projects / mesa.git / blob
? search:


35411e43ff19dc15fac9d10e57e2615cf472fb9c
[mesa.git] / src / intel / vulkan / anv_device.c
1 /*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <assert.h>
25 #include <stdbool.h>
26 #include <string.h>
27 #include <sys/mman.h>
28 #include <sys/sysinfo.h>
29 #include <unistd.h>
30 #include <fcntl.h>
31 #include <xf86drm.h>
32 #include "drm-uapi/drm_fourcc.h"
33
34 #include "anv_private.h"
35 #include "util/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/disk_cache.h"
39 #include "util/mesa-sha1.h"
40 #include "util/os_file.h"
41 #include "util/u_atomic.h"
42 #include "util/u_string.h"
43 #include "git_sha1.h"
44 #include "vk_util.h"
45 #include "common/gen_defines.h"
46 #include "compiler/glsl_types.h"
47
48 #include "genxml/gen7_pack.h"
49
50 /* This is probably far to big but it reflects the max size used for messages
51 * in OpenGLs KHR_debug.
52 */
53 #define MAX_DEBUG_MESSAGE_LENGTH 4096
54
55 static void
56 compiler_debug_log(void *data, const char *fmt, ...)
57 {
58 char str[MAX_DEBUG_MESSAGE_LENGTH];
59 struct anv_device *device = (struct anv_device *)data;
60
61 if (list_empty(&device->instance->debug_report_callbacks.callbacks))
62 return;
63
64 va_list args;
65 va_start(args, fmt);
66 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
67 va_end(args);
68
69 vk_debug_report(&device->instance->debug_report_callbacks,
70 VK_DEBUG_REPORT_DEBUG_BIT_EXT,
71 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,
72 0, 0, 0, "anv", str);
73 }
74
75 static void
76 compiler_perf_log(void *data, const char *fmt, ...)
77 {
78 va_list args;
79 va_start(args, fmt);
80
81 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
82 intel_logd_v(fmt, args);
83
84 va_end(args);
85 }
86
87 static uint64_t
88 anv_compute_heap_size(int fd, uint64_t gtt_size)
89 {
90 /* Query the total ram from the system */
91 struct sysinfo info;
92 sysinfo(&info);
93
94 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
95
96 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
97 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
98 */
99 uint64_t available_ram;
100 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
101 available_ram = total_ram / 2;
102 else
103 available_ram = total_ram * 3 / 4;
104
105 /* We also want to leave some padding for things we allocate in the driver,
106 * so don't go over 3/4 of the GTT either.
107 */
108 uint64_t available_gtt = gtt_size * 3 / 4;
109
110 return MIN2(available_ram, available_gtt);
111 }
112
113 static VkResult
114 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
115 {
116 uint64_t gtt_size;
117 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
118 &gtt_size) == -1) {
119 /* If, for whatever reason, we can't actually get the GTT size from the
120 * kernel (too old?) fall back to the aperture size.
121 */
122 anv_perf_warn(NULL, NULL,
123 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
124
125 if (anv_gem_get_aperture(fd, &gtt_size) == -1) {
126 return vk_errorf(NULL, NULL, VK_ERROR_INITIALIZATION_FAILED,
127 "failed to get aperture size: %m");
128 }
129 }
130
131 device->supports_48bit_addresses = (device->info.gen >= 8) &&
132 gtt_size > (4ULL << 30 /* GiB */);
133
134 uint64_t heap_size = anv_compute_heap_size(fd, gtt_size);
135
136 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
137 /* When running with an overridden PCI ID, we may get a GTT size from
138 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
139 * address support can still fail. Just clamp the address space size to
140 * 2 GiB if we don't have 48-bit support.
141 */
142 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
143 "not support for 48-bit addresses",
144 __FILE__, __LINE__);
145 heap_size = 2ull << 30;
146 }
147
148 if (heap_size <= 3ull * (1ull << 30)) {
149 /* In this case, everything fits nicely into the 32-bit address space,
150 * so there's no need for supporting 48bit addresses on client-allocated
151 * memory objects.
