projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
anv: Implement VK_EXT_conditional_rendering for gen 7.5+
[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_CONDITIONAL_RENDERING_FEATURES_EXT: {
920 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
921 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
922 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
923
924 features->conditionalRendering = pdevice->info.gen >= 8 ||
925 pdevice->info.is_haswell;
926 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
927 pdevice->info.is_haswell;
928 break;
929 }
930
931 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
932 VkPhysicalDeviceMultiviewFeatures *features =
933 (VkPhysicalDeviceMultiviewFeatures *)ext;
934 features->multiview = true;
935 features->multiviewGeometryShader = true;
936 features->multiviewTessellationShader = true;
937 break;
938 }
939
940 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
941 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
942 features->protectedMemory = VK_FALSE;
943 break;
944 }
945
946 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
947 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
948 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
949 features->samplerYcbcrConversion = true;
950 break;
951 }
952
953 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
954 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
955 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
956 features->scalarBlockLayout = true;
957 break;
958 }
959
960 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES: {
961 VkPhysicalDeviceShaderDrawParameterFeatures *features = (void *)ext;
962 features->shaderDrawParameters = true;
963 break;
964 }
965
966 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES: {
967 VkPhysicalDeviceVariablePointerFeatures *features = (void *)ext;
968 features->variablePointersStorageBuffer = true;
969 features->variablePointers = true;
970 break;
971 }
972
973 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
974 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
975 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
976 features->vertexAttributeInstanceRateDivisor = VK_TRUE;
977 features->vertexAttributeInstanceRateZeroDivisor = VK_TRUE;
978 break;
979 }
980
981 default:
982 anv_debug_ignored_stype(ext->sType);
983 break;
984 }
985 }
986 }
987
988 void anv_GetPhysicalDeviceProperties(
989 VkPhysicalDevice physicalDevice,
990 VkPhysicalDeviceProperties* pProperties)
991 {
992 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
993 const struct gen_device_info *devinfo = &pdevice->info;
994
995 /* See assertions made when programming the buffer surface state. */
996 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
997 (1ul << 30) : (1ul << 27);
998
999 const uint32_t max_samplers = (devinfo->gen >= 8 || devinfo->is_haswell) ?
1000 128 : 16;
1001
1002 const uint32_t max_images = devinfo->gen < 9 ? MAX_GEN8_IMAGES : MAX_IMAGES;
1003
1004 VkSampleCountFlags sample_counts =
1005 isl_device_get_sample_counts(&pdevice->isl_dev);
1006
1007
1008 VkPhysicalDeviceLimits limits = {
1009 .maxImageDimension1D = (1 << 14),
1010 .maxImageDimension2D = (1 << 14),
1011 .maxImageDimension3D = (1 << 11),
1012 .maxImageDimensionCube = (1 << 14),
1013 .maxImageArrayLayers = (1 << 11),
1014 .maxTexelBufferElements = 128 * 1024 * 1024,
1015 .maxUniformBufferRange = (1ul << 27),
1016 .maxStorageBufferRange = max_raw_buffer_sz,
1017 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1018 .maxMemoryAllocationCount = UINT32_MAX,
1019 .maxSamplerAllocationCount = 64 * 1024,
1020 .bufferImageGranularity = 64, /* A cache line */
1021 .sparseAddressSpaceSize = 0,
1022 .maxBoundDescriptorSets = MAX_SETS,
1023 .maxPerStageDescriptorSamplers = max_samplers,
1024 .maxPerStageDescriptorUniformBuffers = 64,
1025 .maxPerStageDescriptorStorageBuffers = 64,
1026 .maxPerStageDescriptorSampledImages = max_samplers,
1027 .maxPerStageDescriptorStorageImages = max_images,
1028 .maxPerStageDescriptorInputAttachments = 64,
1029 .maxPerStageResources = 250,
1030 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1031 .maxDescriptorSetUniformBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
1032 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1033 .maxDescriptorSetStorageBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
1034 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1035 .maxDescriptorSetSampledImages = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSampledImages */
1036 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1037 .maxDescriptorSetInputAttachments = 256,
1038 .maxVertexInputAttributes = MAX_VBS,
1039 .maxVertexInputBindings = MAX_VBS,
1040 .maxVertexInputAttributeOffset = 2047,
1041 .maxVertexInputBindingStride = 2048,
1042 .maxVertexOutputComponents = 128,
1043 .maxTessellationGenerationLevel = 64,
1044 .maxTessellationPatchSize = 32,
1045 .maxTessellationControlPerVertexInputComponents = 128,
1046 .maxTessellationControlPerVertexOutputComponents = 128,
1047 .maxTessellationControlPerPatchOutputComponents = 128,
1048 .maxTessellationControlTotalOutputComponents = 2048,
1049 .maxTessellationEvaluationInputComponents = 128,
1050 .maxTessellationEvaluationOutputComponents = 128,
1051 .maxGeometryShaderInvocations = 32,
1052 .maxGeometryInputComponents = 64,
1053 .maxGeometryOutputComponents = 128,
1054 .maxGeometryOutputVertices = 256,
1055 .maxGeometryTotalOutputComponents = 1024,
1056 .maxFragmentInputComponents = 112, /* 128 components - (POS, PSIZ, CLIP_DIST0, CLIP_DIST1) */
1057 .maxFragmentOutputAttachments = 8,
1058 .maxFragmentDualSrcAttachments = 1,
1059 .maxFragmentCombinedOutputResources = 8,
1060 .maxComputeSharedMemorySize = 32768,
1061 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1062 .maxComputeWorkGroupInvocations = 16 * devinfo->max_cs_threads,
1063 .maxComputeWorkGroupSize = {
1064 16 * devinfo->max_cs_threads,
1065 16 * devinfo->max_cs_threads,
1066 16 * devinfo->max_cs_threads,
1067 },
1068 .subPixelPrecisionBits = 4 /* FIXME */,
1069 .subTexelPrecisionBits = 4 /* FIXME */,
1070 .mipmapPrecisionBits = 4 /* FIXME */,
1071 .maxDrawIndexedIndexValue = UINT32_MAX,
1072 .maxDrawIndirectCount = UINT32_MAX,
1073 .maxSamplerLodBias = 16,
1074 .maxSamplerAnisotropy = 16,
1075 .maxViewports = MAX_VIEWPORTS,
1076 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1077 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1078 .viewportSubPixelBits = 13, /* We take a float? */
1079 .minMemoryMapAlignment = 4096, /* A page */
1080 .minTexelBufferOffsetAlignment = 1,
1081 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1082 .minUniformBufferOffsetAlignment = 32,
1083 .minStorageBufferOffsetAlignment = 4,
1084 .minTexelOffset = -8,
1085 .maxTexelOffset = 7,
1086 .minTexelGatherOffset = -32,
1087 .maxTexelGatherOffset = 31,
1088 .minInterpolationOffset = -0.5,
1089 .maxInterpolationOffset = 0.4375,
1090 .subPixelInterpolationOffsetBits = 4,
1091 .maxFramebufferWidth = (1 << 14),
1092 .maxFramebufferHeight = (1 << 14),
1093 .maxFramebufferLayers = (1 << 11),
1094 .framebufferColorSampleCounts = sample_counts,
1095 .framebufferDepthSampleCounts = sample_counts,
1096 .framebufferStencilSampleCounts = sample_counts,
1097 .framebufferNoAttachmentsSampleCounts = sample_counts,
1098 .maxColorAttachments = MAX_RTS,
1099 .sampledImageColorSampleCounts = sample_counts,
1100 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1101 .sampledImageDepthSampleCounts = sample_counts,
1102 .sampledImageStencilSampleCounts = sample_counts,
1103 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1104 .maxSampleMaskWords = 1,
1105 .timestampComputeAndGraphics = false,
1106 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1107 .maxClipDistances = 8,
1108 .maxCullDistances = 8,
1109 .maxCombinedClipAndCullDistances = 8,
1110 .discreteQueuePriorities = 2,
1111 .pointSizeRange = { 0.125, 255.875 },
1112 .lineWidthRange = { 0.0, 7.9921875 },
1113 .pointSizeGranularity = (1.0 / 8.0),
1114 .lineWidthGranularity = (1.0 / 128.0),
1115 .strictLines = false, /* FINISHME */
1116 .standardSampleLocations = true,
1117 .optimalBufferCopyOffsetAlignment = 128,
1118 .optimalBufferCopyRowPitchAlignment = 128,
1119 .nonCoherentAtomSize = 64,
1120 };
1121
1122 *pProperties = (VkPhysicalDeviceProperties) {
1123 .apiVersion = anv_physical_device_api_version(pdevice),
1124 .driverVersion = vk_get_driver_version(),
1125 .vendorID = 0x8086,
1126 .deviceID = pdevice->chipset_id,
1127 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1128 .limits = limits,
1129 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1130 };
1131
1132 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1133 "%s", pdevice->name);
1134 memcpy(pProperties->pipelineCacheUUID,
1135 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1136 }
1137
1138 void anv_GetPhysicalDeviceProperties2(
1139 VkPhysicalDevice physicalDevice,
1140 VkPhysicalDeviceProperties2* pProperties)
1141 {
1142 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1143
1144 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1145
1146 vk_foreach_struct(ext, pProperties->pNext) {
1147 switch (ext->sType) {
1148 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1149 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *props =
1150 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1151
1152 /* We support all of the depth resolve modes */
1153 props->supportedDepthResolveModes =
1154 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1155 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1156 VK_RESOLVE_MODE_MIN_BIT_KHR |
1157 VK_RESOLVE_MODE_MAX_BIT_KHR;
1158
1159 /* Average doesn't make sense for stencil so we don't support that */
1160 props->supportedStencilResolveModes =
1161 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1162 if (pdevice->info.gen >= 8) {
1163 /* The advanced stencil resolve modes currently require stencil
1164 * sampling be supported by the hardware.
