projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
anv/allocator: Rename anv_free_list2 to anv_free_list.
[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_8BIT_STORAGE_FEATURES_KHR: {
897 VkPhysicalDevice8BitStorageFeaturesKHR *features =
898 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
899 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
900
901 features->storageBuffer8BitAccess = pdevice->info.gen >= 8;
902 features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
903 features->storagePushConstant8 = pdevice->info.gen >= 8;
904 break;
905 }
906
907 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
908 VkPhysicalDevice16BitStorageFeatures *features =
909 (VkPhysicalDevice16BitStorageFeatures *)ext;
910 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
911
912 features->storageBuffer16BitAccess = pdevice->info.gen >= 8;
913 features->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
914 features->storagePushConstant16 = pdevice->info.gen >= 8;
915 features->storageInputOutput16 = false;
916 break;
917 }
918
919 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
920 VkPhysicalDeviceMultiviewFeatures *features =
921 (VkPhysicalDeviceMultiviewFeatures *)ext;
922 features->multiview = true;
923 features->multiviewGeometryShader = true;
924 features->multiviewTessellationShader = true;
925 break;
926 }
927
928 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
929 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
930 features->protectedMemory = VK_FALSE;
931 break;
932 }
933
934 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
935 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
936 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
937 features->samplerYcbcrConversion = true;
938 break;
939 }
940
941 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
942 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
943 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
944 features->scalarBlockLayout = true;
945 break;
946 }
947
948 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES: {
949 VkPhysicalDeviceShaderDrawParameterFeatures *features = (void *)ext;
950 features->shaderDrawParameters = true;
951 break;
952 }
953
954 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES: {
955 VkPhysicalDeviceVariablePointerFeatures *features = (void *)ext;
956 features->variablePointersStorageBuffer = true;
957 features->variablePointers = true;
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 const uint32_t max_images = devinfo->gen < 9 ? MAX_GEN8_IMAGES : MAX_IMAGES;
991
992 VkSampleCountFlags sample_counts =
993 isl_device_get_sample_counts(&pdevice->isl_dev);
994
995
996 VkPhysicalDeviceLimits limits = {
997 .maxImageDimension1D = (1 << 14),
998 .maxImageDimension2D = (1 << 14),
999 .maxImageDimension3D = (1 << 11),
1000 .maxImageDimensionCube = (1 << 14),
1001 .maxImageArrayLayers = (1 << 11),
1002 .maxTexelBufferElements = 128 * 1024 * 1024,
1003 .maxUniformBufferRange = (1ul << 27),
1004 .maxStorageBufferRange = max_raw_buffer_sz,
1005 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1006 .maxMemoryAllocationCount = UINT32_MAX,
1007 .maxSamplerAllocationCount = 64 * 1024,
1008 .bufferImageGranularity = 64, /* A cache line */
1009 .sparseAddressSpaceSize = 0,
1010 .maxBoundDescriptorSets = MAX_SETS,
1011 .maxPerStageDescriptorSamplers = max_samplers,
1012 .maxPerStageDescriptorUniformBuffers = 64,
1013 .maxPerStageDescriptorStorageBuffers = 64,
1014 .maxPerStageDescriptorSampledImages = max_samplers,
1015 .maxPerStageDescriptorStorageImages = max_images,
1016 .maxPerStageDescriptorInputAttachments = 64,
1017 .maxPerStageResources = 250,
1018 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1019 .maxDescriptorSetUniformBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
1020 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1021 .maxDescriptorSetStorageBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
1022 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1023 .maxDescriptorSetSampledImages = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSampledImages */
1024 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1025 .maxDescriptorSetInputAttachments = 256,
1026 .maxVertexInputAttributes = MAX_VBS,
1027 .maxVertexInputBindings = MAX_VBS,
1028 .maxVertexInputAttributeOffset = 2047,
1029 .maxVertexInputBindingStride = 2048,
1030 .maxVertexOutputComponents = 128,
1031 .maxTessellationGenerationLevel = 64,
1032 .maxTessellationPatchSize = 32,
1033 .maxTessellationControlPerVertexInputComponents = 128,
1034 .maxTessellationControlPerVertexOutputComponents = 128,
1035 .maxTessellationControlPerPatchOutputComponents = 128,
1036 .maxTessellationControlTotalOutputComponents = 2048,
1037 .maxTessellationEvaluationInputComponents = 128,
1038 .maxTessellationEvaluationOutputComponents = 128,
1039 .maxGeometryShaderInvocations = 32,
1040 .maxGeometryInputComponents = 64,
1041 .maxGeometryOutputComponents = 128,
1042 .maxGeometryOutputVertices = 256,
1043 .maxGeometryTotalOutputComponents = 1024,
1044 .maxFragmentInputComponents = 112, /* 128 components - (POS, PSIZ, CLIP_DIST0, CLIP_DIST1) */
1045 .maxFragmentOutputAttachments = 8,
1046 .maxFragmentDualSrcAttachments = 1,
1047 .maxFragmentCombinedOutputResources = 8,
1048 .maxComputeSharedMemorySize = 32768,
1049 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1050 .maxComputeWorkGroupInvocations = 16 * devinfo->max_cs_threads,
1051 .maxComputeWorkGroupSize = {
1052 16 * devinfo->max_cs_threads,
1053 16 * devinfo->max_cs_threads,
1054 16 * devinfo->max_cs_threads,
1055 },
1056 .subPixelPrecisionBits = 4 /* FIXME */,
1057 .subTexelPrecisionBits = 4 /* FIXME */,
1058 .mipmapPrecisionBits = 4 /* FIXME */,
1059 .maxDrawIndexedIndexValue = UINT32_MAX,
1060 .maxDrawIndirectCount = UINT32_MAX,
1061 .maxSamplerLodBias = 16,
1062 .maxSamplerAnisotropy = 16,
1063 .maxViewports = MAX_VIEWPORTS,
1064 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1065 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1066 .viewportSubPixelBits = 13, /* We take a float? */
1067 .minMemoryMapAlignment = 4096, /* A page */
1068 .minTexelBufferOffsetAlignment = 1,
1069 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1070 .minUniformBufferOffsetAlignment = 32,
1071 .minStorageBufferOffsetAlignment = 4,
1072 .minTexelOffset = -8,
1073 .maxTexelOffset = 7,
1074 .minTexelGatherOffset = -32,
1075 .maxTexelGatherOffset = 31,
1076 .minInterpolationOffset = -0.5,
1077 .maxInterpolationOffset = 0.4375,
1078 .subPixelInterpolationOffsetBits = 4,
1079 .maxFramebufferWidth = (1 << 14),
1080 .maxFramebufferHeight = (1 << 14),
1081 .maxFramebufferLayers = (1 << 11),
1082 .framebufferColorSampleCounts = sample_counts,
1083 .framebufferDepthSampleCounts = sample_counts,
1084 .framebufferStencilSampleCounts = sample_counts,
1085 .framebufferNoAttachmentsSampleCounts = sample_counts,
1086 .maxColorAttachments = MAX_RTS,
1087 .sampledImageColorSampleCounts = sample_counts,
1088 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1089 .sampledImageDepthSampleCounts = sample_counts,
1090 .sampledImageStencilSampleCounts = sample_counts,
1091 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1092 .maxSampleMaskWords = 1,
1093 .timestampComputeAndGraphics = false,
1094 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1095 .maxClipDistances = 8,
1096 .maxCullDistances = 8,
1097 .maxCombinedClipAndCullDistances = 8,
1098 .discreteQueuePriorities = 2,
1099 .pointSizeRange = { 0.125, 255.875 },
1100 .lineWidthRange = { 0.0, 7.9921875 },
1101 .pointSizeGranularity = (1.0 / 8.0),
1102 .lineWidthGranularity = (1.0 / 128.0),
1103 .strictLines = false, /* FINISHME */
1104 .standardSampleLocations = true,
1105 .optimalBufferCopyOffsetAlignment = 128,
1106 .optimalBufferCopyRowPitchAlignment = 128,
1107 .nonCoherentAtomSize = 64,
1108 };
1109
1110 *pProperties = (VkPhysicalDeviceProperties) {
1111 .apiVersion = anv_physical_device_api_version(pdevice),
1112 .driverVersion = vk_get_driver_version(),
1113 .vendorID = 0x8086,
1114 .deviceID = pdevice->chipset_id,
1115 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1116 .limits = limits,
1117 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1118 };
1119
1120 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1121 "%s", pdevice->name);
1122 memcpy(pProperties->pipelineCacheUUID,
1123 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1124 }
1125
1126 void anv_GetPhysicalDeviceProperties2(
1127 VkPhysicalDevice physicalDevice,
1128 VkPhysicalDeviceProperties2* pProperties)
1129 {
1130 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1131
1132 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1133
1134 vk_foreach_struct(ext, pProperties->pNext) {
1135 switch (ext->sType) {
1136 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1137 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *props =
1138 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1139
1140 /* We support all of the depth resolve modes */
1141 props->supportedDepthResolveModes =
1142 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1143 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1144 VK_RESOLVE_MODE_MIN_BIT_KHR |
1145 VK_RESOLVE_MODE_MAX_BIT_KHR;
1146
1147 /* Average doesn't make sense for stencil so we don't support that */
1148 props->supportedStencilResolveModes =
1149 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1150 if (pdevice->info.gen >= 8) {
1151 /* The advanced stencil resolve modes currently require stencil
1152 * sampling be supported by the hardware.