152 */
153 device->memory.heap_count = 1;
154 device->memory.heaps[0] = (struct anv_memory_heap) {
155 .vma_start = LOW_HEAP_MIN_ADDRESS,
156 .vma_size = LOW_HEAP_SIZE,
157 .size = heap_size,
158 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
159 .supports_48bit_addresses = false,
160 };
161 } else {
162 /* Not everything will fit nicely into a 32-bit address space. In this
163 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
164 * larger 48-bit heap. If we're in this case, then we have a total heap
165 * size larger than 3GiB which most likely means they have 8 GiB of
166 * video memory and so carving off 1 GiB for the 32-bit heap should be
167 * reasonable.
168 */
169 const uint64_t heap_size_32bit = 1ull << 30;
170 const uint64_t heap_size_48bit = heap_size - heap_size_32bit;
171
172 assert(device->supports_48bit_addresses);
173
174 device->memory.heap_count = 2;
175 device->memory.heaps[0] = (struct anv_memory_heap) {
176 .vma_start = HIGH_HEAP_MIN_ADDRESS,
177 /* Leave the last 4GiB out of the high vma range, so that no state
178 * base address + size can overflow 48 bits. For more information see
179 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
180 */
181 .vma_size = gtt_size - (1ull << 32) - HIGH_HEAP_MIN_ADDRESS,
182 .size = heap_size_48bit,
183 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
184 .supports_48bit_addresses = true,
185 };
186 device->memory.heaps[1] = (struct anv_memory_heap) {
187 .vma_start = LOW_HEAP_MIN_ADDRESS,
188 .vma_size = LOW_HEAP_SIZE,
189 .size = heap_size_32bit,
190 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
191 .supports_48bit_addresses = false,
192 };
193 }
194
195 uint32_t type_count = 0;
196 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
197 uint32_t valid_buffer_usage = ~0;
198
199 /* There appears to be a hardware issue in the VF cache where it only
200 * considers the bottom 32 bits of memory addresses. If you happen to
201 * have two vertex buffers which get placed exactly 4 GiB apart and use
202 * them in back-to-back draw calls, you can get collisions. In order to
203 * solve this problem, we require vertex and index buffers be bound to
204 * memory allocated out of the 32-bit heap.
205 */
206 if (device->memory.heaps[heap].supports_48bit_addresses) {
207 valid_buffer_usage &= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
208 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
209 }
210
211 if (device->info.has_llc) {
212 /* Big core GPUs share LLC with the CPU and thus one memory type can be
213 * both cached and coherent at the same time.
214 */
215 device->memory.types[type_count++] = (struct anv_memory_type) {
216 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
217 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
218 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
219 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
220 .heapIndex = heap,
221 .valid_buffer_usage = valid_buffer_usage,
222 };
223 } else {
224 /* The spec requires that we expose a host-visible, coherent memory
225 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
226 * to give the application a choice between cached, but not coherent and
227 * coherent but uncached (WC though).
228 */
229 device->memory.types[type_count++] = (struct anv_memory_type) {
230 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
231 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
232 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
233 .heapIndex = heap,
234 .valid_buffer_usage = valid_buffer_usage,
235 };
236 device->memory.types[type_count++] = (struct anv_memory_type) {
237 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
238 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
239 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
240 .heapIndex = heap,
241 .valid_buffer_usage = valid_buffer_usage,
242 };
243 }
244 }
245 device->memory.type_count = type_count;
246
247 return VK_SUCCESS;
248 }
249
250 static VkResult
251 anv_physical_device_init_uuids(struct anv_physical_device *device)
252 {
253 const struct build_id_note *note =
254 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
255 if (!note) {
256 return vk_errorf(device->instance, device,
257 VK_ERROR_INITIALIZATION_FAILED,
258 "Failed to find build-id");
259 }
260
261 unsigned build_id_len = build_id_length(note);
262 if (build_id_len < 20) {
263 return vk_errorf(device->instance, device,
264 VK_ERROR_INITIALIZATION_FAILED,
265 "build-id too short. It needs to be a SHA");
266 }
267
268 memcpy(device->driver_build_sha1, build_id_data(note), 20);
269
270 struct mesa_sha1 sha1_ctx;
271 uint8_t sha1[20];
272 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
273
274 /* The pipeline cache UUID is used for determining when a pipeline cache is
275 * invalid. It needs both a driver build and the PCI ID of the device.
