice device,
2636 VkDeviceMemory memory,
2637 VkDeviceSize* pCommittedMemoryInBytes)
2638 {
2639 *pCommittedMemoryInBytes = 0;
2640 }
2641
2642 static void
2643 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
2644 {
2645 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
2646 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
2647
2648 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
2649
2650 if (mem) {
2651 assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
2652 buffer->address = (struct anv_address) {
2653 .bo = mem->bo,
2654 .offset = pBindInfo->memoryOffset,
2655 };
2656 } else {
2657 buffer->address = ANV_NULL_ADDRESS;
2658 }
2659 }
2660
2661 VkResult anv_BindBufferMemory(
2662 VkDevice device,
2663 VkBuffer buffer,
2664 VkDeviceMemory memory,
2665 VkDeviceSize memoryOffset)
2666 {
2667 anv_bind_buffer_memory(
2668 &(VkBindBufferMemoryInfo) {
2669 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
2670 .buffer = buffer,
2671 .memory = memory,
2672 .memoryOffset = memoryOffset,
2673 });
2674
2675 return VK_SUCCESS;
2676 }
2677
2678 VkResult anv_BindBufferMemory2(
2679 VkDevice device,
2680 uint32_t bindInfoCount,
2681 const VkBindBufferMemoryInfo* pBindInfos)
2682 {
2683 for (uint32_t i = 0; i < bindInfoCount; i++)
2684 anv_bind_buffer_memory(&pBindInfos[i]);
2685
2686 return VK_SUCCESS;
2687 }
2688
2689 VkResult anv_QueueBindSparse(
2690 VkQueue _queue,
2691 uint32_t bindInfoCount,
2692 const VkBindSparseInfo* pBindInfo,
2693 VkFence fence)
2694 {
2695 ANV_FROM_HANDLE(anv_queue, queue, _queue);
2696 if (unlikely(queue->device->lost))
2697 return VK_ERROR_DEVICE_LOST;
2698
2699 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2700 }
2701
2702 // Event functions
2703
2704 VkResult anv_CreateEvent(
2705 VkDevice _device,
2706 const VkEventCreateInfo* pCreateInfo,
2707 const VkAllocationCallbacks* pAllocator,
2708 VkEvent* pEvent)
2709 {
2710 ANV_FROM_HANDLE(anv_device, device, _device);
2711 struct anv_state state;
2712 struct anv_event *event;
2713
2714 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
2715
2716 state = anv_state_pool_alloc(&device->dynamic_state_pool,
2717 sizeof(*event), 8);
2718 event = state.map;
2719 event->state = state;
2720 event->semaphore = VK_EVENT_RESET;
2721
2722 if (!device->info.has_llc) {
2723 /* Make sure the writes we're flushing have landed. */
2724 __builtin_ia32_mfence();
2725 __builtin_ia32_clflush(event);
2726 }
2727
2728 *pEvent = anv_event_to_handle(event);
2729
2730 return VK_SUCCESS;
2731 }
2732
2733 void anv_DestroyEvent(
2734 VkDevice _device,
2735 VkEvent _event,
2736 const VkAllocationCallbacks* pAllocator)
2737 {
2738 ANV_FROM_HANDLE(anv_device, device, _device);
2739 ANV_FROM_HANDLE(anv_event, event, _event);
2740
2741 if (!event)
2742 return;
2743
2744 anv_state_pool_free(&device->dynamic_state_pool, event->state);
2745 }
2746
2747 VkResult anv_GetEventStatus(
2748 VkDevice _device,
2749 VkEvent _event)
2750 {
2751 ANV_FROM_HANDLE(anv_device, device, _device);
2752 ANV_FROM_HANDLE(anv_event, event, _event);
2753
2754 if (unlikely(device->lost))
2755 return VK_ERROR_DEVICE_LOST;
2756
2757 if (!device->info.has_llc) {
2758 /* Invalidate read cache before reading event written by GPU. */
2759 __builtin_ia32_clflush(event);
2760 __builtin_ia32_mfence();
2761
2762 }
2763
2764 return event->semaphore;
2765 }
2766
2767 VkResult anv_SetEvent(
2768 VkDevice _device,
2769 VkEvent _event)
2770 {
2771 ANV_FROM_HANDLE(anv_device, device, _device);
2772 ANV_FROM_HANDLE(anv_event, event, _event);
2773
2774 event->semaphore = VK_EVENT_SET;
2775
2776 if (!device->info.has_llc) {
2777 /* Make sure the writes we're flushing have landed. */
2778 __builtin_ia32_mfence();
2779 __builtin_ia32_clflush(event);
2780 }
2781
2782 return VK_SUCCESS;
2783 }
2784
2785 VkResult anv_ResetEvent(
2786 VkDevice _device,
2787 VkEvent _event)
2788 {
2789 ANV_FROM_HANDLE(anv_device, device, _device);
2790 ANV_FROM_HANDLE(anv_event, event, _event);
2791
2792 event->semaphore = VK_EVENT_RESET;
2793
2794 if (!device->info.has_llc) {
2795 /* Make sure the writes we're flushing have landed. */
2796 __builtin_ia32_mfence();
2797 __builtin_ia32_clflush(event);
2798 }
2799
2800 return VK_SUCCESS;
2801 }
2802
2803 // Buffer functions
2804
2805 VkResult anv_CreateBuffer(
2806 VkDevice _device,
2807 const VkBufferCreateInfo* pCreateInfo,
2808 const VkAllocationCallbacks* pAllocator,
2809 VkBuffer* pBuffer)
2810 {
2811 ANV_FROM_HANDLE(anv_device, device, _device);
2812 struct anv_buffer *buffer;
2813
2814 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2815
2816 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2817 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2818 if (buffer == NULL)