1165 */
1166 props->supportedStencilResolveModes |=
1167 VK_RESOLVE_MODE_MIN_BIT_KHR |
1168 VK_RESOLVE_MODE_MAX_BIT_KHR;
1169 }
1170
1171 props->independentResolveNone = VK_TRUE;
1172 props->independentResolve = VK_TRUE;
1173 break;
1174 }
1175
1176 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1177 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1178 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1179
1180 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1181 util_snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1182 "Intel open-source Mesa driver");
1183
1184 util_snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1185 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1186
1187 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1188 .major = 1,
1189 .minor = 1,
1190 .subminor = 2,
1191 .patch = 0,
1192 };
1193 break;
1194 }
1195
1196 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1197 VkPhysicalDeviceIDProperties *id_props =
1198 (VkPhysicalDeviceIDProperties *)ext;
1199 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1200 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1201 /* The LUID is for Windows. */
1202 id_props->deviceLUIDValid = false;
1203 break;
1204 }
1205
1206 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1207 VkPhysicalDeviceMaintenance3Properties *props =
1208 (VkPhysicalDeviceMaintenance3Properties *)ext;
1209 /* This value doesn't matter for us today as our per-stage
1210 * descriptors are the real limit.
1211 */
1212 props->maxPerSetDescriptors = 1024;
1213 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1214 break;
1215 }
1216
1217 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1218 VkPhysicalDeviceMultiviewProperties *properties =
1219 (VkPhysicalDeviceMultiviewProperties *)ext;
1220 properties->maxMultiviewViewCount = 16;
1221 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1222 break;
1223 }
1224
1225 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1226 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1227 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1228 properties->pciDomain = pdevice->pci_info.domain;
1229 properties->pciBus = pdevice->pci_info.bus;
1230 properties->pciDevice = pdevice->pci_info.device;
1231 properties->pciFunction = pdevice->pci_info.function;
1232 break;
1233 }
1234
1235 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1236 VkPhysicalDevicePointClippingProperties *properties =
1237 (VkPhysicalDevicePointClippingProperties *) ext;
1238 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
1239 anv_finishme("Implement pop-free point clipping");
1240 break;
1241 }
1242
1243 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1244 VkPhysicalDeviceProtectedMemoryProperties *props =
1245 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1246 props->protectedNoFault = false;
1247 break;
1248 }
1249
1250 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1251 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1252 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1253
1254 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1255 break;
1256 }
1257
1258 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1259 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1260 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1261 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1262 properties->filterMinmaxSingleComponentFormats = true;
1263 break;
1264 }
1265
1266 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1267 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1268
1269 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1270
1271 VkShaderStageFlags scalar_stages = 0;
1272 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1273 if (pdevice->compiler->scalar_stage[stage])
1274 scalar_stages |= mesa_to_vk_shader_stage(stage);
1275 }
1276 properties->supportedStages = scalar_stages;
1277
1278 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1279 VK_SUBGROUP_FEATURE_VOTE_BIT |
1280 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1281 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1282 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1283 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1284 VK_SUBGROUP_FEATURE_CLUSTERED_BIT |
1285 VK_SUBGROUP_FEATURE_QUAD_BIT;
1286 properties->quadOperationsInAllStages = VK_TRUE;
1287 break;
1288 }
1289
1290 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1291 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1292 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1293 /* We have to restrict this a bit for multiview */
1294 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1295 break;
1296 }
1297
1298 default:
1299 anv_debug_ignored_stype(ext->sType);
1300 break;
1301 }
1302 }
1303 }
1304
1305 /* We support exactly one queue family. */
1306 static const VkQueueFamilyProperties
1307 anv_queue_family_properties = {
1308 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1309 VK_QUEUE_COMPUTE_BIT |
1310 VK_QUEUE_TRANSFER_BIT,
1311 .queueCount = 1,
1312 .timestampValidBits = 36, /* XXX: Real value here */
1313 .minImageTransferGranularity = { 1, 1, 1 },
1314 };
1315
1316 void anv_GetPhysicalDeviceQueueFamilyProperties(
1317 VkPhysicalDevice physicalDevice,
1318 uint32_t* pCount,
1319 VkQueueFamilyProperties* pQueueFamilyProperties)
1320 {
1321 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1322
1323 vk_outarray_append(&out, p) {
1324 *p = anv_queue_family_properties;
1325 }
1326 }
1327
1328 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1329 VkPhysicalDevice physicalDevice,
1330 uint32_t* pQueueFamilyPropertyCount,
1331 VkQueueFamilyProperties2* pQueueFamilyProperties)
1332 {
1333
1334 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1335
1336 vk_outarray_append(&out, p) {
1337 p->queueFamilyProperties = anv_queue_family_properties;
1338
1339 vk_foreach_struct(s, p->pNext) {
1340 anv_debug_ignored_stype(s->sType);
1341 }
1342 }
1343 }
1344
1345 void anv_GetPhysicalDeviceMemoryProperties(
1346 VkPhysicalDevice physicalDevice,
1347 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1348 {
1349 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1350
1351 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1352 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1353 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1354 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1355 .heapIndex = physical_device->memory.types[i].heapIndex,
1356 };
1357 }
1358
1359 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1360 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1361 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1362 .size = physical_device->memory.heaps[i].size,
1363 .flags = physical_device->memory.heaps[i].flags,
1364 };
1365 }
1366 }
1367
1368 void anv_GetPhysicalDeviceMemoryProperties2(
1369 VkPhysicalDevice physicalDevice,
1370 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1371 {
1372 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1373 &pMemoryProperties->memoryProperties);
1374
1375 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1376 switch (ext->sType) {
1377 default:
1378 anv_debug_ignored_stype(ext->sType);
1379 break;
1380 }
1381 }
1382 }
1383
1384 void
1385 anv_GetDeviceGroupPeerMemoryFeatures(
1386 VkDevice device,
1387 uint32_t heapIndex,
1388 uint32_t localDeviceIndex,
1389 uint32_t remoteDeviceIndex,
1390 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1391 {
1392 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1393 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1394 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1395 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1396 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1397 }
1398
1399 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1400 VkInstance _instance,
1401 const char* pName)
1402 {
1403 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1404
1405 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1406 * when we have to return valid function pointers, NULL, or it's left
1407 * undefined. See the table for exact details.