1153 */
1154 props->supportedStencilResolveModes |=
1155 VK_RESOLVE_MODE_MIN_BIT_KHR |
1156 VK_RESOLVE_MODE_MAX_BIT_KHR;
1157 }
1158
1159 props->independentResolveNone = VK_TRUE;
1160 props->independentResolve = VK_TRUE;
1161 break;
1162 }
1163
1164 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1165 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1166 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1167
1168 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1169 util_snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1170 "Intel open-source Mesa driver");
1171
1172 util_snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1173 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1174
1175 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1176 .major = 1,
1177 .minor = 1,
1178 .subminor = 2,
1179 .patch = 0,
1180 };
1181 break;
1182 }
1183
1184 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1185 VkPhysicalDeviceIDProperties *id_props =
1186 (VkPhysicalDeviceIDProperties *)ext;
1187 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1188 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1189 /* The LUID is for Windows. */
1190 id_props->deviceLUIDValid = false;
1191 break;
1192 }
1193
1194 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1195 VkPhysicalDeviceMaintenance3Properties *props =
1196 (VkPhysicalDeviceMaintenance3Properties *)ext;
1197 /* This value doesn't matter for us today as our per-stage
1198 * descriptors are the real limit.
1199 */
1200 props->maxPerSetDescriptors = 1024;
1201 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1202 break;
1203 }
1204
1205 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1206 VkPhysicalDeviceMultiviewProperties *properties =
1207 (VkPhysicalDeviceMultiviewProperties *)ext;
1208 properties->maxMultiviewViewCount = 16;
1209 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1210 break;
1211 }
1212
1213 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1214 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1215 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1216 properties->pciDomain = pdevice->pci_info.domain;
1217 properties->pciBus = pdevice->pci_info.bus;
1218 properties->pciDevice = pdevice->pci_info.device;
1219 properties->pciFunction = pdevice->pci_info.function;
1220 break;
1221 }
1222
1223 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1224 VkPhysicalDevicePointClippingProperties *properties =
1225 (VkPhysicalDevicePointClippingProperties *) ext;
1226 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
1227 anv_finishme("Implement pop-free point clipping");
1228 break;
1229 }
1230
1231 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1232 VkPhysicalDeviceProtectedMemoryProperties *props =
1233 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1234 props->protectedNoFault = false;
1235 break;
1236 }
1237
1238 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1239 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1240 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1241
1242 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1243 break;
1244 }
1245
1246 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1247 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1248 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1249 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1250 properties->filterMinmaxSingleComponentFormats = true;
1251 break;
1252 }
1253
1254 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1255 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1256
1257 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1258
1259 VkShaderStageFlags scalar_stages = 0;
1260 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1261 if (pdevice->compiler->scalar_stage[stage])
1262 scalar_stages |= mesa_to_vk_shader_stage(stage);
1263 }
1264 properties->supportedStages = scalar_stages;
1265
1266 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1267 VK_SUBGROUP_FEATURE_VOTE_BIT |
1268 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1269 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1270 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1271 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1272 VK_SUBGROUP_FEATURE_CLUSTERED_BIT |
1273 VK_SUBGROUP_FEATURE_QUAD_BIT;
1274 properties->quadOperationsInAllStages = VK_TRUE;
1275 break;
1276 }
1277
1278 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1279 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1280 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1281 /* We have to restrict this a bit for multiview */
1282 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1283 break;
1284 }
1285
1286 default:
1287 anv_debug_ignored_stype(ext->sType);
1288 break;
1289 }
1290 }
1291 }
1292
1293 /* We support exactly one queue family. */
1294 static const VkQueueFamilyProperties
1295 anv_queue_family_properties = {
1296 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1297 VK_QUEUE_COMPUTE_BIT |
1298 VK_QUEUE_TRANSFER_BIT,
1299 .queueCount = 1,
1300 .timestampValidBits = 36, /* XXX: Real value here */
1301 .minImageTransferGranularity = { 1, 1, 1 },
1302 };
1303
1304 void anv_GetPhysicalDeviceQueueFamilyProperties(
1305 VkPhysicalDevice physicalDevice,
1306 uint32_t* pCount,
1307 VkQueueFamilyProperties* pQueueFamilyProperties)
1308 {
1309 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1310
1311 vk_outarray_append(&out, p) {
1312 *p = anv_queue_family_properties;
1313 }
1314 }
1315
1316 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1317 VkPhysicalDevice physicalDevice,
1318 uint32_t* pQueueFamilyPropertyCount,
1319 VkQueueFamilyProperties2* pQueueFamilyProperties)
1320 {
1321
1322 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1323
1324 vk_outarray_append(&out, p) {
1325 p->queueFamilyProperties = anv_queue_family_properties;
1326
1327 vk_foreach_struct(s, p->pNext) {
1328 anv_debug_ignored_stype(s->sType);
1329 }
1330 }
1331 }
1332
1333 void anv_GetPhysicalDeviceMemoryProperties(
1334 VkPhysicalDevice physicalDevice,
1335 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1336 {
1337 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1338
1339 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1340 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1341 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1342 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1343 .heapIndex = physical_device->memory.types[i].heapIndex,
1344 };
1345 }
1346
1347 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1348 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1349 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1350 .size = physical_device->memory.heaps[i].size,
1351 .flags = physical_device->memory.heaps[i].flags,
1352 };
1353 }
1354 }
1355
1356 void anv_GetPhysicalDeviceMemoryProperties2(
1357 VkPhysicalDevice physicalDevice,
1358 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1359 {
1360 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1361 &pMemoryProperties->memoryProperties);
1362
1363 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1364 switch (ext->sType) {
1365 default:
1366 anv_debug_ignored_stype(ext->sType);
1367 break;
1368 }
1369 }
1370 }
1371
1372 void
1373 anv_GetDeviceGroupPeerMemoryFeatures(
1374 VkDevice device,
1375 uint32_t heapIndex,
1376 uint32_t localDeviceIndex,
1377 uint32_t remoteDeviceIndex,
1378 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1379 {
1380 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1381 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1382 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1383 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1384 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1385 }
1386
1387 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1388 VkInstance _instance,
1389 const char* pName)
1390 {
1391 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1392
1393 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1394 * when we have to return valid function pointers, NULL, or it's left
1395 * undefined. See the table for exact details.
1396 */
1397 if (pName == NULL)
1398 return NULL;
1399
1400 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1401 if (strcmp(pName, "vk" #entrypoint) == 0) \
1402 return (PFN_vkVoidFunction)anv_##entrypoint
1403
1404 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1405 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1406 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1407 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1408
1409 #undef LOOKUP_ANV_ENTRYPOINT
1410
1411 if (instance == NULL)
1412 return NULL;
1413
1414 int idx = anv_get_instance_entrypoint_index(pName);
1415 if (idx >= 0)
1416 return instance->dispatch.entrypoints[idx];
1417
1418 idx = anv_get_device_entrypoint_index(pName);
1419 if (idx >= 0)
1420 return instance->device_dispatch.entrypoints[idx];
1421
1422 return NULL;
1423 }
1424
1425 /* With version 1+ of the loader interface the ICD should expose
1426 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1427 */
1428 PUBLIC
1429 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1430 VkInstance instance,
1431 const char* pName);
1432
1433 PUBLIC
1434 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1435 VkInstance instance,
1436 const char* pName)
1437 {
1438 return anv_GetInstanceProcAddr(instance, pName);
1439 }
1440
1441 PFN_vkVoidFunction anv_GetDeviceProcAddr(
1442 VkDevice _device,
1443 const char* pName)
1444 {
1445 ANV_FROM_HANDLE(anv_device, device, _device);
1446
1447 if (!device || !pName)
1448 return NULL;
1449
1450 int idx = anv_get_device_entrypoint_index(pName);
1451 if (idx < 0)
1452 return NULL;
1453
1454 return device->dispatch.entrypoints[idx];
1455 }
1456
1457 VkResult
1458 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
1459 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
1460 const VkAllocationCallbacks* pAllocator,
1461 VkDebugReportCallbackEXT* pCallback)
1462 {
1463 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1464 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
1465 pCreateInfo, pAllocator, &instance->alloc,
1466 pCallback);
1467 }
1468
1469 void
1470 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
1471 VkDebugReportCallbackEXT _callback,
1472 const VkAllocationCallbacks* pAllocator)
1473 {
1474 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1475 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
1476 _callback, pAllocator, &instance->alloc);
1477 }
1478
1479 void
1480 anv_DebugReportMessageEXT(VkInstance _instance,
1481 VkDebugReportFlagsEXT flags,
1482 VkDebugReportObjectTypeEXT objectType,
1483 uint64_t object,
1484 size_t location,
1485 int32_t messageCode,
1486 const char* pLayerPrefix,
1487 const char* pMessage)
1488 {
