projects / mesa.git / blob
? search:


7beb70f9e6b5ace0a117fa0fcd3dfae3cf69d2f7
[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-uapi/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 "util/u_string.h"
41 #include "git_sha1.h"
42 #include "vk_util.h"
43 #include "common/gen_defines.h"
44
45 #include "genxml/gen7_pack.h"
46
47 static void
48 compiler_debug_log(void *data, const char *fmt, ...)
49 { }
50
51 static void
52 compiler_perf_log(void *data, const char *fmt, ...)
53 {
54 va_list args;
55 va_start(args, fmt);
56
57 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
58 intel_logd_v(fmt, args);
59
60 va_end(args);
61 }
62
63 static uint64_t
64 anv_compute_heap_size(int fd, uint64_t gtt_size)
65 {
66 /* Query the total ram from the system */
67 struct sysinfo info;
68 sysinfo(&info);
69
70 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
71
72 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
73 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
74 */
75 uint64_t available_ram;
76 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
77 available_ram = total_ram / 2;
78 else
79 available_ram = total_ram * 3 / 4;
80
81 /* We also want to leave some padding for things we allocate in the driver,
82 * so don't go over 3/4 of the GTT either.
83 */
84 uint64_t available_gtt = gtt_size * 3 / 4;
85
86 return MIN2(available_ram, available_gtt);
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 = anv_compute_heap_size(fd, gtt_size);
111
112 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
113 /* When running with an overridden PCI ID, we may get a GTT size from
114 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
115 * address support can still fail. Just clamp the address space size to
116 * 2 GiB if we don't have 48-bit support.
117 */
118 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
119 "not support for 48-bit addresses",
120 __FILE__, __LINE__);
121 heap_size = 2ull << 30;
122 }
123
124 if (heap_size <= 3ull * (1ull << 30)) {
125 /* In this case, everything fits nicely into the 32-bit address space,
126 * so there's no need for supporting 48bit addresses on client-allocated
127 * memory objects.
128 */
129 device->memory.heap_count = 1;
130 device->memory.heaps[0] = (struct anv_memory_heap) {
131 .size = heap_size,
132 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
133 .supports_48bit_addresses = false,
134 };
135 } else {
136 /* Not everything will fit nicely into a 32-bit address space. In this
137 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
138 * larger 48-bit heap. If we're in this case, then we have a total heap
139 * size larger than 3GiB which most likely means they have 8 GiB of
140 * video memory and so carving off 1 GiB for the 32-bit heap should be
141 * reasonable.
142 */
143 const uint64_t heap_size_32bit = 1ull << 30;
144 const uint64_t heap_size_48bit = heap_size - heap_size_32bit;
145
146 assert(device->supports_48bit_addresses);
147
148 device->memory.heap_count = 2;
149 device->memory.heaps[0] = (struct anv_memory_heap) {
150 .size = heap_size_48bit,
151 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
152 .supports_48bit_addresses = true,
153 };
154 device->memory.heaps[1] = (struct anv_memory_heap) {
155 .size = heap_size_32bit,
156 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
157 .supports_48bit_addresses = false,
158 };
159 }
160
161 uint32_t type_count = 0;
162 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
163 uint32_t valid_buffer_usage = ~0;
164
165 /* There appears to be a hardware issue in the VF cache where it only
166 * considers the bottom 32 bits of memory addresses. If you happen to
167 * have two vertex buffers which get placed exactly 4 GiB apart and use
168 * them in back-to-back draw calls, you can get collisions. In order to
169 * solve this problem, we require vertex and index buffers be bound to
170 * memory allocated out of the 32-bit heap.
171 */
172 if (device->memory.heaps[heap].supports_48bit_addresses) {
173 valid_buffer_usage &= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
174 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
175 }
176
177 if (device->info.has_llc) {
178 /* Big core GPUs share LLC with the CPU and thus one memory type can be
179 * both cached and coherent at the same time.
180 */
181 device->memory.types[type_count++] = (struct anv_memory_type) {
182 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
183 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
184 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
185 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
186 .heapIndex = heap,
187 .valid_buffer_usage = valid_buffer_usage,
188 };
189 } else {
190 /* The spec requires that we expose a host-visible, coherent memory
191 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
192 * to give the application a choice between cached, but not coherent and
193 * coherent but uncached (WC though).
194 */
195 device->memory.types[type_count++] = (struct anv_memory_type) {
196 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
197 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
198 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
199 .heapIndex = heap,
200 .valid_buffer_usage = valid_buffer_usage,
201 };
202 device->memory.types[type_count++] = (struct anv_memory_type) {
203 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
204 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
205 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
206 .heapIndex = heap,
207 .valid_buffer_usage = valid_buffer_usage,
208 };
209 }
210 }
211 device->memory.type_count = type_count;
212
213 return VK_SUCCESS;
214 }
215
216 static VkResult
217 anv_physical_device_init_uuids(struct anv_physical_device *device)
218 {
219 const struct build_id_note *note =
220 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
221 if (!note) {
222 return vk_errorf(device->instance, device,
223 VK_ERROR_INITIALIZATION_FAILED,
224 "Failed to find build-id");
225 }
226
227 unsigned build_id_len = build_id_length(note);
228 if (build_id_len < 20) {
229 return vk_errorf(device->instance, device,
230 VK_ERROR_INITIALIZATION_FAILED,
231 "build-id too short. It needs to be a SHA");
232 }
233
234 memcpy(device->driver_build_sha1, build_id_data(note), 20);
235
236 struct mesa_sha1 sha1_ctx;
237 uint8_t sha1[20];
238 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
239
240 /* The pipeline cache UUID is used for determining when a pipeline cache is
241 * invalid. It needs both a driver build and the PCI ID of the device.
242 */
243 _mesa_sha1_init(&sha1_ctx);
244 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
245 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
246 sizeof(device->chipset_id));
247 _mesa_sha1_final(&sha1_ctx, sha1);
248 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
249
250 /* The driver UUID is used for determining sharability of images and memory
251 * between two Vulkan instances in separate processes. People who want to
252 * share memory need to also check the device UUID (below) so all this
253 * needs to be is the build-id.
254 */
255 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
256
257 /* The device UUID uniquely identifies the given device within the machine.
258 * Since we never have more than one device, this doesn't need to be a real
259 * UUID. However, on the off-chance that someone tries to use this to
260 * cache pre-tiled images or something of the like, we use the PCI ID and
261 * some bits of ISL info to ensure that this is safe.
262 */
263 _mesa_sha1_init(&sha1_ctx);
264 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
265 sizeof(device->chipset_id));
266 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
267 sizeof(device->isl_dev.has_bit6_swizzling));
268 _mesa_sha1_final(&sha1_ctx, sha1);
269 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
270
271 return VK_SUCCESS;
272 }
273
274 static void
275 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
276 {
277 #ifdef ENABLE_SHADER_CACHE
278 char renderer[10];
279 MAYBE_UNUSED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
280 device->chipset_id);
281 assert(len == sizeof(renderer) - 2);
282
283 char timestamp[41];
284 _mesa_sha1_format(timestamp, device->driver_build_sha1);
285
286 const uint64_t driver_flags =
287 brw_get_compiler_config_value(device->compiler);
288 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
289 #else
290 device->disk_cache = NULL;
291 #endif
292 }
293
294 static void
295 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
296 {
297 #ifdef ENABLE_SHADER_CACHE
298 if (device->disk_cache)
299 disk_cache_destroy(device->disk_cache);
300 #else
301 assert(device->disk_cache == NULL);
302 #endif
303 }
304
305 static VkResult
306 anv_physical_device_init(struct anv_physical_device *device,
307 struct anv_instance *instance,
308 drmDevicePtr drm_device)
309 {
310 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
311 const char *path = drm_device->nodes[DRM_NODE_RENDER];
312 VkResult result;
313 int fd;
314 int master_fd = -1;
315
316 brw_process_intel_debug_variable();
317
318 fd = open(path, O_RDWR | O_CLOEXEC);
319 if (fd < 0)
320 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
321
322 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
323 device->instance = instance;
324
325 assert(strlen(path) < ARRAY_SIZE(device->path));
326 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
327
328 device->no_hw = getenv("INTEL_NO_HW") != NULL;
329
330 const int pci_id_override = gen_get_pci_device_id_override();
331 if (pci_id_override < 0) {
332 device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
333 if (!device->chipset_id) {
334 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
335 goto fail;
336 }
337 } else {
338 device->chipset_id = pci_id_override;
339 device->no_hw = true;
340 }
341
342 device->pci_info.domain = drm_device->businfo.pci->domain;
343 device->pci_info.bus = drm_device->businfo.pci->bus;
344 device->pci_info.device = drm_device->businfo.pci->dev;
345 device->pci_info.function = drm_device->businfo.pci->func;
346
347 device->name = gen_get_device_name(device->chipset_id);
348 if (!gen_get_device_info(device->chipset_id, &device->info)) {
349 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
350 goto fail;
351 }
352
353 if (device->info.is_haswell) {
354 intel_logw("Haswell Vulkan support is incomplete");
355 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
356 intel_logw("Ivy Bridge Vulkan support is incomplete");
357 } else if (device->info.gen == 7 && device->info.is_baytrail) {
358 intel_logw("Bay Trail Vulkan support is incomplete");
359 } else if (device->info.gen >= 8 && device->info.gen <= 10) {
360 /* Gen8-10 fully supported */
361 } else if (device->info.gen == 11) {
362 intel_logw("Vulkan is not yet fully supported on gen11.");
363 } else {
364 result = vk_errorf(device->instance, device,
365 VK_ERROR_INCOMPATIBLE_DRIVER,
366 "Vulkan not yet supported on %s", device->name);
367 goto fail;
368 }
369
370 device->cmd_parser_version = -1;
371 if (device->info.gen == 7) {
372 device->cmd_parser_version =
373 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
374 if (device->cmd_parser_version == -1) {
375 result = vk_errorf(device->instance, device,
376 VK_ERROR_INITIALIZATION_FAILED,
377 "failed to get command parser version");
378 goto fail;
379 }
380 }
381
382 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
383 result = vk_errorf(device->instance, device,
384 VK_ERROR_INITIALIZATION_FAILED,
385 "kernel missing gem wait");
386 goto fail;
387 }
388
389 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
390 result = vk_errorf(device->instance, device,
391 VK_ERROR_INITIALIZATION_FAILED,
392 "kernel missing execbuf2");
393 goto fail;
394 }
395
396 if (!device->info.has_llc &&
397 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
398 result = vk_errorf(device->instance, device,
399 VK_ERROR_INITIALIZATION_FAILED,
400 "kernel missing wc mmap");
401 goto fail;
402 }
403
404 result = anv_physical_device_init_heaps(device, fd);
405 if (result != VK_SUCCESS)
406 goto fail;
407
408 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
409 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
410 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
411 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
412 device->has_syncobj_wait = device->has_syncobj &&
413 anv_gem_supports_syncobj_wait(fd);
414 device->has_context_priority = anv_gem_has_context_priority(fd);
415
416 device->use_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN)
417 && device->supports_48bit_addresses;
418
419 device->has_context_isolation =
420 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
421
422 /* Starting with Gen10, the timestamp frequency of the command streamer may
423 * vary from one part to another. We can query the value from the kernel.
424 */
425 if (device->info.gen >= 10) {
426 int timestamp_frequency =
427 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
428
429 if (timestamp_frequency < 0)
430 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
431 else
432 device->info.timestamp_frequency = timestamp_frequency;
433 }
434
435 /* GENs prior to 8 do not support EU/Subslice info */
436 if (device->info.gen >= 8) {
437 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
438 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
439
440 /* Without this information, we cannot get the right Braswell
441 * brandstrings, and we have to use conservative numbers for GPGPU on
442 * many platforms, but otherwise, things will just work.
443 */
444 if (device->subslice_total < 1 || device->eu_total < 1) {
445 intel_logw("Kernel 4.1 required to properly query GPU properties");
446 }
447 } else if (device->info.gen == 7) {
448 device->subslice_total = 1 << (device->info.gt - 1);
449 }
450
451 if (device->info.is_cherryview &&
452 device->subslice_total > 0 && device->eu_total > 0) {
453 /* Logical CS threads = EUs per subslice * num threads per EU */
454 uint32_t max_cs_threads =
455 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
456
457 /* Fuse configurations may give more threads than expected, never less. */
458 if (max_cs_threads > device->info.max_cs_threads)
459 device->info.max_cs_threads = max_cs_threads;
460 }
461
462 device->compiler = brw_compiler_create(NULL, &device->info);
463 if (device->compiler == NULL) {
464 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
465 goto fail;
466 }
467 device->compiler->shader_debug_log = compiler_debug_log;
468 device->compiler->shader_perf_log = compiler_perf_log;
469 device->compiler->supports_pull_constants = false;
470 device->compiler->constant_buffer_0_is_relative =
471 device->info.gen < 8 || !device->has_context_isolation;
472 device->compiler->supports_shader_constants = true;
473
474 /* Broadwell PRM says:
475 *
476 * "Before Gen8, there was a historical configuration control field to
477 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
478 * different places: TILECTL[1:0], ARB_MODE[5:4], and
479 * DISP_ARB_CTL[14:13].