1408 */
1409 if (pName == NULL)
1410 return NULL;
1411
1412 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1413 if (strcmp(pName, "vk" #entrypoint) == 0) \
1414 return (PFN_vkVoidFunction)anv_##entrypoint
1415
1416 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1417 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1418 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1419 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1420
1421 #undef LOOKUP_ANV_ENTRYPOINT
1422
1423 if (instance == NULL)
1424 return NULL;
1425
1426 int idx = anv_get_instance_entrypoint_index(pName);
1427 if (idx >= 0)
1428 return instance->dispatch.entrypoints[idx];
1429
1430 idx = anv_get_device_entrypoint_index(pName);
1431 if (idx >= 0)
1432 return instance->device_dispatch.entrypoints[idx];
1433
1434 return NULL;
1435 }
1436
1437 /* With version 1+ of the loader interface the ICD should expose
1438 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1439 */
1440 PUBLIC
1441 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1442 VkInstance instance,
1443 const char* pName);
1444
1445 PUBLIC
1446 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1447 VkInstance instance,
1448 const char* pName)
1449 {
1450 return anv_GetInstanceProcAddr(instance, pName);
1451 }
1452
1453 PFN_vkVoidFunction anv_GetDeviceProcAddr(
1454 VkDevice _device,
1455 const char* pName)
1456 {
1457 ANV_FROM_HANDLE(anv_device, device, _device);
1458
1459 if (!device || !pName)
1460 return NULL;
1461
1462 int idx = anv_get_device_entrypoint_index(pName);
1463 if (idx < 0)
1464 return NULL;
1465
1466 return device->dispatch.entrypoints[idx];
1467 }
1468
1469 VkResult
1470 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
1471 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
1472 const VkAllocationCallbacks* pAllocator,
1473 VkDebugReportCallbackEXT* pCallback)
1474 {
1475 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1476 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
1477 pCreateInfo, pAllocator, &instance->alloc,
1478 pCallback);
1479 }
1480
1481 void
1482 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
1483 VkDebugReportCallbackEXT _callback,
1484 const VkAllocationCallbacks* pAllocator)
1485 {
1486 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1487 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
1488 _callback, pAllocator, &instance->alloc);
1489 }
1490
1491 void
1492 anv_DebugReportMessageEXT(VkInstance _instance,
1493 VkDebugReportFlagsEXT flags,
1494 VkDebugReportObjectTypeEXT objectType,
1495 uint64_t object,
1496 size_t location,
1497 int32_t messageCode,
1498 const char* pLayerPrefix,
1499 const char* pMessage)
1500 {
1501 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1502 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
1503 object, location, messageCode, pLayerPrefix, pMessage);
1504 }
1505
1506 static void
1507 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
1508 {
1509 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1510 queue->device = device;
1511 queue->flags = 0;
1512 }
1513
1514 static void
1515 anv_queue_finish(struct anv_queue *queue)
1516 {
1517 }
1518
1519 static struct anv_state
1520 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
1521 {
1522 struct anv_state state;
1523
1524 state = anv_state_pool_alloc(pool, size, align);
1525 memcpy(state.map, p, size);
1526
1527 return state;
1528 }
1529
1530 struct gen8_border_color {
1531 union {
1532 float float32[4];
1533 uint32_t uint32[4];
1534 };
1535 /* Pad out to 64 bytes */
1536 uint32_t _pad[12];
1537 };
1538
1539 static void
1540 anv_device_init_border_colors(struct anv_device *device)
1541 {
1542 static const struct gen8_border_color border_colors[] = {
1543 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
1544 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
1545 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
1546 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
1547 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
1548 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
1549 };
1550
1551 device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
1552 sizeof(border_colors), 64,
1553 border_colors);
1554 }
1555
1556 static void
1557 anv_device_init_trivial_batch(struct anv_device *device)
1558 {
1559 anv_bo_init_new(&device->trivial_batch_bo, device, 4096);
1560
1561 if (device->instance->physicalDevice.has_exec_async)
1562 device->trivial_batch_bo.flags |= EXEC_OBJECT_ASYNC;
1563
1564 if (device->instance->physicalDevice.use_softpin)
1565 device->trivial_batch_bo.flags |= EXEC_OBJECT_PINNED;
1566
1567 anv_vma_alloc(device, &device->trivial_batch_bo);
1568
1569 void *map = anv_gem_mmap(device, device->trivial_batch_bo.gem_handle,
1570 0, 4096, 0);
1571
1572 struct anv_batch batch = {
1573 .start = map,
1574 .next = map,
1575 .end = map + 4096,
1576 };
1577
1578 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1579 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1580
1581 if (!device->info.has_llc)
1582 gen_clflush_range(map, batch.next - map);
1583
1584 anv_gem_munmap(map, device->trivial_batch_bo.size);
1585 }
1586
1587 VkResult anv_EnumerateDeviceExtensionProperties(
1588 VkPhysicalDevice physicalDevice,
1589 const char* pLayerName,
1590 uint32_t* pPropertyCount,
1591 VkExtensionProperties* pProperties)
1592 {
1593 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1594 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
1595
1596 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
1597 if (device->supported_extensions.extensions[i]) {
1598 vk_outarray_append(&out, prop) {
1599 *prop = anv_device_extensions[i];
1600 }
1601 }
1602 }
1603
1604 return vk_outarray_status(&out);
1605 }
1606
1607 static void
1608 anv_device_init_dispatch(struct anv_device *device)
1609 {
1610 const struct anv_device_dispatch_table *genX_table;
1611 switch (device->info.gen) {
1612 case 11:
1613 genX_table = &gen11_device_dispatch_table;
1614 break;
1615 case 10:
1616 genX_table = &gen10_device_dispatch_table;
1617 break;
1618 case 9:
1619 genX_table = &gen9_device_dispatch_table;
1620 break;
1621 case 8:
1622 genX_table = &gen8_device_dispatch_table;
1623 break;
1624 case 7:
1625 if (device->info.is_haswell)
1626 genX_table = &gen75_device_dispatch_table;
1627 else
1628 genX_table = &gen7_device_dispatch_table;
1629 break;
1630 default:
1631 unreachable("unsupported gen\n");
1632 }
1633
1634 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
1635 /* Vulkan requires that entrypoints for extensions which have not been
1636 * enabled must not be advertised.
1637 */
1638 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
1639 &device->instance->enabled_extensions,
1640 &device->enabled_extensions)) {
1641 device->dispatch.entrypoints[i] = NULL;
1642 } else if (genX_table->entrypoints[i]) {
1643 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
1644 } else {
1645 device->dispatch.entrypoints[i] =
1646 anv_device_dispatch_table.entrypoints[i];
1647 }
1648 }
1649 }
1650
1651 static int
1652 vk_priority_to_gen(int priority)
1653 {
1654 switch (priority) {
1655 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
1656 return GEN_CONTEXT_LOW_PRIORITY;
1657 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
1658 return GEN_CONTEXT_MEDIUM_PRIORITY;
1659 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
1660 return GEN_CONTEXT_HIGH_PRIORITY;
1661 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
1662 return GEN_CONTEXT_REALTIME_PRIORITY;
1663 default:
1664 unreachable("Invalid priority");
1665 }
1666 }
1667
1668 static void
1669 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
1670 {
1671 anv_bo_init_new(&device->hiz_clear_bo, device, 4096);
1672
1673 if (device->instance->physicalDevice.has_exec_async)
1674 device->hiz_clear_bo.flags |= EXEC_OBJECT_ASYNC;
1675
1676 if (device->instance->physicalDevice.use_softpin)
1677 device->hiz_clear_bo.flags |= EXEC_OBJECT_PINNED;
1678
1679 anv_vma_alloc(device, &device->hiz_clear_bo);
1680
1681 uint32_t *map = anv_gem_mmap(device, device->hiz_clear_bo.gem_handle,
1682 0, 4096, 0);
1683
1684 union isl_color_value hiz_clear = { .u32 = { 0, } };
1685 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
1686
1687 memcpy(map, hiz_clear.u32, sizeof(hiz_clear.u32));
1688 anv_gem_munmap(map, device->hiz_clear_bo.size);
1689 }
1690
1691 VkResult anv_CreateDevice(
1692 VkPhysicalDevice physicalDevice,
1693 const VkDeviceCreateInfo* pCreateInfo,
1694 const VkAllocationCallbacks* pAllocator,
1695 VkDevice* pDevice)
1696 {
1697 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1698 VkResult result;
1699 struct anv_device *device;
1700
1701 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
1702
1703 struct anv_device_extension_table enabled_extensions = { };
1704 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
1705 int idx;
1706 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
1707 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
1708 anv_device_extensions[idx].extensionName) == 0)
1709 break;
1710 }
1711
1712 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
1713 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1714
1715 if (!physical_device->supported_extensions.extensions[idx])
1716 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
1717
1718 enabled_extensions.extensions[idx] = true;
1719 }
1720
1721 /* Check enabled features */
1722 if (pCreateInfo->pEnabledFeatures) {
1723 VkPhysicalDeviceFeatures supported_features;
1724 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
1725 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
1726 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
1727 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
1728 for (uint32_t i = 0; i < num_features; i++) {
1729 if (enabled_feature[i] && !supported_feature[i])
1730 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
1731 }
1732 }
1733
1734 /* Check requested queues and fail if we are requested to create any
1735 * queues with flags we don't support.
1736 */
1737 assert(pCreateInfo->queueCreateInfoCount > 0);
1738 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
1739 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
1740 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
1741 }
1742
1743 /* Check if client specified queue priority. */
1744 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
1745 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
1746 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
1747
1748 VkQueueGlobalPriorityEXT priority =
1749 queue_priority ? queue_priority->globalPriority :
1750 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
1751
1752 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
1753 sizeof(*device), 8,
1754 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1755 if (!device)
1756 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1757
1758 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1759 device->instance = physical_device->instance;
1760 device->chipset_id = physical_device->chipset_id;
1761 device->no_hw = physical_device->no_hw;
1762 device->_lost = false;
1763
1764 if (pAllocator)
1765 device->alloc = *pAllocator;
1766 else
1767 device->alloc = physical_device->instance->alloc;
1768
1769 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1770 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1771 if (device->fd == -1) {
1772 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1773 goto fail_device;
1774 }
1775
1776 device->context_id = anv_gem_create_context(device);
1777 if (device->context_id == -1) {
1778 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1779 goto fail_fd;
1780 }
1781
1782 if (physical_device->use_softpin) {
1783 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
1784 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1785 goto fail_fd;
1786 }
1787
1788 /* keep the page with address zero out of the allocator */
1789 util_vma_heap_init(&device->vma_lo, LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
1790 device->vma_lo_available =
1791 physical_device->memory.heaps[physical_device->memory.heap_count - 1].size;
1792
1793 /* Leave the last 4GiB out of the high vma range, so that no state base
1794 * address + size can overflow 48 bits. For more information see the
1795 * comment about Wa32bitGeneralStateOffset in anv_allocator.c
1796 */
1797 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
1798 HIGH_HEAP_SIZE);
1799 device->vma_hi_available = physical_device->memory.heap_count == 1 ? 0 :
1800 physical_device->memory.heaps[0].size;
1801 }
1802
1803 /* As per spec, the driver implementation may deny requests to acquire
1804 * a priority above the default priority (MEDIUM) if the caller does not
1805 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1806 * is returned.
1807 */
1808 if (physical_device->has_context_priority) {
1809 int err = anv_gem_set_context_param(device->fd, device->context_id,
1810 I915_CONTEXT_PARAM_PRIORITY,
1811 vk_priority_to_gen(priority));
1812 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
1813 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
1814 goto fail_fd;
1815 }
1816 }
1817
1818 device->info = physical_device->info;
1819 device->isl_dev = physical_device->isl_dev;
1820
1821 /* On Broadwell and later, we can use batch chaining to more efficiently
1822 * implement growing command buffers. Prior to Haswell, the kernel
1823 * command parser gets in the way and we have to fall back to growing
1824 * the batch.