1489 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1490 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
1491 object, location, messageCode, pLayerPrefix, pMessage);
1492 }
1493
1494 static void
1495 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
1496 {
1497 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1498 queue->device = device;
1499 queue->flags = 0;
1500 }
1501
1502 static void
1503 anv_queue_finish(struct anv_queue *queue)
1504 {
1505 }
1506
1507 static struct anv_state
1508 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
1509 {
1510 struct anv_state state;
1511
1512 state = anv_state_pool_alloc(pool, size, align);
1513 memcpy(state.map, p, size);
1514
1515 anv_state_flush(pool->block_pool.device, state);
1516
1517 return state;
1518 }
1519
1520 struct gen8_border_color {
1521 union {
1522 float float32[4];
1523 uint32_t uint32[4];
1524 };
1525 /* Pad out to 64 bytes */
1526 uint32_t _pad[12];
1527 };
1528
1529 static void
1530 anv_device_init_border_colors(struct anv_device *device)
1531 {
1532 static const struct gen8_border_color border_colors[] = {
1533 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
1534 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
1535 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
1536 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
1537 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
1538 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
1539 };
1540
1541 device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
1542 sizeof(border_colors), 64,
1543 border_colors);
1544 }
1545
1546 static void
1547 anv_device_init_trivial_batch(struct anv_device *device)
1548 {
1549 anv_bo_init_new(&device->trivial_batch_bo, device, 4096);
1550
1551 if (device->instance->physicalDevice.has_exec_async)
1552 device->trivial_batch_bo.flags |= EXEC_OBJECT_ASYNC;
1553
1554 if (device->instance->physicalDevice.use_softpin)
1555 device->trivial_batch_bo.flags |= EXEC_OBJECT_PINNED;
1556
1557 anv_vma_alloc(device, &device->trivial_batch_bo);
1558
1559 void *map = anv_gem_mmap(device, device->trivial_batch_bo.gem_handle,
1560 0, 4096, 0);
1561
1562 struct anv_batch batch = {
1563 .start = map,
1564 .next = map,
1565 .end = map + 4096,
1566 };
1567
1568 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1569 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1570
1571 if (!device->info.has_llc)
1572 gen_clflush_range(map, batch.next - map);
1573
1574 anv_gem_munmap(map, device->trivial_batch_bo.size);
1575 }
1576
1577 VkResult anv_EnumerateDeviceExtensionProperties(
1578 VkPhysicalDevice physicalDevice,
1579 const char* pLayerName,
1580 uint32_t* pPropertyCount,
1581 VkExtensionProperties* pProperties)
1582 {
1583 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1584 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1585
1586 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
1587 if (device->supported_extensions.extensions[i]) {
1588 vk_outarray_append(&out, prop) {
1589 *prop = anv_device_extensions[i];
1590 }
1591 }
1592 }
1593
1594 return vk_outarray_status(&out);
1595 }
1596
1597 static void
1598 anv_device_init_dispatch(struct anv_device *device)
1599 {
1600 const struct anv_device_dispatch_table *genX_table;
1601 switch (device->info.gen) {
1602 case 11:
1603 genX_table = &gen11_device_dispatch_table;
1604 break;
1605 case 10:
1606 genX_table = &gen10_device_dispatch_table;
1607 break;
1608 case 9:
1609 genX_table = &gen9_device_dispatch_table;
1610 break;
1611 case 8:
1612 genX_table = &gen8_device_dispatch_table;
1613 break;
1614 case 7:
1615 if (device->info.is_haswell)
1616 genX_table = &gen75_device_dispatch_table;
1617 else
1618 genX_table = &gen7_device_dispatch_table;
1619 break;
1620 default:
1621 unreachable("unsupported gen\n");
1622 }
1623
1624 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
1625 /* Vulkan requires that entrypoints for extensions which have not been
1626 * enabled must not be advertised.
1627 */
1628 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
1629 &device->instance->enabled_extensions,
1630 &device->enabled_extensions)) {
1631 device->dispatch.entrypoints[i] = NULL;
1632 } else if (genX_table->entrypoints[i]) {
1633 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
1634 } else {
1635 device->dispatch.entrypoints[i] =
1636 anv_device_dispatch_table.entrypoints[i];
1637 }
1638 }
1639 }
1640
1641 static int
1642 vk_priority_to_gen(int priority)
1643 {
1644 switch (priority) {
1645 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
1646 return GEN_CONTEXT_LOW_PRIORITY;
1647 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
1648 return GEN_CONTEXT_MEDIUM_PRIORITY;
1649 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
1650 return GEN_CONTEXT_HIGH_PRIORITY;
1651 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
1652 return GEN_CONTEXT_REALTIME_PRIORITY;
1653 default:
1654 unreachable("Invalid priority");
1655 }
1656 }
1657
1658 static void
1659 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
1660 {
1661 anv_bo_init_new(&device->hiz_clear_bo, device, 4096);
1662
1663 if (device->instance->physicalDevice.has_exec_async)
1664 device->hiz_clear_bo.flags |= EXEC_OBJECT_ASYNC;
1665
1666 if (device->instance->physicalDevice.use_softpin)
1667 device->hiz_clear_bo.flags |= EXEC_OBJECT_PINNED;
1668
1669 anv_vma_alloc(device, &device->hiz_clear_bo);
1670
1671 uint32_t *map = anv_gem_mmap(device, device->hiz_clear_bo.gem_handle,
1672 0, 4096, 0);
1673
1674 union isl_color_value hiz_clear = { .u32 = { 0, } };
1675 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
1676
1677 memcpy(map, hiz_clear.u32, sizeof(hiz_clear.u32));
1678 anv_gem_munmap(map, device->hiz_clear_bo.size);
1679 }
1680
1681 VkResult anv_CreateDevice(
1682 VkPhysicalDevice physicalDevice,
1683 const VkDeviceCreateInfo* pCreateInfo,
1684 const VkAllocationCallbacks* pAllocator,
1685 VkDevice* pDevice)
1686 {
1687 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1688 VkResult result;
1689 struct anv_device *device;
1690
1691 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
1692
1693 struct anv_device_extension_table enabled_extensions = { };
1694 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1695 int idx;
1696 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
1697 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
1698 anv_device_extensions[idx].extensionName) == 0)
1699 break;
1700 }
1701
1702 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
1703 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1704
1705 if (!physical_device->supported_extensions.extensions[idx])
1706 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1707
1708 enabled_extensions.extensions[idx] = true;
1709 }
1710
1711 /* Check enabled features */
1712 if (pCreateInfo->pEnabledFeatures) {
1713 VkPhysicalDeviceFeatures supported_features;
1714 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1715 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
1716 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
1717 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1718 for (uint32_t i = 0; i < num_features; i++) {
1719 if (enabled_feature[i] && !supported_feature[i])
1720 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1721 }
1722 }
1723
1724 /* Check requested queues and fail if we are requested to create any
1725 * queues with flags we don't support.
1726 */
1727 assert(pCreateInfo->queueCreateInfoCount > 0);
1728 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1729 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
1730 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
1731 }
1732
1733 /* Check if client specified queue priority. */
1734 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
1735 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
1736 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
1737
1738 VkQueueGlobalPriorityEXT priority =
1739 queue_priority ? queue_priority->globalPriority :
1740 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
1741
1742 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
1743 sizeof(*device), 8,
1744 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1745 if (!device)
1746 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1747
1748 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1749 device->instance = physical_device->instance;
1750 device->chipset_id = physical_device->chipset_id;
1751 device->no_hw = physical_device->no_hw;
1752 device->_lost = false;
1753
1754 if (pAllocator)
1755 device->alloc = *pAllocator;
1756 else
1757 device->alloc = physical_device->instance->alloc;
1758
1759 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1760 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1761 if (device->fd == -1) {
1762 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1763 goto fail_device;
1764 }
1765
1766 device->context_id = anv_gem_create_context(device);
1767 if (device->context_id == -1) {
1768 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1769 goto fail_fd;
1770 }
1771
1772 if (physical_device->use_softpin) {
1773 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
1774 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1775 goto fail_fd;
1776 }
1777
1778 /* keep the page with address zero out of the allocator */
1779 util_vma_heap_init(&device->vma_lo, LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
1780 device->vma_lo_available =
1781 physical_device->memory.heaps[physical_device->memory.heap_count - 1].size;
1782
1783 /* Leave the last 4GiB out of the high vma range, so that no state base
1784 * address + size can overflow 48 bits. For more information see the
1785 * comment about Wa32bitGeneralStateOffset in anv_allocator.c
1786 */
1787 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
1788 HIGH_HEAP_SIZE);
1789 device->vma_hi_available = physical_device->memory.heap_count == 1 ? 0 :
1790 physical_device->memory.heaps[0].size;
1791 }
1792
1793 /* As per spec, the driver implementation may deny requests to acquire
1794 * a priority above the default priority (MEDIUM) if the caller does not
1795 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1796 * is returned.
1797 */
1798 if (physical_device->has_context_priority) {
1799 int err = anv_gem_set_context_param(device->fd, device->context_id,
1800 I915_CONTEXT_PARAM_PRIORITY,
1801 vk_priority_to_gen(priority));
1802 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
1803 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
1804 goto fail_fd;
1805 }
1806 }
1807
1808 device->info = physical_device->info;
1809 device->isl_dev = physical_device->isl_dev;
1810
1811 /* On Broadwell and later, we can use batch chaining to more efficiently
1812 * implement growing command buffers. Prior to Haswell, the kernel
1813 * command parser gets in the way and we have to fall back to growing
1814 * the batch.