480 *
481 * For Gen8 and subsequent generations, the swizzle fields are all
482 * reserved, and the CPU's memory controller performs all address
483 * swizzling modifications."
484 */
485 bool swizzled =
486 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
487
488 isl_device_init(&device->isl_dev, &device->info, swizzled);
489
490 result = anv_physical_device_init_uuids(device);
491 if (result != VK_SUCCESS)
492 goto fail;
493
494 anv_physical_device_init_disk_cache(device);
495
496 if (instance->enabled_extensions.KHR_display) {
497 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
498 if (master_fd >= 0) {
499 /* prod the device with a GETPARAM call which will fail if
500 * we don't have permission to even render on this device
501 */
502 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
503 close(master_fd);
504 master_fd = -1;
505 }
506 }
507 }
508 device->master_fd = master_fd;
509
510 result = anv_init_wsi(device);
511 if (result != VK_SUCCESS) {
512 ralloc_free(device->compiler);
513 anv_physical_device_free_disk_cache(device);
514 goto fail;
515 }
516
517 anv_physical_device_get_supported_extensions(device,
518 &device->supported_extensions);
519
520
521 device->local_fd = fd;
522
523 return VK_SUCCESS;
524
525 fail:
526 close(fd);
527 if (master_fd != -1)
528 close(master_fd);
529 return result;
530 }
531
532 static void
533 anv_physical_device_finish(struct anv_physical_device *device)
534 {
535 anv_finish_wsi(device);
536 anv_physical_device_free_disk_cache(device);
537 ralloc_free(device->compiler);
538 close(device->local_fd);
539 if (device->master_fd >= 0)
540 close(device->master_fd);
541 }
542
543 static void *
544 default_alloc_func(void *pUserData, size_t size, size_t align,
545 VkSystemAllocationScope allocationScope)
546 {
547 return malloc(size);
548 }
549
550 static void *
551 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
552 size_t align, VkSystemAllocationScope allocationScope)
553 {
554 return realloc(pOriginal, size);
555 }
556
557 static void
558 default_free_func(void *pUserData, void *pMemory)
559 {
560 free(pMemory);
561 }
562
563 static const VkAllocationCallbacks default_alloc = {
564 .pUserData = NULL,
565 .pfnAllocation = default_alloc_func,
566 .pfnReallocation = default_realloc_func,
567 .pfnFree = default_free_func,
568 };
569
570 VkResult anv_EnumerateInstanceExtensionProperties(
571 const char* pLayerName,
572 uint32_t* pPropertyCount,
573 VkExtensionProperties* pProperties)
574 {
575 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
576
577 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
578 if (anv_instance_extensions_supported.extensions[i]) {
579 vk_outarray_append(&out, prop) {
580 *prop = anv_instance_extensions[i];
581 }
582 }
583 }
584
585 return vk_outarray_status(&out);
586 }
587
588 VkResult anv_CreateInstance(
589 const VkInstanceCreateInfo* pCreateInfo,
590 const VkAllocationCallbacks* pAllocator,
591 VkInstance* pInstance)
592 {
593 struct anv_instance *instance;
594 VkResult result;
595
596 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
597
598 struct anv_instance_extension_table enabled_extensions = {};
599 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
600 int idx;
601 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
602 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
603 anv_instance_extensions[idx].extensionName) == 0)
604 break;
605 }
606
607 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
608 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
609
610 if (!anv_instance_extensions_supported.extensions[idx])
611 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
612
613 enabled_extensions.extensions[idx] = true;
614 }
615
616 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
617 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
618 if (!instance)
619 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
620
621 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
622
623 if (pAllocator)
624 instance->alloc = *pAllocator;
625 else
626 instance->alloc = default_alloc;
627
628 instance->app_info = (struct anv_app_info) { .api_version = 0 };
629 if (pCreateInfo->pApplicationInfo) {
630 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
631
632 instance->app_info.app_name =
633 vk_strdup(&instance->alloc, app->pApplicationName,
634 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
635 instance->app_info.app_version = app->applicationVersion;
636
637 instance->app_info.engine_name =
638 vk_strdup(&instance->alloc, app->pEngineName,
639 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
640 instance->app_info.engine_version = app->engineVersion;
641
642 instance->app_info.api_version = app->apiVersion;
643 }
644
645 if (instance->app_info.api_version == 0)
646 instance->app_info.api_version = VK_API_VERSION_1_0;
647
648 instance->enabled_extensions = enabled_extensions;
649
650 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
651 /* Vulkan requires that entrypoints for extensions which have not been
652 * enabled must not be advertised.
653 */
654 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
655 &instance->enabled_extensions)) {
656 instance->dispatch.entrypoints[i] = NULL;
657 } else {
658 instance->dispatch.entrypoints[i] =
659 anv_instance_dispatch_table.entrypoints[i];
660 }
661 }
662
663 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
664 /* Vulkan requires that entrypoints for extensions which have not been
665 * enabled must not be advertised.
666 */
667 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
668 &instance->enabled_extensions, NULL)) {
669 instance->device_dispatch.entrypoints[i] = NULL;
670 } else {
671 instance->device_dispatch.entrypoints[i] =
672 anv_device_dispatch_table.entrypoints[i];
673 }
674 }
675
676 instance->physicalDeviceCount = -1;
677
678 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
679 if (result != VK_SUCCESS) {
680 vk_free2(&default_alloc, pAllocator, instance);
681 return vk_error(result);
682 }
683
684 instance->pipeline_cache_enabled =
685 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
686
687 _mesa_locale_init();
688
689 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
690
691 *pInstance = anv_instance_to_handle(instance);
692
693 return VK_SUCCESS;
694 }
695
696 void anv_DestroyInstance(
697 VkInstance _instance,
698 const VkAllocationCallbacks* pAllocator)
699 {
700 ANV_FROM_HANDLE(anv_instance, instance, _instance);
701
702 if (!instance)
703 return;
704
705 if (instance->physicalDeviceCount > 0) {
706 /* We support at most one physical device. */
707 assert(instance->physicalDeviceCount == 1);
708 anv_physical_device_finish(&instance->physicalDevice);
709 }
710
711 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
712 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
713
714 VG(VALGRIND_DESTROY_MEMPOOL(instance));
715
716 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
717
718 _mesa_locale_fini();
719
720 vk_free(&instance->alloc, instance);
721 }
722
723 static VkResult
724 anv_enumerate_devices(struct anv_instance *instance)
725 {
726 /* TODO: Check for more devices ? */
727 drmDevicePtr devices[8];
728 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
729 int max_devices;
730
731 instance->physicalDeviceCount = 0;
732
733 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
734 if (max_devices < 1)
735 return VK_ERROR_INCOMPATIBLE_DRIVER;
736
737 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
738 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
739 devices[i]->bustype == DRM_BUS_PCI &&
740 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
741
742 result = anv_physical_device_init(&instance->physicalDevice,
743 instance, devices[i]);
744 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
745 break;
746 }
747 }
748 drmFreeDevices(devices, max_devices);
749
750 if (result == VK_SUCCESS)
751 instance->physicalDeviceCount = 1;
752
753 return result;
754 }
755
756 static VkResult
757 anv_instance_ensure_physical_device(struct anv_instance *instance)
758 {
759 if (instance->physicalDeviceCount < 0) {
760 VkResult result = anv_enumerate_devices(instance);
761 if (result != VK_SUCCESS &&
762 result != VK_ERROR_INCOMPATIBLE_DRIVER)
763 return result;
764 }
765
766 return VK_SUCCESS;
767 }
768
769 VkResult anv_EnumeratePhysicalDevices(
770 VkInstance _instance,
771 uint32_t* pPhysicalDeviceCount,
772 VkPhysicalDevice* pPhysicalDevices)
773 {
774 ANV_FROM_HANDLE(anv_instance, instance, _instance);
775 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
776
777 VkResult result = anv_instance_ensure_physical_device(instance);
778 if (result != VK_SUCCESS)
779 return result;
780
781 if (instance->physicalDeviceCount == 0)
782 return VK_SUCCESS;
783
784 assert(instance->physicalDeviceCount == 1);
785 vk_outarray_append(&out, i) {
786 *i = anv_physical_device_to_handle(&instance->physicalDevice);
787 }
788
789 return vk_outarray_status(&out);
790 }
791
792 VkResult anv_EnumeratePhysicalDeviceGroups(
793 VkInstance _instance,
794 uint32_t* pPhysicalDeviceGroupCount,
795 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
796 {
797 ANV_FROM_HANDLE(anv_instance, instance, _instance);
798 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
799 pPhysicalDeviceGroupCount);
800
801 VkResult result = anv_instance_ensure_physical_device(instance);
802 if (result != VK_SUCCESS)
803 return result;
804
805 if (instance->physicalDeviceCount == 0)
806 return VK_SUCCESS;
807
808 assert(instance->physicalDeviceCount == 1);
809
810 vk_outarray_append(&out, p) {
811 p->physicalDeviceCount = 1;
812 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
813 p->physicalDevices[0] =
814 anv_physical_device_to_handle(&instance->physicalDevice);
815 p->subsetAllocation = VK_FALSE;
816
817 vk_foreach_struct(ext, p->pNext)
818 anv_debug_ignored_stype(ext->sType);
819 }
820
821 return vk_outarray_status(&out);
822 }
823
824 void anv_GetPhysicalDeviceFeatures(
825 VkPhysicalDevice physicalDevice,
826 VkPhysicalDeviceFeatures* pFeatures)
827 {
828 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
829
830 *pFeatures = (VkPhysicalDeviceFeatures) {
831 .robustBufferAccess = true,
832 .fullDrawIndexUint32 = true,
833 .imageCubeArray = true,
834 .independentBlend = true,
835 .geometryShader = true,
836 .tessellationShader = true,
837 .sampleRateShading = true,
838 .dualSrcBlend = true,
839 .logicOp = true,
840 .multiDrawIndirect = true,
841 .drawIndirectFirstInstance = true,
842 .depthClamp = true,
843 .depthBiasClamp = true,
844 .fillModeNonSolid = true,
845 .depthBounds = false,
846 .wideLines = true,
847 .largePoints = true,
848 .alphaToOne = true,
849 .multiViewport = true,
850 .samplerAnisotropy = true,
851 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
852 pdevice->info.is_baytrail,
853 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
854 .textureCompressionBC = true,
855 .occlusionQueryPrecise = true,
856 .pipelineStatisticsQuery = true,
857 .fragmentStoresAndAtomics = true,
858 .shaderTessellationAndGeometryPointSize = true,
859 .shaderImageGatherExtended = true,
860 .shaderStorageImageExtendedFormats = true,
861 .shaderStorageImageMultisample = false,
862 .shaderStorageImageReadWithoutFormat = false,
863 .shaderStorageImageWriteWithoutFormat = true,
864 .shaderUniformBufferArrayDynamicIndexing = true,
865 .shaderSampledImageArrayDynamicIndexing = true,
866 .shaderStorageBufferArrayDynamicIndexing = true,
867 .shaderStorageImageArrayDynamicIndexing = true,
868 .shaderClipDistance = true,
869 .shaderCullDistance = true,
870 .shaderFloat64 = pdevice->info.gen >= 8 &&
871 pdevice->info.has_64bit_types,
872 .shaderInt64 = pdevice->info.gen >= 8 &&
873 pdevice->info.has_64bit_types,
874 .shaderInt16 = pdevice->info.gen >= 8,
875 .shaderResourceMinLod = pdevice->info.gen >= 9,
876 .variableMultisampleRate = true,
877 .inheritedQueries = true,
878 };
879
880 /* We can't do image stores in vec4 shaders */
881 pFeatures->vertexPipelineStoresAndAtomics =
882 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
883 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
884
885 struct anv_app_info *app_info = &pdevice->instance->app_info;
886
887 /* The new DOOM and Wolfenstein games require depthBounds without
888 * checking for it. They seem to run fine without it so just claim it's
889 * there and accept the consequences.