276 */
277 _mesa_sha1_init(&sha1_ctx);
278 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
279 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
280 sizeof(device->chipset_id));
281 _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless,
282 sizeof(device->always_use_bindless));
283 _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
284 sizeof(device->has_a64_buffer_access));
285 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images,
286 sizeof(device->has_bindless_images));
287 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers,
288 sizeof(device->has_bindless_samplers));
289 _mesa_sha1_final(&sha1_ctx, sha1);
290 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
291
292 /* The driver UUID is used for determining sharability of images and memory
293 * between two Vulkan instances in separate processes. People who want to
294 * share memory need to also check the device UUID (below) so all this
295 * needs to be is the build-id.
296 */
297 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
298
299 /* The device UUID uniquely identifies the given device within the machine.
300 * Since we never have more than one device, this doesn't need to be a real
301 * UUID. However, on the off-chance that someone tries to use this to
302 * cache pre-tiled images or something of the like, we use the PCI ID and
303 * some bits of ISL info to ensure that this is safe.
304 */
305 _mesa_sha1_init(&sha1_ctx);
306 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
307 sizeof(device->chipset_id));
308 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
309 sizeof(device->isl_dev.has_bit6_swizzling));
310 _mesa_sha1_final(&sha1_ctx, sha1);
311 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
312
313 return VK_SUCCESS;
314 }
315
316 static void
317 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
318 {
319 #ifdef ENABLE_SHADER_CACHE
320 char renderer[10];
321 MAYBE_UNUSED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
322 device->chipset_id);
323 assert(len == sizeof(renderer) - 2);
324
325 char timestamp[41];
326 _mesa_sha1_format(timestamp, device->driver_build_sha1);
327
328 const uint64_t driver_flags =
329 brw_get_compiler_config_value(device->compiler);
330 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
331 #else
332 device->disk_cache = NULL;
333 #endif
334 }
335
336 static void
337 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
338 {
339 #ifdef ENABLE_SHADER_CACHE
340 if (device->disk_cache)
341 disk_cache_destroy(device->disk_cache);
342 #else
343 assert(device->disk_cache == NULL);
344 #endif
345 }
346
347 static uint64_t
348 get_available_system_memory()
349 {
350 char *meminfo = os_read_file("/proc/meminfo");
351 if (!meminfo)
352 return 0;
353
354 char *str = strstr(meminfo, "MemAvailable:");
355 if (!str) {
356 free(meminfo);
357 return 0;
358 }
359
360 uint64_t kb_mem_available;
361 if (sscanf(str, "MemAvailable: %" PRIx64, &kb_mem_available) == 1) {
362 free(meminfo);
363 return kb_mem_available << 10;
364 }
365
366 free(meminfo);
367 return 0;
368 }
369
370 static VkResult
371 anv_physical_device_init(struct anv_physical_device *device,
372 struct anv_instance *instance,
373 drmDevicePtr drm_device)
374 {
375 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
376 const char *path = drm_device->nodes[DRM_NODE_RENDER];
377 VkResult result;
378 int fd;
379 int master_fd = -1;
380
381 brw_process_intel_debug_variable();
382
383 fd = open(path, O_RDWR | O_CLOEXEC);
384 if (fd < 0)
385 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
386
387 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
388 device->instance = instance;
389
390 assert(strlen(path) < ARRAY_SIZE(device->path));
391 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
392
393 device->no_hw = getenv("INTEL_NO_HW") != NULL;
394
395 const int pci_id_override = gen_get_pci_device_id_override();
396 if (pci_id_override < 0) {
397 device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
398 if (!device->chipset_id) {
399 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
400 goto fail;
401 }
402 } else {
403 device->chipset_id = pci_id_override;
404 device->no_hw = true;
405 }
406
407 device->pci_info.domain = drm_device->businfo.pci->domain;
408 device->pci_info.bus = drm_device->businfo.pci->bus;
409 device->pci_info.device = drm_device->businfo.pci->dev;
410 device->pci_info.function = drm_device->businfo.pci->func;
411
412 device->name = gen_get_device_name(device->chipset_id);
413 if (!gen_get_device_info(device->chipset_id, &device->info)) {
414 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
415 goto fail;
416 }
417
418 if (device->info.is_haswell) {
419 intel_logw("Haswell Vulkan support is incomplete");
420 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
421 intel_logw("Ivy Bridge Vulkan support is incomplete");
422 } else if (device->info.gen == 7 && device->info.is_baytrail) {
423 intel_logw("Bay Trail Vulkan support is incomplete");
424 } else if (device->info.gen >= 8 && device->info.gen <= 11) {
425 /* Gen8-11 fully supported */
426 } else {
427 result = vk_errorf(device->instance, device,
