projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
anv: refactor, remove else block in AllocateMemory
[mesa.git] / src / intel / vulkan / anv_device.c
1 /*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <assert.h>
25 #include <stdbool.h>
26 #include <string.h>
27 #include <sys/mman.h>
28 #include <sys/sysinfo.h>
29 #include <unistd.h>
30 #include <fcntl.h>
31 #include <xf86drm.h>
32 #include <drm_fourcc.h>
33
34 #include "anv_private.h"
35 #include "util/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/disk_cache.h"
39 #include "util/mesa-sha1.h"
40 #include "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 VkResult
64 anv_compute_heap_size(int fd, uint64_t gtt_size, uint64_t *heap_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 *heap_size = MIN2(available_ram, available_gtt);
87
88 return VK_SUCCESS;
89 }
90
91 static VkResult
92 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
93 {
94 uint64_t gtt_size;
95 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
96 &gtt_size) == -1) {
97 /* If, for whatever reason, we can't actually get the GTT size from the
98 * kernel (too old?) fall back to the aperture size.
99 */
100 anv_perf_warn(NULL, NULL,
101 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
102
103 if (anv_gem_get_aperture(fd, &gtt_size) == -1) {
104 return vk_errorf(NULL, NULL, VK_ERROR_INITIALIZATION_FAILED,
105 "failed to get aperture size: %m");
106 }
107 }
108
109 device->supports_48bit_addresses = (device->info.gen >= 8) &&
110 gtt_size > (4ULL << 30 /* GiB */);
111
112 uint64_t heap_size = 0;
113 VkResult result = anv_compute_heap_size(fd, gtt_size, &heap_size);
114 if (result != VK_SUCCESS)
115 return result;
116
117 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
118 /* When running with an overridden PCI ID, we may get a GTT size from
119 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
120 * address support can still fail. Just clamp the address space size to
121 * 2 GiB if we don't have 48-bit support.
122 */
123 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
124 "not support for 48-bit addresses",
125 __FILE__, __LINE__);
126 heap_size = 2ull << 30;
127 }
128
129 if (heap_size <= 3ull * (1ull << 30)) {
130 /* In this case, everything fits nicely into the 32-bit address space,
131 * so there's no need for supporting 48bit addresses on client-allocated
132 * memory objects.
133 */
134 device->memory.heap_count = 1;
135 device->memory.heaps[0] = (struct anv_memory_heap) {
136 .size = heap_size,
137 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
138 .supports_48bit_addresses = false,
139 };
140 } else {
141 /* Not everything will fit nicely into a 32-bit address space. In this
142 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
143 * larger 48-bit heap. If we're in this case, then we have a total heap
144 * size larger than 3GiB which most likely means they have 8 GiB of
145 * video memory and so carving off 1 GiB for the 32-bit heap should be
146 * reasonable.
147 */
148 const uint64_t heap_size_32bit = 1ull << 30;
149 const uint64_t heap_size_48bit = heap_size - heap_size_32bit;
150
151 assert(device->supports_48bit_addresses);
152
153 device->memory.heap_count = 2;
154 device->memory.heaps[0] = (struct anv_memory_heap) {
155 .size = heap_size_48bit,
156 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
157 .supports_48bit_addresses = true,
158 };
159 device->memory.heaps[1] = (struct anv_memory_heap) {
160 .size = heap_size_32bit,
161 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
162 .supports_48bit_addresses = false,
163 };
164 }
165
166 uint32_t type_count = 0;
167 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
168 uint32_t valid_buffer_usage = ~0;
169
170 /* There appears to be a hardware issue in the VF cache where it only
171 * considers the bottom 32 bits of memory addresses. If you happen to
172 * have two vertex buffers which get placed exactly 4 GiB apart and use
173 * them in back-to-back draw calls, you can get collisions. In order to
174 * solve this problem, we require vertex and index buffers be bound to
175 * memory allocated out of the 32-bit heap.
176 */
177 if (device->memory.heaps[heap].supports_48bit_addresses) {
178 valid_buffer_usage &= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
179 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
180 }
181
182 if (device->info.has_llc) {
183 /* Big core GPUs share LLC with the CPU and thus one memory type can be
184 * both cached and coherent at the same time.
185 */
186 device->memory.types[type_count++] = (struct anv_memory_type) {
187 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
188 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
189 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
190 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
191 .heapIndex = heap,
192 .valid_buffer_usage = valid_buffer_usage,
193 };
194 } else {
195 /* The spec requires that we expose a host-visible, coherent memory
196 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
197 * to give the application a choice between cached, but not coherent and
198 * coherent but uncached (WC though).
199 */
200 device->memory.types[type_count++] = (struct anv_memory_type) {
201 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
202 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
203 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
204 .heapIndex = heap,
205 .valid_buffer_usage = valid_buffer_usage,
206 };
207 device->memory.types[type_count++] = (struct anv_memory_type) {
208 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
209 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
210 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
211 .heapIndex = heap,
212 .valid_buffer_usage = valid_buffer_usage,
213 };
214 }
215 }
216 device->memory.type_count = type_count;
217
218 return VK_SUCCESS;
219 }
220
221 static VkResult
222 anv_physical_device_init_uuids(struct anv_physical_device *device)
223 {
224 const struct build_id_note *note =
225 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
226 if (!note) {
227 return vk_errorf(device->instance, device,
228 VK_ERROR_INITIALIZATION_FAILED,
229 "Failed to find build-id");
230 }
231
232 unsigned build_id_len = build_id_length(note);
233 if (build_id_len < 20) {
234 return vk_errorf(device->instance, device,
235 VK_ERROR_INITIALIZATION_FAILED,
236 "build-id too short. It needs to be a SHA");
237 }
238
239 memcpy(device->driver_build_sha1, build_id_data(note), 20);
240
241 struct mesa_sha1 sha1_ctx;
242 uint8_t sha1[20];
243 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
244
245 /* The pipeline cache UUID is used for determining when a pipeline cache is
246 * invalid. It needs both a driver build and the PCI ID of the device.
247 */
248 _mesa_sha1_init(&sha1_ctx);
249 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
250 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
251 sizeof(device->chipset_id));
252 _mesa_sha1_final(&sha1_ctx, sha1);
253 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
254
255 /* The driver UUID is used for determining sharability of images and memory
256 * between two Vulkan instances in separate processes. People who want to
257 * share memory need to also check the device UUID (below) so all this
258 * needs to be is the build-id.
259 */
260 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
261
262 /* The device UUID uniquely identifies the given device within the machine.
263 * Since we never have more than one device, this doesn't need to be a real
264 * UUID. However, on the off-chance that someone tries to use this to
265 * cache pre-tiled images or something of the like, we use the PCI ID and
266 * some bits of ISL info to ensure that this is safe.
267 */
268 _mesa_sha1_init(&sha1_ctx);
269 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
270 sizeof(device->chipset_id));
271 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
272 sizeof(device->isl_dev.has_bit6_swizzling));
273 _mesa_sha1_final(&sha1_ctx, sha1);
274 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
275
276 return VK_SUCCESS;
277 }
278
279 static void
280 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
281 {
282 #ifdef ENABLE_SHADER_CACHE
283 char renderer[10];
284 MAYBE_UNUSED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
285 device->chipset_id);
286 assert(len == sizeof(renderer) - 2);
287
288 char timestamp[41];
289 _mesa_sha1_format(timestamp, device->driver_build_sha1);
290
291 const uint64_t driver_flags =
292 brw_get_compiler_config_value(device->compiler);
293 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
294 #else
295 device->disk_cache = NULL;
296 #endif
297 }
298
299 static void
300 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
301 {
302 #ifdef ENABLE_SHADER_CACHE
303 if (device->disk_cache)
304 disk_cache_destroy(device->disk_cache);
305 #else
306 assert(device->disk_cache == NULL);
307 #endif
308 }
309
310 static VkResult
311 anv_physical_device_init(struct anv_physical_device *device,
312 struct anv_instance *instance,
313 drmDevicePtr drm_device)
314 {
315 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
316 const char *path = drm_device->nodes[DRM_NODE_RENDER];
317 VkResult result;
318 int fd;
319 int master_fd = -1;
320
321 brw_process_intel_debug_variable();
322
323 fd = open(path, O_RDWR | O_CLOEXEC);
324 if (fd < 0)
325 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
326
327 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
328 device->instance = instance;
329
330 assert(strlen(path) < ARRAY_SIZE(device->path));
331 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
332
333 device->no_hw = getenv("INTEL_NO_HW") != NULL;
334
335 const int pci_id_override = gen_get_pci_device_id_override();
336 if (pci_id_override < 0) {
337 device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
338 if (!device->chipset_id) {
339 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
340 goto fail;
341 }
342 } else {
343 device->chipset_id = pci_id_override;
344 device->no_hw = true;
345 }
346
347 device->pci_info.domain = drm_device->businfo.pci->domain;
348 device->pci_info.bus = drm_device->businfo.pci->bus;
349 device->pci_info.device = drm_device->businfo.pci->dev;
350 device->pci_info.function = drm_device->businfo.pci->func;
351
352 device->name = gen_get_device_name(device->chipset_id);
353 if (!gen_get_device_info(device->chipset_id, &device->info)) {
354 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
355 goto fail;
356 }
357
358 if (device->info.is_haswell) {
359 intel_logw("Haswell Vulkan support is incomplete");
360 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
361 intel_logw("Ivy Bridge Vulkan support is incomplete");
362 } else if (device->info.gen == 7 && device->info.is_baytrail) {
363 intel_logw("Bay Trail Vulkan support is incomplete");
364 } else if (device->info.gen >= 8 && device->info.gen <= 10) {
365 /* Gen8-10 fully supported */
366 } else if (device->info.gen == 11) {
367 intel_logw("Vulkan is not yet fully supported on gen11.");
368 } else {
369 result = vk_errorf(device->instance, device,
370 VK_ERROR_INCOMPATIBLE_DRIVER,
371 "Vulkan not yet supported on %s", device->name);
372 goto fail;
373 }
374
375 device->cmd_parser_version = -1;
376 if (device->info.gen == 7) {
377 device->cmd_parser_version =
378 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
379 if (device->cmd_parser_version == -1) {
380 result = vk_errorf(device->instance, device,
381 VK_ERROR_INITIALIZATION_FAILED,
382 "failed to get command parser version");
383 goto fail;
384 }
385 }
386
387 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
388 result = vk_errorf(device->instance, device,
389 VK_ERROR_INITIALIZATION_FAILED,
390 "kernel missing gem wait");
391 goto fail;
392 }
393
394 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
395 result = vk_errorf(device->instance, device,
396 VK_ERROR_INITIALIZATION_FAILED,
397 "kernel missing execbuf2");
398 goto fail;
399 }
400
401 if (!device->info.has_llc &&
402 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
403 result = vk_errorf(device->instance, device,
404 VK_ERROR_INITIALIZATION_FAILED,
405 "kernel missing wc mmap");
406 goto fail;
407 }
408
409 result = anv_physical_device_init_heaps(device, fd);
410 if (result != VK_SUCCESS)
411 goto fail;
412
413 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
414 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
415 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
416 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
417 device->has_syncobj_wait = device->has_syncobj &&
418 anv_gem_supports_syncobj_wait(fd);
419 device->has_context_priority = anv_gem_has_context_priority(fd);
420
421 device->use_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN)
422 && device->supports_48bit_addresses;
423
424 device->has_context_isolation =
425 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
426
427 bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
428
429 /* Starting with Gen10, the timestamp frequency of the command streamer may
430 * vary from one part to another. We can query the value from the kernel.
431 */
432 if (device->info.gen >= 10) {
433 int timestamp_frequency =
434 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
435
436 if (timestamp_frequency < 0)
437 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
438 else
439 device->info.timestamp_frequency = timestamp_frequency;
440 }
441
442 /* GENs prior to 8 do not support EU/Subslice info */
443 if (device->info.gen >= 8) {
444 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
445 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
446
447 /* Without this information, we cannot get the right Braswell
448 * brandstrings, and we have to use conservative numbers for GPGPU on
449 * many platforms, but otherwise, things will just work.