890 */
891 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
892 pFeatures->depthBounds = true;
893 }
894
895 void anv_GetPhysicalDeviceFeatures2(
896 VkPhysicalDevice physicalDevice,
897 VkPhysicalDeviceFeatures2* pFeatures)
898 {
899 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
900 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
901
902 vk_foreach_struct(ext, pFeatures->pNext) {
903 switch (ext->sType) {
904 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
905 VkPhysicalDevice8BitStorageFeaturesKHR *features =
906 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
907 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
908
909 features->storageBuffer8BitAccess = pdevice->info.gen >= 8;
910 features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
911 features->storagePushConstant8 = pdevice->info.gen >= 8;
912 break;
913 }
914
915 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
916 VkPhysicalDevice16BitStorageFeatures *features =
917 (VkPhysicalDevice16BitStorageFeatures *)ext;
918 features->storageBuffer16BitAccess = pdevice->info.gen >= 8;
919 features->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
920 features->storagePushConstant16 = pdevice->info.gen >= 8;
921 features->storageInputOutput16 = false;
922 break;
923 }
924
925 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_ADDRESS_FEATURES_EXT: {
926 VkPhysicalDeviceBufferAddressFeaturesEXT *features = (void *)ext;
927 features->bufferDeviceAddress = pdevice->use_softpin &&
928 pdevice->info.gen >= 8;
929 features->bufferDeviceAddressCaptureReplay = false;
930 features->bufferDeviceAddressMultiDevice = false;
931 break;
932 }
933
934 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
935 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
936 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
937 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
938
939 features->conditionalRendering = pdevice->info.gen >= 8 ||
940 pdevice->info.is_haswell;
941 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
942 pdevice->info.is_haswell;
943 break;
944 }
945
946 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
947 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
948 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
949 features->depthClipEnable = VK_TRUE;
950 break;
951 }
952
953 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
954 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
955 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
956 features->inlineUniformBlock = true;
957 features->descriptorBindingInlineUniformBlockUpdateAfterBind = false;
958 break;
959 }
960
961 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
962 VkPhysicalDeviceMultiviewFeatures *features =
963 (VkPhysicalDeviceMultiviewFeatures *)ext;
964 features->multiview = true;
965 features->multiviewGeometryShader = true;
966 features->multiviewTessellationShader = true;
967 break;
968 }
969
970 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
971 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
972 features->protectedMemory = VK_FALSE;
973 break;
974 }
975
976 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
977 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
978 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
979 features->samplerYcbcrConversion = true;
980 break;
981 }
982
983 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
984 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
985 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
986 features->scalarBlockLayout = true;
987 break;
988 }
989
990 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES: {
991 VkPhysicalDeviceShaderDrawParameterFeatures *features = (void *)ext;
992 features->shaderDrawParameters = true;
993 break;
994 }
995
996 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES: {
997 VkPhysicalDeviceVariablePointerFeatures *features = (void *)ext;
998 features->variablePointersStorageBuffer = true;
999 features->variablePointers = true;
1000 break;
1001 }
1002
1003 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1004 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1005 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1006 features->transformFeedback = VK_TRUE;
1007 features->geometryStreams = VK_TRUE;
1008 break;
1009 }
1010
1011 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1012 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1013 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1014 features->vertexAttributeInstanceRateDivisor = VK_TRUE;
1015 features->vertexAttributeInstanceRateZeroDivisor = VK_TRUE;
1016 break;
1017 }
1018
1019 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1020 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1021 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1022 features->ycbcrImageArrays = VK_TRUE;
1023 break;
1024 }
1025
1026 default:
1027 anv_debug_ignored_stype(ext->sType);
1028 break;
1029 }
1030 }
1031 }
1032
1033 void anv_GetPhysicalDeviceProperties(
1034 VkPhysicalDevice physicalDevice,
1035 VkPhysicalDeviceProperties* pProperties)
1036 {
1037 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1038 const struct gen_device_info *devinfo = &pdevice->info;
1039
1040 /* See assertions made when programming the buffer surface state. */
1041 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1042 (1ul << 30) : (1ul << 27);
1043
1044 const uint32_t max_samplers = (devinfo->gen >= 8 || devinfo->is_haswell) ?
1045 128 : 16;
1046
1047 const uint32_t max_images = devinfo->gen < 9 ? MAX_GEN8_IMAGES : MAX_IMAGES;
1048
1049 VkSampleCountFlags sample_counts =
1050 isl_device_get_sample_counts(&pdevice->isl_dev);
1051
1052
1053 VkPhysicalDeviceLimits limits = {
1054 .maxImageDimension1D = (1 << 14),
1055 .maxImageDimension2D = (1 << 14),
1056 .maxImageDimension3D = (1 << 11),
1057 .maxImageDimensionCube = (1 << 14),
1058 .maxImageArrayLayers = (1 << 11),
1059 .maxTexelBufferElements = 128 * 1024 * 1024,
1060 .maxUniformBufferRange = (1ul << 27),
1061 .maxStorageBufferRange = max_raw_buffer_sz,
1062 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1063 .maxMemoryAllocationCount = UINT32_MAX,
1064 .maxSamplerAllocationCount = 64 * 1024,
1065 .bufferImageGranularity = 64, /* A cache line */
1066 .sparseAddressSpaceSize = 0,
1067 .maxBoundDescriptorSets = MAX_SETS,
1068 .maxPerStageDescriptorSamplers = max_samplers,
1069 .maxPerStageDescriptorUniformBuffers = 64,
1070 .maxPerStageDescriptorStorageBuffers = 64,
1071 .maxPerStageDescriptorSampledImages = max_samplers,
1072 .maxPerStageDescriptorStorageImages = max_images,
1073 .maxPerStageDescriptorInputAttachments = 64,
1074 .maxPerStageResources = 250,
1075 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1076 .maxDescriptorSetUniformBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
1077 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1078 .maxDescriptorSetStorageBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
1079 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1080 .maxDescriptorSetSampledImages = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSampledImages */
1081 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1082 .maxDescriptorSetInputAttachments = 256,
1083 .maxVertexInputAttributes = MAX_VBS,
1084 .maxVertexInputBindings = MAX_VBS,
1085 .maxVertexInputAttributeOffset = 2047,
1086 .maxVertexInputBindingStride = 2048,
1087 .maxVertexOutputComponents = 128,
1088 .maxTessellationGenerationLevel = 64,
1089 .maxTessellationPatchSize = 32,
1090 .maxTessellationControlPerVertexInputComponents = 128,
1091 .maxTessellationControlPerVertexOutputComponents = 128,
1092 .maxTessellationControlPerPatchOutputComponents = 128,
1093 .maxTessellationControlTotalOutputComponents = 2048,
1094 .maxTessellationEvaluationInputComponents = 128,
1095 .maxTessellationEvaluationOutputComponents = 128,
1096 .maxGeometryShaderInvocations = 32,
1097 .maxGeometryInputComponents = 64,
1098 .maxGeometryOutputComponents = 128,
1099 .maxGeometryOutputVertices = 256,
1100 .maxGeometryTotalOutputComponents = 1024,
1101 .maxFragmentInputComponents = 112, /* 128 components - (POS, PSIZ, CLIP_DIST0, CLIP_DIST1) */
1102 .maxFragmentOutputAttachments = 8,
1103 .maxFragmentDualSrcAttachments = 1,
1104 .maxFragmentCombinedOutputResources = 8,
1105 .maxComputeSharedMemorySize = 32768,
1106 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1107 .maxComputeWorkGroupInvocations = 16 * devinfo->max_cs_threads,
1108 .maxComputeWorkGroupSize = {
1109 16 * devinfo->max_cs_threads,
1110 16 * devinfo->max_cs_threads,
1111 16 * devinfo->max_cs_threads,
1112 },
1113 .subPixelPrecisionBits = 8,
1114 .subTexelPrecisionBits = 4 /* FIXME */,
1115 .mipmapPrecisionBits = 4 /* FIXME */,
1116 .maxDrawIndexedIndexValue = UINT32_MAX,
1117 .maxDrawIndirectCount = UINT32_MAX,
1118 .maxSamplerLodBias = 16,
1119 .maxSamplerAnisotropy = 16,
1120 .maxViewports = MAX_VIEWPORTS,
1121 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1122 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1123 .viewportSubPixelBits = 13, /* We take a float? */
1124 .minMemoryMapAlignment = 4096, /* A page */
1125 .minTexelBufferOffsetAlignment = 1,
1126 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1127 .minUniformBufferOffsetAlignment = 32,
1128 .minStorageBufferOffsetAlignment = 4,
1129 .minTexelOffset = -8,
1130 .maxTexelOffset = 7,
1131 .minTexelGatherOffset = -32,
1132 .maxTexelGatherOffset = 31,
1133 .minInterpolationOffset = -0.5,
1134 .maxInterpolationOffset = 0.4375,
1135 .subPixelInterpolationOffsetBits = 4,
1136 .maxFramebufferWidth = (1 << 14),
1137 .maxFramebufferHeight = (1 << 14),
1138 .maxFramebufferLayers = (1 << 11),
1139 .framebufferColorSampleCounts = sample_counts,
1140 .framebufferDepthSampleCounts = sample_counts,
1141 .framebufferStencilSampleCounts = sample_counts,
1142 .framebufferNoAttachmentsSampleCounts = sample_counts,
1143 .maxColorAttachments = MAX_RTS,
1144 .sampledImageColorSampleCounts = sample_counts,
1145 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1146 .sampledImageDepthSampleCounts = sample_counts,
1147 .sampledImageStencilSampleCounts = sample_counts,
1148 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1149 .maxSampleMaskWords = 1,
1150 .timestampComputeAndGraphics = false,
1151 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1152 .maxClipDistances = 8,
1153 .maxCullDistances = 8,
1154 .maxCombinedClipAndCullDistances = 8,
1155 .discreteQueuePriorities = 2,
1156 .pointSizeRange = { 0.125, 255.875 },
1157 .lineWidthRange = { 0.0, 7.9921875 },
1158 .pointSizeGranularity = (1.0 / 8.0),
1159 .lineWidthGranularity = (1.0 / 128.0),
1160 .strictLines = false, /* FINISHME */
1161 .standardSampleLocations = true,
1162 .optimalBufferCopyOffsetAlignment = 128,
1163 .optimalBufferCopyRowPitchAlignment = 128,
1164 .nonCoherentAtomSize = 64,
1165 };
1166
1167 *pProperties = (VkPhysicalDeviceProperties) {
1168 .apiVersion = anv_physical_device_api_version(pdevice),
1169 .driverVersion = vk_get_driver_version(),
1170 .vendorID = 0x8086,
1171 .deviceID = pdevice->chipset_id,
1172 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1173 .limits = limits,
1174 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1175 };
1176
1177 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1178 "%s", pdevice->name);
1179 memcpy(pProperties->pipelineCacheUUID,
1180 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1181 }
1182
1183 void anv_GetPhysicalDeviceProperties2(
1184 VkPhysicalDevice physicalDevice,
1185 VkPhysicalDeviceProperties2* pProperties)
1186 {
1187 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1188
1189 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1190
1191 vk_foreach_struct(ext, pProperties->pNext) {
1192 switch (ext->sType) {
1193 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1194 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *props =
1195 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1196
1197 /* We support all of the depth resolve modes */
1198 props->supportedDepthResolveModes =
1199 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1200 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1201 VK_RESOLVE_MODE_MIN_BIT_KHR |
1202 VK_RESOLVE_MODE_MAX_BIT_KHR;
1203
1204 /* Average doesn't make sense for stencil so we don't support that */
1205 props->supportedStencilResolveModes =
1206 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1207 if (pdevice->info.gen >= 8) {
1208 /* The advanced stencil resolve modes currently require stencil
1209 * sampling be supported by the hardware.
1210 */
1211 props->supportedStencilResolveModes |=
1212 VK_RESOLVE_MODE_MIN_BIT_KHR |
1213 VK_RESOLVE_MODE_MAX_BIT_KHR;
1214 }
1215
1216 props->independentResolveNone = VK_TRUE;
1217 props->independentResolve = VK_TRUE;
1218 break;
1219 }
1220
1221 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1222 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1223 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1224
1225 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1226 util_snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1227 "Intel open-source Mesa driver");
1228
1229 util_snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1230 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1231
1232 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1233 .major = 1,
1234 .minor = 1,
1235 .subminor = 2,
1236 .patch = 0,
1237 };
1238 break;
1239 }
1240
1241 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1242 VkPhysicalDeviceIDProperties *id_props =
1243 (VkPhysicalDeviceIDProperties *)ext;
1244 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1245 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1246 /* The LUID is for Windows. */
1247 id_props->deviceLUIDValid = false;
1248 break;
1249 }
1250
1251 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1252 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1253 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1254 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1255 props->maxPerStageDescriptorInlineUniformBlocks =
1256 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1257 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1258 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1259 props->maxDescriptorSetInlineUniformBlocks =
1260 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1261 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1262 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1263 break;
1264 }
1265
1266 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1267 VkPhysicalDeviceMaintenance3Properties *props =
1268 (VkPhysicalDeviceMaintenance3Properties *)ext;
1269 /* This value doesn't matter for us today as our per-stage
1270 * descriptors are the real limit.