2819 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2820
2821 buffer->size = pCreateInfo->size;
2822 buffer->usage = pCreateInfo->usage;
2823 buffer->address = ANV_NULL_ADDRESS;
2824
2825 *pBuffer = anv_buffer_to_handle(buffer);
2826
2827 return VK_SUCCESS;
2828 }
2829
2830 void anv_DestroyBuffer(
2831 VkDevice _device,
2832 VkBuffer _buffer,
2833 const VkAllocationCallbacks* pAllocator)
2834 {
2835 ANV_FROM_HANDLE(anv_device, device, _device);
2836 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2837
2838 if (!buffer)
2839 return;
2840
2841 vk_free2(&device->alloc, pAllocator, buffer);
2842 }
2843
2844 void
2845 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
2846 enum isl_format format,
2847 struct anv_address address,
2848 uint32_t range, uint32_t stride)
2849 {
2850 isl_buffer_fill_state(&device->isl_dev, state.map,
2851 .address = anv_address_physical(address),
2852 .mocs = device->default_mocs,
2853 .size = range,
2854 .format = format,
2855 .stride = stride);
2856
2857 anv_state_flush(device, state);
2858 }
2859
2860 void anv_DestroySampler(
2861 VkDevice _device,
2862 VkSampler _sampler,
2863 const VkAllocationCallbacks* pAllocator)
2864 {
2865 ANV_FROM_HANDLE(anv_device, device, _device);
2866 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
2867
2868 if (!sampler)
2869 return;
2870
2871 vk_free2(&device->alloc, pAllocator, sampler);
2872 }
2873
2874 VkResult anv_CreateFramebuffer(
2875 VkDevice _device,
2876 const VkFramebufferCreateInfo* pCreateInfo,
2877 const VkAllocationCallbacks* pAllocator,
2878 VkFramebuffer* pFramebuffer)
2879 {
2880 ANV_FROM_HANDLE(anv_device, device, _device);
2881 struct anv_framebuffer *framebuffer;
2882
2883 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2884
2885 size_t size = sizeof(*framebuffer) +
2886 sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2887 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2888 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2889 if (framebuffer == NULL)
2890 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2891
2892 framebuffer->attachment_count = pCreateInfo->attachmentCount;
2893 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2894 VkImageView _iview = pCreateInfo->pAttachments[i];
2895 framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2896 }
2897
2898 framebuffer->width = pCreateInfo->width;
2899 framebuffer->height = pCreateInfo->height;
2900 framebuffer->layers = pCreateInfo->layers;
2901
2902 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2903
2904 return VK_SUCCESS;
2905 }
2906
2907 void anv_DestroyFramebuffer(
2908 VkDevice _device,
2909 VkFramebuffer _fb,
2910 const VkAllocationCallbacks* pAllocator)
2911 {
2912 ANV_FROM_HANDLE(anv_device, device, _device);
2913 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2914
2915 if (!fb)
2916 return;
2917
2918 vk_free2(&device->alloc, pAllocator, fb);
2919 }
2920
2921 /* vk_icd.h does not declare this function, so we declare it here to
2922 * suppress Wmissing-prototypes.
2923 */
2924 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2925 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2926
2927 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2928 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2929 {
2930 /* For the full details on loader interface versioning, see
2931 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2932 * What follows is a condensed summary, to help you navigate the large and
2933 * confusing official doc.
2934 *
2935 * - Loader interface v0 is incompatible with later versions. We don't
2936 * support it.
2937 *
2938 * - In loader interface v1:
2939 * - The first ICD entrypoint called by the loader is
2940 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2941 * entrypoint.
2942 * - The ICD must statically expose no other Vulkan symbol unless it is
2943 * linked with -Bsymbolic.
2944 * - Each dispatchable Vulkan handle created by the ICD must be
2945 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2946 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2947 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2948 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2949 * such loader-managed surfaces.
2950 *
2951 * - Loader interface v2 differs from v1 in:
2952 * - The first ICD entrypoint called by the loader is
2953 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2954 * statically expose this entrypoint.
2955 *
2956 * - Loader interface v3 differs from v2 in:
2957 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2958 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2959 * because the loader no longer does so.
2960 */
2961 *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2962 return VK_SUCCESS;
2963 }