450 */
451 if (device->subslice_total < 1 || device->eu_total < 1) {
452 intel_logw("Kernel 4.1 required to properly query GPU properties");
453 }
454 } else if (device->info.gen == 7) {
455 device->subslice_total = 1 << (device->info.gt - 1);
456 }
457
458 if (device->info.is_cherryview &&
459 device->subslice_total > 0 && device->eu_total > 0) {
460 /* Logical CS threads = EUs per subslice * num threads per EU */
461 uint32_t max_cs_threads =
462 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
463
464 /* Fuse configurations may give more threads than expected, never less. */
465 if (max_cs_threads > device->info.max_cs_threads)
466 device->info.max_cs_threads = max_cs_threads;
467 }
468
469 device->compiler = brw_compiler_create(NULL, &device->info);
470 if (device->compiler == NULL) {
471 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
472 goto fail;
473 }
474 device->compiler->shader_debug_log = compiler_debug_log;
475 device->compiler->shader_perf_log = compiler_perf_log;
476 device->compiler->supports_pull_constants = false;
477 device->compiler->constant_buffer_0_is_relative =
478 device->info.gen < 8 || !device->has_context_isolation;
479 device->compiler->supports_shader_constants = true;
480
481 isl_device_init(&device->isl_dev, &device->info, swizzled);
482
483 result = anv_physical_device_init_uuids(device);
484 if (result != VK_SUCCESS)
485 goto fail;
486
487 anv_physical_device_init_disk_cache(device);
488
489 if (instance->enabled_extensions.KHR_display) {
490 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
491 if (master_fd >= 0) {
492 /* prod the device with a GETPARAM call which will fail if
493 * we don't have permission to even render on this device
494 */
495 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
496 close(master_fd);
497 master_fd = -1;
498 }
499 }
500 }
501 device->master_fd = master_fd;
502
503 result = anv_init_wsi(device);
504 if (result != VK_SUCCESS) {
505 ralloc_free(device->compiler);
506 anv_physical_device_free_disk_cache(device);
507 goto fail;
508 }
509
510 anv_physical_device_get_supported_extensions(device,
511 &device->supported_extensions);
512
513
514 device->local_fd = fd;
515
516 return VK_SUCCESS;
517
518 fail:
519 close(fd);
520 if (master_fd != -1)
521 close(master_fd);
522 return result;
523 }
524
525 static void
526 anv_physical_device_finish(struct anv_physical_device *device)
527 {
528 anv_finish_wsi(device);
529 anv_physical_device_free_disk_cache(device);
530 ralloc_free(device->compiler);
531 close(device->local_fd);
532 if (device->master_fd >= 0)
533 close(device->master_fd);
534 }
535
536 static void *
537 default_alloc_func(void *pUserData, size_t size, size_t align,
538 VkSystemAllocationScope allocationScope)
539 {
540 return malloc(size);
541 }
542
543 static void *
544 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
545 size_t align, VkSystemAllocationScope allocationScope)
546 {
547 return realloc(pOriginal, size);
548 }
549
550 static void
551 default_free_func(void *pUserData, void *pMemory)
552 {
553 free(pMemory);
554 }
555
556 static const VkAllocationCallbacks default_alloc = {
557 .pUserData = NULL,
558 .pfnAllocation = default_alloc_func,
559 .pfnReallocation = default_realloc_func,
560 .pfnFree = default_free_func,
561 };
562
563 VkResult anv_EnumerateInstanceExtensionProperties(
564 const char* pLayerName,
565 uint32_t* pPropertyCount,
566 VkExtensionProperties* pProperties)
567 {
568 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
569
570 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
571 if (anv_instance_extensions_supported.extensions[i]) {
572 vk_outarray_append(&out, prop) {
573 *prop = anv_instance_extensions[i];
574 }
575 }
576 }
577
578 return vk_outarray_status(&out);
579 }
580
581 VkResult anv_CreateInstance(
582 const VkInstanceCreateInfo* pCreateInfo,
583 const VkAllocationCallbacks* pAllocator,
584 VkInstance* pInstance)
585 {
586 struct anv_instance *instance;
587 VkResult result;
588
589 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
590
591 struct anv_instance_extension_table enabled_extensions = {};
592 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
593 int idx;
594 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
595 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
596 anv_instance_extensions[idx].extensionName) == 0)
597 break;
598 }
599
600 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
601 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
602
603 if (!anv_instance_extensions_supported.extensions[idx])
604 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
605
606 enabled_extensions.extensions[idx] = true;
607 }
608
609 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
610 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
611 if (!instance)
612 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
613
614 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
615
616 if (pAllocator)
617 instance->alloc = *pAllocator;
618 else
619 instance->alloc = default_alloc;
620
621 instance->app_info = (struct anv_app_info) { .api_version = 0 };
622 if (pCreateInfo->pApplicationInfo) {
623 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
624
625 instance->app_info.app_name =
626 vk_strdup(&instance->alloc, app->pApplicationName,
627 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
628 instance->app_info.app_version = app->applicationVersion;
629
630 instance->app_info.engine_name =
631 vk_strdup(&instance->alloc, app->pEngineName,
632 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
633 instance->app_info.engine_version = app->engineVersion;
634
635 instance->app_info.api_version = app->apiVersion;
636 }
637
638 if (instance->app_info.api_version == 0)
639 instance->app_info.api_version = VK_API_VERSION_1_0;
640
641 instance->enabled_extensions = enabled_extensions;
642
643 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
644 /* Vulkan requires that entrypoints for extensions which have not been
645 * enabled must not be advertised.
646 */
647 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
648 &instance->enabled_extensions)) {
649 instance->dispatch.entrypoints[i] = NULL;
650 } else {
651 instance->dispatch.entrypoints[i] =
652 anv_instance_dispatch_table.entrypoints[i];
653 }
654 }
655
656 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
657 /* Vulkan requires that entrypoints for extensions which have not been
658 * enabled must not be advertised.
659 */
660 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
661 &instance->enabled_extensions, NULL)) {
662 instance->device_dispatch.entrypoints[i] = NULL;
663 } else {
664 instance->device_dispatch.entrypoints[i] =
665 anv_device_dispatch_table.entrypoints[i];
666 }
667 }
668
669 instance->physicalDeviceCount = -1;
670
671 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
672 if (result != VK_SUCCESS) {
673 vk_free2(&default_alloc, pAllocator, instance);
674 return vk_error(result);
675 }
676
677 instance->pipeline_cache_enabled =
678 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
679
680 _mesa_locale_init();
681
682 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
683
684 *pInstance = anv_instance_to_handle(instance);
685
686 return VK_SUCCESS;
687 }
688
689 void anv_DestroyInstance(
690 VkInstance _instance,
691 const VkAllocationCallbacks* pAllocator)
692 {
693 ANV_FROM_HANDLE(anv_instance, instance, _instance);
694
695 if (!instance)
696 return;
697
698 if (instance->physicalDeviceCount > 0) {
699 /* We support at most one physical device. */
700 assert(instance->physicalDeviceCount == 1);
701 anv_physical_device_finish(&instance->physicalDevice);
702 }
703
704 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
705 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
706
707 VG(VALGRIND_DESTROY_MEMPOOL(instance));
708
709 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
710
711 _mesa_locale_fini();
712
713 vk_free(&instance->alloc, instance);
714 }
715
716 static VkResult
717 anv_enumerate_devices(struct anv_instance *instance)
718 {
719 /* TODO: Check for more devices ? */
720 drmDevicePtr devices[8];
721 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
722 int max_devices;
723
724 instance->physicalDeviceCount = 0;
725
726 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
727 if (max_devices < 1)
728 return VK_ERROR_INCOMPATIBLE_DRIVER;
729
730 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
731 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
732 devices[i]->bustype == DRM_BUS_PCI &&
733 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
734
735 result = anv_physical_device_init(&instance->physicalDevice,
736 instance, devices[i]);
737 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
738 break;
739 }
740 }
741 drmFreeDevices(devices, max_devices);
742
743 if (result == VK_SUCCESS)
744 instance->physicalDeviceCount = 1;
745
746 return result;
747 }
748
749 static VkResult
750 anv_instance_ensure_physical_device(struct anv_instance *instance)
751 {
752 if (instance->physicalDeviceCount < 0) {
753 VkResult result = anv_enumerate_devices(instance);
754 if (result != VK_SUCCESS &&
755 result != VK_ERROR_INCOMPATIBLE_DRIVER)
756 return result;
757 }
758
759 return VK_SUCCESS;
760 }
761
762 VkResult anv_EnumeratePhysicalDevices(
763 VkInstance _instance,
764 uint32_t* pPhysicalDeviceCount,
765 VkPhysicalDevice* pPhysicalDevices)
766 {
767 ANV_FROM_HANDLE(anv_instance, instance, _instance);
768 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
769
770 VkResult result = anv_instance_ensure_physical_device(instance);
771 if (result != VK_SUCCESS)
772 return result;
773
774 if (instance->physicalDeviceCount == 0)
775 return VK_SUCCESS;
776
777 assert(instance->physicalDeviceCount == 1);
778 vk_outarray_append(&out, i) {
779 *i = anv_physical_device_to_handle(&instance->physicalDevice);
780 }
781
782 return vk_outarray_status(&out);
783 }
784
785 VkResult anv_EnumeratePhysicalDeviceGroups(
786 VkInstance _instance,
787 uint32_t* pPhysicalDeviceGroupCount,
788 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
789 {
790 ANV_FROM_HANDLE(anv_instance, instance, _instance);
791 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
792 pPhysicalDeviceGroupCount);
793
794 VkResult result = anv_instance_ensure_physical_device(instance);
795 if (result != VK_SUCCESS)
796 return result;
797
798 if (instance->physicalDeviceCount == 0)
799 return VK_SUCCESS;
800
801 assert(instance->physicalDeviceCount == 1);
802
803 vk_outarray_append(&out, p) {
804 p->physicalDeviceCount = 1;
805 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
806 p->physicalDevices[0] =
807 anv_physical_device_to_handle(&instance->physicalDevice);
808 p->subsetAllocation = VK_FALSE;
809
810 vk_foreach_struct(ext, p->pNext)
811 anv_debug_ignored_stype(ext->sType);
812 }
813
814 return vk_outarray_status(&out);
815 }
816
817 void anv_GetPhysicalDeviceFeatures(
818 VkPhysicalDevice physicalDevice,
819 VkPhysicalDeviceFeatures* pFeatures)
820 {
821 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
822
823 *pFeatures = (VkPhysicalDeviceFeatures) {
824 .robustBufferAccess = true,
825 .fullDrawIndexUint32 = true,
826 .imageCubeArray = true,
827 .independentBlend = true,
828 .geometryShader = true,
829 .tessellationShader = true,
830 .sampleRateShading = true,
831 .dualSrcBlend = true,
832 .logicOp = true,
833 .multiDrawIndirect = true,
834 .drawIndirectFirstInstance = true,
835 .depthClamp = true,
836 .depthBiasClamp = true,
837 .fillModeNonSolid = true,
838 .depthBounds = false,
839 .wideLines = true,
840 .largePoints = true,
841 .alphaToOne = true,
842 .multiViewport = true,
843 .samplerAnisotropy = true,
844 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
845 pdevice->info.is_baytrail,
846 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
847 .textureCompressionBC = true,
848 .occlusionQueryPrecise = true,
849 .pipelineStatisticsQuery = true,
850 .fragmentStoresAndAtomics = true,
851 .shaderTessellationAndGeometryPointSize = true,
852 .shaderImageGatherExtended = true,
853 .shaderStorageImageExtendedFormats = true,
854 .shaderStorageImageMultisample = false,
855 .shaderStorageImageReadWithoutFormat = false,
856 .shaderStorageImageWriteWithoutFormat = true,
857 .shaderUniformBufferArrayDynamicIndexing = true,
858 .shaderSampledImageArrayDynamicIndexing = true,
859 .shaderStorageBufferArrayDynamicIndexing = true,
860 .shaderStorageImageArrayDynamicIndexing = true,
861 .shaderClipDistance = true,
862 .shaderCullDistance = true,
863 .shaderFloat64 = pdevice->info.gen >= 8 &&
864 pdevice->info.has_64bit_types,
865 .shaderInt64 = pdevice->info.gen >= 8 &&
866 pdevice->info.has_64bit_types,
867 .shaderInt16 = pdevice->info.gen >= 8,
868 .shaderResourceMinLod = pdevice->info.gen >= 9,
869 .variableMultisampleRate = true,
870 .inheritedQueries = true,
871 };
872
873 /* We can't do image stores in vec4 shaders */
874 pFeatures->vertexPipelineStoresAndAtomics =
875 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
876 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
877
878 struct anv_app_info *app_info = &pdevice->instance->app_info;
879
880 /* The new DOOM and Wolfenstein games require depthBounds without
881 * checking for it. They seem to run fine without it so just claim it's
882 * there and accept the consequences.