1271 */
1272 props->maxPerSetDescriptors = 1024;
1273 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1274 break;
1275 }
1276
1277 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1278 VkPhysicalDeviceMultiviewProperties *properties =
1279 (VkPhysicalDeviceMultiviewProperties *)ext;
1280 properties->maxMultiviewViewCount = 16;
1281 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1282 break;
1283 }
1284
1285 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1286 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1287 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1288 properties->pciDomain = pdevice->pci_info.domain;
1289 properties->pciBus = pdevice->pci_info.bus;
1290 properties->pciDevice = pdevice->pci_info.device;
1291 properties->pciFunction = pdevice->pci_info.function;
1292 break;
1293 }
1294
1295 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1296 VkPhysicalDevicePointClippingProperties *properties =
1297 (VkPhysicalDevicePointClippingProperties *) ext;
1298 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
1299 anv_finishme("Implement pop-free point clipping");
1300 break;
1301 }
1302
1303 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1304 VkPhysicalDeviceProtectedMemoryProperties *props =
1305 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1306 props->protectedNoFault = false;
1307 break;
1308 }
1309
1310 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1311 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1312 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1313
1314 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1315 break;
1316 }
1317
1318 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1319 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1320 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1321 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1322 properties->filterMinmaxSingleComponentFormats = true;
1323 break;
1324 }
1325
1326 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1327 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1328
1329 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1330
1331 VkShaderStageFlags scalar_stages = 0;
1332 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1333 if (pdevice->compiler->scalar_stage[stage])
1334 scalar_stages |= mesa_to_vk_shader_stage(stage);
1335 }
1336 properties->supportedStages = scalar_stages;
1337
1338 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1339 VK_SUBGROUP_FEATURE_VOTE_BIT |
1340 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1341 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1342 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1343 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1344 VK_SUBGROUP_FEATURE_CLUSTERED_BIT |
1345 VK_SUBGROUP_FEATURE_QUAD_BIT;
1346 properties->quadOperationsInAllStages = VK_TRUE;
1347 break;
1348 }
1349
1350 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1351 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1352 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1353
1354 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1355 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1356 props->maxTransformFeedbackBufferSize = (1ull << 32);
1357 props->maxTransformFeedbackStreamDataSize = 128 * 4;
1358 props->maxTransformFeedbackBufferDataSize = 128 * 4;
1359 props->maxTransformFeedbackBufferDataStride = 2048;
1360 props->transformFeedbackQueries = VK_TRUE;
1361 props->transformFeedbackStreamsLinesTriangles = VK_FALSE;
1362 props->transformFeedbackRasterizationStreamSelect = VK_FALSE;
1363 props->transformFeedbackDraw = VK_TRUE;
1364 break;
1365 }
1366
1367 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1368 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1369 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1370 /* We have to restrict this a bit for multiview */
1371 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1372 break;
1373 }
1374
1375 default:
1376 anv_debug_ignored_stype(ext->sType);
1377 break;
1378 }
1379 }
1380 }
1381
1382 /* We support exactly one queue family. */
1383 static const VkQueueFamilyProperties
1384 anv_queue_family_properties = {
1385 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1386 VK_QUEUE_COMPUTE_BIT |
1387 VK_QUEUE_TRANSFER_BIT,
1388 .queueCount = 1,
1389 .timestampValidBits = 36, /* XXX: Real value here */
1390 .minImageTransferGranularity = { 1, 1, 1 },
1391 };
1392
1393 void anv_GetPhysicalDeviceQueueFamilyProperties(
1394 VkPhysicalDevice physicalDevice,
1395 uint32_t* pCount,
1396 VkQueueFamilyProperties* pQueueFamilyProperties)
1397 {
1398 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1399
1400 vk_outarray_append(&out, p) {
1401 *p = anv_queue_family_properties;
1402 }
1403 }
1404
1405 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1406 VkPhysicalDevice physicalDevice,
1407 uint32_t* pQueueFamilyPropertyCount,
1408 VkQueueFamilyProperties2* pQueueFamilyProperties)
1409 {
1410
1411 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1412
1413 vk_outarray_append(&out, p) {
1414 p->queueFamilyProperties = anv_queue_family_properties;
1415
1416 vk_foreach_struct(s, p->pNext) {
1417 anv_debug_ignored_stype(s->sType);
1418 }
1419 }
1420 }
1421
1422 void anv_GetPhysicalDeviceMemoryProperties(
1423 VkPhysicalDevice physicalDevice,
1424 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1425 {
1426 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1427
1428 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1429 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1430 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1431 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1432 .heapIndex = physical_device->memory.types[i].heapIndex,
1433 };
1434 }
1435
1436 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1437 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1438 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1439 .size = physical_device->memory.heaps[i].size,
1440 .flags = physical_device->memory.heaps[i].flags,
1441 };
1442 }
1443 }
1444
1445 void anv_GetPhysicalDeviceMemoryProperties2(
1446 VkPhysicalDevice physicalDevice,
1447 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1448 {
1449 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1450 &pMemoryProperties->memoryProperties);
1451
1452 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1453 switch (ext->sType) {
1454 default:
1455 anv_debug_ignored_stype(ext->sType);
1456 break;
1457 }
1458 }
1459 }
1460
1461 void
1462 anv_GetDeviceGroupPeerMemoryFeatures(
1463 VkDevice device,
1464 uint32_t heapIndex,
1465 uint32_t localDeviceIndex,
1466 uint32_t remoteDeviceIndex,
1467 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1468 {
1469 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1470 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1471 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1472 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1473 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1474 }
1475
1476 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1477 VkInstance _instance,
1478 const char* pName)
1479 {
1480 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1481
1482 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1483 * when we have to return valid function pointers, NULL, or it's left
1484 * undefined. See the table for exact details.
1485 */
1486 if (pName == NULL)
1487 return NULL;
1488
1489 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1490 if (strcmp(pName, "vk" #entrypoint) == 0) \
1491 return (PFN_vkVoidFunction)anv_##entrypoint
1492
1493 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1494 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1495 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1496 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1497
1498 #undef LOOKUP_ANV_ENTRYPOINT
1499
1500 if (instance == NULL)
1501 return NULL;
1502
1503 int idx = anv_get_instance_entrypoint_index(pName);
1504 if (idx >= 0)
1505 return instance->dispatch.entrypoints[idx];
1506
1507 idx = anv_get_device_entrypoint_index(pName);
1508 if (idx >= 0)
1509 return instance->device_dispatch.entrypoints[idx];
1510
1511 return NULL;
1512 }
1513
1514 /* With version 1+ of the loader interface the ICD should expose
1515 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1516 */
1517 PUBLIC
1518 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1519 VkInstance instance,
1520 const char* pName);
1521
1522 PUBLIC
1523 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1524 VkInstance instance,
1525 const char* pName)
1526 {
1527 return anv_GetInstanceProcAddr(instance, pName);
1528 }
1529
1530 PFN_vkVoidFunction anv_GetDeviceProcAddr(
1531 VkDevice _device,
1532 const char* pName)
1533 {
1534 ANV_FROM_HANDLE(anv_device, device, _device);
1535
1536 if (!device || !pName)
1537 return NULL;
1538
1539 int idx = anv_get_device_entrypoint_index(pName);
1540 if (idx < 0)
1541 return NULL;
1542
1543 return device->dispatch.entrypoints[idx];
1544 }
1545
1546 VkResult
1547 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
1548 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
1549 const VkAllocationCallbacks* pAllocator,
1550 VkDebugReportCallbackEXT* pCallback)
1551 {
1552 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1553 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
1554 pCreateInfo, pAllocator, &instance->alloc,
1555 pCallback);
1556 }
1557
1558 void
1559 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
1560 VkDebugReportCallbackEXT _callback,
1561 const VkAllocationCallbacks* pAllocator)
1562 {
1563 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1564 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
1565 _callback, pAllocator, &instance->alloc);
1566 }
1567
1568 void
1569 anv_DebugReportMessageEXT(VkInstance _instance,
1570 VkDebugReportFlagsEXT flags,
1571 VkDebugReportObjectTypeEXT objectType,
1572 uint64_t object,
1573 size_t location,
1574 int32_t messageCode,
1575 const char* pLayerPrefix,
1576 const char* pMessage)
1577 {
1578 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1579 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
1580 object, location, messageCode, pLayerPrefix, pMessage);
1581 }
1582
1583 static void
1584 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
1585 {
1586 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1587 queue->device = device;
1588 queue->flags = 0;
1589 }
1590
1591 static void
1592 anv_queue_finish(struct anv_queue *queue)
1593 {
1594 }
1595
1596 static struct anv_state
1597 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
1598 {
1599 struct anv_state state;
1600
1601 state = anv_state_pool_alloc(pool, size, align);
1602 memcpy(state.map, p, size);
1603
1604 return state;
1605 }
1606
1607 struct gen8_border_color {
1608 union {
1609 float float32[4];
1610 uint32_t uint32[4];
1611 };
1612 /* Pad out to 64 bytes */
1613 uint32_t _pad[12];
1614 };
1615
1616 static void
1617 anv_device_init_border_colors(struct anv_device *device)
1618 {
1619 static const struct gen8_border_color border_colors[] = {
1620 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
1621 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
1622 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
1623 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
1624 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
1625 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
1626 };
1627
1628 device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
1629 sizeof(border_colors), 64,
1630 border_colors);
1631 }
1632
1633 static void
1634 anv_device_init_trivial_batch(struct anv_device *device)
1635 {
1636 anv_bo_init_new(&device->trivial_batch_bo, device, 4096);
1637
1638 if (device->instance->physicalDevice.has_exec_async)
1639 device->trivial_batch_bo.flags |= EXEC_OBJECT_ASYNC;
1640
1641 if (device->instance->physicalDevice.use_softpin)
1642 device->trivial_batch_bo.flags |= EXEC_OBJECT_PINNED;
1643
1644 anv_vma_alloc(device, &device->trivial_batch_bo);
1645
1646 void *map = anv_gem_mmap(device, device->trivial_batch_bo.gem_handle,
1647 0, 4096, 0);
1648
1649 struct anv_batch batch = {
1650 .start = map,
1651 .next = map,
1652 .end = map + 4096,
1653 };
1654
1655 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1656 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1657
1658 if (!device->info.has_llc)
1659 gen_clflush_range(map, batch.next - map);
1660
1661 anv_gem_munmap(map, device->trivial_batch_bo.size);
1662 }
1663
1664 VkResult anv_EnumerateDeviceExtensionProperties(
1665 VkPhysicalDevice physicalDevice,
1666 const char* pLayerName,
1667 uint32_t* pPropertyCount,
1668 VkExtensionProperties* pProperties)
1669 {
1670 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1671 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1672
1673 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
1674 if (device->supported_extensions.extensions[i]) {
1675 vk_outarray_append(&out, prop) {
1676 *prop = anv_device_extensions[i];
1677 }
1678 }
1679 }
1680
1681 return vk_outarray_status(&out);
1682 }
1683
1684 static void
1685 anv_device_init_dispatch(struct anv_device *device)
1686 {
1687 const struct anv_device_dispatch_table *genX_table;
1688 switch (device->info.gen) {
1689 case 11:
1690 genX_table = &gen11_device_dispatch_table;
1691 break;
1692 case 10:
1693 genX_table = &gen10_device_dispatch_table;
1694 break;
1695 case 9:
1696 genX_table = &gen9_device_dispatch_table;
1697 break;
1698 case 8:
1699 genX_table = &gen8_device_dispatch_table;
1700 break;
1701 case 7:
1702 if (device->info.is_haswell)
1703 genX_table = &gen75_device_dispatch_table;
1704 else
1705 genX_table = &gen7_device_dispatch_table;
1706 break;
1707 default:
1708 unreachable("unsupported gen\n");
1709 }
1710
1711 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
1712 /* Vulkan requires that entrypoints for extensions which have not been
1713 * enabled must not be advertised.
1714 */
1715 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
1716 &device->instance->enabled_extensions,
1717 &device->enabled_extensions)) {
1718 device->dispatch.entrypoints[i] = NULL;
1719 } else if (genX_table->entrypoints[i]) {
1720 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
1721 } else {
1722 device->dispatch.entrypoints[i] =
1723 anv_device_dispatch_table.entrypoints[i];
1724 }
1725 }
1726 }
1727
1728 static int
1729 vk_priority_to_gen(int priority)
1730 {
1731 switch (priority) {
1732 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
1733 return GEN_CONTEXT_LOW_PRIORITY;
1734 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
1735 return GEN_CONTEXT_MEDIUM_PRIORITY;
1736 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
1737 return GEN_CONTEXT_HIGH_PRIORITY;
1738 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
1739 return GEN_CONTEXT_REALTIME_PRIORITY;
1740 default:
1741 unreachable("Invalid priority");
1742 }
1743 }
1744
1745 static void
1746 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
1747 {
1748 anv_bo_init_new(&device->hiz_clear_bo, device, 4096);
1749
1750 if (device->instance->physicalDevice.has_exec_async)
1751 device->hiz_clear_bo.flags |= EXEC_OBJECT_ASYNC;
1752
1753 if (device->instance->physicalDevice.use_softpin)
1754 device->hiz_clear_bo.flags |= EXEC_OBJECT_PINNED;
1755
1756 anv_vma_alloc(device, &device->hiz_clear_bo);
1757
1758 uint32_t *map = anv_gem_mmap(device, device->hiz_clear_bo.gem_handle,
1759 0, 4096, 0);
1760
1761 union isl_color_value hiz_clear = { .u32 = { 0, } };
1762 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
1763
1764 memcpy(map, hiz_clear.u32, sizeof(hiz_clear.u32));
1765 anv_gem_munmap(map, device->hiz_clear_bo.size);
1766 }
1767
1768 static bool
1769 get_bo_from_pool(struct gen_batch_decode_bo *ret,
1770 struct anv_block_pool *pool,
1771 uint64_t address)
1772 {
1773 for (uint32_t i = 0; i < pool->nbos; i++) {
1774 uint64_t bo_address = pool->bos[i].offset & (~0ull >> 16);
1775 uint32_t bo_size = pool->bos[i].size;
1776 if (address >= bo_address && address < (bo_address + bo_size)) {
1777 *ret = (struct gen_batch_decode_bo) {
1778 .addr = bo_address,
1779 .size = bo_size,
1780 .map = pool->bos[i].map,
1781 };
1782 return true;
1783 }
1784 }
1785 return false;
1786 }
1787
1788 /* Finding a buffer for batch decoding */
1789 static struct gen_batch_decode_bo
1790 decode_get_bo(void *v_batch, uint64_t address)
1791 {
1792 struct anv_device *device = v_batch;
1793 struct gen_batch_decode_bo ret_bo = {};
1794
1795 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
1796 return ret_bo;
1797 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
1798 return ret_bo;
1799 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
1800 return ret_bo;
1801 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
1802 return ret_bo;
1803
1804 if (!device->cmd_buffer_being_decoded)
1805 return (struct gen_batch_decode_bo) { };
1806
1807 struct anv_batch_bo **bo;
1808
1809 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
1810 /* The decoder zeroes out the top 16 bits, so we need to as well */
1811 uint64_t bo_address = (*bo)->bo.offset & (~0ull >> 16);
1812
1813 if (address >= bo_address && address < bo_address + (*bo)->bo.size) {
1814 return (struct gen_batch_decode_bo) {
1815 .addr = bo_address,
1816 .size = (*bo)->bo.size,
1817 .map = (*bo)->bo.map,
1818 };
1819 }
1820 }
1821
1822 return (struct gen_batch_decode_bo) { };
1823 }
1824
1825 VkResult anv_CreateDevice(
1826 VkPhysicalDevice physicalDevice,
1827 const VkDeviceCreateInfo* pCreateInfo,
1828 const VkAllocationCallbacks* pAllocator,
1829 VkDevice* pDevice)
1830 {
1831 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1832 VkResult result;
1833 struct anv_device *device;
1834
1835 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
1836
1837 struct anv_device_extension_table enabled_extensions = { };
1838 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1839 int idx;
1840 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
1841 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
1842 anv_device_extensions[idx].extensionName) == 0)
1843 break;
1844 }
1845
1846 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
1847 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1848
1849 if (!physical_device->supported_extensions.extensions[idx])
1850 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1851
1852 enabled_extensions.extensions[idx] = true;
1853 }
1854
1855 /* Check enabled features */
1856 if (pCreateInfo->pEnabledFeatures) {
1857 VkPhysicalDeviceFeatures supported_features;
1858 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1859 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
1860 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
1861 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1862 for (uint32_t i = 0; i < num_features; i++) {
1863 if (enabled_feature[i] && !supported_feature[i])
1864 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1865 }
1866 }
1867
1868 /* Check requested queues and fail if we are requested to create any
1869 * queues with flags we don't support.