1815 */
1816 device->can_chain_batches = device->info.gen >= 8;
1817
1818 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1819 pCreateInfo->pEnabledFeatures->robustBufferAccess;
1820 device->enabled_extensions = enabled_extensions;
1821
1822 anv_device_init_dispatch(device);
1823
1824 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
1825 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1826 goto fail_context_id;
1827 }
1828
1829 pthread_condattr_t condattr;
1830 if (pthread_condattr_init(&condattr) != 0) {
1831 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1832 goto fail_mutex;
1833 }
1834 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
1835 pthread_condattr_destroy(&condattr);
1836 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1837 goto fail_mutex;
1838 }
1839 if (pthread_cond_init(&device->queue_submit, NULL) != 0) {
1840 pthread_condattr_destroy(&condattr);
1841 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1842 goto fail_mutex;
1843 }
1844 pthread_condattr_destroy(&condattr);
1845
1846 uint64_t bo_flags =
1847 (physical_device->supports_48bit_addresses ? EXEC_OBJECT_SUPPORTS_48B_ADDRESS : 0) |
1848 (physical_device->has_exec_async ? EXEC_OBJECT_ASYNC : 0) |
1849 (physical_device->has_exec_capture ? EXEC_OBJECT_CAPTURE : 0) |
1850 (physical_device->use_softpin ? EXEC_OBJECT_PINNED : 0);
1851
1852 anv_bo_pool_init(&device->batch_bo_pool, device, bo_flags);
1853
1854 result = anv_bo_cache_init(&device->bo_cache);
1855 if (result != VK_SUCCESS)
1856 goto fail_batch_bo_pool;
1857
1858 if (!physical_device->use_softpin)
1859 bo_flags &= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
1860
1861 result = anv_state_pool_init(&device->dynamic_state_pool, device,
1862 DYNAMIC_STATE_POOL_MIN_ADDRESS,
1863 16384,
1864 bo_flags);
1865 if (result != VK_SUCCESS)
1866 goto fail_bo_cache;
1867
1868 result = anv_state_pool_init(&device->instruction_state_pool, device,
1869 INSTRUCTION_STATE_POOL_MIN_ADDRESS,
1870 16384,
1871 bo_flags);
1872 if (result != VK_SUCCESS)
1873 goto fail_dynamic_state_pool;
1874
1875 result = anv_state_pool_init(&device->surface_state_pool, device,
1876 SURFACE_STATE_POOL_MIN_ADDRESS,
1877 4096,
1878 bo_flags);
1879 if (result != VK_SUCCESS)
1880 goto fail_instruction_state_pool;
1881
1882 if (physical_device->use_softpin) {
1883 result = anv_state_pool_init(&device->binding_table_pool, device,
1884 BINDING_TABLE_POOL_MIN_ADDRESS,
1885 4096,
1886 bo_flags);
1887 if (result != VK_SUCCESS)
1888 goto fail_surface_state_pool;
1889 }
1890
1891 result = anv_bo_init_new(&device->workaround_bo, device, 1024);
1892 if (result != VK_SUCCESS)
1893 goto fail_binding_table_pool;
1894
1895 if (physical_device->use_softpin)
1896 device->workaround_bo.flags |= EXEC_OBJECT_PINNED;
1897
1898 if (!anv_vma_alloc(device, &device->workaround_bo))
1899 goto fail_workaround_bo;
1900
1901 anv_device_init_trivial_batch(device);
1902
1903 if (device->info.gen >= 10)
1904 anv_device_init_hiz_clear_value_bo(device);
1905
1906 anv_scratch_pool_init(device, &device->scratch_pool);
1907
1908 anv_queue_init(device, &device->queue);
1909
1910 switch (device->info.gen) {
1911 case 7:
1912 if (!device->info.is_haswell)
1913 result = gen7_init_device_state(device);
1914 else
1915 result = gen75_init_device_state(device);
1916 break;
1917 case 8:
1918 result = gen8_init_device_state(device);
1919 break;
1920 case 9:
1921 result = gen9_init_device_state(device);
1922 break;
1923 case 10:
1924 result = gen10_init_device_state(device);
1925 break;
1926 case 11:
1927 result = gen11_init_device_state(device);
1928 break;
1929 default:
1930 /* Shouldn't get here as we don't create physical devices for any other
1931 * gens. */
1932 unreachable("unhandled gen");
1933 }
1934 if (result != VK_SUCCESS)
1935 goto fail_workaround_bo;
1936
1937 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
1938
1939 anv_device_init_blorp(device);
1940
1941 anv_device_init_border_colors(device);
1942
1943 *pDevice = anv_device_to_handle(device);
1944
1945 return VK_SUCCESS;
1946
1947 fail_workaround_bo:
1948 anv_queue_finish(&device->queue);
1949 anv_scratch_pool_finish(device, &device->scratch_pool);
1950 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1951 anv_gem_close(device, device->workaround_bo.gem_handle);
1952 fail_binding_table_pool:
1953 if (physical_device->use_softpin)
1954 anv_state_pool_finish(&device->binding_table_pool);
1955 fail_surface_state_pool:
1956 anv_state_pool_finish(&device->surface_state_pool);
1957 fail_instruction_state_pool:
1958 anv_state_pool_finish(&device->instruction_state_pool);
1959 fail_dynamic_state_pool:
1960 anv_state_pool_finish(&device->dynamic_state_pool);
1961 fail_bo_cache:
1962 anv_bo_cache_finish(&device->bo_cache);
1963 fail_batch_bo_pool:
1964 anv_bo_pool_finish(&device->batch_bo_pool);
1965 pthread_cond_destroy(&device->queue_submit);
1966 fail_mutex:
1967 pthread_mutex_destroy(&device->mutex);
1968 fail_context_id:
1969 anv_gem_destroy_context(device, device->context_id);
1970 fail_fd:
1971 close(device->fd);
1972 fail_device:
1973 vk_free(&device->alloc, device);
1974
1975 return result;
1976 }
1977
1978 void anv_DestroyDevice(
1979 VkDevice _device,
1980 const VkAllocationCallbacks* pAllocator)
1981 {
1982 ANV_FROM_HANDLE(anv_device, device, _device);
1983 struct anv_physical_device *physical_device;
1984
1985 if (!device)
1986 return;
1987
1988 physical_device = &device->instance->physicalDevice;
1989
1990 anv_device_finish_blorp(device);
1991
1992 anv_pipeline_cache_finish(&device->default_pipeline_cache);
1993
1994 anv_queue_finish(&device->queue);
1995
1996 #ifdef HAVE_VALGRIND
1997 /* We only need to free these to prevent valgrind errors. The backing
1998 * BO will go away in a couple of lines so we don't actually leak.
1999 */
2000 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2001 #endif
2002
2003 anv_scratch_pool_finish(device, &device->scratch_pool);
2004
2005 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
2006 anv_vma_free(device, &device->workaround_bo);
2007 anv_gem_close(device, device->workaround_bo.gem_handle);
2008
2009 anv_vma_free(device, &device->trivial_batch_bo);
2010 anv_gem_close(device, device->trivial_batch_bo.gem_handle);
2011 if (device->info.gen >= 10)
2012 anv_gem_close(device, device->hiz_clear_bo.gem_handle);
2013
2014 if (physical_device->use_softpin)
2015 anv_state_pool_finish(&device->binding_table_pool);
2016 anv_state_pool_finish(&device->surface_state_pool);
2017 anv_state_pool_finish(&device->instruction_state_pool);
2018 anv_state_pool_finish(&device->dynamic_state_pool);
2019
2020 anv_bo_cache_finish(&device->bo_cache);
2021
2022 anv_bo_pool_finish(&device->batch_bo_pool);
2023
2024 pthread_cond_destroy(&device->queue_submit);
2025 pthread_mutex_destroy(&device->mutex);
2026
2027 anv_gem_destroy_context(device, device->context_id);
2028
2029 close(device->fd);
2030
2031 vk_free(&device->alloc, device);
2032 }
2033
2034 VkResult anv_EnumerateInstanceLayerProperties(
2035 uint32_t* pPropertyCount,
2036 VkLayerProperties* pProperties)
2037 {
2038 if (pProperties == NULL) {
2039 *pPropertyCount = 0;
2040 return VK_SUCCESS;
2041 }
2042
2043 /* None supported at this time */
2044 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2045 }
2046
2047 VkResult anv_EnumerateDeviceLayerProperties(
2048 VkPhysicalDevice physicalDevice,
2049 uint32_t* pPropertyCount,
2050 VkLayerProperties* pProperties)
2051 {
2052 if (pProperties == NULL) {
2053 *pPropertyCount = 0;
2054 return VK_SUCCESS;
2055 }
2056
2057 /* None supported at this time */
2058 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2059 }
2060
2061 void anv_GetDeviceQueue(
2062 VkDevice _device,
2063 uint32_t queueNodeIndex,
2064 uint32_t queueIndex,
2065 VkQueue* pQueue)
2066 {
2067 ANV_FROM_HANDLE(anv_device, device, _device);
2068
2069 assert(queueIndex == 0);
2070
2071 *pQueue = anv_queue_to_handle(&device->queue);
2072 }
2073
2074 void anv_GetDeviceQueue2(
2075 VkDevice _device,
2076 const VkDeviceQueueInfo2* pQueueInfo,
2077 VkQueue* pQueue)
2078 {
2079 ANV_FROM_HANDLE(anv_device, device, _device);
2080
2081 assert(pQueueInfo->queueIndex == 0);
2082
2083 if (pQueueInfo->flags == device->queue.flags)
2084 *pQueue = anv_queue_to_handle(&device->queue);
2085 else
2086 *pQueue = NULL;
2087 }
2088
2089 VkResult
2090 _anv_device_set_lost(struct anv_device *device,
2091 const char *file, int line,
2092 const char *msg, ...)