883 */
884 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
885 pFeatures->depthBounds = true;
886 }
887
888 void anv_GetPhysicalDeviceFeatures2(
889 VkPhysicalDevice physicalDevice,
890 VkPhysicalDeviceFeatures2* pFeatures)
891 {
892 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
893
894 vk_foreach_struct(ext, pFeatures->pNext) {
895 switch (ext->sType) {
896 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
897 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
898 features->protectedMemory = VK_FALSE;
899 break;
900 }
901
902 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
903 VkPhysicalDeviceMultiviewFeatures *features =
904 (VkPhysicalDeviceMultiviewFeatures *)ext;
905 features->multiview = true;
906 features->multiviewGeometryShader = true;
907 features->multiviewTessellationShader = true;
908 break;
909 }
910
911 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES: {
912 VkPhysicalDeviceVariablePointerFeatures *features = (void *)ext;
913 features->variablePointersStorageBuffer = true;
914 features->variablePointers = true;
915 break;
916 }
917
918 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
919 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
920 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
921 features->samplerYcbcrConversion = true;
922 break;
923 }
924
925 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
926 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
927 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
928 features->scalarBlockLayout = true;
929 break;
930 }
931
932 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES: {
933 VkPhysicalDeviceShaderDrawParameterFeatures *features = (void *)ext;
934 features->shaderDrawParameters = true;
935 break;
936 }
937
938 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR: {
939 VkPhysicalDevice16BitStorageFeaturesKHR *features =
940 (VkPhysicalDevice16BitStorageFeaturesKHR *)ext;
941 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
942
943 features->storageBuffer16BitAccess = pdevice->info.gen >= 8;
944 features->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
945 features->storagePushConstant16 = pdevice->info.gen >= 8;
946 features->storageInputOutput16 = false;
947 break;
948 }
949
950 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
951 VkPhysicalDevice8BitStorageFeaturesKHR *features =
952 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
953 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
954
955 features->storageBuffer8BitAccess = pdevice->info.gen >= 8;
956 features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
957 features->storagePushConstant8 = pdevice->info.gen >= 8;
958 break;
959 }
960
961 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
962 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
963 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
964 features->vertexAttributeInstanceRateDivisor = VK_TRUE;
965 features->vertexAttributeInstanceRateZeroDivisor = VK_TRUE;
966 break;
967 }
968
969 default:
970 anv_debug_ignored_stype(ext->sType);
971 break;
972 }
973 }
974 }
975
976 void anv_GetPhysicalDeviceProperties(
977 VkPhysicalDevice physicalDevice,
978 VkPhysicalDeviceProperties* pProperties)
979 {
980 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
981 const struct gen_device_info *devinfo = &pdevice->info;
982
983 /* See assertions made when programming the buffer surface state. */
984 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
985 (1ul << 30) : (1ul << 27);
986
987 const uint32_t max_samplers = (devinfo->gen >= 8 || devinfo->is_haswell) ?
988 128 : 16;
989
990 VkSampleCountFlags sample_counts =
991 isl_device_get_sample_counts(&pdevice->isl_dev);
992
993 VkPhysicalDeviceLimits limits = {
994 .maxImageDimension1D = (1 << 14),
995 .maxImageDimension2D = (1 << 14),
996 .maxImageDimension3D = (1 << 11),
997 .maxImageDimensionCube = (1 << 14),
998 .maxImageArrayLayers = (1 << 11),
999 .maxTexelBufferElements = 128 * 1024 * 1024,
1000 .maxUniformBufferRange = (1ul << 27),
1001 .maxStorageBufferRange = max_raw_buffer_sz,
1002 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1003 .maxMemoryAllocationCount = UINT32_MAX,
1004 .maxSamplerAllocationCount = 64 * 1024,
1005 .bufferImageGranularity = 64, /* A cache line */
1006 .sparseAddressSpaceSize = 0,
1007 .maxBoundDescriptorSets = MAX_SETS,
1008 .maxPerStageDescriptorSamplers = max_samplers,
1009 .maxPerStageDescriptorUniformBuffers = 64,
1010 .maxPerStageDescriptorStorageBuffers = 64,
1011 .maxPerStageDescriptorSampledImages = max_samplers,
1012 .maxPerStageDescriptorStorageImages = 64,
1013 .maxPerStageDescriptorInputAttachments = 64,
1014 .maxPerStageResources = 250,
1015 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1016 .maxDescriptorSetUniformBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
1017 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1018 .maxDescriptorSetStorageBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
1019 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1020 .maxDescriptorSetSampledImages = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSampledImages */
1021 .maxDescriptorSetStorageImages = 6 * 64, /* number of stages * maxPerStageDescriptorStorageImages */
1022 .maxDescriptorSetInputAttachments = 256,
1023 .maxVertexInputAttributes = MAX_VBS,
1024 .maxVertexInputBindings = MAX_VBS,
1025 .maxVertexInputAttributeOffset = 2047,
1026 .maxVertexInputBindingStride = 2048,
1027 .maxVertexOutputComponents = 128,
1028 .maxTessellationGenerationLevel = 64,
1029 .maxTessellationPatchSize = 32,
1030 .maxTessellationControlPerVertexInputComponents = 128,
1031 .maxTessellationControlPerVertexOutputComponents = 128,
1032 .maxTessellationControlPerPatchOutputComponents = 128,
1033 .maxTessellationControlTotalOutputComponents = 2048,
1034 .maxTessellationEvaluationInputComponents = 128,
1035 .maxTessellationEvaluationOutputComponents = 128,
1036 .maxGeometryShaderInvocations = 32,
1037 .maxGeometryInputComponents = 64,
1038 .maxGeometryOutputComponents = 128,
1039 .maxGeometryOutputVertices = 256,
1040 .maxGeometryTotalOutputComponents = 1024,
1041 .maxFragmentInputComponents = 112, /* 128 components - (POS, PSIZ, CLIP_DIST0, CLIP_DIST1) */
1042 .maxFragmentOutputAttachments = 8,
1043 .maxFragmentDualSrcAttachments = 1,
1044 .maxFragmentCombinedOutputResources = 8,
1045 .maxComputeSharedMemorySize = 32768,
1046 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1047 .maxComputeWorkGroupInvocations = 16 * devinfo->max_cs_threads,
1048 .maxComputeWorkGroupSize = {
1049 16 * devinfo->max_cs_threads,
1050 16 * devinfo->max_cs_threads,
1051 16 * devinfo->max_cs_threads,
1052 },
1053 .subPixelPrecisionBits = 4 /* FIXME */,
1054 .subTexelPrecisionBits = 4 /* FIXME */,
1055 .mipmapPrecisionBits = 4 /* FIXME */,
1056 .maxDrawIndexedIndexValue = UINT32_MAX,
1057 .maxDrawIndirectCount = UINT32_MAX,
1058 .maxSamplerLodBias = 16,
1059 .maxSamplerAnisotropy = 16,
1060 .maxViewports = MAX_VIEWPORTS,
1061 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1062 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1063 .viewportSubPixelBits = 13, /* We take a float? */
1064 .minMemoryMapAlignment = 4096, /* A page */
1065 .minTexelBufferOffsetAlignment = 1,
1066 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1067 .minUniformBufferOffsetAlignment = 32,
1068 .minStorageBufferOffsetAlignment = 4,
1069 .minTexelOffset = -8,
1070 .maxTexelOffset = 7,
1071 .minTexelGatherOffset = -32,
1072 .maxTexelGatherOffset = 31,
1073 .minInterpolationOffset = -0.5,
1074 .maxInterpolationOffset = 0.4375,
1075 .subPixelInterpolationOffsetBits = 4,
1076 .maxFramebufferWidth = (1 << 14),
1077 .maxFramebufferHeight = (1 << 14),
1078 .maxFramebufferLayers = (1 << 11),
1079 .framebufferColorSampleCounts = sample_counts,
1080 .framebufferDepthSampleCounts = sample_counts,
1081 .framebufferStencilSampleCounts = sample_counts,
1082 .framebufferNoAttachmentsSampleCounts = sample_counts,
1083 .maxColorAttachments = MAX_RTS,
1084 .sampledImageColorSampleCounts = sample_counts,
1085 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1086 .sampledImageDepthSampleCounts = sample_counts,
1087 .sampledImageStencilSampleCounts = sample_counts,
1088 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1089 .maxSampleMaskWords = 1,
1090 .timestampComputeAndGraphics = false,
1091 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1092 .maxClipDistances = 8,
1093 .maxCullDistances = 8,
1094 .maxCombinedClipAndCullDistances = 8,
1095 .discreteQueuePriorities = 2,
1096 .pointSizeRange = { 0.125, 255.875 },
1097 .lineWidthRange = { 0.0, 7.9921875 },
1098 .pointSizeGranularity = (1.0 / 8.0),
1099 .lineWidthGranularity = (1.0 / 128.0),
1100 .strictLines = false, /* FINISHME */
1101 .standardSampleLocations = true,
1102 .optimalBufferCopyOffsetAlignment = 128,
1103 .optimalBufferCopyRowPitchAlignment = 128,
1104 .nonCoherentAtomSize = 64,
1105 };
1106
1107 *pProperties = (VkPhysicalDeviceProperties) {
1108 .apiVersion = anv_physical_device_api_version(pdevice),
1109 .driverVersion = vk_get_driver_version(),
1110 .vendorID = 0x8086,
1111 .deviceID = pdevice->chipset_id,
1112 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1113 .limits = limits,
1114 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1115 };
1116
1117 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1118 "%s", pdevice->name);
1119 memcpy(pProperties->pipelineCacheUUID,
1120 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1121 }
1122
1123 void anv_GetPhysicalDeviceProperties2(
1124 VkPhysicalDevice physicalDevice,
1125 VkPhysicalDeviceProperties2* pProperties)
1126 {
1127 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1128
1129 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1130
1131 vk_foreach_struct(ext, pProperties->pNext) {
1132 switch (ext->sType) {
1133 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1134 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1135 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1136
1137 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1138 break;
1139 }
1140
1141 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1142 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1143 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1144
1145 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1146 util_snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1147 "Intel open-source Mesa driver");
1148
1149 util_snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1150 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1151
1152 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1153 .major = 1,
1154 .minor = 1,
1155 .subminor = 2,
1156 .patch = 0,
1157 };
1158 break;
1159 }
1160
1161 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1162 VkPhysicalDeviceIDProperties *id_props =
1163 (VkPhysicalDeviceIDProperties *)ext;
1164 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1165 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1166 /* The LUID is for Windows. */
1167 id_props->deviceLUIDValid = false;
1168 break;
1169 }
1170
1171 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1172 VkPhysicalDeviceMaintenance3Properties *props =
1173 (VkPhysicalDeviceMaintenance3Properties *)ext;
1174 /* This value doesn't matter for us today as our per-stage
1175 * descriptors are the real limit.