1870 */
1871 assert(pCreateInfo->queueCreateInfoCount > 0);
1872 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1873 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
1874 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
1875 }
1876
1877 /* Check if client specified queue priority. */
1878 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
1879 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
1880 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
1881
1882 VkQueueGlobalPriorityEXT priority =
1883 queue_priority ? queue_priority->globalPriority :
1884 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
1885
1886 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
1887 sizeof(*device), 8,
1888 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1889 if (!device)
1890 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1891
1892 const unsigned decode_flags =
1893 GEN_BATCH_DECODE_FULL |
1894 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
1895 GEN_BATCH_DECODE_OFFSETS |
1896 GEN_BATCH_DECODE_FLOATS;
1897
1898 gen_batch_decode_ctx_init(&device->decoder_ctx,
1899 &physical_device->info,
1900 stderr, decode_flags, NULL,
1901 decode_get_bo, NULL, device);
1902
1903 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1904 device->instance = physical_device->instance;
1905 device->chipset_id = physical_device->chipset_id;
1906 device->no_hw = physical_device->no_hw;
1907 device->_lost = false;
1908
1909 if (pAllocator)
1910 device->alloc = *pAllocator;
1911 else
1912 device->alloc = physical_device->instance->alloc;
1913
1914 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1915 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1916 if (device->fd == -1) {
1917 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1918 goto fail_device;
1919 }
1920
1921 device->context_id = anv_gem_create_context(device);
1922 if (device->context_id == -1) {
1923 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1924 goto fail_fd;
1925 }
1926
1927 if (physical_device->use_softpin) {
1928 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
1929 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1930 goto fail_fd;
1931 }
1932
1933 /* keep the page with address zero out of the allocator */
1934 util_vma_heap_init(&device->vma_lo, LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
1935 device->vma_lo_available =
1936 physical_device->memory.heaps[physical_device->memory.heap_count - 1].size;
1937
1938 /* Leave the last 4GiB out of the high vma range, so that no state base
1939 * address + size can overflow 48 bits. For more information see the
1940 * comment about Wa32bitGeneralStateOffset in anv_allocator.c
1941 */
1942 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
1943 HIGH_HEAP_SIZE);
1944 device->vma_hi_available = physical_device->memory.heap_count == 1 ? 0 :
1945 physical_device->memory.heaps[0].size;
1946 }
1947
1948 /* As per spec, the driver implementation may deny requests to acquire
1949 * a priority above the default priority (MEDIUM) if the caller does not
1950 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1951 * is returned.
1952 */
1953 if (physical_device->has_context_priority) {
1954 int err = anv_gem_set_context_param(device->fd, device->context_id,
1955 I915_CONTEXT_PARAM_PRIORITY,
1956 vk_priority_to_gen(priority));
1957 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
1958 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
1959 goto fail_fd;
1960 }
1961 }
1962
1963 device->info = physical_device->info;
1964 device->isl_dev = physical_device->isl_dev;
1965
1966 /* On Broadwell and later, we can use batch chaining to more efficiently
1967 * implement growing command buffers. Prior to Haswell, the kernel
1968 * command parser gets in the way and we have to fall back to growing
1969 * the batch.
1970 */
1971 device->can_chain_batches = device->info.gen >= 8;
1972
1973 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1974 pCreateInfo->pEnabledFeatures->robustBufferAccess;
1975 device->enabled_extensions = enabled_extensions;
1976
1977 anv_device_init_dispatch(device);
1978
1979 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
1980 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1981 goto fail_context_id;
1982 }
1983
1984 pthread_condattr_t condattr;
1985 if (pthread_condattr_init(&condattr) != 0) {
1986 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1987 goto fail_mutex;
1988 }
1989 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
1990 pthread_condattr_destroy(&condattr);
1991 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1992 goto fail_mutex;
1993 }
1994 if (pthread_cond_init(&device->queue_submit, NULL) != 0) {
1995 pthread_condattr_destroy(&condattr);
1996 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1997 goto fail_mutex;
1998 }
1999 pthread_condattr_destroy(&condattr);
2000
2001 uint64_t bo_flags =
2002 (physical_device->supports_48bit_addresses ? EXEC_OBJECT_SUPPORTS_48B_ADDRESS : 0) |
2003 (physical_device->has_exec_async ? EXEC_OBJECT_ASYNC : 0) |
2004 (physical_device->has_exec_capture ? EXEC_OBJECT_CAPTURE : 0) |
2005 (physical_device->use_softpin ? EXEC_OBJECT_PINNED : 0);
2006
2007 anv_bo_pool_init(&device->batch_bo_pool, device, bo_flags);
2008
2009 result = anv_bo_cache_init(&device->bo_cache);
2010 if (result != VK_SUCCESS)
2011 goto fail_batch_bo_pool;
2012
2013 if (!physical_device->use_softpin)
2014 bo_flags &= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2015
2016 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2017 DYNAMIC_STATE_POOL_MIN_ADDRESS,
2018 16384,
2019 bo_flags);
2020 if (result != VK_SUCCESS)
2021 goto fail_bo_cache;
2022
2023 result = anv_state_pool_init(&device->instruction_state_pool, device,
2024 INSTRUCTION_STATE_POOL_MIN_ADDRESS,
2025 16384,
2026 bo_flags);
2027 if (result != VK_SUCCESS)
2028 goto fail_dynamic_state_pool;
2029
2030 result = anv_state_pool_init(&device->surface_state_pool, device,
2031 SURFACE_STATE_POOL_MIN_ADDRESS,
2032 4096,
2033 bo_flags);
2034 if (result != VK_SUCCESS)
2035 goto fail_instruction_state_pool;
2036
2037 if (physical_device->use_softpin) {
2038 result = anv_state_pool_init(&device->binding_table_pool, device,
2039 BINDING_TABLE_POOL_MIN_ADDRESS,
2040 4096,
2041 bo_flags);
2042 if (result != VK_SUCCESS)
2043 goto fail_surface_state_pool;
2044 }
2045
2046 result = anv_bo_init_new(&device->workaround_bo, device, 1024);
2047 if (result != VK_SUCCESS)
2048 goto fail_binding_table_pool;
2049
2050 if (physical_device->use_softpin)
2051 device->workaround_bo.flags |= EXEC_OBJECT_PINNED;
2052
2053 if (!anv_vma_alloc(device, &device->workaround_bo))
2054 goto fail_workaround_bo;
2055
2056 anv_device_init_trivial_batch(device);
2057
2058 if (device->info.gen >= 10)
2059 anv_device_init_hiz_clear_value_bo(device);
2060
2061 if (physical_device->use_softpin)
2062 device->pinned_buffers = _mesa_pointer_set_create(NULL);
2063
2064 anv_scratch_pool_init(device, &device->scratch_pool);
2065
2066 anv_queue_init(device, &device->queue);
2067
2068 switch (device->info.gen) {
2069 case 7:
2070 if (!device->info.is_haswell)
2071 result = gen7_init_device_state(device);
2072 else
2073 result = gen75_init_device_state(device);
2074 break;
2075 case 8:
2076 result = gen8_init_device_state(device);
2077 break;
2078 case 9:
2079 result = gen9_init_device_state(device);
2080 break;
2081 case 10:
2082 result = gen10_init_device_state(device);
2083 break;
2084 case 11:
2085 result = gen11_init_device_state(device);
2086 break;
2087 default:
2088 /* Shouldn't get here as we don't create physical devices for any other
2089 * gens. */
2090 unreachable("unhandled gen");
2091 }
2092 if (result != VK_SUCCESS)
2093 goto fail_workaround_bo;
2094
2095 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2096
2097 anv_device_init_blorp(device);
2098
2099 anv_device_init_border_colors(device);
2100
2101 *pDevice = anv_device_to_handle(device);
2102
2103 return VK_SUCCESS;
2104
2105 fail_workaround_bo:
2106 anv_queue_finish(&device->queue);
2107 anv_scratch_pool_finish(device, &device->scratch_pool);
2108 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
2109 anv_gem_close(device, device->workaround_bo.gem_handle);
2110 fail_binding_table_pool:
2111 if (physical_device->use_softpin)
2112 anv_state_pool_finish(&device->binding_table_pool);
2113 fail_surface_state_pool:
2114 anv_state_pool_finish(&device->surface_state_pool);
2115 fail_instruction_state_pool:
2116 anv_state_pool_finish(&device->instruction_state_pool);
2117 fail_dynamic_state_pool:
2118 anv_state_pool_finish(&device->dynamic_state_pool);
2119 fail_bo_cache:
2120 anv_bo_cache_finish(&device->bo_cache);
2121 fail_batch_bo_pool:
2122 anv_bo_pool_finish(&device->batch_bo_pool);
2123 pthread_cond_destroy(&device->queue_submit);
2124 fail_mutex:
2125 pthread_mutex_destroy(&device->mutex);
2126 fail_context_id:
2127 anv_gem_destroy_context(device, device->context_id);
2128 fail_fd:
2129 close(device->fd);
2130 fail_device:
2131 vk_free(&device->alloc, device);
2132
2133 return result;
2134 }
2135
2136 void anv_DestroyDevice(
2137 VkDevice _device,
2138 const VkAllocationCallbacks* pAllocator)
2139 {
2140 ANV_FROM_HANDLE(anv_device, device, _device);
2141 struct anv_physical_device *physical_device;
2142
2143 if (!device)
2144 return;
2145
2146 physical_device = &device->instance->physicalDevice;
2147
2148 anv_device_finish_blorp(device);
2149
2150 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2151
2152 anv_queue_finish(&device->queue);
2153
2154 if (physical_device->use_softpin)
2155 _mesa_set_destroy(device->pinned_buffers, NULL);
2156
2157 #ifdef HAVE_VALGRIND
2158 /* We only need to free these to prevent valgrind errors. The backing
2159 * BO will go away in a couple of lines so we don't actually leak.