428 VK_ERROR_INCOMPATIBLE_DRIVER,
429 "Vulkan not yet supported on %s", device->name);
430 goto fail;
431 }
432
433 device->cmd_parser_version = -1;
434 if (device->info.gen == 7) {
435 device->cmd_parser_version =
436 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
437 if (device->cmd_parser_version == -1) {
438 result = vk_errorf(device->instance, device,
439 VK_ERROR_INITIALIZATION_FAILED,
440 "failed to get command parser version");
441 goto fail;
442 }
443 }
444
445 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
446 result = vk_errorf(device->instance, device,
447 VK_ERROR_INITIALIZATION_FAILED,
448 "kernel missing gem wait");
449 goto fail;
450 }
451
452 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
453 result = vk_errorf(device->instance, device,
454 VK_ERROR_INITIALIZATION_FAILED,
455 "kernel missing execbuf2");
456 goto fail;
457 }
458
459 if (!device->info.has_llc &&
460 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
461 result = vk_errorf(device->instance, device,
462 VK_ERROR_INITIALIZATION_FAILED,
463 "kernel missing wc mmap");
464 goto fail;
465 }
466
467 result = anv_physical_device_init_heaps(device, fd);
468 if (result != VK_SUCCESS)
469 goto fail;
470
471 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
472 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
473 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
474 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
475 device->has_syncobj_wait = device->has_syncobj &&
476 anv_gem_supports_syncobj_wait(fd);
477 device->has_context_priority = anv_gem_has_context_priority(fd);
478
479 device->use_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN)
480 && device->supports_48bit_addresses;
481
482 device->has_context_isolation =
483 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
484
485 device->always_use_bindless =
486 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
487
488 /* We first got the A64 messages on broadwell and we can only use them if
489 * we can pass addresses directly into the shader which requires softpin.
490 */
491 device->has_a64_buffer_access = device->info.gen >= 8 &&
492 device->use_softpin;
493
494 /* We first get bindless image access on Skylake and we can only really do
495 * it if we don't have any relocations so we need softpin.
496 */
497 device->has_bindless_images = device->info.gen >= 9 &&
498 device->use_softpin;
499
500 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
501 * because it's just a matter of setting the sampler address in the sample
502 * message header. However, we've not bothered to wire it up for vec4 so
503 * we leave it disabled on gen7.
504 */
505 device->has_bindless_samplers = device->info.gen >= 8;
506
507 device->has_mem_available = get_available_system_memory() != 0;
508
509 /* Starting with Gen10, the timestamp frequency of the command streamer may
510 * vary from one part to another. We can query the value from the kernel.
511 */
512 if (device->info.gen >= 10) {
513 int timestamp_frequency =
514 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
515
516 if (timestamp_frequency < 0)
517 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
518 else
519 device->info.timestamp_frequency = timestamp_frequency;
520 }
521
522 /* GENs prior to 8 do not support EU/Subslice info */
523 if (device->info.gen >= 8) {
524 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
525 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
526
527 /* Without this information, we cannot get the right Braswell
528 * brandstrings, and we have to use conservative numbers for GPGPU on
529 * many platforms, but otherwise, things will just work.
530 */
531 if (device->subslice_total < 1 || device->eu_total < 1) {
532 intel_logw("Kernel 4.1 required to properly query GPU properties");
533 }
534 } else if (device->info.gen == 7) {
535 device->subslice_total = 1 << (device->info.gt - 1);
536 }
537
538 if (device->info.is_cherryview &&
539 device->subslice_total > 0 && device->eu_total > 0) {
540 /* Logical CS threads = EUs per subslice * num threads per EU */
541 uint32_t max_cs_threads =
542 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
543
544 /* Fuse configurations may give more threads than expected, never less. */
545 if (max_cs_threads > device->info.max_cs_threads)
546 device->info.max_cs_threads = max_cs_threads;
547 }
548
549 device->compiler = brw_compiler_create(NULL, &device->info);
550 if (device->compiler == NULL) {
551 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
552 goto fail;
553 }
554 device->compiler->shader_debug_log = compiler_debug_log;
555 device->compiler->shader_perf_log = compiler_perf_log;
556 device->compiler->supports_pull_constants = false;
557 device->compiler->constant_buffer_0_is_relative =
558 device->info.gen < 8 || !device->has_context_isolation;
559 device->compiler->supports_shader_constants = true;
560
561 /* Broadwell PRM says:
562 *
563 * "Before Gen8, there was a historical configuration control field to
564 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
565 * different places: TILECTL[1:0], ARB_MODE[5:4], and
566 * DISP_ARB_CTL[14:13].