1176 */
1177 props->maxPerSetDescriptors = 1024;
1178 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1179 break;
1180 }
1181
1182 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1183 VkPhysicalDeviceMultiviewProperties *properties =
1184 (VkPhysicalDeviceMultiviewProperties *)ext;
1185 properties->maxMultiviewViewCount = 16;
1186 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1187 break;
1188 }
1189
1190 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1191 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1192 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1193 properties->pciDomain = pdevice->pci_info.domain;
1194 properties->pciBus = pdevice->pci_info.bus;
1195 properties->pciDevice = pdevice->pci_info.device;
1196 properties->pciFunction = pdevice->pci_info.function;
1197 break;
1198 }
1199
1200 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1201 VkPhysicalDevicePointClippingProperties *properties =
1202 (VkPhysicalDevicePointClippingProperties *) ext;
1203 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
1204 anv_finishme("Implement pop-free point clipping");
1205 break;
1206 }
1207
1208 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1209 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1210 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1211 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1212 properties->filterMinmaxSingleComponentFormats = true;
1213 break;
1214 }
1215
1216 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1217 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1218
1219 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1220
1221 VkShaderStageFlags scalar_stages = 0;
1222 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1223 if (pdevice->compiler->scalar_stage[stage])
1224 scalar_stages |= mesa_to_vk_shader_stage(stage);
1225 }
1226 properties->supportedStages = scalar_stages;
1227
1228 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1229 VK_SUBGROUP_FEATURE_VOTE_BIT |
1230 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1231 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1232 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1233 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1234 VK_SUBGROUP_FEATURE_CLUSTERED_BIT |
1235 VK_SUBGROUP_FEATURE_QUAD_BIT;
1236 properties->quadOperationsInAllStages = VK_TRUE;
1237 break;
1238 }
1239
1240 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1241 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1242 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1243 /* We have to restrict this a bit for multiview */
1244 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1245 break;
1246 }
1247
1248 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1249 VkPhysicalDeviceProtectedMemoryProperties *props =
1250 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1251 props->protectedNoFault = false;
1252 break;
1253 }
1254
1255 default:
1256 anv_debug_ignored_stype(ext->sType);
1257 break;
1258 }
1259 }
1260 }
1261
1262 /* We support exactly one queue family. */
1263 static const VkQueueFamilyProperties
1264 anv_queue_family_properties = {
1265 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1266 VK_QUEUE_COMPUTE_BIT |
1267 VK_QUEUE_TRANSFER_BIT,
1268 .queueCount = 1,
1269 .timestampValidBits = 36, /* XXX: Real value here */
1270 .minImageTransferGranularity = { 1, 1, 1 },
1271 };
1272
1273 void anv_GetPhysicalDeviceQueueFamilyProperties(
1274 VkPhysicalDevice physicalDevice,
1275 uint32_t* pCount,
1276 VkQueueFamilyProperties* pQueueFamilyProperties)
1277 {
1278 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1279
1280 vk_outarray_append(&out, p) {
1281 *p = anv_queue_family_properties;
1282 }
1283 }
1284
1285 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1286 VkPhysicalDevice physicalDevice,
1287 uint32_t* pQueueFamilyPropertyCount,
1288 VkQueueFamilyProperties2* pQueueFamilyProperties)
1289 {
1290
1291 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1292
1293 vk_outarray_append(&out, p) {
1294 p->queueFamilyProperties = anv_queue_family_properties;
1295
1296 vk_foreach_struct(s, p->pNext) {
1297 anv_debug_ignored_stype(s->sType);
1298 }
1299 }
1300 }
1301
1302 void anv_GetPhysicalDeviceMemoryProperties(
1303 VkPhysicalDevice physicalDevice,
1304 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1305 {
1306 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1307
1308 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1309 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1310 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1311 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1312 .heapIndex = physical_device->memory.types[i].heapIndex,
1313 };
1314 }
1315
1316 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1317 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1318 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1319 .size = physical_device->memory.heaps[i].size,
1320 .flags = physical_device->memory.heaps[i].flags,
1321 };
1322 }
1323 }
1324
1325 void anv_GetPhysicalDeviceMemoryProperties2(
1326 VkPhysicalDevice physicalDevice,
1327 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1328 {
1329 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1330 &pMemoryProperties->memoryProperties);
1331
1332 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1333 switch (ext->sType) {
1334 default:
1335 anv_debug_ignored_stype(ext->sType);
1336 break;
1337 }
1338 }
1339 }
1340
1341 void
1342 anv_GetDeviceGroupPeerMemoryFeatures(
1343 VkDevice device,
1344 uint32_t heapIndex,
1345 uint32_t localDeviceIndex,
1346 uint32_t remoteDeviceIndex,
1347 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1348 {
1349 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1350 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1351 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1352 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1353 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1354 }
1355
1356 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1357 VkInstance _instance,
1358 const char* pName)
1359 {
1360 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1361
1362 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1363 * when we have to return valid function pointers, NULL, or it's left
1364 * undefined. See the table for exact details.
1365 */
1366 if (pName == NULL)
1367 return NULL;
1368
1369 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1370 if (strcmp(pName, "vk" #entrypoint) == 0) \
1371 return (PFN_vkVoidFunction)anv_##entrypoint
1372
1373 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1374 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1375 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1376 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1377
1378 #undef LOOKUP_ANV_ENTRYPOINT
1379
1380 if (instance == NULL)
1381 return NULL;
1382
1383 int idx = anv_get_instance_entrypoint_index(pName);
1384 if (idx >= 0)
1385 return instance->dispatch.entrypoints[idx];
1386
1387 idx = anv_get_device_entrypoint_index(pName);
1388 if (idx >= 0)
1389 return instance->device_dispatch.entrypoints[idx];
1390
1391 return NULL;
1392 }
1393
1394 /* With version 1+ of the loader interface the ICD should expose
1395 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1396 */
1397 PUBLIC
1398 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1399 VkInstance instance,
1400 const char* pName);
1401
1402 PUBLIC
1403 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1404 VkInstance instance,
1405 const char* pName)
1406 {
1407 return anv_GetInstanceProcAddr(instance, pName);
1408 }
1409
1410 PFN_vkVoidFunction anv_GetDeviceProcAddr(
1411 VkDevice _device,
1412 const char* pName)
1413 {
1414 ANV_FROM_HANDLE(anv_device, device, _device);
1415
1416 if (!device || !pName)
1417 return NULL;
1418
1419 int idx = anv_get_device_entrypoint_index(pName);
1420 if (idx < 0)
1421 return NULL;
1422
1423 return device->dispatch.entrypoints[idx];
1424 }
1425
1426 VkResult
1427 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
1428 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
1429 const VkAllocationCallbacks* pAllocator,
1430 VkDebugReportCallbackEXT* pCallback)
1431 {
1432 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1433 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
1434 pCreateInfo, pAllocator, &instance->alloc,
1435 pCallback);
1436 }
1437
1438 void
1439 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
1440 VkDebugReportCallbackEXT _callback,
1441 const VkAllocationCallbacks* pAllocator)
1442 {
1443 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1444 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
1445 _callback, pAllocator, &instance->alloc);
1446 }
1447
1448 void
1449 anv_DebugReportMessageEXT(VkInstance _instance,
1450 VkDebugReportFlagsEXT flags,
1451 VkDebugReportObjectTypeEXT objectType,
1452 uint64_t object,
1453 size_t location,
1454 int32_t messageCode,
1455 const char* pLayerPrefix,
1456 const char* pMessage)
1457 {
1458 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1459 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
1460 object, location, messageCode, pLayerPrefix, pMessage);
1461 }
1462
1463 static void
1464 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
1465 {
1466 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1467 queue->device = device;
1468 queue->flags = 0;
1469 }
1470
1471 static void
1472 anv_queue_finish(struct anv_queue *queue)
1473 {
1474 }
1475
1476 static struct anv_state
1477 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
1478 {
1479 struct anv_state state;
1480
1481 state = anv_state_pool_alloc(pool, size, align);
1482 memcpy(state.map, p, size);
1483
1484 anv_state_flush(pool->block_pool.device, state);
1485
1486 return state;
1487 }
1488
1489 struct gen8_border_color {
1490 union {
1491 float float32[4];
1492 uint32_t uint32[4];
1493 };
1494 /* Pad out to 64 bytes */
1495 uint32_t _pad[12];
1496 };
1497
1498 static void
1499 anv_device_init_border_colors(struct anv_device *device)
1500 {
1501 static const struct gen8_border_color border_colors[] = {
1502 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
1503 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
1504 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
1505 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
1506 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
1507 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
1508 };
1509
1510 device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
1511 sizeof(border_colors), 64,
1512 border_colors);
1513 }
1514
1515 static void
1516 anv_device_init_trivial_batch(struct anv_device *device)
1517 {
1518 anv_bo_init_new(&device->trivial_batch_bo, device, 4096);
1519
1520 if (device->instance->physicalDevice.has_exec_async)
1521 device->trivial_batch_bo.flags |= EXEC_OBJECT_ASYNC;
1522
1523 if (device->instance->physicalDevice.use_softpin)
1524 device->trivial_batch_bo.flags |= EXEC_OBJECT_PINNED;
1525
1526 anv_vma_alloc(device, &device->trivial_batch_bo);
1527
1528 void *map = anv_gem_mmap(device, device->trivial_batch_bo.gem_handle,
1529 0, 4096, 0);
1530
1531 struct anv_batch batch = {
1532 .start = map,
1533 .next = map,
1534 .end = map + 4096,
1535 };
1536
1537 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1538 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1539
1540 if (!device->info.has_llc)
1541 gen_clflush_range(map, batch.next - map);
1542
1543 anv_gem_munmap(map, device->trivial_batch_bo.size);
1544 }
1545
1546 VkResult anv_EnumerateDeviceExtensionProperties(
1547 VkPhysicalDevice physicalDevice,
1548 const char* pLayerName,
1549 uint32_t* pPropertyCount,
1550 VkExtensionProperties* pProperties)
1551 {
1552 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1553 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1554
1555 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
1556 if (device->supported_extensions.extensions[i]) {
1557 vk_outarray_append(&out, prop) {
1558 *prop = anv_device_extensions[i];
1559 }
1560 }
1561 }
1562
1563 return vk_outarray_status(&out);
1564 }
1565
1566 static void
1567 anv_device_init_dispatch(struct anv_device *device)
1568 {
1569 const struct anv_device_dispatch_table *genX_table;
1570 switch (device->info.gen) {
1571 case 11:
1572 genX_table = &gen11_device_dispatch_table;
1573 break;
1574 case 10:
1575 genX_table = &gen10_device_dispatch_table;
1576 break;
1577 case 9:
1578 genX_table = &gen9_device_dispatch_table;
1579 break;
1580 case 8:
1581 genX_table = &gen8_device_dispatch_table;
1582 break;
1583 case 7:
1584 if (device->info.is_haswell)
1585 genX_table = &gen75_device_dispatch_table;
1586 else
1587 genX_table = &gen7_device_dispatch_table;
1588 break;
1589 default:
1590 unreachable("unsupported gen\n");
1591 }
1592
1593 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
1594 /* Vulkan requires that entrypoints for extensions which have not been
1595 * enabled must not be advertised.
1596 */
1597 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
1598 &device->instance->enabled_extensions,
1599 &device->enabled_extensions)) {
1600 device->dispatch.entrypoints[i] = NULL;
1601 } else if (genX_table->entrypoints[i]) {
1602 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
1603 } else {
1604 device->dispatch.entrypoints[i] =
1605 anv_device_dispatch_table.entrypoints[i];
1606 }
1607 }
1608 }
1609
1610 static int
1611 vk_priority_to_gen(int priority)
1612 {
1613 switch (priority) {
1614 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
1615 return GEN_CONTEXT_LOW_PRIORITY;
1616 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
1617 return GEN_CONTEXT_MEDIUM_PRIORITY;
1618 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
1619 return GEN_CONTEXT_HIGH_PRIORITY;
1620 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
1621 return GEN_CONTEXT_REALTIME_PRIORITY;
1622 default:
1623 unreachable("Invalid priority");
1624 }
1625 }
1626
1627 static void
1628 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
1629 {
1630 anv_bo_init_new(&device->hiz_clear_bo, device, 4096);
1631
1632 if (device->instance->physicalDevice.has_exec_async)
1633 device->hiz_clear_bo.flags |= EXEC_OBJECT_ASYNC;
1634
1635 if (device->instance->physicalDevice.use_softpin)
1636 device->hiz_clear_bo.flags |= EXEC_OBJECT_PINNED;
1637
1638 anv_vma_alloc(device, &device->hiz_clear_bo);
1639
1640 uint32_t *map = anv_gem_mmap(device, device->hiz_clear_bo.gem_handle,
1641 0, 4096, 0);
1642
1643 union isl_color_value hiz_clear = { .u32 = { 0, } };
1644 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
1645
1646 memcpy(map, hiz_clear.u32, sizeof(hiz_clear.u32));
1647 anv_gem_munmap(map, device->hiz_clear_bo.size);
1648 }
1649
1650 VkResult anv_CreateDevice(
1651 VkPhysicalDevice physicalDevice,
1652 const VkDeviceCreateInfo* pCreateInfo,
1653 const VkAllocationCallbacks* pAllocator,
1654 VkDevice* pDevice)
1655 {
1656 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1657 VkResult result;
1658 struct anv_device *device;
1659
1660 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
1661
1662 struct anv_device_extension_table enabled_extensions = { };
1663 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1664 int idx;
1665 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
1666 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
1667 anv_device_extensions[idx].extensionName) == 0)
1668 break;
1669 }
1670
1671 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
1672 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1673
1674 if (!physical_device->supported_extensions.extensions[idx])
1675 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1676
1677 enabled_extensions.extensions[idx] = true;
1678 }
1679
1680 /* Check enabled features */
1681 if (pCreateInfo->pEnabledFeatures) {
1682 VkPhysicalDeviceFeatures supported_features;
1683 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1684 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
1685 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
1686 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1687 for (uint32_t i = 0; i < num_features; i++) {
1688 if (enabled_feature[i] && !supported_feature[i])
1689 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1690 }
1691 }
1692
1693 /* Check requested queues and fail if we are requested to create any
1694 * queues with flags we don't support.