2160 */
2161 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2162 #endif
2163
2164 anv_scratch_pool_finish(device, &device->scratch_pool);
2165
2166 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
2167 anv_vma_free(device, &device->workaround_bo);
2168 anv_gem_close(device, device->workaround_bo.gem_handle);
2169
2170 anv_vma_free(device, &device->trivial_batch_bo);
2171 anv_gem_close(device, device->trivial_batch_bo.gem_handle);
2172 if (device->info.gen >= 10)
2173 anv_gem_close(device, device->hiz_clear_bo.gem_handle);
2174
2175 if (physical_device->use_softpin)
2176 anv_state_pool_finish(&device->binding_table_pool);
2177 anv_state_pool_finish(&device->surface_state_pool);
2178 anv_state_pool_finish(&device->instruction_state_pool);
2179 anv_state_pool_finish(&device->dynamic_state_pool);
2180
2181 anv_bo_cache_finish(&device->bo_cache);
2182
2183 anv_bo_pool_finish(&device->batch_bo_pool);
2184
2185 pthread_cond_destroy(&device->queue_submit);
2186 pthread_mutex_destroy(&device->mutex);
2187
2188 anv_gem_destroy_context(device, device->context_id);
2189
2190 gen_batch_decode_ctx_finish(&device->decoder_ctx);
2191
2192 close(device->fd);
2193
2194 vk_free(&device->alloc, device);
2195 }
2196
2197 VkResult anv_EnumerateInstanceLayerProperties(
2198 uint32_t* pPropertyCount,
2199 VkLayerProperties* pProperties)
2200 {
2201 if (pProperties == NULL) {
2202 *pPropertyCount = 0;
2203 return VK_SUCCESS;
2204 }
2205
2206 /* None supported at this time */
2207 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2208 }
2209
2210 VkResult anv_EnumerateDeviceLayerProperties(
2211 VkPhysicalDevice physicalDevice,
2212 uint32_t* pPropertyCount,
2213 VkLayerProperties* pProperties)
2214 {
2215 if (pProperties == NULL) {
2216 *pPropertyCount = 0;
2217 return VK_SUCCESS;
2218 }
2219
2220 /* None supported at this time */
2221 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2222 }
2223
2224 void anv_GetDeviceQueue(
2225 VkDevice _device,
2226 uint32_t queueNodeIndex,
2227 uint32_t queueIndex,
2228 VkQueue* pQueue)
2229 {
2230 ANV_FROM_HANDLE(anv_device, device, _device);
2231
2232 assert(queueIndex == 0);
2233
2234 *pQueue = anv_queue_to_handle(&device->queue);
2235 }
2236
2237 void anv_GetDeviceQueue2(
2238 VkDevice _device,
2239 const VkDeviceQueueInfo2* pQueueInfo,
2240 VkQueue* pQueue)
2241 {
2242 ANV_FROM_HANDLE(anv_device, device, _device);
2243
2244 assert(pQueueInfo->queueIndex == 0);
2245
2246 if (pQueueInfo->flags == device->queue.flags)
2247 *pQueue = anv_queue_to_handle(&device->queue);
2248 else
2249 *pQueue = NULL;
2250 }
2251
2252 VkResult
2253 _anv_device_set_lost(struct anv_device *device,
2254 const char *file, int line,
2255 const char *msg, ...)
2256 {
2257 VkResult err;
2258 va_list ap;
2259
2260 device->_lost = true;
2261
2262 va_start(ap, msg);
2263 err = __vk_errorv(device->instance, device,
2264 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2265 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2266 va_end(ap);
2267
2268 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2269 abort();
2270
2271 return err;
2272 }
2273
2274 VkResult
2275 anv_device_query_status(struct anv_device *device)
2276 {
2277 /* This isn't likely as most of the callers of this function already check
2278 * for it. However, it doesn't hurt to check and it potentially lets us
2279 * avoid an ioctl.
2280 */
2281 if (anv_device_is_lost(device))
2282 return VK_ERROR_DEVICE_LOST;
2283
2284 uint32_t active, pending;
2285 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2286 if (ret == -1) {
2287 /* We don't know the real error. */
2288 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2289 }
2290
2291 if (active) {
2292 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2293 } else if (pending) {
2294 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2295 }
2296
2297 return VK_SUCCESS;
2298 }
2299
2300 VkResult
2301 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2302 {
2303 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2304 * Other usages of the BO (such as on different hardware) will not be
2305 * flagged as "busy" by this ioctl. Use with care.
2306 */
2307 int ret = anv_gem_busy(device, bo->gem_handle);
2308 if (ret == 1) {
2309 return VK_NOT_READY;
2310 } else if (ret == -1) {
2311 /* We don't know the real error. */
2312 return anv_device_set_lost(device, "gem wait failed: %m");
2313 }
2314
2315 /* Query for device status after the busy call. If the BO we're checking
2316 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2317 * client because it clearly doesn't have valid data. Yes, this most
2318 * likely means an ioctl, but we just did an ioctl to query the busy status
2319 * so it's no great loss.
2320 */
2321 return anv_device_query_status(device);
2322 }
2323
2324 VkResult
2325 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2326 int64_t timeout)
2327 {
2328 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2329 if (ret == -1 && errno == ETIME) {
2330 return VK_TIMEOUT;
2331 } else if (ret == -1) {
2332 /* We don't know the real error. */
2333 return anv_device_set_lost(device, "gem wait failed: %m");
2334 }
2335
2336 /* Query for device status after the wait. If the BO we're waiting on got
2337 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2338 * because it clearly doesn't have valid data. Yes, this most likely means
2339 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2340 */
2341 return anv_device_query_status(device);
2342 }
2343
2344 VkResult anv_DeviceWaitIdle(
2345 VkDevice _device)
2346 {
2347 ANV_FROM_HANDLE(anv_device, device, _device);
2348 if (anv_device_is_lost(device))
2349 return VK_ERROR_DEVICE_LOST;
2350
2351 struct anv_batch batch;
2352
2353 uint32_t cmds[8];
2354 batch.start = batch.next = cmds;
2355 batch.end = (void *) cmds + sizeof(cmds);
2356
2357 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2358 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2359
2360 return anv_device_submit_simple_batch(device, &batch);
2361 }
2362
2363 bool
2364 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo)
2365 {
2366 if (!(bo->flags & EXEC_OBJECT_PINNED))
2367 return true;
2368
2369 pthread_mutex_lock(&device->vma_mutex);
2370
2371 bo->offset = 0;
2372
2373 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS &&
2374 device->vma_hi_available >= bo->size) {
2375 uint64_t addr = util_vma_heap_alloc(&device->vma_hi, bo->size, 4096);
2376 if (addr) {
2377 bo->offset = gen_canonical_address(addr);
2378 assert(addr == gen_48b_address(bo->offset));
2379 device->vma_hi_available -= bo->size;
2380 }
2381 }
2382
2383 if (bo->offset == 0 && device->vma_lo_available >= bo->size) {
2384 uint64_t addr = util_vma_heap_alloc(&device->vma_lo, bo->size, 4096);
2385 if (addr) {
2386 bo->offset = gen_canonical_address(addr);
2387 assert(addr == gen_48b_address(bo->offset));
2388 device->vma_lo_available -= bo->size;
2389 }
2390 }
2391
2392 pthread_mutex_unlock(&device->vma_mutex);
2393
2394 return bo->offset != 0;
2395 }
2396
2397 void
2398 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
2399 {
2400 if (!(bo->flags & EXEC_OBJECT_PINNED))
2401 return;
2402
2403 const uint64_t addr_48b = gen_48b_address(bo->offset);
2404
2405 pthread_mutex_lock(&device->vma_mutex);
2406
2407 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
2408 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
2409 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
2410 device->vma_lo_available += bo->size;
2411 } else {
2412 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS &&
2413 addr_48b <= HIGH_HEAP_MAX_ADDRESS);
2414 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
2415 device->vma_hi_available += bo->size;
2416 }
2417
2418 pthread_mutex_unlock(&device->vma_mutex);
2419
2420 bo->offset = 0;
2421 }
2422
2423 VkResult
2424 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
2425 {
2426 uint32_t gem_handle = anv_gem_create(device, size);
2427 if (!gem_handle)
2428 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
2429
2430 anv_bo_init(bo, gem_handle, size);
2431
2432 return VK_SUCCESS;
2433 }
2434
2435 VkResult anv_AllocateMemory(
2436 VkDevice _device,
2437 const VkMemoryAllocateInfo* pAllocateInfo,
2438 const VkAllocationCallbacks* pAllocator,
2439 VkDeviceMemory* pMem)
2440 {
2441 ANV_FROM_HANDLE(anv_device, device, _device);
2442 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2443 struct anv_device_memory *mem;
2444 VkResult result = VK_SUCCESS;
2445
2446 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
2447
2448 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2449 assert(pAllocateInfo->allocationSize > 0);
2450
2451 if (pAllocateInfo->allocationSize > MAX_MEMORY_ALLOCATION_SIZE)
2452 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
2453
2454 /* FINISHME: Fail if allocation request exceeds heap size. */
2455
2456 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
2457 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2458 if (mem == NULL)
2459 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2460
2461 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
2462 mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
2463 mem->map = NULL;
2464 mem->map_size = 0;
2465 mem->ahw = NULL;
2466
2467 uint64_t bo_flags = 0;
2468
2469 assert(mem->type->heapIndex < pdevice->memory.heap_count);
2470 if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
2471 bo_flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2472
2473 const struct wsi_memory_allocate_info *wsi_info =
2474 vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
2475 if (wsi_info && wsi_info->implicit_sync) {
2476 /* We need to set the WRITE flag on window system buffers so that GEM
2477 * will know we're writing to them and synchronize uses on other rings
2478 * (eg if the display server uses the blitter ring).
2479 */
2480 bo_flags |= EXEC_OBJECT_WRITE;
2481 } else if (pdevice->has_exec_async) {
2482 bo_flags |= EXEC_OBJECT_ASYNC;
2483 }
2484
2485 if (pdevice->use_softpin)
2486 bo_flags |= EXEC_OBJECT_PINNED;
2487
2488 const VkExportMemoryAllocateInfo *export_info =
2489 vk_find_struct_const(pAllocateInfo->pNext, EXPORT_MEMORY_ALLOCATE_INFO);
2490
2491 /* Check if we need to support Android HW buffer export. If so,
2492 * create AHardwareBuffer and import memory from it.
2493 */
2494 bool android_export = false;
2495 if (export_info && export_info->handleTypes &
2496 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
2497 android_export = true;
2498
2499 /* Android memory import. */
2500 const struct VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info =
2501 vk_find_struct_const(pAllocateInfo->pNext,
2502 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID);
2503
2504 if (ahw_import_info) {
2505 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
2506 if (result != VK_SUCCESS)
2507 goto fail;
2508
2509 goto success;
2510 } else if (android_export) {
2511 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
2512 if (result != VK_SUCCESS)
2513 goto fail;
2514
2515 const struct VkImportAndroidHardwareBufferInfoANDROID import_info = {
2516 .buffer = mem->ahw,
2517 };
2518 result = anv_import_ahw_memory(_device, mem, &import_info);
2519 if (result != VK_SUCCESS)
2520 goto fail;
2521
2522 goto success;
2523 }
2524
2525 const VkImportMemoryFdInfoKHR *fd_info =
2526 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
2527
2528 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2529 * ignored.
2530 */
2531 if (fd_info && fd_info->handleType) {
2532 /* At the moment, we support only the below handle types. */
2533 assert(fd_info->handleType ==
2534 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2535 fd_info->handleType ==
2536 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2537
2538 result = anv_bo_cache_import(device, &device->bo_cache, fd_info->fd,
2539 bo_flags | ANV_BO_EXTERNAL, &mem->bo);
2540 if (result != VK_SUCCESS)
2541 goto fail;
2542
2543 VkDeviceSize aligned_alloc_size =
2544 align_u64(pAllocateInfo->allocationSize, 4096);
2545
2546 /* For security purposes, we reject importing the bo if it's smaller
2547 * than the requested allocation size. This prevents a malicious client
2548 * from passing a buffer to a trusted client, lying about the size, and
2549 * telling the trusted client to try and texture from an image that goes
2550 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2551 * in the trusted client. The trusted client can protect itself against
2552 * this sort of attack but only if it can trust the buffer size.
2553 */
2554 if (mem->bo->size < aligned_alloc_size) {
2555 result = vk_errorf(device->instance, device,
2556 VK_ERROR_INVALID_EXTERNAL_HANDLE,
2557 "aligned allocationSize too large for "
2558 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2559 "%"PRIu64"B > %"PRIu64"B",
2560 aligned_alloc_size, mem->bo->size);
2561 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2562 goto fail;
2563 }
2564
2565 /* From the Vulkan spec:
2566 *
2567 * "Importing memory from a file descriptor transfers ownership of
2568 * the file descriptor from the application to the Vulkan
2569 * implementation. The application must not perform any operations on
2570 * the file descriptor after a successful import."
2571 *
2572 * If the import fails, we leave the file descriptor open.
2573 */
2574 close(fd_info->fd);
2575 goto success;
2576 }
2577
2578 /* Regular allocate (not importing memory). */
2579
2580 if (export_info && export_info->handleTypes)
2581 bo_flags |= ANV_BO_EXTERNAL;
2582
2583 result = anv_bo_cache_alloc(device, &device->bo_cache,
2584 pAllocateInfo->allocationSize, bo_flags,
2585 &mem->bo);
2586 if (result != VK_SUCCESS)
2587 goto fail;
2588
2589 const VkMemoryDedicatedAllocateInfo *dedicated_info =
2590 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_DEDICATED_ALLOCATE_INFO);
2591 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
2592 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
2593
2594 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2595 * the BO. In this case, we have a dedicated allocation.