1825 */
1826 device->can_chain_batches = device->info.gen >= 8;
1827
1828 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1829 pCreateInfo->pEnabledFeatures->robustBufferAccess;
1830 device->enabled_extensions = enabled_extensions;
1831
1832 anv_device_init_dispatch(device);
1833
1834 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
1835 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1836 goto fail_context_id;
1837 }
1838
1839 pthread_condattr_t condattr;
1840 if (pthread_condattr_init(&condattr) != 0) {
1841 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1842 goto fail_mutex;
1843 }
1844 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
1845 pthread_condattr_destroy(&condattr);
1846 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1847 goto fail_mutex;
1848 }
1849 if (pthread_cond_init(&device->queue_submit, NULL) != 0) {
1850 pthread_condattr_destroy(&condattr);
1851 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1852 goto fail_mutex;
1853 }
1854 pthread_condattr_destroy(&condattr);
1855
1856 uint64_t bo_flags =
1857 (physical_device->supports_48bit_addresses ? EXEC_OBJECT_SUPPORTS_48B_ADDRESS : 0) |
1858 (physical_device->has_exec_async ? EXEC_OBJECT_ASYNC : 0) |
1859 (physical_device->has_exec_capture ? EXEC_OBJECT_CAPTURE : 0) |
1860 (physical_device->use_softpin ? EXEC_OBJECT_PINNED : 0);
1861
1862 anv_bo_pool_init(&device->batch_bo_pool, device, bo_flags);
1863
1864 result = anv_bo_cache_init(&device->bo_cache);
1865 if (result != VK_SUCCESS)
1866 goto fail_batch_bo_pool;
1867
1868 if (!physical_device->use_softpin)
1869 bo_flags &= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
1870
1871 result = anv_state_pool_init(&device->dynamic_state_pool, device,
1872 DYNAMIC_STATE_POOL_MIN_ADDRESS,
1873 16384,
1874 bo_flags);
1875 if (result != VK_SUCCESS)
1876 goto fail_bo_cache;
1877
1878 result = anv_state_pool_init(&device->instruction_state_pool, device,
1879 INSTRUCTION_STATE_POOL_MIN_ADDRESS,
1880 16384,
1881 bo_flags);
1882 if (result != VK_SUCCESS)
1883 goto fail_dynamic_state_pool;
1884
1885 result = anv_state_pool_init(&device->surface_state_pool, device,
1886 SURFACE_STATE_POOL_MIN_ADDRESS,
1887 4096,
1888 bo_flags);
1889 if (result != VK_SUCCESS)
1890 goto fail_instruction_state_pool;
1891
1892 if (physical_device->use_softpin) {
1893 result = anv_state_pool_init(&device->binding_table_pool, device,
1894 BINDING_TABLE_POOL_MIN_ADDRESS,
1895 4096,
1896 bo_flags);
1897 if (result != VK_SUCCESS)
1898 goto fail_surface_state_pool;
1899 }
1900
1901 result = anv_bo_init_new(&device->workaround_bo, device, 1024);
1902 if (result != VK_SUCCESS)
1903 goto fail_binding_table_pool;
1904
1905 if (physical_device->use_softpin)
1906 device->workaround_bo.flags |= EXEC_OBJECT_PINNED;
1907
1908 if (!anv_vma_alloc(device, &device->workaround_bo))
1909 goto fail_workaround_bo;
1910
1911 anv_device_init_trivial_batch(device);
1912
1913 if (device->info.gen >= 10)
1914 anv_device_init_hiz_clear_value_bo(device);
1915
1916 anv_scratch_pool_init(device, &device->scratch_pool);
1917
1918 anv_queue_init(device, &device->queue);
1919
1920 switch (device->info.gen) {
1921 case 7:
1922 if (!device->info.is_haswell)
1923 result = gen7_init_device_state(device);
1924 else
1925 result = gen75_init_device_state(device);
1926 break;
1927 case 8:
1928 result = gen8_init_device_state(device);
1929 break;
1930 case 9:
1931 result = gen9_init_device_state(device);
1932 break;
1933 case 10:
1934 result = gen10_init_device_state(device);
1935 break;
1936 case 11:
1937 result = gen11_init_device_state(device);
1938 break;
1939 default:
1940 /* Shouldn't get here as we don't create physical devices for any other
1941 * gens. */
1942 unreachable("unhandled gen");
1943 }
1944 if (result != VK_SUCCESS)
1945 goto fail_workaround_bo;
1946
1947 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
1948
1949 anv_device_init_blorp(device);
1950
1951 anv_device_init_border_colors(device);
1952
1953 *pDevice = anv_device_to_handle(device);
1954
1955 return VK_SUCCESS;
1956
1957 fail_workaround_bo:
1958 anv_queue_finish(&device->queue);
1959 anv_scratch_pool_finish(device, &device->scratch_pool);
1960 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1961 anv_gem_close(device, device->workaround_bo.gem_handle);
1962 fail_binding_table_pool:
1963 if (physical_device->use_softpin)
1964 anv_state_pool_finish(&device->binding_table_pool);
1965 fail_surface_state_pool:
1966 anv_state_pool_finish(&device->surface_state_pool);
1967 fail_instruction_state_pool:
1968 anv_state_pool_finish(&device->instruction_state_pool);
1969 fail_dynamic_state_pool:
1970 anv_state_pool_finish(&device->dynamic_state_pool);
1971 fail_bo_cache:
1972 anv_bo_cache_finish(&device->bo_cache);
1973 fail_batch_bo_pool:
1974 anv_bo_pool_finish(&device->batch_bo_pool);
1975 pthread_cond_destroy(&device->queue_submit);
1976 fail_mutex:
1977 pthread_mutex_destroy(&device->mutex);
1978 fail_context_id:
1979 anv_gem_destroy_context(device, device->context_id);
1980 fail_fd:
1981 close(device->fd);
1982 fail_device:
1983 vk_free(&device->alloc, device);
1984
1985 return result;
1986 }
1987
1988 void anv_DestroyDevice(
1989 VkDevice _device,
1990 const VkAllocationCallbacks* pAllocator)
1991 {
1992 ANV_FROM_HANDLE(anv_device, device, _device);
1993 struct anv_physical_device *physical_device;
1994
1995 if (!device)
1996 return;
1997
1998 physical_device = &device->instance->physicalDevice;
1999
2000 anv_device_finish_blorp(device);
2001
2002 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2003
2004 anv_queue_finish(&device->queue);
2005
2006 #ifdef HAVE_VALGRIND
2007 /* We only need to free these to prevent valgrind errors. The backing
2008 * BO will go away in a couple of lines so we don't actually leak.
2009 */
2010 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2011 #endif
2012
2013 anv_scratch_pool_finish(device, &device->scratch_pool);
2014
2015 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
2016 anv_vma_free(device, &device->workaround_bo);
2017 anv_gem_close(device, device->workaround_bo.gem_handle);
2018
2019 anv_vma_free(device, &device->trivial_batch_bo);
2020 anv_gem_close(device, device->trivial_batch_bo.gem_handle);
2021 if (device->info.gen >= 10)
2022 anv_gem_close(device, device->hiz_clear_bo.gem_handle);
2023
2024 if (physical_device->use_softpin)
2025 anv_state_pool_finish(&device->binding_table_pool);
2026 anv_state_pool_finish(&device->surface_state_pool);
2027 anv_state_pool_finish(&device->instruction_state_pool);
2028 anv_state_pool_finish(&device->dynamic_state_pool);
2029
2030 anv_bo_cache_finish(&device->bo_cache);
2031
2032 anv_bo_pool_finish(&device->batch_bo_pool);
2033
2034 pthread_cond_destroy(&device->queue_submit);
2035 pthread_mutex_destroy(&device->mutex);
2036
2037 anv_gem_destroy_context(device, device->context_id);
2038
2039 close(device->fd);
2040
2041 vk_free(&device->alloc, device);
2042 }
2043
2044 VkResult anv_EnumerateInstanceLayerProperties(
2045 uint32_t* pPropertyCount,
2046 VkLayerProperties* pProperties)
2047 {
2048 if (pProperties == NULL) {
2049 *pPropertyCount = 0;
2050 return VK_SUCCESS;
2051 }
2052
2053 /* None supported at this time */
2054 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2055 }
2056
2057 VkResult anv_EnumerateDeviceLayerProperties(
2058 VkPhysicalDevice physicalDevice,
2059 uint32_t* pPropertyCount,
2060 VkLayerProperties* pProperties)
2061 {
2062 if (pProperties == NULL) {
2063 *pPropertyCount = 0;
2064 return VK_SUCCESS;
2065 }
2066
2067 /* None supported at this time */
2068 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2069 }
2070
2071 void anv_GetDeviceQueue(
2072 VkDevice _device,
2073 uint32_t queueNodeIndex,
2074 uint32_t queueIndex,
2075 VkQueue* pQueue)
2076 {
2077 ANV_FROM_HANDLE(anv_device, device, _device);
2078
2079 assert(queueIndex == 0);
2080
2081 *pQueue = anv_queue_to_handle(&device->queue);
2082 }
2083
2084 void anv_GetDeviceQueue2(
2085 VkDevice _device,
2086 const VkDeviceQueueInfo2* pQueueInfo,
2087 VkQueue* pQueue)
2088 {
2089 ANV_FROM_HANDLE(anv_device, device, _device);
2090
2091 assert(pQueueInfo->queueIndex == 0);
2092
2093 if (pQueueInfo->flags == device->queue.flags)
2094 *pQueue = anv_queue_to_handle(&device->queue);
2095 else
2096 *pQueue = NULL;
2097 }
2098
2099 VkResult
2100 _anv_device_set_lost(struct anv_device *device,
2101 const char *file, int line,
2102 const char *msg, ...)