1695 */
1696 assert(pCreateInfo->queueCreateInfoCount > 0);
1697 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1698 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
1699 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
1700 }
1701
1702 /* Check if client specified queue priority. */
1703 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
1704 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
1705 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
1706
1707 VkQueueGlobalPriorityEXT priority =
1708 queue_priority ? queue_priority->globalPriority :
1709 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
1710
1711 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
1712 sizeof(*device), 8,
1713 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1714 if (!device)
1715 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1716
1717 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1718 device->instance = physical_device->instance;
1719 device->chipset_id = physical_device->chipset_id;
1720 device->no_hw = physical_device->no_hw;
1721 device->_lost = false;
1722
1723 if (pAllocator)
1724 device->alloc = *pAllocator;
1725 else
1726 device->alloc = physical_device->instance->alloc;
1727
1728 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1729 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1730 if (device->fd == -1) {
1731 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1732 goto fail_device;
1733 }
1734
1735 device->context_id = anv_gem_create_context(device);
1736 if (device->context_id == -1) {
1737 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1738 goto fail_fd;
1739 }
1740
1741 if (physical_device->use_softpin) {
1742 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
1743 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1744 goto fail_fd;
1745 }
1746
1747 /* keep the page with address zero out of the allocator */
1748 util_vma_heap_init(&device->vma_lo, LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
1749 device->vma_lo_available =
1750 physical_device->memory.heaps[physical_device->memory.heap_count - 1].size;
1751
1752 /* Leave the last 4GiB out of the high vma range, so that no state base
1753 * address + size can overflow 48 bits. For more information see the
1754 * comment about Wa32bitGeneralStateOffset in anv_allocator.c
1755 */
1756 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
1757 HIGH_HEAP_SIZE);
1758 device->vma_hi_available = physical_device->memory.heap_count == 1 ? 0 :
1759 physical_device->memory.heaps[0].size;
1760 }
1761
1762 /* As per spec, the driver implementation may deny requests to acquire
1763 * a priority above the default priority (MEDIUM) if the caller does not
1764 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1765 * is returned.
1766 */
1767 if (physical_device->has_context_priority) {
1768 int err = anv_gem_set_context_param(device->fd, device->context_id,
1769 I915_CONTEXT_PARAM_PRIORITY,
1770 vk_priority_to_gen(priority));
1771 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
1772 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
1773 goto fail_fd;
1774 }
1775 }
1776
1777 device->info = physical_device->info;
1778 device->isl_dev = physical_device->isl_dev;
1779
1780 /* On Broadwell and later, we can use batch chaining to more efficiently
1781 * implement growing command buffers. Prior to Haswell, the kernel
1782 * command parser gets in the way and we have to fall back to growing
1783 * the batch.
1784 */
1785 device->can_chain_batches = device->info.gen >= 8;
1786
1787 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1788 pCreateInfo->pEnabledFeatures->robustBufferAccess;
1789 device->enabled_extensions = enabled_extensions;
1790
1791 anv_device_init_dispatch(device);
1792
1793 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
1794 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1795 goto fail_context_id;
1796 }
1797
1798 pthread_condattr_t condattr;
1799 if (pthread_condattr_init(&condattr) != 0) {
1800 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1801 goto fail_mutex;
1802 }
1803 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
1804 pthread_condattr_destroy(&condattr);
1805 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1806 goto fail_mutex;
1807 }
1808 if (pthread_cond_init(&device->queue_submit, NULL) != 0) {
1809 pthread_condattr_destroy(&condattr);
1810 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1811 goto fail_mutex;
1812 }
1813 pthread_condattr_destroy(&condattr);
1814
1815 uint64_t bo_flags =
1816 (physical_device->supports_48bit_addresses ? EXEC_OBJECT_SUPPORTS_48B_ADDRESS : 0) |
1817 (physical_device->has_exec_async ? EXEC_OBJECT_ASYNC : 0) |
1818 (physical_device->has_exec_capture ? EXEC_OBJECT_CAPTURE : 0) |
1819 (physical_device->use_softpin ? EXEC_OBJECT_PINNED : 0);
1820
1821 anv_bo_pool_init(&device->batch_bo_pool, device, bo_flags);
1822
1823 result = anv_bo_cache_init(&device->bo_cache);
1824 if (result != VK_SUCCESS)
1825 goto fail_batch_bo_pool;
1826
1827 if (!physical_device->use_softpin)
1828 bo_flags &= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
1829
1830 result = anv_state_pool_init(&device->dynamic_state_pool, device,
1831 DYNAMIC_STATE_POOL_MIN_ADDRESS,
1832 16384,
1833 bo_flags);
1834 if (result != VK_SUCCESS)
1835 goto fail_bo_cache;
1836
1837 result = anv_state_pool_init(&device->instruction_state_pool, device,
1838 INSTRUCTION_STATE_POOL_MIN_ADDRESS,
1839 16384,
1840 bo_flags);
1841 if (result != VK_SUCCESS)
1842 goto fail_dynamic_state_pool;
1843
1844 result = anv_state_pool_init(&device->surface_state_pool, device,
1845 SURFACE_STATE_POOL_MIN_ADDRESS,
1846 4096,
1847 bo_flags);
1848 if (result != VK_SUCCESS)
1849 goto fail_instruction_state_pool;
1850
1851 if (physical_device->use_softpin) {
1852 result = anv_state_pool_init(&device->binding_table_pool, device,
1853 BINDING_TABLE_POOL_MIN_ADDRESS,
1854 4096,
1855 bo_flags);
1856 if (result != VK_SUCCESS)
1857 goto fail_surface_state_pool;
1858 }
1859
1860 result = anv_bo_init_new(&device->workaround_bo, device, 1024);
1861 if (result != VK_SUCCESS)
1862 goto fail_binding_table_pool;
1863
1864 if (physical_device->use_softpin)
1865 device->workaround_bo.flags |= EXEC_OBJECT_PINNED;
1866
1867 if (!anv_vma_alloc(device, &device->workaround_bo))
1868 goto fail_workaround_bo;
1869
1870 anv_device_init_trivial_batch(device);
1871
1872 if (device->info.gen >= 10)
1873 anv_device_init_hiz_clear_value_bo(device);
1874
1875 anv_scratch_pool_init(device, &device->scratch_pool);
1876
1877 anv_queue_init(device, &device->queue);
1878
1879 switch (device->info.gen) {
1880 case 7:
1881 if (!device->info.is_haswell)
1882 result = gen7_init_device_state(device);
1883 else
1884 result = gen75_init_device_state(device);
1885 break;
1886 case 8:
1887 result = gen8_init_device_state(device);
1888 break;
1889 case 9:
1890 result = gen9_init_device_state(device);
1891 break;
1892 case 10:
1893 result = gen10_init_device_state(device);
1894 break;
1895 case 11:
1896 result = gen11_init_device_state(device);
1897 break;
1898 default:
1899 /* Shouldn't get here as we don't create physical devices for any other
1900 * gens. */
1901 unreachable("unhandled gen");
1902 }
1903 if (result != VK_SUCCESS)
1904 goto fail_workaround_bo;
1905
1906 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
1907
1908 anv_device_init_blorp(device);
1909
1910 anv_device_init_border_colors(device);
1911
1912 *pDevice = anv_device_to_handle(device);
1913
1914 return VK_SUCCESS;
1915
1916 fail_workaround_bo:
1917 anv_queue_finish(&device->queue);
1918 anv_scratch_pool_finish(device, &device->scratch_pool);
1919 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1920 anv_gem_close(device, device->workaround_bo.gem_handle);
1921 fail_binding_table_pool:
1922 if (physical_device->use_softpin)
1923 anv_state_pool_finish(&device->binding_table_pool);
1924 fail_surface_state_pool:
1925 anv_state_pool_finish(&device->surface_state_pool);
1926 fail_instruction_state_pool:
1927 anv_state_pool_finish(&device->instruction_state_pool);
1928 fail_dynamic_state_pool:
1929 anv_state_pool_finish(&device->dynamic_state_pool);
1930 fail_bo_cache:
1931 anv_bo_cache_finish(&device->bo_cache);
1932 fail_batch_bo_pool:
1933 anv_bo_pool_finish(&device->batch_bo_pool);
1934 pthread_cond_destroy(&device->queue_submit);
1935 fail_mutex:
1936 pthread_mutex_destroy(&device->mutex);
1937 fail_context_id:
1938 anv_gem_destroy_context(device, device->context_id);
1939 fail_fd:
1940 close(device->fd);
1941 fail_device:
1942 vk_free(&device->alloc, device);
1943
1944 return result;
1945 }
1946
1947 void anv_DestroyDevice(
1948 VkDevice _device,
1949 const VkAllocationCallbacks* pAllocator)
1950 {
1951 ANV_FROM_HANDLE(anv_device, device, _device);
1952 struct anv_physical_device *physical_device;
1953
1954 if (!device)
1955 return;
1956
1957 physical_device = &device->instance->physicalDevice;
1958
1959 anv_device_finish_blorp(device);
1960
1961 anv_pipeline_cache_finish(&device->default_pipeline_cache);
1962
1963 anv_queue_finish(&device->queue);
1964
1965 #ifdef HAVE_VALGRIND
1966 /* We only need to free these to prevent valgrind errors. The backing
1967 * BO will go away in a couple of lines so we don't actually leak.
1968 */
1969 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1970 #endif
1971
1972 anv_scratch_pool_finish(device, &device->scratch_pool);
1973
1974 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1975 anv_vma_free(device, &device->workaround_bo);
1976 anv_gem_close(device, device->workaround_bo.gem_handle);
1977
1978 anv_vma_free(device, &device->trivial_batch_bo);
1979 anv_gem_close(device, device->trivial_batch_bo.gem_handle);
1980 if (device->info.gen >= 10)
1981 anv_gem_close(device, device->hiz_clear_bo.gem_handle);
1982
1983 if (physical_device->use_softpin)
1984 anv_state_pool_finish(&device->binding_table_pool);
1985 anv_state_pool_finish(&device->surface_state_pool);
1986 anv_state_pool_finish(&device->instruction_state_pool);
1987 anv_state_pool_finish(&device->dynamic_state_pool);
1988
1989 anv_bo_cache_finish(&device->bo_cache);
1990
1991 anv_bo_pool_finish(&device->batch_bo_pool);
1992
1993 pthread_cond_destroy(&device->queue_submit);
1994 pthread_mutex_destroy(&device->mutex);
1995
1996 anv_gem_destroy_context(device, device->context_id);
1997
1998 close(device->fd);
1999
2000 vk_free(&device->alloc, device);
2001 }
2002
2003 VkResult anv_EnumerateInstanceLayerProperties(
2004 uint32_t* pPropertyCount,
2005 VkLayerProperties* pProperties)
2006 {
2007 if (pProperties == NULL) {
2008 *pPropertyCount = 0;
2009 return VK_SUCCESS;
2010 }
2011
2012 /* None supported at this time */
2013 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2014 }
2015
2016 VkResult anv_EnumerateDeviceLayerProperties(
2017 VkPhysicalDevice physicalDevice,
2018 uint32_t* pPropertyCount,
2019 VkLayerProperties* pProperties)
2020 {
2021 if (pProperties == NULL) {
2022 *pPropertyCount = 0;
2023 return VK_SUCCESS;
2024 }
2025
2026 /* None supported at this time */
2027 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2028 }
2029
2030 void anv_GetDeviceQueue(
2031 VkDevice _device,
2032 uint32_t queueNodeIndex,
2033 uint32_t queueIndex,
2034 VkQueue* pQueue)
2035 {
2036 ANV_FROM_HANDLE(anv_device, device, _device);
2037
2038 assert(queueIndex == 0);
2039
2040 *pQueue = anv_queue_to_handle(&device->queue);
2041 }
2042
2043 void anv_GetDeviceQueue2(
2044 VkDevice _device,
2045 const VkDeviceQueueInfo2* pQueueInfo,
2046 VkQueue* pQueue)
2047 {
2048 ANV_FROM_HANDLE(anv_device, device, _device);
2049
2050 assert(pQueueInfo->queueIndex == 0);
2051
2052 if (pQueueInfo->flags == device->queue.flags)
2053 *pQueue = anv_queue_to_handle(&device->queue);
2054 else
2055 *pQueue = NULL;
2056 }
2057
2058 VkResult
2059 _anv_device_set_lost(struct anv_device *device,
2060 const char *file, int line,
2061 const char *msg, ...)