2596 */
2597 if (image->needs_set_tiling) {
2598 const uint32_t i915_tiling =
2599 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
2600 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
2601 image->planes[0].surface.isl.row_pitch_B,
2602 i915_tiling);
2603 if (ret) {
2604 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2605 return vk_errorf(device->instance, NULL,
2606 VK_ERROR_OUT_OF_DEVICE_MEMORY,
2607 "failed to set BO tiling: %m");
2608 }
2609 }
2610 }
2611
2612 success:
2613 *pMem = anv_device_memory_to_handle(mem);
2614
2615 return VK_SUCCESS;
2616
2617 fail:
2618 vk_free2(&device->alloc, pAllocator, mem);
2619
2620 return result;
2621 }
2622
2623 VkResult anv_GetMemoryFdKHR(
2624 VkDevice device_h,
2625 const VkMemoryGetFdInfoKHR* pGetFdInfo,
2626 int* pFd)
2627 {
2628 ANV_FROM_HANDLE(anv_device, dev, device_h);
2629 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
2630
2631 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2632
2633 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2634 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2635
2636 return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
2637 }
2638
2639 VkResult anv_GetMemoryFdPropertiesKHR(
2640 VkDevice _device,
2641 VkExternalMemoryHandleTypeFlagBits handleType,
2642 int fd,
2643 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
2644 {
2645 ANV_FROM_HANDLE(anv_device, device, _device);
2646 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2647
2648 switch (handleType) {
2649 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
2650 /* dma-buf can be imported as any memory type */
2651 pMemoryFdProperties->memoryTypeBits =
2652 (1 << pdevice->memory.type_count) - 1;
2653 return VK_SUCCESS;
2654
2655 default:
2656 /* The valid usage section for this function says:
2657 *
2658 * "handleType must not be one of the handle types defined as
2659 * opaque."
2660 *
2661 * So opaque handle types fall into the default "unsupported" case.
2662 */
2663 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
2664 }
2665 }
2666
2667 void anv_FreeMemory(
2668 VkDevice _device,
2669 VkDeviceMemory _mem,
2670 const VkAllocationCallbacks* pAllocator)
2671 {
2672 ANV_FROM_HANDLE(anv_device, device, _device);
2673 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
2674
2675 if (mem == NULL)
2676 return;
2677
2678 if (mem->map)
2679 anv_UnmapMemory(_device, _mem);
2680
2681 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2682
2683 #ifdef ANDROID
2684 if (mem->ahw)
2685 AHardwareBuffer_release(mem->ahw);
2686 #endif
2687
2688 vk_free2(&device->alloc, pAllocator, mem);
2689 }
2690
2691 VkResult anv_MapMemory(
2692 VkDevice _device,
2693 VkDeviceMemory _memory,
2694 VkDeviceSize offset,
2695 VkDeviceSize size,
2696 VkMemoryMapFlags flags,
2697 void** ppData)
2698 {
2699 ANV_FROM_HANDLE(anv_device, device, _device);
2700 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
2701
2702 if (mem == NULL) {
2703 *ppData = NULL;
2704 return VK_SUCCESS;
2705 }
2706
2707 if (size == VK_WHOLE_SIZE)
2708 size = mem->bo->size - offset;
2709
2710 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2711 *
2712 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2713 * assert(size != 0);
2714 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2715 * equal to the size of the memory minus offset
2716 */
2717 assert(size > 0);
2718 assert(offset + size <= mem->bo->size);
2719
2720 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2721 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2722 * at a time is valid. We could just mmap up front and return an offset
2723 * pointer here, but that may exhaust virtual memory on 32 bit
2724 * userspace. */
2725
2726 uint32_t gem_flags = 0;
2727
2728 if (!device->info.has_llc &&
2729 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
2730 gem_flags |= I915_MMAP_WC;
2731
2732 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2733 uint64_t map_offset = offset & ~4095ull;
2734 assert(offset >= map_offset);
2735 uint64_t map_size = (offset + size) - map_offset;
2736
2737 /* Let's map whole pages */
2738 map_size = align_u64(map_size, 4096);
2739
2740 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
2741 map_offset, map_size, gem_flags);
2742 if (map == MAP_FAILED)
2743 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
2744
2745 mem->map = map;
2746 mem->map_size = map_size;
2747
2748 *ppData = mem->map + (offset - map_offset);
2749
2750 return VK_SUCCESS;
2751 }
2752
2753 void anv_UnmapMemory(
2754 VkDevice _device,
2755 VkDeviceMemory _memory)
2756 {
2757 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
2758
2759 if (mem == NULL)
2760 return;
2761
2762 anv_gem_munmap(mem->map, mem->map_size);
2763
2764 mem->map = NULL;
2765 mem->map_size = 0;
2766 }
2767
2768 static void
2769 clflush_mapped_ranges(struct anv_device *device,
2770 uint32_t count,
2771 const VkMappedMemoryRange *ranges)
2772 {
2773 for (uint32_t i = 0; i < count; i++) {
2774 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
2775 if (ranges[i].offset >= mem->map_size)
2776 continue;
2777
2778 gen_clflush_range(mem->map + ranges[i].offset,
2779 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
2780 }
2781 }
2782
2783 VkResult anv_FlushMappedMemoryRanges(
2784 VkDevice _device,
2785 uint32_t memoryRangeCount,
2786 const VkMappedMemoryRange* pMemoryRanges)
2787 {
2788 ANV_FROM_HANDLE(anv_device, device, _device);
2789
2790 if (device->info.has_llc)
2791 return VK_SUCCESS;
2792
2793 /* Make sure the writes we're flushing have landed. */
2794 __builtin_ia32_mfence();
2795
2796 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
2797
2798 return VK_SUCCESS;
2799 }
2800
2801 VkResult anv_InvalidateMappedMemoryRanges(
2802 VkDevice _device,
2803 uint32_t memoryRangeCount,
2804 const VkMappedMemoryRange* pMemoryRanges)
2805 {
2806 ANV_FROM_HANDLE(anv_device, device, _device);
2807
2808 if (device->info.has_llc)
2809 return VK_SUCCESS;
2810
2811 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
2812
2813 /* Make sure no reads get moved up above the invalidate. */
2814 __builtin_ia32_mfence();
2815
2816 return VK_SUCCESS;
2817 }
2818
2819 void anv_GetBufferMemoryRequirements(
2820 VkDevice _device,
2821 VkBuffer _buffer,
2822 VkMemoryRequirements* pMemoryRequirements)
2823 {
2824 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2825 ANV_FROM_HANDLE(anv_device, device, _device);
2826 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2827
2828 /* The Vulkan spec (git aaed022) says:
2829 *
2830 * memoryTypeBits is a bitfield and contains one bit set for every
2831 * supported memory type for the resource. The bit `1<<i` is set if and
2832 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2833 * structure for the physical device is supported.
2834 */
2835 uint32_t memory_types = 0;
2836 for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
2837 uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
2838 if ((valid_usage & buffer->usage) == buffer->usage)
2839 memory_types |= (1u << i);
2840 }
2841
2842 /* Base alignment requirement of a cache line */
2843 uint32_t alignment = 16;
2844
2845 /* We need an alignment of 32 for pushing UBOs */
2846 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
2847 alignment = MAX2(alignment, 32);
2848
2849 pMemoryRequirements->size = buffer->size;
2850 pMemoryRequirements->alignment = alignment;
2851
2852 /* Storage and Uniform buffers should have their size aligned to
2853 * 32-bits to avoid boundary checks when last DWord is not complete.
2854 * This would ensure that not internal padding would be needed for
2855 * 16-bit types.
2856 */
2857 if (device->robust_buffer_access &&
2858 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
2859 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
2860 pMemoryRequirements->size = align_u64(buffer->size, 4);
2861
2862 pMemoryRequirements->memoryTypeBits = memory_types;
2863 }
2864
2865 void anv_GetBufferMemoryRequirements2(
2866 VkDevice _device,
2867 const VkBufferMemoryRequirementsInfo2* pInfo,
2868 VkMemoryRequirements2* pMemoryRequirements)
2869 {
2870 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
2871 &pMemoryRequirements->memoryRequirements);
2872
2873 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
2874 switch (ext->sType) {
2875 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
2876 VkMemoryDedicatedRequirements *requirements = (void *)ext;
2877 requirements->prefersDedicatedAllocation = VK_FALSE;
2878 requirements->requiresDedicatedAllocation = VK_FALSE;
2879 break;
2880 }
2881
2882 default:
2883 anv_debug_ignored_stype(ext->sType);
2884 break;
2885 }
2886 }
2887 }
2888
2889 void anv_GetImageMemoryRequirements(
2890 VkDevice _device,
2891 VkImage _image,
2892 VkMemoryRequirements* pMemoryRequirements)
2893 {
2894 ANV_FROM_HANDLE(anv_image, image, _image);
2895 ANV_FROM_HANDLE(anv_device, device, _device);
2896 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2897
2898 /* The Vulkan spec (git aaed022) says:
2899 *
2900 * memoryTypeBits is a bitfield and contains one bit set for every
2901 * supported memory type for the resource. The bit `1<<i` is set if and
2902 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2903 * structure for the physical device is supported.
2904 *
2905 * All types are currently supported for images.
2906 */
2907 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
2908
2909 /* We must have image allocated or imported at this point. According to the
2910 * specification, external images must have been bound to memory before
2911 * calling GetImageMemoryRequirements.
2912 */
2913 assert(image->size > 0);
2914
2915 pMemoryRequirements->size = image->size;
2916 pMemoryRequirements->alignment = image->alignment;
2917 pMemoryRequirements->memoryTypeBits = memory_types;
2918 }
2919
2920 void anv_GetImageMemoryRequirements2(
2921 VkDevice _device,
2922 const VkImageMemoryRequirementsInfo2* pInfo,
2923 VkMemoryRequirements2* pMemoryRequirements)
2924 {
2925 ANV_FROM_HANDLE(anv_device, device, _device);
2926 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
2927
2928 anv_GetImageMemoryRequirements(_device, pInfo->image,
2929 &pMemoryRequirements->memoryRequirements);
2930
2931 vk_foreach_struct_const(ext, pInfo->pNext) {
2932 switch (ext->sType) {
2933 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
2934 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2935 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
2936 (const VkImagePlaneMemoryRequirementsInfo *) ext;
2937 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
2938 plane_reqs->planeAspect);
2939
2940 assert(image->planes[plane].offset == 0);
2941
2942 /* The Vulkan spec (git aaed022) says:
2943 *
2944 * memoryTypeBits is a bitfield and contains one bit set for every
2945 * supported memory type for the resource. The bit `1<<i` is set
2946 * if and only if the memory type `i` in the
2947 * VkPhysicalDeviceMemoryProperties structure for the physical
2948 * device is supported.
2949 *
2950 * All types are currently supported for images.
2951 */
2952 pMemoryRequirements->memoryRequirements.memoryTypeBits =
2953 (1ull << pdevice->memory.type_count) - 1;
2954
2955 /* We must have image allocated or imported at this point. According to the
2956 * specification, external images must have been bound to memory before
2957 * calling GetImageMemoryRequirements.
2958 */
2959 assert(image->planes[plane].size > 0);
2960
2961 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
2962 pMemoryRequirements->memoryRequirements.alignment =
2963 image->planes[plane].alignment;
2964 break;
2965 }
2966
2967 default:
2968 anv_debug_ignored_stype(ext->sType);
2969 break;
2970 }
2971 }
2972
2973 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
2974 switch (ext->sType) {
2975 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
2976 VkMemoryDedicatedRequirements *requirements = (void *)ext;
2977 if (image->needs_set_tiling || image->external_format) {
2978 /* If we need to set the tiling for external consumers, we need a
2979 * dedicated allocation.
2980 *
2981 * See also anv_AllocateMemory.