2103 {
2104 VkResult err;
2105 va_list ap;
2106
2107 device->_lost = true;
2108
2109 va_start(ap, msg);
2110 err = __vk_errorv(device->instance, device,
2111 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2112 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2113 va_end(ap);
2114
2115 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2116 abort();
2117
2118 return err;
2119 }
2120
2121 VkResult
2122 anv_device_query_status(struct anv_device *device)
2123 {
2124 /* This isn't likely as most of the callers of this function already check
2125 * for it. However, it doesn't hurt to check and it potentially lets us
2126 * avoid an ioctl.
2127 */
2128 if (anv_device_is_lost(device))
2129 return VK_ERROR_DEVICE_LOST;
2130
2131 uint32_t active, pending;
2132 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2133 if (ret == -1) {
2134 /* We don't know the real error. */
2135 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2136 }
2137
2138 if (active) {
2139 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2140 } else if (pending) {
2141 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2142 }
2143
2144 return VK_SUCCESS;
2145 }
2146
2147 VkResult
2148 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2149 {
2150 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2151 * Other usages of the BO (such as on different hardware) will not be
2152 * flagged as "busy" by this ioctl. Use with care.
2153 */
2154 int ret = anv_gem_busy(device, bo->gem_handle);
2155 if (ret == 1) {
2156 return VK_NOT_READY;
2157 } else if (ret == -1) {
2158 /* We don't know the real error. */
2159 return anv_device_set_lost(device, "gem wait failed: %m");
2160 }
2161
2162 /* Query for device status after the busy call. If the BO we're checking
2163 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2164 * client because it clearly doesn't have valid data. Yes, this most
2165 * likely means an ioctl, but we just did an ioctl to query the busy status
2166 * so it's no great loss.
2167 */
2168 return anv_device_query_status(device);
2169 }
2170
2171 VkResult
2172 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2173 int64_t timeout)
2174 {
2175 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2176 if (ret == -1 && errno == ETIME) {
2177 return VK_TIMEOUT;
2178 } else if (ret == -1) {
2179 /* We don't know the real error. */
2180 return anv_device_set_lost(device, "gem wait failed: %m");
2181 }
2182
2183 /* Query for device status after the wait. If the BO we're waiting on got
2184 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2185 * because it clearly doesn't have valid data. Yes, this most likely means
2186 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2187 */
2188 return anv_device_query_status(device);
2189 }
2190
2191 VkResult anv_DeviceWaitIdle(
2192 VkDevice _device)
2193 {
2194 ANV_FROM_HANDLE(anv_device, device, _device);
2195 if (anv_device_is_lost(device))
2196 return VK_ERROR_DEVICE_LOST;
2197
2198 struct anv_batch batch;
2199
2200 uint32_t cmds[8];
2201 batch.start = batch.next = cmds;
2202 batch.end = (void *) cmds + sizeof(cmds);
2203
2204 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2205 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2206
2207 return anv_device_submit_simple_batch(device, &batch);
2208 }
2209
2210 bool
2211 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo)
2212 {
2213 if (!(bo->flags & EXEC_OBJECT_PINNED))
2214 return true;
2215
2216 pthread_mutex_lock(&device->vma_mutex);
2217
2218 bo->offset = 0;
2219
2220 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS &&
2221 device->vma_hi_available >= bo->size) {
2222 uint64_t addr = util_vma_heap_alloc(&device->vma_hi, bo->size, 4096);
2223 if (addr) {
2224 bo->offset = gen_canonical_address(addr);
2225 assert(addr == gen_48b_address(bo->offset));
2226 device->vma_hi_available -= bo->size;
2227 }
2228 }
2229
2230 if (bo->offset == 0 && device->vma_lo_available >= bo->size) {
2231 uint64_t addr = util_vma_heap_alloc(&device->vma_lo, bo->size, 4096);
2232 if (addr) {
2233 bo->offset = gen_canonical_address(addr);
2234 assert(addr == gen_48b_address(bo->offset));
2235 device->vma_lo_available -= bo->size;
2236 }
2237 }
2238
2239 pthread_mutex_unlock(&device->vma_mutex);
2240
2241 return bo->offset != 0;
2242 }
2243
2244 void
2245 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
2246 {
2247 if (!(bo->flags & EXEC_OBJECT_PINNED))
2248 return;
2249
2250 const uint64_t addr_48b = gen_48b_address(bo->offset);
2251
2252 pthread_mutex_lock(&device->vma_mutex);
2253
2254 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
2255 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
2256 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
2257 device->vma_lo_available += bo->size;
2258 } else {
2259 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS &&
2260 addr_48b <= HIGH_HEAP_MAX_ADDRESS);
2261 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
2262 device->vma_hi_available += bo->size;
2263 }
2264
2265 pthread_mutex_unlock(&device->vma_mutex);
2266
2267 bo->offset = 0;
2268 }
2269
2270 VkResult
2271 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
2272 {
2273 uint32_t gem_handle = anv_gem_create(device, size);
2274 if (!gem_handle)
2275 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
2276
2277 anv_bo_init(bo, gem_handle, size);
2278
2279 return VK_SUCCESS;
2280 }
2281
2282 VkResult anv_AllocateMemory(
2283 VkDevice _device,
2284 const VkMemoryAllocateInfo* pAllocateInfo,
2285 const VkAllocationCallbacks* pAllocator,
2286 VkDeviceMemory* pMem)
2287 {
2288 ANV_FROM_HANDLE(anv_device, device, _device);
2289 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2290 struct anv_device_memory *mem;
2291 VkResult result = VK_SUCCESS;
2292
2293 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
2294
2295 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2296 assert(pAllocateInfo->allocationSize > 0);
2297
2298 if (pAllocateInfo->allocationSize > MAX_MEMORY_ALLOCATION_SIZE)
2299 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
2300
2301 /* FINISHME: Fail if allocation request exceeds heap size. */
2302
2303 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
2304 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2305 if (mem == NULL)
2306 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2307
2308 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
2309 mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
2310 mem->map = NULL;
2311 mem->map_size = 0;
2312 mem->ahw = NULL;
2313
2314 uint64_t bo_flags = 0;
2315
2316 assert(mem->type->heapIndex < pdevice->memory.heap_count);
2317 if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
2318 bo_flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2319
2320 const struct wsi_memory_allocate_info *wsi_info =
2321 vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
2322 if (wsi_info && wsi_info->implicit_sync) {
2323 /* We need to set the WRITE flag on window system buffers so that GEM
2324 * will know we're writing to them and synchronize uses on other rings
2325 * (eg if the display server uses the blitter ring).
2326 */
2327 bo_flags |= EXEC_OBJECT_WRITE;
2328 } else if (pdevice->has_exec_async) {
2329 bo_flags |= EXEC_OBJECT_ASYNC;
2330 }
2331
2332 if (pdevice->use_softpin)
2333 bo_flags |= EXEC_OBJECT_PINNED;
2334
2335 const VkExportMemoryAllocateInfo *export_info =
2336 vk_find_struct_const(pAllocateInfo->pNext, EXPORT_MEMORY_ALLOCATE_INFO);
2337
2338 /* Check if we need to support Android HW buffer export. If so,
2339 * create AHardwareBuffer and import memory from it.
2340 */
2341 bool android_export = false;
2342 if (export_info && export_info->handleTypes &
2343 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
2344 android_export = true;
2345
2346 /* Android memory import. */
2347 const struct VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info =
2348 vk_find_struct_const(pAllocateInfo->pNext,
2349 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID);
2350
2351 if (ahw_import_info) {
2352 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
2353 if (result != VK_SUCCESS)
2354 goto fail;
2355
2356 goto success;
2357 } else if (android_export) {
2358 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
2359 if (result != VK_SUCCESS)
2360 goto fail;
2361
2362 const struct VkImportAndroidHardwareBufferInfoANDROID import_info = {
2363 .buffer = mem->ahw,
2364 };
2365 result = anv_import_ahw_memory(_device, mem, &import_info);
2366 if (result != VK_SUCCESS)
2367 goto fail;
2368
2369 goto success;
2370 }
2371
2372 const VkImportMemoryFdInfoKHR *fd_info =
2373 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
2374
2375 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2376 * ignored.
2377 */
2378 if (fd_info && fd_info->handleType) {
2379 /* At the moment, we support only the below handle types. */
2380 assert(fd_info->handleType ==
2381 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2382 fd_info->handleType ==
2383 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2384
2385 result = anv_bo_cache_import(device, &device->bo_cache, fd_info->fd,
2386 bo_flags | ANV_BO_EXTERNAL, &mem->bo);
2387 if (result != VK_SUCCESS)
2388 goto fail;
2389
2390 VkDeviceSize aligned_alloc_size =
2391 align_u64(pAllocateInfo->allocationSize, 4096);
2392
2393 /* For security purposes, we reject importing the bo if it's smaller
2394 * than the requested allocation size. This prevents a malicious client
2395 * from passing a buffer to a trusted client, lying about the size, and
2396 * telling the trusted client to try and texture from an image that goes
2397 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2398 * in the trusted client. The trusted client can protect itself against
2399 * this sort of attack but only if it can trust the buffer size.