2062 {
2063 VkResult err;
2064 va_list ap;
2065
2066 device->_lost = true;
2067
2068 va_start(ap, msg);
2069 err = __vk_errorv(device->instance, device,
2070 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2071 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2072 va_end(ap);
2073
2074 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2075 abort();
2076
2077 return err;
2078 }
2079
2080 VkResult
2081 anv_device_query_status(struct anv_device *device)
2082 {
2083 /* This isn't likely as most of the callers of this function already check
2084 * for it. However, it doesn't hurt to check and it potentially lets us
2085 * avoid an ioctl.
2086 */
2087 if (anv_device_is_lost(device))
2088 return VK_ERROR_DEVICE_LOST;
2089
2090 uint32_t active, pending;
2091 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2092 if (ret == -1) {
2093 /* We don't know the real error. */
2094 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2095 }
2096
2097 if (active) {
2098 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2099 } else if (pending) {
2100 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2101 }
2102
2103 return VK_SUCCESS;
2104 }
2105
2106 VkResult
2107 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2108 {
2109 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2110 * Other usages of the BO (such as on different hardware) will not be
2111 * flagged as "busy" by this ioctl. Use with care.
2112 */
2113 int ret = anv_gem_busy(device, bo->gem_handle);
2114 if (ret == 1) {
2115 return VK_NOT_READY;
2116 } else if (ret == -1) {
2117 /* We don't know the real error. */
2118 return anv_device_set_lost(device, "gem wait failed: %m");
2119 }
2120
2121 /* Query for device status after the busy call. If the BO we're checking
2122 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2123 * client because it clearly doesn't have valid data. Yes, this most
2124 * likely means an ioctl, but we just did an ioctl to query the busy status
2125 * so it's no great loss.
2126 */
2127 return anv_device_query_status(device);
2128 }
2129
2130 VkResult
2131 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2132 int64_t timeout)
2133 {
2134 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2135 if (ret == -1 && errno == ETIME) {
2136 return VK_TIMEOUT;
2137 } else if (ret == -1) {
2138 /* We don't know the real error. */
2139 return anv_device_set_lost(device, "gem wait failed: %m");
2140 }
2141
2142 /* Query for device status after the wait. If the BO we're waiting on got
2143 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2144 * because it clearly doesn't have valid data. Yes, this most likely means
2145 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2146 */
2147 return anv_device_query_status(device);
2148 }
2149
2150 VkResult anv_DeviceWaitIdle(
2151 VkDevice _device)
2152 {
2153 ANV_FROM_HANDLE(anv_device, device, _device);
2154 if (anv_device_is_lost(device))
2155 return VK_ERROR_DEVICE_LOST;
2156
2157 struct anv_batch batch;
2158
2159 uint32_t cmds[8];
2160 batch.start = batch.next = cmds;
2161 batch.end = (void *) cmds + sizeof(cmds);
2162
2163 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2164 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2165
2166 return anv_device_submit_simple_batch(device, &batch);
2167 }
2168
2169 bool
2170 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo)
2171 {
2172 if (!(bo->flags & EXEC_OBJECT_PINNED))
2173 return true;
2174
2175 pthread_mutex_lock(&device->vma_mutex);
2176
2177 bo->offset = 0;
2178
2179 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS &&
2180 device->vma_hi_available >= bo->size) {
2181 uint64_t addr = util_vma_heap_alloc(&device->vma_hi, bo->size, 4096);
2182 if (addr) {
2183 bo->offset = gen_canonical_address(addr);
2184 assert(addr == gen_48b_address(bo->offset));
2185 device->vma_hi_available -= bo->size;
2186 }
2187 }
2188
2189 if (bo->offset == 0 && device->vma_lo_available >= bo->size) {
2190 uint64_t addr = util_vma_heap_alloc(&device->vma_lo, bo->size, 4096);
2191 if (addr) {
2192 bo->offset = gen_canonical_address(addr);
2193 assert(addr == gen_48b_address(bo->offset));
2194 device->vma_lo_available -= bo->size;
2195 }
2196 }
2197
2198 pthread_mutex_unlock(&device->vma_mutex);
2199
2200 return bo->offset != 0;
2201 }
2202
2203 void
2204 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
2205 {
2206 if (!(bo->flags & EXEC_OBJECT_PINNED))
2207 return;
2208
2209 const uint64_t addr_48b = gen_48b_address(bo->offset);
2210
2211 pthread_mutex_lock(&device->vma_mutex);
2212
2213 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
2214 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
2215 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
2216 device->vma_lo_available += bo->size;
2217 } else {
2218 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS &&
2219 addr_48b <= HIGH_HEAP_MAX_ADDRESS);
2220 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
2221 device->vma_hi_available += bo->size;
2222 }
2223
2224 pthread_mutex_unlock(&device->vma_mutex);
2225
2226 bo->offset = 0;
2227 }
2228
2229 VkResult
2230 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
2231 {
2232 uint32_t gem_handle = anv_gem_create(device, size);
2233 if (!gem_handle)
2234 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
2235
2236 anv_bo_init(bo, gem_handle, size);
2237
2238 return VK_SUCCESS;
2239 }
2240
2241 VkResult anv_AllocateMemory(
2242 VkDevice _device,
2243 const VkMemoryAllocateInfo* pAllocateInfo,
2244 const VkAllocationCallbacks* pAllocator,
2245 VkDeviceMemory* pMem)
2246 {
2247 ANV_FROM_HANDLE(anv_device, device, _device);
2248 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2249 struct anv_device_memory *mem;
2250 VkResult result = VK_SUCCESS;
2251
2252 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
2253
2254 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2255 assert(pAllocateInfo->allocationSize > 0);
2256
2257 if (pAllocateInfo->allocationSize > MAX_MEMORY_ALLOCATION_SIZE)
2258 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
2259
2260 /* FINISHME: Fail if allocation request exceeds heap size. */
2261
2262 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
2263 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2264 if (mem == NULL)
2265 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2266
2267 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
2268 mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
2269 mem->map = NULL;
2270 mem->map_size = 0;
2271
2272 uint64_t bo_flags = 0;
2273
2274 assert(mem->type->heapIndex < pdevice->memory.heap_count);
2275 if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
2276 bo_flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2277
2278 const struct wsi_memory_allocate_info *wsi_info =
2279 vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
2280 if (wsi_info && wsi_info->implicit_sync) {
2281 /* We need to set the WRITE flag on window system buffers so that GEM
2282 * will know we're writing to them and synchronize uses on other rings
2283 * (eg if the display server uses the blitter ring).
2284 */
2285 bo_flags |= EXEC_OBJECT_WRITE;
2286 } else if (pdevice->has_exec_async) {
2287 bo_flags |= EXEC_OBJECT_ASYNC;
2288 }
2289
2290 if (pdevice->use_softpin)
2291 bo_flags |= EXEC_OBJECT_PINNED;
2292
2293 const VkImportMemoryFdInfoKHR *fd_info =
2294 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
2295
2296 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2297 * ignored.
2298 */
2299 if (fd_info && fd_info->handleType) {
2300 /* At the moment, we support only the below handle types. */
2301 assert(fd_info->handleType ==
2302 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2303 fd_info->handleType ==
2304 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2305
2306 result = anv_bo_cache_import(device, &device->bo_cache, fd_info->fd,
2307 bo_flags | ANV_BO_EXTERNAL, &mem->bo);
2308 if (result != VK_SUCCESS)
2309 goto fail;
2310
2311 VkDeviceSize aligned_alloc_size =
2312 align_u64(pAllocateInfo->allocationSize, 4096);
2313
2314 /* For security purposes, we reject importing the bo if it's smaller
2315 * than the requested allocation size. This prevents a malicious client
2316 * from passing a buffer to a trusted client, lying about the size, and
2317 * telling the trusted client to try and texture from an image that goes
2318 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2319 * in the trusted client. The trusted client can protect itself against
2320 * this sort of attack but only if it can trust the buffer size.
2321 */
2322 if (mem->bo->size < aligned_alloc_size) {
2323 result = vk_errorf(device->instance, device,
2324 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR,
2325 "aligned allocationSize too large for "
2326 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
2327 "%"PRIu64"B > %"PRIu64"B",
2328 aligned_alloc_size, mem->bo->size);
2329 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2330 goto fail;
2331 }
2332
2333 /* From the Vulkan spec:
2334 *
2335 * "Importing memory from a file descriptor transfers ownership of
2336 * the file descriptor from the application to the Vulkan
2337 * implementation. The application must not perform any operations on
2338 * the file descriptor after a successful import."
2339 *
2340 * If the import fails, we leave the file descriptor open.
2341 */
2342 close(fd_info->fd);
2343 goto success;
2344 }
2345
2346 /* Regular allocate (not importing memory). */
2347
2348 const VkExportMemoryAllocateInfoKHR *export_info =
2349 vk_find_struct_const(pAllocateInfo->pNext, EXPORT_MEMORY_ALLOCATE_INFO_KHR);
2350 if (export_info && export_info->handleTypes)
2351 bo_flags |= ANV_BO_EXTERNAL;
2352
2353 result = anv_bo_cache_alloc(device, &device->bo_cache,
2354 pAllocateInfo->allocationSize, bo_flags,
2355 &mem->bo);
2356 if (result != VK_SUCCESS)
2357 goto fail;
2358
2359 const VkMemoryDedicatedAllocateInfoKHR *dedicated_info =
2360 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_DEDICATED_ALLOCATE_INFO_KHR);
2361 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
2362 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
2363
2364 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2365 * the BO. In this case, we have a dedicated allocation.
2366 */
2367 if (image->needs_set_tiling) {
2368 const uint32_t i915_tiling =
2369 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
2370 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
2371 image->planes[0].surface.isl.row_pitch_B,
2372 i915_tiling);
2373 if (ret) {
2374 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2375 return vk_errorf(device->instance, NULL,
2376 VK_ERROR_OUT_OF_DEVICE_MEMORY,
2377 "failed to set BO tiling: %m");
2378 }
2379 }
2380 }
2381
2382 success:
2383 *pMem = anv_device_memory_to_handle(mem);
2384
2385 return VK_SUCCESS;
2386
2387 fail:
2388 vk_free2(&device->alloc, pAllocator, mem);
2389
2390 return result;
2391 }
2392
2393 VkResult anv_GetMemoryFdKHR(
2394 VkDevice device_h,
2395 const VkMemoryGetFdInfoKHR* pGetFdInfo,
2396 int* pFd)
2397 {
2398 ANV_FROM_HANDLE(anv_device, dev, device_h);
2399 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
2400
2401 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2402
2403 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2404 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2405
2406 return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
2407 }
2408
2409 VkResult anv_GetMemoryFdPropertiesKHR(
2410 VkDevice _device,
2411 VkExternalMemoryHandleTypeFlagBitsKHR handleType,
2412 int fd,
2413 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
2414 {
2415 ANV_FROM_HANDLE(anv_device, device, _device);
2416 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2417
2418 switch (handleType) {
2419 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
2420 /* dma-buf can be imported as any memory type */
2421 pMemoryFdProperties->memoryTypeBits =
2422 (1 << pdevice->memory.type_count) - 1;
2423 return VK_SUCCESS;
2424
2425 default:
2426 /* The valid usage section for this function says:
2427 *
2428 * "handleType must not be one of the handle types defined as
2429 * opaque."