2093 {
2094 VkResult err;
2095 va_list ap;
2096
2097 device->_lost = true;
2098
2099 va_start(ap, msg);
2100 err = __vk_errorv(device->instance, device,
2101 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2102 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2103 va_end(ap);
2104
2105 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2106 abort();
2107
2108 return err;
2109 }
2110
2111 VkResult
2112 anv_device_query_status(struct anv_device *device)
2113 {
2114 /* This isn't likely as most of the callers of this function already check
2115 * for it. However, it doesn't hurt to check and it potentially lets us
2116 * avoid an ioctl.
2117 */
2118 if (anv_device_is_lost(device))
2119 return VK_ERROR_DEVICE_LOST;
2120
2121 uint32_t active, pending;
2122 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2123 if (ret == -1) {
2124 /* We don't know the real error. */
2125 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2126 }
2127
2128 if (active) {
2129 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2130 } else if (pending) {
2131 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2132 }
2133
2134 return VK_SUCCESS;
2135 }
2136
2137 VkResult
2138 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2139 {
2140 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2141 * Other usages of the BO (such as on different hardware) will not be
2142 * flagged as "busy" by this ioctl. Use with care.
2143 */
2144 int ret = anv_gem_busy(device, bo->gem_handle);
2145 if (ret == 1) {
2146 return VK_NOT_READY;
2147 } else if (ret == -1) {
2148 /* We don't know the real error. */
2149 return anv_device_set_lost(device, "gem wait failed: %m");
2150 }
2151
2152 /* Query for device status after the busy call. If the BO we're checking
2153 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2154 * client because it clearly doesn't have valid data. Yes, this most
2155 * likely means an ioctl, but we just did an ioctl to query the busy status
2156 * so it's no great loss.
2157 */
2158 return anv_device_query_status(device);
2159 }
2160
2161 VkResult
2162 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2163 int64_t timeout)
2164 {
2165 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2166 if (ret == -1 && errno == ETIME) {
2167 return VK_TIMEOUT;
2168 } else if (ret == -1) {
2169 /* We don't know the real error. */
2170 return anv_device_set_lost(device, "gem wait failed: %m");
2171 }
2172
2173 /* Query for device status after the wait. If the BO we're waiting on got
2174 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2175 * because it clearly doesn't have valid data. Yes, this most likely means
2176 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2177 */
2178 return anv_device_query_status(device);
2179 }
2180
2181 VkResult anv_DeviceWaitIdle(
2182 VkDevice _device)
2183 {
2184 ANV_FROM_HANDLE(anv_device, device, _device);
2185 if (anv_device_is_lost(device))
2186 return VK_ERROR_DEVICE_LOST;
2187
2188 struct anv_batch batch;
2189
2190 uint32_t cmds[8];
2191 batch.start = batch.next = cmds;
2192 batch.end = (void *) cmds + sizeof(cmds);
2193
2194 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2195 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2196
2197 return anv_device_submit_simple_batch(device, &batch);
2198 }
2199
2200 bool
2201 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo)
2202 {
2203 if (!(bo->flags & EXEC_OBJECT_PINNED))
2204 return true;
2205
2206 pthread_mutex_lock(&device->vma_mutex);
2207
2208 bo->offset = 0;
2209
2210 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS &&
2211 device->vma_hi_available >= bo->size) {
2212 uint64_t addr = util_vma_heap_alloc(&device->vma_hi, bo->size, 4096);
2213 if (addr) {
2214 bo->offset = gen_canonical_address(addr);
2215 assert(addr == gen_48b_address(bo->offset));
2216 device->vma_hi_available -= bo->size;
2217 }
2218 }
2219
2220 if (bo->offset == 0 && device->vma_lo_available >= bo->size) {
2221 uint64_t addr = util_vma_heap_alloc(&device->vma_lo, bo->size, 4096);
2222 if (addr) {
2223 bo->offset = gen_canonical_address(addr);
2224 assert(addr == gen_48b_address(bo->offset));
2225 device->vma_lo_available -= bo->size;
2226 }
2227 }
2228
2229 pthread_mutex_unlock(&device->vma_mutex);
2230
2231 return bo->offset != 0;
2232 }
2233
2234 void
2235 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
2236 {
2237 if (!(bo->flags & EXEC_OBJECT_PINNED))
2238 return;
2239
2240 const uint64_t addr_48b = gen_48b_address(bo->offset);
2241
2242 pthread_mutex_lock(&device->vma_mutex);
2243
2244 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
2245 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
2246 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
2247 device->vma_lo_available += bo->size;
2248 } else {
2249 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS &&
2250 addr_48b <= HIGH_HEAP_MAX_ADDRESS);
2251 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
2252 device->vma_hi_available += bo->size;
2253 }
2254
2255 pthread_mutex_unlock(&device->vma_mutex);
2256
2257 bo->offset = 0;
2258 }
2259
2260 VkResult
2261 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
2262 {
2263 uint32_t gem_handle = anv_gem_create(device, size);
2264 if (!gem_handle)
2265 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
2266
2267 anv_bo_init(bo, gem_handle, size);
2268
2269 return VK_SUCCESS;
2270 }
2271
2272 VkResult anv_AllocateMemory(
2273 VkDevice _device,
2274 const VkMemoryAllocateInfo* pAllocateInfo,
2275 const VkAllocationCallbacks* pAllocator,
2276 VkDeviceMemory* pMem)
2277 {
2278 ANV_FROM_HANDLE(anv_device, device, _device);
2279 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2280 struct anv_device_memory *mem;
2281 VkResult result = VK_SUCCESS;
2282
2283 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
2284
2285 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2286 assert(pAllocateInfo->allocationSize > 0);
2287
2288 if (pAllocateInfo->allocationSize > MAX_MEMORY_ALLOCATION_SIZE)
2289 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
2290
2291 /* FINISHME: Fail if allocation request exceeds heap size. */
2292
2293 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
2294 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2295 if (mem == NULL)
2296 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2297
2298 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
2299 mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
2300 mem->map = NULL;
2301 mem->map_size = 0;
2302 mem->ahw = NULL;
2303
2304 uint64_t bo_flags = 0;
2305
2306 assert(mem->type->heapIndex < pdevice->memory.heap_count);
2307 if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
2308 bo_flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2309
2310 const struct wsi_memory_allocate_info *wsi_info =
2311 vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
2312 if (wsi_info && wsi_info->implicit_sync) {
2313 /* We need to set the WRITE flag on window system buffers so that GEM
2314 * will know we're writing to them and synchronize uses on other rings
2315 * (eg if the display server uses the blitter ring).
2316 */
2317 bo_flags |= EXEC_OBJECT_WRITE;
2318 } else if (pdevice->has_exec_async) {
2319 bo_flags |= EXEC_OBJECT_ASYNC;
2320 }
2321
2322 if (pdevice->use_softpin)
2323 bo_flags |= EXEC_OBJECT_PINNED;
2324
2325 const VkExportMemoryAllocateInfo *export_info =
2326 vk_find_struct_const(pAllocateInfo->pNext, EXPORT_MEMORY_ALLOCATE_INFO);
2327
2328 /* Check if we need to support Android HW buffer export. If so,
2329 * create AHardwareBuffer and import memory from it.
2330 */
2331 bool android_export = false;
2332 if (export_info && export_info->handleTypes &
2333 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
2334 android_export = true;
2335
2336 /* Android memory import. */
2337 const struct VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info =
2338 vk_find_struct_const(pAllocateInfo->pNext,
2339 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID);
2340
2341 if (ahw_import_info) {
2342 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
2343 if (result != VK_SUCCESS)
2344 goto fail;
2345
2346 goto success;
2347 } else if (android_export) {
2348 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
2349 if (result != VK_SUCCESS)
2350 goto fail;
2351
2352 const struct VkImportAndroidHardwareBufferInfoANDROID import_info = {
2353 .buffer = mem->ahw,
2354 };
2355 result = anv_import_ahw_memory(_device, mem, &import_info);
2356 if (result != VK_SUCCESS)
2357 goto fail;
2358
2359 goto success;
2360 }
2361
2362 const VkImportMemoryFdInfoKHR *fd_info =
2363 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
2364
2365 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2366 * ignored.
2367 */
2368 if (fd_info && fd_info->handleType) {
2369 /* At the moment, we support only the below handle types. */
2370 assert(fd_info->handleType ==
2371 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2372 fd_info->handleType ==
2373 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2374
2375 result = anv_bo_cache_import(device, &device->bo_cache, fd_info->fd,
2376 bo_flags | ANV_BO_EXTERNAL, &mem->bo);
2377 if (result != VK_SUCCESS)
2378 goto fail;
2379
2380 VkDeviceSize aligned_alloc_size =
2381 align_u64(pAllocateInfo->allocationSize, 4096);
2382
2383 /* For security purposes, we reject importing the bo if it's smaller
2384 * than the requested allocation size. This prevents a malicious client
2385 * from passing a buffer to a trusted client, lying about the size, and
2386 * telling the trusted client to try and texture from an image that goes
2387 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2388 * in the trusted client. The trusted client can protect itself against
2389 * this sort of attack but only if it can trust the buffer size.