2982 */
2983 requirements->prefersDedicatedAllocation = VK_TRUE;
2984 requirements->requiresDedicatedAllocation = VK_TRUE;
2985 } else {
2986 requirements->prefersDedicatedAllocation = VK_FALSE;
2987 requirements->requiresDedicatedAllocation = VK_FALSE;
2988 }
2989 break;
2990 }
2991
2992 default:
2993 anv_debug_ignored_stype(ext->sType);
2994 break;
2995 }
2996 }
2997 }
2998
2999 void anv_GetImageSparseMemoryRequirements(
3000 VkDevice device,
3001 VkImage image,
3002 uint32_t* pSparseMemoryRequirementCount,
3003 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3004 {
3005 *pSparseMemoryRequirementCount = 0;
3006 }
3007
3008 void anv_GetImageSparseMemoryRequirements2(
3009 VkDevice device,
3010 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3011 uint32_t* pSparseMemoryRequirementCount,
3012 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3013 {
3014 *pSparseMemoryRequirementCount = 0;
3015 }
3016
3017 void anv_GetDeviceMemoryCommitment(
3018 VkDevice device,
3019 VkDeviceMemory memory,
3020 VkDeviceSize* pCommittedMemoryInBytes)
3021 {
3022 *pCommittedMemoryInBytes = 0;
3023 }
3024
3025 static void
3026 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3027 {
3028 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3029 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3030
3031 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3032
3033 if (mem) {
3034 assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
3035 buffer->address = (struct anv_address) {
3036 .bo = mem->bo,
3037 .offset = pBindInfo->memoryOffset,
3038 };
3039 } else {
3040 buffer->address = ANV_NULL_ADDRESS;
3041 }
3042 }
3043
3044 VkResult anv_BindBufferMemory(
3045 VkDevice device,
3046 VkBuffer buffer,
3047 VkDeviceMemory memory,
3048 VkDeviceSize memoryOffset)
3049 {
3050 anv_bind_buffer_memory(
3051 &(VkBindBufferMemoryInfo) {
3052 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
3053 .buffer = buffer,
3054 .memory = memory,
3055 .memoryOffset = memoryOffset,
3056 });
3057
3058 return VK_SUCCESS;
3059 }
3060
3061 VkResult anv_BindBufferMemory2(
3062 VkDevice device,
3063 uint32_t bindInfoCount,
3064 const VkBindBufferMemoryInfo* pBindInfos)
3065 {
3066 for (uint32_t i = 0; i < bindInfoCount; i++)
3067 anv_bind_buffer_memory(&pBindInfos[i]);
3068
3069 return VK_SUCCESS;
3070 }
3071
3072 VkResult anv_QueueBindSparse(
3073 VkQueue _queue,
3074 uint32_t bindInfoCount,
3075 const VkBindSparseInfo* pBindInfo,
3076 VkFence fence)
3077 {
3078 ANV_FROM_HANDLE(anv_queue, queue, _queue);
3079 if (anv_device_is_lost(queue->device))
3080 return VK_ERROR_DEVICE_LOST;
3081
3082 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
3083 }
3084
3085 // Event functions
3086
3087 VkResult anv_CreateEvent(
3088 VkDevice _device,
3089 const VkEventCreateInfo* pCreateInfo,
3090 const VkAllocationCallbacks* pAllocator,
3091 VkEvent* pEvent)
3092 {
3093 ANV_FROM_HANDLE(anv_device, device, _device);
3094 struct anv_state state;
3095 struct anv_event *event;
3096
3097 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
3098
3099 state = anv_state_pool_alloc(&device->dynamic_state_pool,
3100 sizeof(*event), 8);
3101 event = state.map;
3102 event->state = state;
3103 event->semaphore = VK_EVENT_RESET;
3104
3105 if (!device->info.has_llc) {
3106 /* Make sure the writes we're flushing have landed. */
3107 __builtin_ia32_mfence();
3108 __builtin_ia32_clflush(event);
3109 }
3110
3111 *pEvent = anv_event_to_handle(event);
3112
3113 return VK_SUCCESS;
3114 }
3115
3116 void anv_DestroyEvent(
3117 VkDevice _device,
3118 VkEvent _event,
3119 const VkAllocationCallbacks* pAllocator)
3120 {
3121 ANV_FROM_HANDLE(anv_device, device, _device);
3122 ANV_FROM_HANDLE(anv_event, event, _event);
3123
3124 if (!event)
3125 return;
3126
3127 anv_state_pool_free(&device->dynamic_state_pool, event->state);
3128 }
3129
3130 VkResult anv_GetEventStatus(
3131 VkDevice _device,
3132 VkEvent _event)
3133 {
3134 ANV_FROM_HANDLE(anv_device, device, _device);
3135 ANV_FROM_HANDLE(anv_event, event, _event);
3136
3137 if (anv_device_is_lost(device))
3138 return VK_ERROR_DEVICE_LOST;
3139
3140 if (!device->info.has_llc) {
3141 /* Invalidate read cache before reading event written by GPU. */
3142 __builtin_ia32_clflush(event);
3143 __builtin_ia32_mfence();
3144
3145 }
3146
3147 return event->semaphore;
3148 }
3149
3150 VkResult anv_SetEvent(
3151 VkDevice _device,
3152 VkEvent _event)
3153 {
3154 ANV_FROM_HANDLE(anv_device, device, _device);
3155 ANV_FROM_HANDLE(anv_event, event, _event);
3156
3157 event->semaphore = VK_EVENT_SET;
3158
3159 if (!device->info.has_llc) {
3160 /* Make sure the writes we're flushing have landed. */
3161 __builtin_ia32_mfence();
3162 __builtin_ia32_clflush(event);
3163 }
3164
3165 return VK_SUCCESS;
3166 }
3167
3168 VkResult anv_ResetEvent(
3169 VkDevice _device,
3170 VkEvent _event)
3171 {
3172 ANV_FROM_HANDLE(anv_device, device, _device);
3173 ANV_FROM_HANDLE(anv_event, event, _event);
3174
3175 event->semaphore = VK_EVENT_RESET;
3176
3177 if (!device->info.has_llc) {
3178 /* Make sure the writes we're flushing have landed. */
3179 __builtin_ia32_mfence();
3180 __builtin_ia32_clflush(event);
3181 }
3182
3183 return VK_SUCCESS;
3184 }
3185
3186 // Buffer functions
3187
3188 VkResult anv_CreateBuffer(
3189 VkDevice _device,
3190 const VkBufferCreateInfo* pCreateInfo,
3191 const VkAllocationCallbacks* pAllocator,
3192 VkBuffer* pBuffer)
3193 {
3194 ANV_FROM_HANDLE(anv_device, device, _device);
3195 struct anv_buffer *buffer;
3196
3197 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
3198
3199 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
3200 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3201 if (buffer == NULL)
3202 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3203
3204 buffer->size = pCreateInfo->size;
3205 buffer->usage = pCreateInfo->usage;
3206 buffer->address = ANV_NULL_ADDRESS;
3207
3208 if (buffer->usage & VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT_EXT) {
3209 pthread_mutex_lock(&device->mutex);
3210 _mesa_set_add(device->pinned_buffers, buffer);
3211 pthread_mutex_unlock(&device->mutex);
3212 }
3213
3214 *pBuffer = anv_buffer_to_handle(buffer);
3215
3216 return VK_SUCCESS;
3217 }
3218
3219 void anv_DestroyBuffer(
3220 VkDevice _device,
3221 VkBuffer _buffer,
3222 const VkAllocationCallbacks* pAllocator)
3223 {
3224 ANV_FROM_HANDLE(anv_device, device, _device);
3225 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3226
3227 if (!buffer)
3228 return;
3229
3230 if (buffer->usage & VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT_EXT) {
3231 pthread_mutex_lock(&device->mutex);
3232 _mesa_set_remove_key(device->pinned_buffers, buffer);
3233 pthread_mutex_unlock(&device->mutex);
3234 }
3235
3236 vk_free2(&device->alloc, pAllocator, buffer);
3237 }
3238
3239 VkDeviceAddress anv_GetBufferDeviceAddressEXT(
3240 VkDevice device,
3241 const VkBufferDeviceAddressInfoEXT* pInfo)
3242 {
3243 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
3244
3245 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
3246
3247 return anv_address_physical(buffer->address);
3248 }
3249
3250 void
3251 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
3252 enum isl_format format,
3253 struct anv_address address,
3254 uint32_t range, uint32_t stride)
3255 {
3256 isl_buffer_fill_state(&device->isl_dev, state.map,
3257 .address = anv_address_physical(address),
3258 .mocs = device->default_mocs,
3259 .size_B = range,
3260 .format = format,
3261 .stride_B = stride);
3262 }
3263
3264 void anv_DestroySampler(
3265 VkDevice _device,
3266 VkSampler _sampler,
3267 const VkAllocationCallbacks* pAllocator)
3268 {
3269 ANV_FROM_HANDLE(anv_device, device, _device);
3270 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
3271
3272 if (!sampler)
3273 return;
3274
3275 vk_free2(&device->alloc, pAllocator, sampler);
3276 }
3277
3278 VkResult anv_CreateFramebuffer(
3279 VkDevice _device,
3280 const VkFramebufferCreateInfo* pCreateInfo,
3281 const VkAllocationCallbacks* pAllocator,
3282 VkFramebuffer* pFramebuffer)
3283 {
3284 ANV_FROM_HANDLE(anv_device, device, _device);
3285 struct anv_framebuffer *framebuffer;
3286
3287 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
3288
3289 size_t size = sizeof(*framebuffer) +
3290 sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
3291 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
3292 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3293 if (framebuffer == NULL)
3294 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3295
3296 framebuffer->attachment_count = pCreateInfo->attachmentCount;
3297 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
3298 VkImageView _iview = pCreateInfo->pAttachments[i];
3299 framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
3300 }
3301
3302 framebuffer->width = pCreateInfo->width;
3303 framebuffer->height = pCreateInfo->height;
3304 framebuffer->layers = pCreateInfo->layers;
3305
3306 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
3307
3308 return VK_SUCCESS;
3309 }
3310
3311 void anv_DestroyFramebuffer(
3312 VkDevice _device,
3313 VkFramebuffer _fb,
3314 const VkAllocationCallbacks* pAllocator)
3315 {
3316 ANV_FROM_HANDLE(anv_device, device, _device);
3317 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
3318
3319 if (!fb)
3320 return;
3321
3322 vk_free2(&device->alloc, pAllocator, fb);
3323 }
3324
3325 static const VkTimeDomainEXT anv_time_domains[] = {
3326 VK_TIME_DOMAIN_DEVICE_EXT,
3327 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
3328 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
3329 };
3330
3331 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3332 VkPhysicalDevice physicalDevice,
3333 uint32_t *pTimeDomainCount,
3334 VkTimeDomainEXT *pTimeDomains)
3335 {
3336 int d;
3337 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
3338
3339 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
3340 vk_outarray_append(&out, i) {
3341 *i = anv_time_domains[d];
3342 }
3343 }
3344
3345 return vk_outarray_status(&out);
3346 }
3347
3348 static uint64_t
3349 anv_clock_gettime(clockid_t clock_id)
3350 {
3351 struct timespec current;
3352 int ret;
3353
3354 ret = clock_gettime(clock_id, &current);
3355 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
3356 ret = clock_gettime(CLOCK_MONOTONIC, &current);
3357 if (ret < 0)
3358 return 0;
3359
3360 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
3361 }
3362
3363 #define TIMESTAMP 0x2358
3364
3365 VkResult anv_GetCalibratedTimestampsEXT(
3366 VkDevice _device,
3367 uint32_t timestampCount,
3368 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
3369 uint64_t *pTimestamps,
3370 uint64_t *pMaxDeviation)
3371 {
3372 ANV_FROM_HANDLE(anv_device, device, _device);
3373 uint64_t timestamp_frequency = device->info.timestamp_frequency;
3374 int ret;
3375 int d;
3376 uint64_t begin, end;
3377 uint64_t max_clock_period = 0;
3378
3379 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3380
3381 for (d = 0; d < timestampCount; d++) {
3382 switch (pTimestampInfos[d].timeDomain) {
3383 case VK_TIME_DOMAIN_DEVICE_EXT:
3384 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
3385 &pTimestamps[d]);
3386
3387 if (ret != 0) {
3388 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
3389 "register: %m");
3390 }
3391 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
3392 max_clock_period = MAX2(max_clock_period, device_period);
3393 break;
3394 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
3395 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
3396 max_clock_period = MAX2(max_clock_period, 1);
3397 break;
3398
3399 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
3400 pTimestamps[d] = begin;
3401 break;
3402 default:
3403 pTimestamps[d] = 0;
3404 break;
3405 }
3406 }
3407
3408 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3409
3410 /*
3411 * The maximum deviation is the sum of the interval over which we
3412 * perform the sampling and the maximum period of any sampled
3413 * clock. That's because the maximum skew between any two sampled
3414 * clock edges is when the sampled clock with the largest period is
3415 * sampled at the end of that period but right at the beginning of the
3416 * sampling interval and some other clock is sampled right at the
3417 * begining of its sampling period and right at the end of the
3418 * sampling interval. Let's assume the GPU has the longest clock
3419 * period and that the application is sampling GPU and monotonic:
3420 *
3421 * s e
3422 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3423 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3424 *
3425 * g
3426 * 0 1 2 3
3427 * GPU -----_____-----_____-----_____-----_____
3428 *
3429 * m
3430 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3431 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3432 *
3433 * Interval <----------------->
3434 * Deviation <-------------------------->
3435 *
3436 * s = read(raw) 2
3437 * g = read(GPU) 1
3438 * m = read(monotonic) 2
3439 * e = read(raw) b
3440 *
3441 * We round the sample interval up by one tick to cover sampling error
3442 * in the interval clock
3443 */
3444
3445 uint64_t sample_interval = end - begin + 1;
3446
3447 *pMaxDeviation = sample_interval + max_clock_period;
3448
3449 return VK_SUCCESS;
3450 }
3451
3452 /* vk_icd.h does not declare this function, so we declare it here to
3453 * suppress Wmissing-prototypes.
3454 */
3455 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3456 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
3457
3458 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3459 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
3460 {
3461 /* For the full details on loader interface versioning, see
3462 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3463 * What follows is a condensed summary, to help you navigate the large and
3464 * confusing official doc.
3465 *
3466 * - Loader interface v0 is incompatible with later versions. We don't
3467 * support it.
3468 *
3469 * - In loader interface v1:
3470 * - The first ICD entrypoint called by the loader is
3471 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3472 * entrypoint.
3473 * - The ICD must statically expose no other Vulkan symbol unless it is
3474 * linked with -Bsymbolic.
3475 * - Each dispatchable Vulkan handle created by the ICD must be
3476 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3477 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3478 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3479 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3480 * such loader-managed surfaces.
3481 *
3482 * - Loader interface v2 differs from v1 in:
3483 * - The first ICD entrypoint called by the loader is
3484 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3485 * statically expose this entrypoint.
3486 *
3487 * - Loader interface v3 differs from v2 in:
3488 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3489 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3490 * because the loader no longer does so.
3491 */
3492 *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
3493 return VK_SUCCESS;
3494 }