2430 *
2431 * So opaque handle types fall into the default "unsupported" case.
2432 */
2433 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
2434 }
2435 }
2436
2437 void anv_FreeMemory(
2438 VkDevice _device,
2439 VkDeviceMemory _mem,
2440 const VkAllocationCallbacks* pAllocator)
2441 {
2442 ANV_FROM_HANDLE(anv_device, device, _device);
2443 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
2444
2445 if (mem == NULL)
2446 return;
2447
2448 if (mem->map)
2449 anv_UnmapMemory(_device, _mem);
2450
2451 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2452
2453 vk_free2(&device->alloc, pAllocator, mem);
2454 }
2455
2456 VkResult anv_MapMemory(
2457 VkDevice _device,
2458 VkDeviceMemory _memory,
2459 VkDeviceSize offset,
2460 VkDeviceSize size,
2461 VkMemoryMapFlags flags,
2462 void** ppData)
2463 {
2464 ANV_FROM_HANDLE(anv_device, device, _device);
2465 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
2466
2467 if (mem == NULL) {
2468 *ppData = NULL;
2469 return VK_SUCCESS;
2470 }
2471
2472 if (size == VK_WHOLE_SIZE)
2473 size = mem->bo->size - offset;
2474
2475 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2476 *
2477 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2478 * assert(size != 0);
2479 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2480 * equal to the size of the memory minus offset
2481 */
2482 assert(size > 0);
2483 assert(offset + size <= mem->bo->size);
2484
2485 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2486 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2487 * at a time is valid. We could just mmap up front and return an offset
2488 * pointer here, but that may exhaust virtual memory on 32 bit
2489 * userspace. */
2490
2491 uint32_t gem_flags = 0;
2492
2493 if (!device->info.has_llc &&
2494 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
2495 gem_flags |= I915_MMAP_WC;
2496
2497 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2498 uint64_t map_offset = offset & ~4095ull;
2499 assert(offset >= map_offset);
2500 uint64_t map_size = (offset + size) - map_offset;
2501
2502 /* Let's map whole pages */
2503 map_size = align_u64(map_size, 4096);
2504
2505 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
2506 map_offset, map_size, gem_flags);
2507 if (map == MAP_FAILED)
2508 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
2509
2510 mem->map = map;
2511 mem->map_size = map_size;
2512
2513 *ppData = mem->map + (offset - map_offset);
2514
2515 return VK_SUCCESS;
2516 }
2517
2518 void anv_UnmapMemory(
2519 VkDevice _device,
2520 VkDeviceMemory _memory)
2521 {
2522 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
2523
2524 if (mem == NULL)
2525 return;
2526
2527 anv_gem_munmap(mem->map, mem->map_size);
2528
2529 mem->map = NULL;
2530 mem->map_size = 0;
2531 }
2532
2533 static void
2534 clflush_mapped_ranges(struct anv_device *device,
2535 uint32_t count,
2536 const VkMappedMemoryRange *ranges)
2537 {
2538 for (uint32_t i = 0; i < count; i++) {
2539 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
2540 if (ranges[i].offset >= mem->map_size)
2541 continue;
2542
2543 gen_clflush_range(mem->map + ranges[i].offset,
2544 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
2545 }
2546 }
2547
2548 VkResult anv_FlushMappedMemoryRanges(
2549 VkDevice _device,
2550 uint32_t memoryRangeCount,
2551 const VkMappedMemoryRange* pMemoryRanges)
2552 {
2553 ANV_FROM_HANDLE(anv_device, device, _device);
2554
2555 if (device->info.has_llc)
2556 return VK_SUCCESS;
2557
2558 /* Make sure the writes we're flushing have landed. */
2559 __builtin_ia32_mfence();
2560
2561 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
2562
2563 return VK_SUCCESS;
2564 }
2565
2566 VkResult anv_InvalidateMappedMemoryRanges(
2567 VkDevice _device,
2568 uint32_t memoryRangeCount,
2569 const VkMappedMemoryRange* pMemoryRanges)
2570 {
2571 ANV_FROM_HANDLE(anv_device, device, _device);
2572
2573 if (device->info.has_llc)
2574 return VK_SUCCESS;
2575
2576 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
2577
2578 /* Make sure no reads get moved up above the invalidate. */
2579 __builtin_ia32_mfence();
2580
2581 return VK_SUCCESS;
2582 }
2583
2584 void anv_GetBufferMemoryRequirements(
2585 VkDevice _device,
2586 VkBuffer _buffer,
2587 VkMemoryRequirements* pMemoryRequirements)
2588 {
2589 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2590 ANV_FROM_HANDLE(anv_device, device, _device);
2591 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2592
2593 /* The Vulkan spec (git aaed022) says:
2594 *
2595 * memoryTypeBits is a bitfield and contains one bit set for every
2596 * supported memory type for the resource. The bit `1<<i` is set if and
2597 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2598 * structure for the physical device is supported.
2599 */
2600 uint32_t memory_types = 0;
2601 for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
2602 uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
2603 if ((valid_usage & buffer->usage) == buffer->usage)
2604 memory_types |= (1u << i);
2605 }
2606
2607 /* Base alignment requirement of a cache line */
2608 uint32_t alignment = 16;
2609
2610 /* We need an alignment of 32 for pushing UBOs */
2611 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
2612 alignment = MAX2(alignment, 32);
2613
2614 pMemoryRequirements->size = buffer->size;
2615 pMemoryRequirements->alignment = alignment;
2616
2617 /* Storage and Uniform buffers should have their size aligned to
2618 * 32-bits to avoid boundary checks when last DWord is not complete.
2619 * This would ensure that not internal padding would be needed for
2620 * 16-bit types.
2621 */
2622 if (device->robust_buffer_access &&
2623 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
2624 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
2625 pMemoryRequirements->size = align_u64(buffer->size, 4);
2626
2627 pMemoryRequirements->memoryTypeBits = memory_types;
2628 }
2629
2630 void anv_GetBufferMemoryRequirements2(
2631 VkDevice _device,
2632 const VkBufferMemoryRequirementsInfo2* pInfo,
2633 VkMemoryRequirements2* pMemoryRequirements)
2634 {
2635 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
2636 &pMemoryRequirements->memoryRequirements);
2637
2638 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
2639 switch (ext->sType) {
2640 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
2641 VkMemoryDedicatedRequirements *requirements = (void *)ext;
2642 requirements->prefersDedicatedAllocation = VK_FALSE;
2643 requirements->requiresDedicatedAllocation = VK_FALSE;
2644 break;
2645 }
2646
2647 default:
2648 anv_debug_ignored_stype(ext->sType);
2649 break;
2650 }
2651 }
2652 }
2653
2654 void anv_GetImageMemoryRequirements(
2655 VkDevice _device,
2656 VkImage _image,
2657 VkMemoryRequirements* pMemoryRequirements)
2658 {
2659 ANV_FROM_HANDLE(anv_image, image, _image);
2660 ANV_FROM_HANDLE(anv_device, device, _device);
2661 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2662
2663 /* The Vulkan spec (git aaed022) says:
2664 *
2665 * memoryTypeBits is a bitfield and contains one bit set for every
2666 * supported memory type for the resource. The bit `1<<i` is set if and
2667 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2668 * structure for the physical device is supported.
2669 *
2670 * All types are currently supported for images.
2671 */
2672 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
2673
2674 pMemoryRequirements->size = image->size;
2675 pMemoryRequirements->alignment = image->alignment;
2676 pMemoryRequirements->memoryTypeBits = memory_types;
2677 }
2678
2679 void anv_GetImageMemoryRequirements2(
2680 VkDevice _device,
2681 const VkImageMemoryRequirementsInfo2* pInfo,
2682 VkMemoryRequirements2* pMemoryRequirements)
2683 {
2684 ANV_FROM_HANDLE(anv_device, device, _device);
2685 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
2686
2687 anv_GetImageMemoryRequirements(_device, pInfo->image,
2688 &pMemoryRequirements->memoryRequirements);
2689
2690 vk_foreach_struct_const(ext, pInfo->pNext) {
2691 switch (ext->sType) {
2692 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
2693 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2694 const VkImagePlaneMemoryRequirementsInfoKHR *plane_reqs =
2695 (const VkImagePlaneMemoryRequirementsInfoKHR *) ext;
2696 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
2697 plane_reqs->planeAspect);
2698
2699 assert(image->planes[plane].offset == 0);
2700
2701 /* The Vulkan spec (git aaed022) says:
2702 *
2703 * memoryTypeBits is a bitfield and contains one bit set for every
2704 * supported memory type for the resource. The bit `1<<i` is set
2705 * if and only if the memory type `i` in the
2706 * VkPhysicalDeviceMemoryProperties structure for the physical
2707 * device is supported.
2708 *
2709 * All types are currently supported for images.
2710 */
2711 pMemoryRequirements->memoryRequirements.memoryTypeBits =
2712 (1ull << pdevice->memory.type_count) - 1;
2713
2714 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
2715 pMemoryRequirements->memoryRequirements.alignment =
2716 image->planes[plane].alignment;
2717 break;
2718 }
2719
2720 default:
2721 anv_debug_ignored_stype(ext->sType);
2722 break;
2723 }
2724 }
2725
2726 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
2727 switch (ext->sType) {
2728 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
2729 VkMemoryDedicatedRequirements *requirements = (void *)ext;
2730 if (image->needs_set_tiling) {
2731 /* If we need to set the tiling for external consumers, we need a
2732 * dedicated allocation.
2733 *
2734 * See also anv_AllocateMemory.