567 *
568 * For Gen8 and subsequent generations, the swizzle fields are all
569 * reserved, and the CPU's memory controller performs all address
570 * swizzling modifications."
571 */
572 bool swizzled =
573 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
574
575 isl_device_init(&device->isl_dev, &device->info, swizzled);
576
577 result = anv_physical_device_init_uuids(device);
578 if (result != VK_SUCCESS)
579 goto fail;
580
581 anv_physical_device_init_disk_cache(device);
582
583 if (instance->enabled_extensions.KHR_display) {
584 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
585 if (master_fd >= 0) {
586 /* prod the device with a GETPARAM call which will fail if
587 * we don't have permission to even render on this device
588 */
589 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
590 close(master_fd);
591 master_fd = -1;
592 }
593 }
594 }
595 device->master_fd = master_fd;
596
597 result = anv_init_wsi(device);
598 if (result != VK_SUCCESS) {
599 ralloc_free(device->compiler);
600 anv_physical_device_free_disk_cache(device);
601 goto fail;
602 }
603
604 anv_physical_device_get_supported_extensions(device,
605 &device->supported_extensions);
606
607
608 device->local_fd = fd;
609
610 return VK_SUCCESS;
611
612 fail:
613 close(fd);
614 if (master_fd != -1)
615 close(master_fd);
616 return result;
617 }
618
619 static void
620 anv_physical_device_finish(struct anv_physical_device *device)
621 {
622 anv_finish_wsi(device);
623 anv_physical_device_free_disk_cache(device);
624 ralloc_free(device->compiler);
625 close(device->local_fd);
626 if (device->master_fd >= 0)
627 close(device->master_fd);
628 }
629
630 static void *
631 default_alloc_func(void *pUserData, size_t size, size_t align,
632 VkSystemAllocationScope allocationScope)
633 {
634 return malloc(size);
635 }
636
637 static void *
638 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
639 size_t align, VkSystemAllocationScope allocationScope)
640 {
641 return realloc(pOriginal, size);
642 }
643
644 static void
645 default_free_func(void *pUserData, void *pMemory)
646 {
647 free(pMemory);
648 }
649
650 static const VkAllocationCallbacks default_alloc = {
651 .pUserData = NULL,
652 .pfnAllocation = default_alloc_func,
653 .pfnReallocation = default_realloc_func,
654 .pfnFree = default_free_func,
655 };
656
657 VkResult anv_EnumerateInstanceExtensionProperties(
658 const char* pLayerName,
659 uint32_t* pPropertyCount,
660 VkExtensionProperties* pProperties)
661 {
662 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
663
664 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
665 if (anv_instance_extensions_supported.extensions[i]) {
666 vk_outarray_append(&out, prop) {
667 *prop = anv_instance_extensions[i];
668 }
669 }
670 }
671
672 return vk_outarray_status(&out);
673 }
674
675 VkResult anv_CreateInstance(
676 const VkInstanceCreateInfo* pCreateInfo,
677 const VkAllocationCallbacks* pAllocator,
678 VkInstance* pInstance)
679 {
680 struct anv_instance *instance;
681 VkResult result;
682
683 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
684
685 struct anv_instance_extension_table enabled_extensions = {};
686 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
687 int idx;
688 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
689 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
690 anv_instance_extensions[idx].extensionName) == 0)
691 break;
692 }
693
694 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
695 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
696
697 if (!anv_instance_extensions_supported.extensions[idx])
698 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
699
700 enabled_extensions.extensions[idx] = true;
701 }
702
703 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
704 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
705 if (!instance)
706 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
707
708 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
709
710 if (pAllocator)
711 instance->alloc = *pAllocator;
712 else
713 instance->alloc = default_alloc;
714
715 instance->app_info = (struct anv_app_info) { .api_version = 0 };
716 if (pCreateInfo->pApplicationInfo) {
717 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
718
719 instance->app_info.app_name =
720 vk_strdup(&instance->alloc, app->pApplicationName,
721 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
722 instance->app_info.app_version = app->applicationVersion;
723
724 instance->app_info.engine_name =
725 vk_strdup(&instance->alloc, app->pEngineName,
726 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
727 instance->app_info.engine_version = app->engineVersion;
728
729 instance->app_info.api_version = app->apiVersion;
730 }
731
732 if (instance->app_info.api_version == 0)
733 instance->app_info.api_version = VK_API_VERSION_1_0;
734
735 instance->enabled_extensions = enabled_extensions;
736
737 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
738 /* Vulkan requires that entrypoints for extensions which have not been
739 * enabled must not be advertised.