2390 */
2391 if (mem->bo->size < aligned_alloc_size) {
2392 result = vk_errorf(device->instance, device,
2393 VK_ERROR_INVALID_EXTERNAL_HANDLE,
2394 "aligned allocationSize too large for "
2395 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2396 "%"PRIu64"B > %"PRIu64"B",
2397 aligned_alloc_size, mem->bo->size);
2398 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2399 goto fail;
2400 }
2401
2402 /* From the Vulkan spec:
2403 *
2404 * "Importing memory from a file descriptor transfers ownership of
2405 * the file descriptor from the application to the Vulkan
2406 * implementation. The application must not perform any operations on
2407 * the file descriptor after a successful import."
2408 *
2409 * If the import fails, we leave the file descriptor open.
2410 */
2411 close(fd_info->fd);
2412 goto success;
2413 }
2414
2415 /* Regular allocate (not importing memory). */
2416
2417 if (export_info && export_info->handleTypes)
2418 bo_flags |= ANV_BO_EXTERNAL;
2419
2420 result = anv_bo_cache_alloc(device, &device->bo_cache,
2421 pAllocateInfo->allocationSize, bo_flags,
2422 &mem->bo);
2423 if (result != VK_SUCCESS)
2424 goto fail;
2425
2426 const VkMemoryDedicatedAllocateInfo *dedicated_info =
2427 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_DEDICATED_ALLOCATE_INFO);
2428 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
2429 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
2430
2431 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2432 * the BO. In this case, we have a dedicated allocation.
2433 */
2434 if (image->needs_set_tiling) {
2435 const uint32_t i915_tiling =
2436 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
2437 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
2438 image->planes[0].surface.isl.row_pitch_B,
2439 i915_tiling);
2440 if (ret) {
2441 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2442 return vk_errorf(device->instance, NULL,
2443 VK_ERROR_OUT_OF_DEVICE_MEMORY,
2444 "failed to set BO tiling: %m");
2445 }
2446 }
2447 }
2448
2449 success:
2450 *pMem = anv_device_memory_to_handle(mem);
2451
2452 return VK_SUCCESS;
2453
2454 fail:
2455 vk_free2(&device->alloc, pAllocator, mem);
2456
2457 return result;
2458 }
2459
2460 VkResult anv_GetMemoryFdKHR(
2461 VkDevice device_h,
2462 const VkMemoryGetFdInfoKHR* pGetFdInfo,
2463 int* pFd)
2464 {
2465 ANV_FROM_HANDLE(anv_device, dev, device_h);
2466 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
2467
2468 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2469
2470 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2471 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2472
2473 return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
2474 }
2475
2476 VkResult anv_GetMemoryFdPropertiesKHR(
2477 VkDevice _device,
2478 VkExternalMemoryHandleTypeFlagBits handleType,
2479 int fd,
2480 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
2481 {
2482 ANV_FROM_HANDLE(anv_device, device, _device);
2483 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2484
2485 switch (handleType) {
2486 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
2487 /* dma-buf can be imported as any memory type */
2488 pMemoryFdProperties->memoryTypeBits =
2489 (1 << pdevice->memory.type_count) - 1;
2490 return VK_SUCCESS;
2491
2492 default:
2493 /* The valid usage section for this function says:
2494 *
2495 * "handleType must not be one of the handle types defined as
2496 * opaque."
2497 *
2498 * So opaque handle types fall into the default "unsupported" case.
2499 */
2500 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
2501 }
2502 }
2503
2504 void anv_FreeMemory(
2505 VkDevice _device,
2506 VkDeviceMemory _mem,
2507 const VkAllocationCallbacks* pAllocator)
2508 {
2509 ANV_FROM_HANDLE(anv_device, device, _device);
2510 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
2511
2512 if (mem == NULL)
2513 return;
2514
2515 if (mem->map)
2516 anv_UnmapMemory(_device, _mem);
2517
2518 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2519
2520 #ifdef ANDROID
2521 if (mem->ahw)
2522 AHardwareBuffer_release(mem->ahw);
2523 #endif
2524
2525 vk_free2(&device->alloc, pAllocator, mem);
2526 }
2527
2528 VkResult anv_MapMemory(
2529 VkDevice _device,
2530 VkDeviceMemory _memory,
2531 VkDeviceSize offset,
2532 VkDeviceSize size,
2533 VkMemoryMapFlags flags,
2534 void** ppData)
2535 {
2536 ANV_FROM_HANDLE(anv_device, device, _device);
2537 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
2538
2539 if (mem == NULL) {
2540 *ppData = NULL;
2541 return VK_SUCCESS;
2542 }
2543
2544 if (size == VK_WHOLE_SIZE)
2545 size = mem->bo->size - offset;
2546
2547 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2548 *
2549 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2550 * assert(size != 0);
2551 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2552 * equal to the size of the memory minus offset
2553 */
2554 assert(size > 0);
2555 assert(offset + size <= mem->bo->size);
2556
2557 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2558 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2559 * at a time is valid. We could just mmap up front and return an offset
2560 * pointer here, but that may exhaust virtual memory on 32 bit
2561 * userspace. */
2562
2563 uint32_t gem_flags = 0;
2564
2565 if (!device->info.has_llc &&
2566 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
2567 gem_flags |= I915_MMAP_WC;
2568
2569 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2570 uint64_t map_offset = offset & ~4095ull;
2571 assert(offset >= map_offset);
2572 uint64_t map_size = (offset + size) - map_offset;
2573
2574 /* Let's map whole pages */
2575 map_size = align_u64(map_size, 4096);
2576
2577 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
2578 map_offset, map_size, gem_flags);
2579 if (map == MAP_FAILED)
2580 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
2581
2582 mem->map = map;
2583 mem->map_size = map_size;
2584
2585 *ppData = mem->map + (offset - map_offset);
2586
2587 return VK_SUCCESS;
2588 }
2589
2590 void anv_UnmapMemory(
2591 VkDevice _device,
2592 VkDeviceMemory _memory)
2593 {
2594 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
2595
2596 if (mem == NULL)
2597 return;
2598
2599 anv_gem_munmap(mem->map, mem->map_size);
2600
2601 mem->map = NULL;
2602 mem->map_size = 0;
2603 }
2604
2605 static void
2606 clflush_mapped_ranges(struct anv_device *device,
2607 uint32_t count,
2608 const VkMappedMemoryRange *ranges)
2609 {
2610 for (uint32_t i = 0; i < count; i++) {
2611 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
2612 if (ranges[i].offset >= mem->map_size)
2613 continue;
2614
2615 gen_clflush_range(mem->map + ranges[i].offset,
2616 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
2617 }
2618 }
2619
2620 VkResult anv_FlushMappedMemoryRanges(
2621 VkDevice _device,
2622 uint32_t memoryRangeCount,
2623 const VkMappedMemoryRange* pMemoryRanges)
2624 {
2625 ANV_FROM_HANDLE(anv_device, device, _device);
2626
2627 if (device->info.has_llc)
2628 return VK_SUCCESS;
2629
2630 /* Make sure the writes we're flushing have landed. */
2631 __builtin_ia32_mfence();
2632
2633 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
2634
2635 return VK_SUCCESS;
2636 }
2637
2638 VkResult anv_InvalidateMappedMemoryRanges(
2639 VkDevice _device,
2640 uint32_t memoryRangeCount,
2641 const VkMappedMemoryRange* pMemoryRanges)
2642 {
2643 ANV_FROM_HANDLE(anv_device, device, _device);
2644
2645 if (device->info.has_llc)
2646 return VK_SUCCESS;
2647
2648 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
2649
2650 /* Make sure no reads get moved up above the invalidate. */
2651 __builtin_ia32_mfence();
2652
2653 return VK_SUCCESS;
2654 }
2655
2656 void anv_GetBufferMemoryRequirements(
2657 VkDevice _device,
2658 VkBuffer _buffer,
2659 VkMemoryRequirements* pMemoryRequirements)
2660 {
2661 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2662 ANV_FROM_HANDLE(anv_device, device, _device);
2663 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2664
2665 /* The Vulkan spec (git aaed022) says:
2666 *
2667 * memoryTypeBits is a bitfield and contains one bit set for every
2668 * supported memory type for the resource. The bit `1<<i` is set if and
2669 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2670 * structure for the physical device is supported.
2671 */
2672 uint32_t memory_types = 0;
2673 for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
2674 uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
2675 if ((valid_usage & buffer->usage) == buffer->usage)
2676 memory_types |= (1u << i);
2677 }
2678
2679 /* Base alignment requirement of a cache line */
2680 uint32_t alignment = 16;
2681
2682 /* We need an alignment of 32 for pushing UBOs */
2683 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
2684 alignment = MAX2(alignment, 32);
2685
2686 pMemoryRequirements->size = buffer->size;
2687 pMemoryRequirements->alignment = alignment;
2688
2689 /* Storage and Uniform buffers should have their size aligned to
2690 * 32-bits to avoid boundary checks when last DWord is not complete.
2691 * This would ensure that not internal padding would be needed for
2692 * 16-bit types.