2400 */
2401 if (mem->bo->size < aligned_alloc_size) {
2402 result = vk_errorf(device->instance, device,
2403 VK_ERROR_INVALID_EXTERNAL_HANDLE,
2404 "aligned allocationSize too large for "
2405 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2406 "%"PRIu64"B > %"PRIu64"B",
2407 aligned_alloc_size, mem->bo->size);
2408 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2409 goto fail;
2410 }
2411
2412 /* From the Vulkan spec:
2413 *
2414 * "Importing memory from a file descriptor transfers ownership of
2415 * the file descriptor from the application to the Vulkan
2416 * implementation. The application must not perform any operations on
2417 * the file descriptor after a successful import."
2418 *
2419 * If the import fails, we leave the file descriptor open.
2420 */
2421 close(fd_info->fd);
2422 goto success;
2423 }
2424
2425 /* Regular allocate (not importing memory). */
2426
2427 if (export_info && export_info->handleTypes)
2428 bo_flags |= ANV_BO_EXTERNAL;
2429
2430 result = anv_bo_cache_alloc(device, &device->bo_cache,
2431 pAllocateInfo->allocationSize, bo_flags,
2432 &mem->bo);
2433 if (result != VK_SUCCESS)
2434 goto fail;
2435
2436 const VkMemoryDedicatedAllocateInfo *dedicated_info =
2437 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_DEDICATED_ALLOCATE_INFO);
2438 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
2439 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
2440
2441 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2442 * the BO. In this case, we have a dedicated allocation.
2443 */
2444 if (image->needs_set_tiling) {
2445 const uint32_t i915_tiling =
2446 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
2447 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
2448 image->planes[0].surface.isl.row_pitch_B,
2449 i915_tiling);
2450 if (ret) {
2451 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2452 return vk_errorf(device->instance, NULL,
2453 VK_ERROR_OUT_OF_DEVICE_MEMORY,
2454 "failed to set BO tiling: %m");
2455 }
2456 }
2457 }
2458
2459 success:
2460 *pMem = anv_device_memory_to_handle(mem);
2461
2462 return VK_SUCCESS;
2463
2464 fail:
2465 vk_free2(&device->alloc, pAllocator, mem);
2466
2467 return result;
2468 }
2469
2470 VkResult anv_GetMemoryFdKHR(
2471 VkDevice device_h,
2472 const VkMemoryGetFdInfoKHR* pGetFdInfo,
2473 int* pFd)
2474 {
2475 ANV_FROM_HANDLE(anv_device, dev, device_h);
2476 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
2477
2478 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
2479
2480 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2481 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2482
2483 return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
2484 }
2485
2486 VkResult anv_GetMemoryFdPropertiesKHR(
2487 VkDevice _device,
2488 VkExternalMemoryHandleTypeFlagBits handleType,
2489 int fd,
2490 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
2491 {
2492 ANV_FROM_HANDLE(anv_device, device, _device);
2493 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2494
2495 switch (handleType) {
2496 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
2497 /* dma-buf can be imported as any memory type */
2498 pMemoryFdProperties->memoryTypeBits =
2499 (1 << pdevice->memory.type_count) - 1;
2500 return VK_SUCCESS;
2501
2502 default:
2503 /* The valid usage section for this function says:
2504 *
2505 * "handleType must not be one of the handle types defined as
2506 * opaque."
2507 *
2508 * So opaque handle types fall into the default "unsupported" case.
2509 */
2510 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
2511 }
2512 }
2513
2514 void anv_FreeMemory(
2515 VkDevice _device,
2516 VkDeviceMemory _mem,
2517 const VkAllocationCallbacks* pAllocator)
2518 {
2519 ANV_FROM_HANDLE(anv_device, device, _device);
2520 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
2521
2522 if (mem == NULL)
2523 return;
2524
2525 if (mem->map)
2526 anv_UnmapMemory(_device, _mem);
2527
2528 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2529
2530 #ifdef ANDROID
2531 if (mem->ahw)
2532 AHardwareBuffer_release(mem->ahw);
2533 #endif
2534
2535 vk_free2(&device->alloc, pAllocator, mem);
2536 }
2537
2538 VkResult anv_MapMemory(
2539 VkDevice _device,
2540 VkDeviceMemory _memory,
2541 VkDeviceSize offset,
2542 VkDeviceSize size,
2543 VkMemoryMapFlags flags,
2544 void** ppData)
2545 {
2546 ANV_FROM_HANDLE(anv_device, device, _device);
2547 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
2548
2549 if (mem == NULL) {
2550 *ppData = NULL;
2551 return VK_SUCCESS;
2552 }
2553
2554 if (size == VK_WHOLE_SIZE)
2555 size = mem->bo->size - offset;
2556
2557 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2558 *
2559 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2560 * assert(size != 0);
2561 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2562 * equal to the size of the memory minus offset
2563 */
2564 assert(size > 0);
2565 assert(offset + size <= mem->bo->size);
2566
2567 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2568 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2569 * at a time is valid. We could just mmap up front and return an offset
2570 * pointer here, but that may exhaust virtual memory on 32 bit
2571 * userspace. */
2572
2573 uint32_t gem_flags = 0;
2574
2575 if (!device->info.has_llc &&
2576 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
2577 gem_flags |= I915_MMAP_WC;
2578
2579 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2580 uint64_t map_offset = offset & ~4095ull;
2581 assert(offset >= map_offset);
2582 uint64_t map_size = (offset + size) - map_offset;
2583
2584 /* Let's map whole pages */
2585 map_size = align_u64(map_size, 4096);
2586
2587 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
2588 map_offset, map_size, gem_flags);
2589 if (map == MAP_FAILED)
2590 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
2591
2592 mem->map = map;
2593 mem->map_size = map_size;
2594
2595 *ppData = mem->map + (offset - map_offset);
2596
2597 return VK_SUCCESS;
2598 }
2599
2600 void anv_UnmapMemory(
2601 VkDevice _device,
2602 VkDeviceMemory _memory)
2603 {
2604 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
2605
2606 if (mem == NULL)
2607 return;
2608
2609 anv_gem_munmap(mem->map, mem->map_size);
2610
2611 mem->map = NULL;
2612 mem->map_size = 0;
2613 }
2614
2615 static void
2616 clflush_mapped_ranges(struct anv_device *device,
2617 uint32_t count,
2618 const VkMappedMemoryRange *ranges)
2619 {
2620 for (uint32_t i = 0; i < count; i++) {
2621 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
2622 if (ranges[i].offset >= mem->map_size)
2623 continue;
2624
2625 gen_clflush_range(mem->map + ranges[i].offset,
2626 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
2627 }
2628 }
2629
2630 VkResult anv_FlushMappedMemoryRanges(
2631 VkDevice _device,
2632 uint32_t memoryRangeCount,
2633 const VkMappedMemoryRange* pMemoryRanges)
2634 {
2635 ANV_FROM_HANDLE(anv_device, device, _device);
2636
2637 if (device->info.has_llc)
2638 return VK_SUCCESS;
2639
2640 /* Make sure the writes we're flushing have landed. */
2641 __builtin_ia32_mfence();
2642
2643 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
2644
2645 return VK_SUCCESS;
2646 }
2647
2648 VkResult anv_InvalidateMappedMemoryRanges(
2649 VkDevice _device,
2650 uint32_t memoryRangeCount,
2651 const VkMappedMemoryRange* pMemoryRanges)
2652 {
2653 ANV_FROM_HANDLE(anv_device, device, _device);
2654
2655 if (device->info.has_llc)
2656 return VK_SUCCESS;
2657
2658 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
2659
2660 /* Make sure no reads get moved up above the invalidate. */
2661 __builtin_ia32_mfence();
2662
2663 return VK_SUCCESS;
2664 }
2665
2666 void anv_GetBufferMemoryRequirements(
2667 VkDevice _device,
2668 VkBuffer _buffer,
2669 VkMemoryRequirements* pMemoryRequirements)
2670 {
2671 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2672 ANV_FROM_HANDLE(anv_device, device, _device);
2673 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2674
2675 /* The Vulkan spec (git aaed022) says:
2676 *
2677 * memoryTypeBits is a bitfield and contains one bit set for every
2678 * supported memory type for the resource. The bit `1<<i` is set if and
2679 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2680 * structure for the physical device is supported.
2681 */
2682 uint32_t memory_types = 0;
2683 for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
2684 uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
2685 if ((valid_usage & buffer->usage) == buffer->usage)
2686 memory_types |= (1u << i);
2687 }
2688
2689 /* Base alignment requirement of a cache line */
2690 uint32_t alignment = 16;
2691
2692 /* We need an alignment of 32 for pushing UBOs */
2693 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
2694 alignment = MAX2(alignment, 32);
2695
2696 pMemoryRequirements->size = buffer->size;
2697 pMemoryRequirements->alignment = alignment;
2698
2699 /* Storage and Uniform buffers should have their size aligned to
2700 * 32-bits to avoid boundary checks when last DWord is not complete.
2701 * This would ensure that not internal padding would be needed for
2702 * 16-bit types.