740 */
741 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
742 &instance->enabled_extensions)) {
743 instance->dispatch.entrypoints[i] = NULL;
744 } else {
745 instance->dispatch.entrypoints[i] =
746 anv_instance_dispatch_table.entrypoints[i];
747 }
748 }
749
750 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
751 /* Vulkan requires that entrypoints for extensions which have not been
752 * enabled must not be advertised.
753 */
754 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
755 &instance->enabled_extensions, NULL)) {
756 instance->device_dispatch.entrypoints[i] = NULL;
757 } else {
758 instance->device_dispatch.entrypoints[i] =
759 anv_device_dispatch_table.entrypoints[i];
760 }
761 }
762
763 instance->physicalDeviceCount = -1;
764
765 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
766 if (result != VK_SUCCESS) {
767 vk_free2(&default_alloc, pAllocator, instance);
768 return vk_error(result);
769 }
770
771 instance->pipeline_cache_enabled =
772 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
773
774 _mesa_locale_init();
775 glsl_type_singleton_init_or_ref();
776
777 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
778
779 *pInstance = anv_instance_to_handle(instance);
780
781 return VK_SUCCESS;
782 }
783
784 void anv_DestroyInstance(
785 VkInstance _instance,
786 const VkAllocationCallbacks* pAllocator)
787 {
788 ANV_FROM_HANDLE(anv_instance, instance, _instance);
789
790 if (!instance)
791 return;
792
793 if (instance->physicalDeviceCount > 0) {
794 /* We support at most one physical device. */
795 assert(instance->physicalDeviceCount == 1);
796 anv_physical_device_finish(&instance->physicalDevice);
797 }
798
799 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
800 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
801
802 VG(VALGRIND_DESTROY_MEMPOOL(instance));
803
804 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
805
806 glsl_type_singleton_decref();
807 _mesa_locale_fini();
808
809 vk_free(&instance->alloc, instance);
810 }
811
812 static VkResult
813 anv_enumerate_devices(struct anv_instance *instance)
814 {
815 /* TODO: Check for more devices ? */
816 drmDevicePtr devices[8];
817 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
818 int max_devices;
819
820 instance->physicalDeviceCount = 0;
821
822 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
823 if (max_devices < 1)
824 return VK_ERROR_INCOMPATIBLE_DRIVER;
825
826 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
827 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
828 devices[i]->bustype == DRM_BUS_PCI &&
829 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
830
831 result = anv_physical_device_init(&instance->physicalDevice,
832 instance, devices[i]);
833 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
834 break;
835 }
836 }
837 drmFreeDevices(devices, max_devices);
838
839 if (result == VK_SUCCESS)
840 instance->physicalDeviceCount = 1;
841
842 return result;
843 }
844
845 static VkResult
846 anv_instance_ensure_physical_device(struct anv_instance *instance)
847 {
848 if (instance->physicalDeviceCount < 0) {
849 VkResult result = anv_enumerate_devices(instance);
850 if (result != VK_SUCCESS &&
851 result != VK_ERROR_INCOMPATIBLE_DRIVER)
852 return result;
853 }
854
855 return VK_SUCCESS;
856 }
857
858 VkResult anv_EnumeratePhysicalDevices(
859 VkInstance _instance,
860 uint32_t* pPhysicalDeviceCount,
861 VkPhysicalDevice* pPhysicalDevices)
862 {
863 ANV_FROM_HANDLE(anv_instance, instance, _instance);
864 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
865
866 VkResult result = anv_instance_ensure_physical_device(instance);
867 if (result != VK_SUCCESS)