2693 */
2694 if (device->robust_buffer_access &&
2695 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
2696 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
2697 pMemoryRequirements->size = align_u64(buffer->size, 4);
2698
2699 pMemoryRequirements->memoryTypeBits = memory_types;
2700 }
2701
2702 void anv_GetBufferMemoryRequirements2(
2703 VkDevice _device,
2704 const VkBufferMemoryRequirementsInfo2* pInfo,
2705 VkMemoryRequirements2* pMemoryRequirements)
2706 {
2707 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
2708 &pMemoryRequirements->memoryRequirements);
2709
2710 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
2711 switch (ext->sType) {
2712 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
2713 VkMemoryDedicatedRequirements *requirements = (void *)ext;
2714 requirements->prefersDedicatedAllocation = VK_FALSE;
2715 requirements->requiresDedicatedAllocation = VK_FALSE;
2716 break;
2717 }
2718
2719 default:
2720 anv_debug_ignored_stype(ext->sType);
2721 break;
2722 }
2723 }
2724 }
2725
2726 void anv_GetImageMemoryRequirements(
2727 VkDevice _device,
2728 VkImage _image,
2729 VkMemoryRequirements* pMemoryRequirements)
2730 {
2731 ANV_FROM_HANDLE(anv_image, image, _image);
2732 ANV_FROM_HANDLE(anv_device, device, _device);
2733 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2734
2735 /* The Vulkan spec (git aaed022) says:
2736 *
2737 * memoryTypeBits is a bitfield and contains one bit set for every
2738 * supported memory type for the resource. The bit `1<<i` is set if and
2739 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2740 * structure for the physical device is supported.
2741 *
2742 * All types are currently supported for images.
2743 */
2744 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
2745
2746 /* We must have image allocated or imported at this point. According to the
2747 * specification, external images must have been bound to memory before
2748 * calling GetImageMemoryRequirements.
2749 */
2750 assert(image->size > 0);
2751
2752 pMemoryRequirements->size = image->size;
2753 pMemoryRequirements->alignment = image->alignment;
2754 pMemoryRequirements->memoryTypeBits = memory_types;
2755 }
2756
2757 void anv_GetImageMemoryRequirements2(
2758 VkDevice _device,
2759 const VkImageMemoryRequirementsInfo2* pInfo,
2760 VkMemoryRequirements2* pMemoryRequirements)
2761 {
2762 ANV_FROM_HANDLE(anv_device, device, _device);
2763 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
2764
2765 anv_GetImageMemoryRequirements(_device, pInfo->image,
2766 &pMemoryRequirements->memoryRequirements);
2767
2768 vk_foreach_struct_const(ext, pInfo->pNext) {
2769 switch (ext->sType) {
2770 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
2771 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2772 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
2773 (const VkImagePlaneMemoryRequirementsInfo *) ext;
2774 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
2775 plane_reqs->planeAspect);
2776
2777 assert(image->planes[plane].offset == 0);
2778
2779 /* The Vulkan spec (git aaed022) says:
2780 *
2781 * memoryTypeBits is a bitfield and contains one bit set for every
2782 * supported memory type for the resource. The bit `1<<i` is set
2783 * if and only if the memory type `i` in the
2784 * VkPhysicalDeviceMemoryProperties structure for the physical
2785 * device is supported.
2786 *
2787 * All types are currently supported for images.
2788 */
2789 pMemoryRequirements->memoryRequirements.memoryTypeBits =
2790 (1ull << pdevice->memory.type_count) - 1;
2791
2792 /* We must have image allocated or imported at this point. According to the
2793 * specification, external images must have been bound to memory before
2794 * calling GetImageMemoryRequirements.
2795 */
2796 assert(image->planes[plane].size > 0);
2797
2798 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
2799 pMemoryRequirements->memoryRequirements.alignment =
2800 image->planes[plane].alignment;
2801 break;
2802 }
2803
2804 default:
2805 anv_debug_ignored_stype(ext->sType);
2806 break;
2807 }
2808 }
2809
2810 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
2811 switch (ext->sType) {
2812 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
2813 VkMemoryDedicatedRequirements *requirements = (void *)ext;
2814 if (image->needs_set_tiling || image->external_format) {
2815 /* If we need to set the tiling for external consumers, we need a
2816 * dedicated allocation.
2817 *
2818 * See also anv_AllocateMemory.
2819 */
2820 requirements->prefersDedicatedAllocation = VK_TRUE;
2821 requirements->requiresDedicatedAllocation = VK_TRUE;
2822 } else {
2823 requirements->prefersDedicatedAllocation = VK_FALSE;
2824 requirements->requiresDedicatedAllocation = VK_FALSE;
2825 }
2826 break;
2827 }
2828
2829 default:
2830 anv_debug_ignored_stype(ext->sType);
2831 break;
2832 }
2833 }
2834 }
2835
2836 void anv_GetImageSparseMemoryRequirements(
2837 VkDevice device,
2838 VkImage image,
2839 uint32_t* pSparseMemoryRequirementCount,
2840 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
2841 {
2842 *pSparseMemoryRequirementCount = 0;
2843 }
2844
2845 void anv_GetImageSparseMemoryRequirements2(
2846 VkDevice device,
2847 const VkImageSparseMemoryRequirementsInfo2* pInfo,
2848 uint32_t* pSparseMemoryRequirementCount,
2849 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
2850 {
2851 *pSparseMemoryRequirementCount = 0;
2852 }
2853
2854 void anv_GetDeviceMemoryCommitment(
2855 VkDevice device,
2856 VkDeviceMemory memory,
2857 VkDeviceSize* pCommittedMemoryInBytes)
2858 {
2859 *pCommittedMemoryInBytes = 0;
2860 }
2861
2862 static void
2863 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
2864 {
2865 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
2866 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
2867
2868 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
2869
2870 if (mem) {
2871 assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
2872 buffer->address = (struct anv_address) {
2873 .bo = mem->bo,
2874 .offset = pBindInfo->memoryOffset,
2875 };
2876 } else {
2877 buffer->address = ANV_NULL_ADDRESS;
2878 }
2879 }
2880
2881 VkResult anv_BindBufferMemory(
2882 VkDevice device,
2883 VkBuffer buffer,
2884 VkDeviceMemory memory,
2885 VkDeviceSize memoryOffset)
2886 {
2887 anv_bind_buffer_memory(
2888 &(VkBindBufferMemoryInfo) {
2889 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
2890 .buffer = buffer,
2891 .memory = memory,
2892 .memoryOffset = memoryOffset,
2893 });
2894
2895 return VK_SUCCESS;
2896 }
2897
2898 VkResult anv_BindBufferMemory2(
2899 VkDevice device,
2900 uint32_t bindInfoCount,
2901 const VkBindBufferMemoryInfo* pBindInfos)
2902 {
2903 for (uint32_t i = 0; i < bindInfoCount; i++)
2904 anv_bind_buffer_memory(&pBindInfos[i]);
2905
2906 return VK_SUCCESS;
2907 }
2908
2909 VkResult anv_QueueBindSparse(
2910 VkQueue _queue,
2911 uint32_t bindInfoCount,
2912 const VkBindSparseInfo* pBindInfo,
2913 VkFence fence)
2914 {
2915 ANV_FROM_HANDLE(anv_queue, queue, _queue);
2916 if (anv_device_is_lost(queue->device))
2917 return VK_ERROR_DEVICE_LOST;
2918
2919 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2920 }
2921
2922 // Event functions
2923
2924 VkResult anv_CreateEvent(
2925 VkDevice _device,
2926 const VkEventCreateInfo* pCreateInfo,
2927 const VkAllocationCallbacks* pAllocator,
2928 VkEvent* pEvent)
2929 {
2930 ANV_FROM_HANDLE(anv_device, device, _device);
2931 struct anv_state state;
2932 struct anv_event *event;
2933
2934 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
2935
2936 state = anv_state_pool_alloc(&device->dynamic_state_pool,
2937 sizeof(*event), 8);
2938 event = state.map;
2939 event->state = state;
2940 event->semaphore = VK_EVENT_RESET;
2941
2942 if (!device->info.has_llc) {
2943 /* Make sure the writes we're flushing have landed. */
2944 __builtin_ia32_mfence();
2945 __builtin_ia32_clflush(event);
2946 }
2947
2948 *pEvent = anv_event_to_handle(event);
2949
2950 return VK_SUCCESS;
2951 }
2952
2953 void anv_DestroyEvent(
2954 VkDevice _device,
2955 VkEvent _event,
2956 const VkAllocationCallbacks* pAllocator)
2957 {
2958 ANV_FROM_HANDLE(anv_device, device, _device);
2959 ANV_FROM_HANDLE(anv_event, event, _event);
2960
2961 if (!event)
2962 return;
2963
2964 anv_state_pool_free(&device->dynamic_state_pool, event->state);
2965 }
2966
2967 VkResult anv_GetEventStatus(
2968 VkDevice _device,
2969 VkEvent _event)
2970 {
2971 ANV_FROM_HANDLE(anv_device, device, _device);
2972 ANV_FROM_HANDLE(anv_event, event, _event);
2973
2974 if (anv_device_is_lost(device))
2975 return VK_ERROR_DEVICE_LOST;
2976
2977 if (!device->info.has_llc) {
2978 /* Invalidate read cache before reading event written by GPU. */
2979 __builtin_ia32_clflush(event);
2980 __builtin_ia32_mfence();
2981
2982 }
2983
2984 return event->semaphore;
2985 }
2986
2987 VkResult anv_SetEvent(
2988 VkDevice _device,
2989 VkEvent _event)
2990 {
2991 ANV_FROM_HANDLE(anv_device, device, _device);
2992 ANV_FROM_HANDLE(anv_event, event, _event);
2993
2994 event->semaphore = VK_EVENT_SET;
2995
2996 if (!device->info.has_llc) {
2997 /* Make sure the writes we're flushing have landed. */
2998 __builtin_ia32_mfence();
2999 __builtin_ia32_clflush(event);
3000 }
3001
3002 return VK_SUCCESS;
3003 }
3004
3005 VkResult anv_ResetEvent(
3006 VkDevice _device,
3007 VkEvent _event)
3008 {
3009 ANV_FROM_HANDLE(anv_device, device, _device);
3010 ANV_FROM_HANDLE(anv_event, event, _event);