2703 */
2704 if (device->robust_buffer_access &&
2705 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
2706 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
2707 pMemoryRequirements->size = align_u64(buffer->size, 4);
2708
2709 pMemoryRequirements->memoryTypeBits = memory_types;
2710 }
2711
2712 void anv_GetBufferMemoryRequirements2(
2713 VkDevice _device,
2714 const VkBufferMemoryRequirementsInfo2* pInfo,
2715 VkMemoryRequirements2* pMemoryRequirements)
2716 {
2717 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
2718 &pMemoryRequirements->memoryRequirements);
2719
2720 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
2721 switch (ext->sType) {
2722 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
2723 VkMemoryDedicatedRequirements *requirements = (void *)ext;
2724 requirements->prefersDedicatedAllocation = VK_FALSE;
2725 requirements->requiresDedicatedAllocation = VK_FALSE;
2726 break;
2727 }
2728
2729 default:
2730 anv_debug_ignored_stype(ext->sType);
2731 break;
2732 }
2733 }
2734 }
2735
2736 void anv_GetImageMemoryRequirements(
2737 VkDevice _device,
2738 VkImage _image,
2739 VkMemoryRequirements* pMemoryRequirements)
2740 {
2741 ANV_FROM_HANDLE(anv_image, image, _image);
2742 ANV_FROM_HANDLE(anv_device, device, _device);
2743 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2744
2745 /* The Vulkan spec (git aaed022) says:
2746 *
2747 * memoryTypeBits is a bitfield and contains one bit set for every
2748 * supported memory type for the resource. The bit `1<<i` is set if and
2749 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2750 * structure for the physical device is supported.
2751 *
2752 * All types are currently supported for images.
2753 */
2754 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
2755
2756 /* We must have image allocated or imported at this point. According to the
2757 * specification, external images must have been bound to memory before
2758 * calling GetImageMemoryRequirements.
2759 */
2760 assert(image->size > 0);
2761
2762 pMemoryRequirements->size = image->size;
2763 pMemoryRequirements->alignment = image->alignment;
2764 pMemoryRequirements->memoryTypeBits = memory_types;
2765 }
2766
2767 void anv_GetImageMemoryRequirements2(
2768 VkDevice _device,
2769 const VkImageMemoryRequirementsInfo2* pInfo,
2770 VkMemoryRequirements2* pMemoryRequirements)
2771 {
2772 ANV_FROM_HANDLE(anv_device, device, _device);
2773 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
2774
2775 anv_GetImageMemoryRequirements(_device, pInfo->image,
2776 &pMemoryRequirements->memoryRequirements);
2777
2778 vk_foreach_struct_const(ext, pInfo->pNext) {
2779 switch (ext->sType) {
2780 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
2781 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2782 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
2783 (const VkImagePlaneMemoryRequirementsInfo *) ext;
2784 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
2785 plane_reqs->planeAspect);
2786
2787 assert(image->planes[plane].offset == 0);
2788
2789 /* The Vulkan spec (git aaed022) says:
2790 *
2791 * memoryTypeBits is a bitfield and contains one bit set for every
2792 * supported memory type for the resource. The bit `1<<i` is set
2793 * if and only if the memory type `i` in the
2794 * VkPhysicalDeviceMemoryProperties structure for the physical
2795 * device is supported.
2796 *
2797 * All types are currently supported for images.
2798 */
2799 pMemoryRequirements->memoryRequirements.memoryTypeBits =
2800 (1ull << pdevice->memory.type_count) - 1;
2801
2802 /* We must have image allocated or imported at this point. According to the
2803 * specification, external images must have been bound to memory before
2804 * calling GetImageMemoryRequirements.
2805 */
2806 assert(image->planes[plane].size > 0);
2807
2808 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
2809 pMemoryRequirements->memoryRequirements.alignment =
2810 image->planes[plane].alignment;
2811 break;
2812 }
2813
2814 default:
2815 anv_debug_ignored_stype(ext->sType);
2816 break;
2817 }
2818 }
2819
2820 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
2821 switch (ext->sType) {
2822 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
2823 VkMemoryDedicatedRequirements *requirements = (void *)ext;
2824 if (image->needs_set_tiling || image->external_format) {
2825 /* If we need to set the tiling for external consumers, we need a
2826 * dedicated allocation.
2827 *
2828 * See also anv_AllocateMemory.
2829 */
2830 requirements->prefersDedicatedAllocation = VK_TRUE;
2831 requirements->requiresDedicatedAllocation = VK_TRUE;
2832 } else {
2833 requirements->prefersDedicatedAllocation = VK_FALSE;
2834 requirements->requiresDedicatedAllocation = VK_FALSE;
2835 }
2836 break;
2837 }
2838
2839 default:
2840 anv_debug_ignored_stype(ext->sType);
2841 break;
2842 }
2843 }
2844 }
2845
2846 void anv_GetImageSparseMemoryRequirements(
2847 VkDevice device,
2848 VkImage image,
2849 uint32_t* pSparseMemoryRequirementCount,
2850 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
2851 {
2852 *pSparseMemoryRequirementCount = 0;
2853 }
2854
2855 void anv_GetImageSparseMemoryRequirements2(
2856 VkDevice device,
2857 const VkImageSparseMemoryRequirementsInfo2* pInfo,
2858 uint32_t* pSparseMemoryRequirementCount,
2859 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
2860 {
2861 *pSparseMemoryRequirementCount = 0;
2862 }
2863
2864 void anv_GetDeviceMemoryCommitment(
2865 VkDevice device,
2866 VkDeviceMemory memory,
2867 VkDeviceSize* pCommittedMemoryInBytes)
2868 {
2869 *pCommittedMemoryInBytes = 0;
2870 }
2871
2872 static void
2873 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
2874 {
2875 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
2876 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
2877
2878 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
2879
2880 if (mem) {
2881 assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
2882 buffer->address = (struct anv_address) {
2883 .bo = mem->bo,
2884 .offset = pBindInfo->memoryOffset,
2885 };
2886 } else {
2887 buffer->address = ANV_NULL_ADDRESS;
2888 }
2889 }
2890
2891 VkResult anv_BindBufferMemory(
2892 VkDevice device,
2893 VkBuffer buffer,
2894 VkDeviceMemory memory,
2895 VkDeviceSize memoryOffset)
2896 {
2897 anv_bind_buffer_memory(
2898 &(VkBindBufferMemoryInfo) {
2899 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
2900 .buffer = buffer,
2901 .memory = memory,
2902 .memoryOffset = memoryOffset,
2903 });
2904
2905 return VK_SUCCESS;
2906 }
2907
2908 VkResult anv_BindBufferMemory2(
2909 VkDevice device,
2910 uint32_t bindInfoCount,
2911 const VkBindBufferMemoryInfo* pBindInfos)
2912 {
2913 for (uint32_t i = 0; i < bindInfoCount; i++)
2914 anv_bind_buffer_memory(&pBindInfos[i]);
2915
2916 return VK_SUCCESS;
2917 }
2918
2919 VkResult anv_QueueBindSparse(
2920 VkQueue _queue,
2921 uint32_t bindInfoCount,
2922 const VkBindSparseInfo* pBindInfo,
2923 VkFence fence)
2924 {
2925 ANV_FROM_HANDLE(anv_queue, queue, _queue);
2926 if (anv_device_is_lost(queue->device))
2927 return VK_ERROR_DEVICE_LOST;
2928
2929 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2930 }
2931
2932 // Event functions
2933
2934 VkResult anv_CreateEvent(
2935 VkDevice _device,
2936 const VkEventCreateInfo* pCreateInfo,
2937 const VkAllocationCallbacks* pAllocator,
2938 VkEvent* pEvent)
2939 {
2940 ANV_FROM_HANDLE(anv_device, device, _device);
2941 struct anv_state state;
2942 struct anv_event *event;
2943
2944 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
2945
2946 state = anv_state_pool_alloc(&device->dynamic_state_pool,
2947 sizeof(*event), 8);
2948 event = state.map;
2949 event->state = state;
2950 event->semaphore = VK_EVENT_RESET;
2951
2952 if (!device->info.has_llc) {
2953 /* Make sure the writes we're flushing have landed. */
2954 __builtin_ia32_mfence();
2955 __builtin_ia32_clflush(event);
2956 }
2957
2958 *pEvent = anv_event_to_handle(event);
2959
2960 return VK_SUCCESS;
2961 }
2962
2963 void anv_DestroyEvent(
2964 VkDevice _device,
2965 VkEvent _event,
2966 const VkAllocationCallbacks* pAllocator)
2967 {
2968 ANV_FROM_HANDLE(anv_device, device, _device);
2969 ANV_FROM_HANDLE(anv_event, event, _event);
2970
2971 if (!event)
2972 return;
2973
2974 anv_state_pool_free(&device->dynamic_state_pool, event->state);
2975 }
2976
2977 VkResult anv_GetEventStatus(
2978 VkDevice _device,
2979 VkEvent _event)
2980 {
2981 ANV_FROM_HANDLE(anv_device, device, _device);
2982 ANV_FROM_HANDLE(anv_event, event, _event);
2983
2984 if (anv_device_is_lost(device))
2985 return VK_ERROR_DEVICE_LOST;
2986
2987 if (!device->info.has_llc) {
2988 /* Invalidate read cache before reading event written by GPU. */
2989 __builtin_ia32_clflush(event);
2990 __builtin_ia32_mfence();
2991
2992 }
2993
2994 return event->semaphore;
2995 }
2996
2997 VkResult anv_SetEvent(
2998 VkDevice _device,
2999 VkEvent _event)
3000 {
3001 ANV_FROM_HANDLE(anv_device, device, _device);
3002 ANV_FROM_HANDLE(anv_event, event, _event);
3003
3004 event->semaphore = VK_EVENT_SET;
3005
3006 if (!device->info.has_llc) {
3007 /* Make sure the writes we're flushing have landed. */
3008 __builtin_ia32_mfence();
3009 __builtin_ia32_clflush(event);
3010 }
3011
3012 return VK_SUCCESS;
3013 }
3014
3015 VkResult anv_ResetEvent(
3016 VkDevice _device,
3017 VkEvent _event)
3018 {
3019 ANV_FROM_HANDLE(anv_device, device, _device);