2735 */
2736 requirements->prefersDedicatedAllocation = VK_TRUE;
2737 requirements->requiresDedicatedAllocation = VK_TRUE;
2738 } else {
2739 requirements->prefersDedicatedAllocation = VK_FALSE;
2740 requirements->requiresDedicatedAllocation = VK_FALSE;
2741 }
2742 break;
2743 }
2744
2745 default:
2746 anv_debug_ignored_stype(ext->sType);
2747 break;
2748 }
2749 }
2750 }
2751
2752 void anv_GetImageSparseMemoryRequirements(
2753 VkDevice device,
2754 VkImage image,
2755 uint32_t* pSparseMemoryRequirementCount,
2756 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
2757 {
2758 *pSparseMemoryRequirementCount = 0;
2759 }
2760
2761 void anv_GetImageSparseMemoryRequirements2(
2762 VkDevice device,
2763 const VkImageSparseMemoryRequirementsInfo2* pInfo,
2764 uint32_t* pSparseMemoryRequirementCount,
2765 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
2766 {
2767 *pSparseMemoryRequirementCount = 0;
2768 }
2769
2770 void anv_GetDeviceMemoryCommitment(
2771 VkDevice device,
2772 VkDeviceMemory memory,
2773 VkDeviceSize* pCommittedMemoryInBytes)
2774 {
2775 *pCommittedMemoryInBytes = 0;
2776 }
2777
2778 static void
2779 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
2780 {
2781 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
2782 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
2783
2784 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
2785
2786 if (mem) {
2787 assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
2788 buffer->address = (struct anv_address) {
2789 .bo = mem->bo,
2790 .offset = pBindInfo->memoryOffset,
2791 };
2792 } else {
2793 buffer->address = ANV_NULL_ADDRESS;
2794 }
2795 }
2796
2797 VkResult anv_BindBufferMemory(
2798 VkDevice device,
2799 VkBuffer buffer,
2800 VkDeviceMemory memory,
2801 VkDeviceSize memoryOffset)
2802 {
2803 anv_bind_buffer_memory(
2804 &(VkBindBufferMemoryInfo) {
2805 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
2806 .buffer = buffer,
2807 .memory = memory,
2808 .memoryOffset = memoryOffset,
2809 });
2810
2811 return VK_SUCCESS;
2812 }
2813
2814 VkResult anv_BindBufferMemory2(
2815 VkDevice device,
2816 uint32_t bindInfoCount,
2817 const VkBindBufferMemoryInfo* pBindInfos)
2818 {
2819 for (uint32_t i = 0; i < bindInfoCount; i++)
2820 anv_bind_buffer_memory(&pBindInfos[i]);
2821
2822 return VK_SUCCESS;
2823 }
2824
2825 VkResult anv_QueueBindSparse(
2826 VkQueue _queue,
2827 uint32_t bindInfoCount,
2828 const VkBindSparseInfo* pBindInfo,
2829 VkFence fence)
2830 {
2831 ANV_FROM_HANDLE(anv_queue, queue, _queue);
2832 if (anv_device_is_lost(queue->device))
2833 return VK_ERROR_DEVICE_LOST;
2834
2835 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2836 }
2837
2838 // Event functions
2839
2840 VkResult anv_CreateEvent(
2841 VkDevice _device,
2842 const VkEventCreateInfo* pCreateInfo,
2843 const VkAllocationCallbacks* pAllocator,
2844 VkEvent* pEvent)
2845 {
2846 ANV_FROM_HANDLE(anv_device, device, _device);
2847 struct anv_state state;
2848 struct anv_event *event;
2849
2850 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
2851
2852 state = anv_state_pool_alloc(&device->dynamic_state_pool,
2853 sizeof(*event), 8);
2854 event = state.map;
2855 event->state = state;
2856 event->semaphore = VK_EVENT_RESET;
2857
2858 if (!device->info.has_llc) {
2859 /* Make sure the writes we're flushing have landed. */
2860 __builtin_ia32_mfence();
2861 __builtin_ia32_clflush(event);
2862 }
2863
2864 *pEvent = anv_event_to_handle(event);
2865
2866 return VK_SUCCESS;
2867 }
2868
2869 void anv_DestroyEvent(
2870 VkDevice _device,
2871 VkEvent _event,
2872 const VkAllocationCallbacks* pAllocator)
2873 {
2874 ANV_FROM_HANDLE(anv_device, device, _device);
2875 ANV_FROM_HANDLE(anv_event, event, _event);
2876
2877 if (!event)
2878 return;
2879
2880 anv_state_pool_free(&device->dynamic_state_pool, event->state);
2881 }
2882
2883 VkResult anv_GetEventStatus(
2884 VkDevice _device,
2885 VkEvent _event)
2886 {
2887 ANV_FROM_HANDLE(anv_device, device, _device);
2888 ANV_FROM_HANDLE(anv_event, event, _event);
2889
2890 if (anv_device_is_lost(device))
2891 return VK_ERROR_DEVICE_LOST;
2892
2893 if (!device->info.has_llc) {
2894 /* Invalidate read cache before reading event written by GPU. */
2895 __builtin_ia32_clflush(event);
2896 __builtin_ia32_mfence();
2897
2898 }
2899
2900 return event->semaphore;
2901 }
2902
2903 VkResult anv_SetEvent(
2904 VkDevice _device,
2905 VkEvent _event)
2906 {
2907 ANV_FROM_HANDLE(anv_device, device, _device);
2908 ANV_FROM_HANDLE(anv_event, event, _event);
2909
2910 event->semaphore = VK_EVENT_SET;
2911
2912 if (!device->info.has_llc) {
2913 /* Make sure the writes we're flushing have landed. */
2914 __builtin_ia32_mfence();
2915 __builtin_ia32_clflush(event);
2916 }
2917
2918 return VK_SUCCESS;
2919 }
2920
2921 VkResult anv_ResetEvent(
2922 VkDevice _device,
2923 VkEvent _event)
2924 {
2925 ANV_FROM_HANDLE(anv_device, device, _device);
2926 ANV_FROM_HANDLE(anv_event, event, _event);
2927
2928 event->semaphore = VK_EVENT_RESET;
2929
2930 if (!device->info.has_llc) {
2931 /* Make sure the writes we're flushing have landed. */
2932 __builtin_ia32_mfence();
2933 __builtin_ia32_clflush(event);
2934 }
2935
2936 return VK_SUCCESS;
2937 }
2938
2939 // Buffer functions
2940
2941 VkResult anv_CreateBuffer(
2942 VkDevice _device,
2943 const VkBufferCreateInfo* pCreateInfo,
2944 const VkAllocationCallbacks* pAllocator,
2945 VkBuffer* pBuffer)
2946 {
2947 ANV_FROM_HANDLE(anv_device, device, _device);
2948 struct anv_buffer *buffer;
2949
2950 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2951
2952 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2953 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2954 if (buffer == NULL)
2955 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2956
2957 buffer->size = pCreateInfo->size;
2958 buffer->usage = pCreateInfo->usage;
2959 buffer->address = ANV_NULL_ADDRESS;
2960
2961 *pBuffer = anv_buffer_to_handle(buffer);
2962
2963 return VK_SUCCESS;
2964 }
2965
2966 void anv_DestroyBuffer(
2967 VkDevice _device,
2968 VkBuffer _buffer,
2969 const VkAllocationCallbacks* pAllocator)
2970 {
2971 ANV_FROM_HANDLE(anv_device, device, _device);
2972 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2973
2974 if (!buffer)
2975 return;
2976
2977 vk_free2(&device->alloc, pAllocator, buffer);
2978 }
2979
2980 void
2981 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
2982 enum isl_format format,
2983 struct anv_address address,
2984 uint32_t range, uint32_t stride)
2985 {
2986 isl_buffer_fill_state(&device->isl_dev, state.map,
2987 .address = anv_address_physical(address),
2988 .mocs = device->default_mocs,
2989 .size_B = range,
2990 .format = format,
2991 .stride_B = stride);
2992
2993 anv_state_flush(device, state);
2994 }
2995
2996 void anv_DestroySampler(
2997 VkDevice _device,
2998 VkSampler _sampler,
2999 const VkAllocationCallbacks* pAllocator)
3000 {
3001 ANV_FROM_HANDLE(anv_device, device, _device);
3002 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
3003
3004 if (!sampler)
3005 return;
3006
3007 vk_free2(&device->alloc, pAllocator, sampler);
3008 }
3009
3010 VkResult anv_CreateFramebuffer(
3011 VkDevice _device,
3012 const VkFramebufferCreateInfo* pCreateInfo,
3013 const VkAllocationCallbacks* pAllocator,
3014 VkFramebuffer* pFramebuffer)
3015 {
3016 ANV_FROM_HANDLE(anv_device, device, _device);
3017 struct anv_framebuffer *framebuffer;
3018
3019 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
3020
3021 size_t size = sizeof(*framebuffer) +
3022 sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
3023 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
3024 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3025 if (framebuffer == NULL)
3026 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3027
3028 framebuffer->attachment_count = pCreateInfo->attachmentCount;
3029 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
3030 VkImageView _iview = pCreateInfo->pAttachments[i];
3031 framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
3032 }
3033
3034 framebuffer->width = pCreateInfo->width;
3035 framebuffer->height = pCreateInfo->height;
3036 framebuffer->layers = pCreateInfo->layers;
3037
3038 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
3039
3040 return VK_SUCCESS;
3041 }
3042
3043 void anv_DestroyFramebuffer(
3044 VkDevice _device,
3045 VkFramebuffer _fb,
3046 const VkAllocationCallbacks* pAllocator)
3047 {
3048 ANV_FROM_HANDLE(anv_device, device, _device);
3049 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
3050
3051 if (!fb)
3052 return;
3053
3054 vk_free2(&device->alloc, pAllocator, fb);
3055 }
3056
3057 static const VkTimeDomainEXT anv_time_domains[] = {
3058 VK_TIME_DOMAIN_DEVICE_EXT,
3059 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
3060 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
3061 };
3062
3063 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3064 VkPhysicalDevice physicalDevice,
3065 uint32_t *pTimeDomainCount,
3066 VkTimeDomainEXT *pTimeDomains)
3067 {
3068 int d;
3069 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
3070
3071 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
3072 vk_outarray_append(&out, i) {
3073 *i = anv_time_domains[d];
3074 }
3075 }
3076
3077 return vk_outarray_status(&out);
3078 }
3079
3080 static uint64_t
3081 anv_clock_gettime(clockid_t clock_id)
3082 {
3083 struct timespec current;
3084 int ret;
3085
3086 ret = clock_gettime(clock_id, &current);
3087 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
3088 ret = clock_gettime(CLOCK_MONOTONIC, &current);
3089 if (ret < 0)
3090 return 0;
3091
3092 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
3093 }
3094
3095 #define TIMESTAMP 0x2358
3096
3097 VkResult anv_GetCalibratedTimestampsEXT(
3098 VkDevice _device,
3099 uint32_t timestampCount,
3100 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
3101 uint64_t *pTimestamps,
3102 uint64_t *pMaxDeviation)
3103 {
3104 ANV_FROM_HANDLE(anv_device, device, _device);
3105 uint64_t timestamp_frequency = device->info.timestamp_frequency;
3106 int ret;
3107 int d;
3108 uint64_t begin, end;
3109 uint64_t max_clock_period = 0;
3110
3111 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3112
3113 for (d = 0; d < timestampCount; d++) {
3114 switch (pTimestampInfos[d].timeDomain) {
3115 case VK_TIME_DOMAIN_DEVICE_EXT:
3116 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
3117 &pTimestamps[d]);
3118
3119 if (ret != 0) {
3120 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
3121 "register: %m");
3122 }
3123 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
3124 max_clock_period = MAX2(max_clock_period, device_period);
3125 break;
3126 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
3127 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
3128 max_clock_period = MAX2(max_clock_period, 1);
3129 break;
3130
3131 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
3132 pTimestamps[d] = begin;
3133 break;
3134 default:
3135 pTimestamps[d] = 0;
3136 break;
3137 }
3138 }
3139
3140 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3141
3142 /*
3143 * The maximum deviation is the sum of the interval over which we
3144 * perform the sampling and the maximum period of any sampled
3145 * clock. That's because the maximum skew between any two sampled
3146 * clock edges is when the sampled clock with the largest period is
3147 * sampled at the end of that period but right at the beginning of the
3148 * sampling interval and some other clock is sampled right at the
3149 * begining of its sampling period and right at the end of the
3150 * sampling interval. Let's assume the GPU has the longest clock
3151 * period and that the application is sampling GPU and monotonic:
3152 *
3153 * s e
3154 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3155 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3156 *
3157 * g
3158 * 0 1 2 3
3159 * GPU -----_____-----_____-----_____-----_____
3160 *
3161 * m
3162 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3163 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3164 *
3165 * Interval <----------------->
3166 * Deviation <-------------------------->
3167 *
3168 * s = read(raw) 2
3169 * g = read(GPU) 1
3170 * m = read(monotonic) 2
3171 * e = read(raw) b
3172 *
3173 * We round the sample interval up by one tick to cover sampling error
3174 * in the interval clock
3175 */
3176
3177 uint64_t sample_interval = end - begin + 1;
3178
3179 *pMaxDeviation = sample_interval + max_clock_period;
3180
3181 return VK_SUCCESS;
3182 }
3183
3184 /* vk_icd.h does not declare this function, so we declare it here to
3185 * suppress Wmissing-prototypes.
3186 */
3187 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3188 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
3189
3190 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3191 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
3192 {
3193 /* For the full details on loader interface versioning, see
3194 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3195 * What follows is a condensed summary, to help you navigate the large and
3196 * confusing official doc.
3197 *
3198 * - Loader interface v0 is incompatible with later versions. We don't
3199 * support it.
3200 *
3201 * - In loader interface v1:
3202 * - The first ICD entrypoint called by the loader is
3203 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3204 * entrypoint.
3205 * - The ICD must statically expose no other Vulkan symbol unless it is
3206 * linked with -Bsymbolic.
3207 * - Each dispatchable Vulkan handle created by the ICD must be
3208 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3209 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3210 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3211 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3212 * such loader-managed surfaces.
3213 *
3214 * - Loader interface v2 differs from v1 in:
3215 * - The first ICD entrypoint called by the loader is
3216 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3217 * statically expose this entrypoint.
3218 *
3219 * - Loader interface v3 differs from v2 in:
3220 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3221 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3222 * because the loader no longer does so.
3223 */
3224 *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
3225 return VK_SUCCESS;
3226 }