3011
3012 event->semaphore = VK_EVENT_RESET;
3013
3014 if (!device->info.has_llc) {
3015 /* Make sure the writes we're flushing have landed. */
3016 __builtin_ia32_mfence();
3017 __builtin_ia32_clflush(event);
3018 }
3019
3020 return VK_SUCCESS;
3021 }
3022
3023 // Buffer functions
3024
3025 VkResult anv_CreateBuffer(
3026 VkDevice _device,
3027 const VkBufferCreateInfo* pCreateInfo,
3028 const VkAllocationCallbacks* pAllocator,
3029 VkBuffer* pBuffer)
3030 {
3031 ANV_FROM_HANDLE(anv_device, device, _device);
3032 struct anv_buffer *buffer;
3033
3034 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
3035
3036 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
3037 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3038 if (buffer == NULL)
3039 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3040
3041 buffer->size = pCreateInfo->size;
3042 buffer->usage = pCreateInfo->usage;
3043 buffer->address = ANV_NULL_ADDRESS;
3044
3045 *pBuffer = anv_buffer_to_handle(buffer);
3046
3047 return VK_SUCCESS;
3048 }
3049
3050 void anv_DestroyBuffer(
3051 VkDevice _device,
3052 VkBuffer _buffer,
3053 const VkAllocationCallbacks* pAllocator)
3054 {
3055 ANV_FROM_HANDLE(anv_device, device, _device);
3056 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3057
3058 if (!buffer)
3059 return;
3060
3061 vk_free2(&device->alloc, pAllocator, buffer);
3062 }
3063
3064 void
3065 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
3066 enum isl_format format,
3067 struct anv_address address,
3068 uint32_t range, uint32_t stride)
3069 {
3070 isl_buffer_fill_state(&device->isl_dev, state.map,
3071 .address = anv_address_physical(address),
3072 .mocs = device->default_mocs,
3073 .size_B = range,
3074 .format = format,
3075 .stride_B = stride);
3076
3077 anv_state_flush(device, state);
3078 }
3079
3080 void anv_DestroySampler(
3081 VkDevice _device,
3082 VkSampler _sampler,
3083 const VkAllocationCallbacks* pAllocator)
3084 {
3085 ANV_FROM_HANDLE(anv_device, device, _device);
3086 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
3087
3088 if (!sampler)
3089 return;
3090
3091 vk_free2(&device->alloc, pAllocator, sampler);
3092 }
3093
3094 VkResult anv_CreateFramebuffer(
3095 VkDevice _device,
3096 const VkFramebufferCreateInfo* pCreateInfo,
3097 const VkAllocationCallbacks* pAllocator,
3098 VkFramebuffer* pFramebuffer)
3099 {
3100 ANV_FROM_HANDLE(anv_device, device, _device);
3101 struct anv_framebuffer *framebuffer;
3102
3103 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
3104
3105 size_t size = sizeof(*framebuffer) +
3106 sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
3107 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
3108 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3109 if (framebuffer == NULL)
3110 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3111
3112 framebuffer->attachment_count = pCreateInfo->attachmentCount;
3113 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
3114 VkImageView _iview = pCreateInfo->pAttachments[i];
3115 framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
3116 }
3117
3118 framebuffer->width = pCreateInfo->width;
3119 framebuffer->height = pCreateInfo->height;
3120 framebuffer->layers = pCreateInfo->layers;
3121
3122 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
3123
3124 return VK_SUCCESS;
3125 }
3126
3127 void anv_DestroyFramebuffer(
3128 VkDevice _device,
3129 VkFramebuffer _fb,
3130 const VkAllocationCallbacks* pAllocator)
3131 {
3132 ANV_FROM_HANDLE(anv_device, device, _device);
3133 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
3134
3135 if (!fb)
3136 return;
3137
3138 vk_free2(&device->alloc, pAllocator, fb);
3139 }
3140
3141 static const VkTimeDomainEXT anv_time_domains[] = {
3142 VK_TIME_DOMAIN_DEVICE_EXT,
3143 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
3144 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
3145 };
3146
3147 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3148 VkPhysicalDevice physicalDevice,
3149 uint32_t *pTimeDomainCount,
3150 VkTimeDomainEXT *pTimeDomains)
3151 {
3152 int d;
3153 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
3154
3155 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
3156 vk_outarray_append(&out, i) {
3157 *i = anv_time_domains[d];
3158 }
3159 }
3160
3161 return vk_outarray_status(&out);
3162 }
3163
3164 static uint64_t
3165 anv_clock_gettime(clockid_t clock_id)
3166 {
3167 struct timespec current;
3168 int ret;
3169
3170 ret = clock_gettime(clock_id, &current);
3171 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
3172 ret = clock_gettime(CLOCK_MONOTONIC, &current);
3173 if (ret < 0)
3174 return 0;
3175
3176 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
3177 }
3178
3179 #define TIMESTAMP 0x2358
3180
3181 VkResult anv_GetCalibratedTimestampsEXT(
3182 VkDevice _device,
3183 uint32_t timestampCount,
3184 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
3185 uint64_t *pTimestamps,
3186 uint64_t *pMaxDeviation)
3187 {
3188 ANV_FROM_HANDLE(anv_device, device, _device);
3189 uint64_t timestamp_frequency = device->info.timestamp_frequency;
3190 int ret;
3191 int d;
3192 uint64_t begin, end;
3193 uint64_t max_clock_period = 0;
3194
3195 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3196
3197 for (d = 0; d < timestampCount; d++) {
3198 switch (pTimestampInfos[d].timeDomain) {
3199 case VK_TIME_DOMAIN_DEVICE_EXT:
3200 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
3201 &pTimestamps[d]);
3202
3203 if (ret != 0) {
3204 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
3205 "register: %m");
3206 }
3207 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
3208 max_clock_period = MAX2(max_clock_period, device_period);
3209 break;
3210 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
3211 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
3212 max_clock_period = MAX2(max_clock_period, 1);
3213 break;
3214
3215 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
3216 pTimestamps[d] = begin;
3217 break;
3218 default:
3219 pTimestamps[d] = 0;
3220 break;
3221 }
3222 }
3223
3224 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3225
3226 /*
3227 * The maximum deviation is the sum of the interval over which we
3228 * perform the sampling and the maximum period of any sampled
3229 * clock. That's because the maximum skew between any two sampled
3230 * clock edges is when the sampled clock with the largest period is
3231 * sampled at the end of that period but right at the beginning of the
3232 * sampling interval and some other clock is sampled right at the
3233 * begining of its sampling period and right at the end of the
3234 * sampling interval. Let's assume the GPU has the longest clock
3235 * period and that the application is sampling GPU and monotonic:
3236 *
3237 * s e
3238 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3239 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3240 *
3241 * g
3242 * 0 1 2 3
3243 * GPU -----_____-----_____-----_____-----_____
3244 *
3245 * m
3246 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3247 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3248 *
3249 * Interval <----------------->
3250 * Deviation <-------------------------->
3251 *
3252 * s = read(raw) 2
3253 * g = read(GPU) 1
3254 * m = read(monotonic) 2
3255 * e = read(raw) b
3256 *
3257 * We round the sample interval up by one tick to cover sampling error
3258 * in the interval clock
3259 */
3260
3261 uint64_t sample_interval = end - begin + 1;
3262
3263 *pMaxDeviation = sample_interval + max_clock_period;
3264
3265 return VK_SUCCESS;
3266 }
3267
3268 /* vk_icd.h does not declare this function, so we declare it here to
3269 * suppress Wmissing-prototypes.
3270 */
3271 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3272 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
3273
3274 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3275 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
3276 {
3277 /* For the full details on loader interface versioning, see
3278 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3279 * What follows is a condensed summary, to help you navigate the large and
3280 * confusing official doc.
3281 *
3282 * - Loader interface v0 is incompatible with later versions. We don't
3283 * support it.
3284 *
3285 * - In loader interface v1:
3286 * - The first ICD entrypoint called by the loader is
3287 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3288 * entrypoint.
3289 * - The ICD must statically expose no other Vulkan symbol unless it is
3290 * linked with -Bsymbolic.
3291 * - Each dispatchable Vulkan handle created by the ICD must be
3292 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3293 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3294 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3295 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3296 * such loader-managed surfaces.
3297 *
3298 * - Loader interface v2 differs from v1 in:
3299 * - The first ICD entrypoint called by the loader is
3300 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3301 * statically expose this entrypoint.
3302 *
3303 * - Loader interface v3 differs from v2 in:
3304 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3305 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3306 * because the loader no longer does so.
3307 */
3308 *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
3309 return VK_SUCCESS;
3310 }