3020 ANV_FROM_HANDLE(anv_event, event, _event);
3021
3022 event->semaphore = VK_EVENT_RESET;
3023
3024 if (!device->info.has_llc) {
3025 /* Make sure the writes we're flushing have landed. */
3026 __builtin_ia32_mfence();
3027 __builtin_ia32_clflush(event);
3028 }
3029
3030 return VK_SUCCESS;
3031 }
3032
3033 // Buffer functions
3034
3035 VkResult anv_CreateBuffer(
3036 VkDevice _device,
3037 const VkBufferCreateInfo* pCreateInfo,
3038 const VkAllocationCallbacks* pAllocator,
3039 VkBuffer* pBuffer)
3040 {
3041 ANV_FROM_HANDLE(anv_device, device, _device);
3042 struct anv_buffer *buffer;
3043
3044 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
3045
3046 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
3047 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3048 if (buffer == NULL)
3049 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3050
3051 buffer->size = pCreateInfo->size;
3052 buffer->usage = pCreateInfo->usage;
3053 buffer->address = ANV_NULL_ADDRESS;
3054
3055 *pBuffer = anv_buffer_to_handle(buffer);
3056
3057 return VK_SUCCESS;
3058 }
3059
3060 void anv_DestroyBuffer(
3061 VkDevice _device,
3062 VkBuffer _buffer,
3063 const VkAllocationCallbacks* pAllocator)
3064 {
3065 ANV_FROM_HANDLE(anv_device, device, _device);
3066 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3067
3068 if (!buffer)
3069 return;
3070
3071 vk_free2(&device->alloc, pAllocator, buffer);
3072 }
3073
3074 void
3075 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
3076 enum isl_format format,
3077 struct anv_address address,
3078 uint32_t range, uint32_t stride)
3079 {
3080 isl_buffer_fill_state(&device->isl_dev, state.map,
3081 .address = anv_address_physical(address),
3082 .mocs = device->default_mocs,
3083 .size_B = range,
3084 .format = format,
3085 .stride_B = stride);
3086 }
3087
3088 void anv_DestroySampler(
3089 VkDevice _device,
3090 VkSampler _sampler,
3091 const VkAllocationCallbacks* pAllocator)
3092 {
3093 ANV_FROM_HANDLE(anv_device, device, _device);
3094 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
3095
3096 if (!sampler)
3097 return;
3098
3099 vk_free2(&device->alloc, pAllocator, sampler);
3100 }
3101
3102 VkResult anv_CreateFramebuffer(
3103 VkDevice _device,
3104 const VkFramebufferCreateInfo* pCreateInfo,
3105 const VkAllocationCallbacks* pAllocator,
3106 VkFramebuffer* pFramebuffer)
3107 {
3108 ANV_FROM_HANDLE(anv_device, device, _device);
3109 struct anv_framebuffer *framebuffer;
3110
3111 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
3112
3113 size_t size = sizeof(*framebuffer) +
3114 sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
3115 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
3116 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3117 if (framebuffer == NULL)
3118 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3119
3120 framebuffer->attachment_count = pCreateInfo->attachmentCount;
3121 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
3122 VkImageView _iview = pCreateInfo->pAttachments[i];
3123 framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
3124 }
3125
3126 framebuffer->width = pCreateInfo->width;
3127 framebuffer->height = pCreateInfo->height;
3128 framebuffer->layers = pCreateInfo->layers;
3129
3130 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
3131
3132 return VK_SUCCESS;
3133 }
3134
3135 void anv_DestroyFramebuffer(
3136 VkDevice _device,
3137 VkFramebuffer _fb,
3138 const VkAllocationCallbacks* pAllocator)
3139 {
3140 ANV_FROM_HANDLE(anv_device, device, _device);
3141 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
3142
3143 if (!fb)
3144 return;
3145
3146 vk_free2(&device->alloc, pAllocator, fb);
3147 }
3148
3149 static const VkTimeDomainEXT anv_time_domains[] = {
3150 VK_TIME_DOMAIN_DEVICE_EXT,
3151 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
3152 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
3153 };
3154
3155 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3156 VkPhysicalDevice physicalDevice,
3157 uint32_t *pTimeDomainCount,
3158 VkTimeDomainEXT *pTimeDomains)
3159 {
3160 int d;
3161 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
3162
3163 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
3164 vk_outarray_append(&out, i) {
3165 *i = anv_time_domains[d];
3166 }
3167 }
3168
3169 return vk_outarray_status(&out);
3170 }
3171
3172 static uint64_t
3173 anv_clock_gettime(clockid_t clock_id)
3174 {
3175 struct timespec current;
3176 int ret;
3177
3178 ret = clock_gettime(clock_id, &current);
3179 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
3180 ret = clock_gettime(CLOCK_MONOTONIC, &current);
3181 if (ret < 0)
3182 return 0;
3183
3184 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
3185 }
3186
3187 #define TIMESTAMP 0x2358
3188
3189 VkResult anv_GetCalibratedTimestampsEXT(
3190 VkDevice _device,
3191 uint32_t timestampCount,
3192 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
3193 uint64_t *pTimestamps,
3194 uint64_t *pMaxDeviation)
3195 {
3196 ANV_FROM_HANDLE(anv_device, device, _device);
3197 uint64_t timestamp_frequency = device->info.timestamp_frequency;
3198 int ret;
3199 int d;
3200 uint64_t begin, end;
3201 uint64_t max_clock_period = 0;
3202
3203 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3204
3205 for (d = 0; d < timestampCount; d++) {
3206 switch (pTimestampInfos[d].timeDomain) {
3207 case VK_TIME_DOMAIN_DEVICE_EXT:
3208 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
3209 &pTimestamps[d]);
3210
3211 if (ret != 0) {
3212 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
3213 "register: %m");
3214 }
3215 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
3216 max_clock_period = MAX2(max_clock_period, device_period);
3217 break;
3218 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
3219 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
3220 max_clock_period = MAX2(max_clock_period, 1);
3221 break;
3222
3223 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
3224 pTimestamps[d] = begin;
3225 break;
3226 default:
3227 pTimestamps[d] = 0;
3228 break;
3229 }
3230 }
3231
3232 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3233
3234 /*
3235 * The maximum deviation is the sum of the interval over which we
3236 * perform the sampling and the maximum period of any sampled
3237 * clock. That's because the maximum skew between any two sampled
3238 * clock edges is when the sampled clock with the largest period is
3239 * sampled at the end of that period but right at the beginning of the
3240 * sampling interval and some other clock is sampled right at the
3241 * begining of its sampling period and right at the end of the
3242 * sampling interval. Let's assume the GPU has the longest clock
3243 * period and that the application is sampling GPU and monotonic:
3244 *
3245 * s e
3246 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3247 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3248 *
3249 * g
3250 * 0 1 2 3
3251 * GPU -----_____-----_____-----_____-----_____
3252 *
3253 * m
3254 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3255 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3256 *
3257 * Interval <----------------->
3258 * Deviation <-------------------------->
3259 *
3260 * s = read(raw) 2
3261 * g = read(GPU) 1
3262 * m = read(monotonic) 2
3263 * e = read(raw) b
3264 *
3265 * We round the sample interval up by one tick to cover sampling error
3266 * in the interval clock
3267 */
3268
3269 uint64_t sample_interval = end - begin + 1;
3270
3271 *pMaxDeviation = sample_interval + max_clock_period;
3272
3273 return VK_SUCCESS;
3274 }
3275
3276 /* vk_icd.h does not declare this function, so we declare it here to
3277 * suppress Wmissing-prototypes.
3278 */
3279 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3280 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
3281
3282 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3283 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
3284 {
3285 /* For the full details on loader interface versioning, see
3286 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3287 * What follows is a condensed summary, to help you navigate the large and
3288 * confusing official doc.
3289 *
3290 * - Loader interface v0 is incompatible with later versions. We don't
3291 * support it.
3292 *
3293 * - In loader interface v1:
3294 * - The first ICD entrypoint called by the loader is
3295 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3296 * entrypoint.
3297 * - The ICD must statically expose no other Vulkan symbol unless it is
3298 * linked with -Bsymbolic.
3299 * - Each dispatchable Vulkan handle created by the ICD must be
3300 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3301 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3302 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3303 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3304 * such loader-managed surfaces.
3305 *
3306 * - Loader interface v2 differs from v1 in:
3307 * - The first ICD entrypoint called by the loader is
3308 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3309 * statically expose this entrypoint.
3310 *
3311 * - Loader interface v3 differs from v2 in:
3312 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3313 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3314 * because the loader no longer does so.
3315 */
3316 *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
3317 return VK_SUCCESS;
3318 }