868 return result;
869
870 if (instance->physicalDeviceCount == 0)
871 return VK_SUCCESS;
872
873 assert(instance->physicalDeviceCount == 1);
874 vk_outarray_append(&out, i) {
875 *i = anv_physical_device_to_handle(&instance->physicalDevice);
876 }
877
878 return vk_outarray_status(&out);
879 }
880
881 VkResult anv_EnumeratePhysicalDeviceGroups(
882 VkInstance _instance,
883 uint32_t* pPhysicalDeviceGroupCount,
884 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
885 {
886 ANV_FROM_HANDLE(anv_instance, instance, _instance);
887 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
888 pPhysicalDeviceGroupCount);
889
890 VkResult result = anv_instance_ensure_physical_device(instance);
891 if (result != VK_SUCCESS)
892 return result;
893
894 if (instance->physicalDeviceCount == 0)
895 return VK_SUCCESS;
896
897 assert(instance->physicalDeviceCount == 1);
898
899 vk_outarray_append(&out, p) {
900 p->physicalDeviceCount = 1;
901 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
902 p->physicalDevices[0] =
903 anv_physical_device_to_handle(&instance->physicalDevice);
904 p->subsetAllocation = false;
905
906 vk_foreach_struct(ext, p->pNext)
907 anv_debug_ignored_stype(ext->sType);
908 }
909
910 return vk_outarray_status(&out);
911 }
912
913 void anv_GetPhysicalDeviceFeatures(
914 VkPhysicalDevice physicalDevice,
915 VkPhysicalDeviceFeatures* pFeatures)
916 {
917 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
918
919 *pFeatures = (VkPhysicalDeviceFeatures) {
920 .robustBufferAccess = true,
921 .fullDrawIndexUint32 = true,
922 .imageCubeArray = true,
923 .independentBlend = true,
924 .geometryShader = true,
925 .tessellationShader = true,
926 .sampleRateShading = true,
927 .dualSrcBlend = true,
928 .logicOp = true,
929 .multiDrawIndirect = true,
930 .drawIndirectFirstInstance = true,
931 .depthClamp = true,
932 .depthBiasClamp = true,
933 .fillModeNonSolid = true,
934 .depthBounds = false,
935 .wideLines = true,
936 .largePoints = true,
937 .alphaToOne = true,
938 .multiViewport = true,
939 .samplerAnisotropy = true,
940 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
941 pdevice->info.is_baytrail,
942 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
943 .textureCompressionBC = true,
944 .occlusionQueryPrecise = true,
945 .pipelineStatisticsQuery = true,
946 .fragmentStoresAndAtomics = true,
947 .shaderTessellationAndGeometryPointSize = true,
948 .shaderImageGatherExtended = true,
949 .shaderStorageImageExtendedFormats = true,
950 .shaderStorageImageMultisample = false,
951 .shaderStorageImageReadWithoutFormat = false,
952 .shaderStorageImageWriteWithoutFormat = true,
953 .shaderUniformBufferArrayDynamicIndexing = true,
954 .shaderSampledImageArrayDynamicIndexing = true,
955 .shaderStorageBufferArrayDynamicIndexing = true,
956 .shaderStorageImageArrayDynamicIndexing = true,
957 .shaderClipDistance = true,
958 .shaderCullDistance = true,
959 .shaderFloat64 = pdevice->info.gen >= 8 &&
960 pdevice->info.has_64bit_types,
961 .shaderInt64 = pdevice->info.gen >= 8 &&
962 pdevice->info.has_64bit_types,
963 .shaderInt16 = pdevice->info.gen >= 8,
964 .shaderResourceMinLod = pdevice->info.gen >= 9,
965 .variableMultisampleRate = true,
966 .inheritedQueries = true,
967 };
968
969 /* We can't do image stores in vec4 shaders */
970 pFeatures->vertexPipelineStoresAndAtomics =
971 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
972 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
973
974 struct anv_app_info *app_info = &pdevice->instance->app_info;
975
976 /* The new DOOM and Wolfenstein games require depthBounds without
977 * checking for it. They seem to run fine without it so just claim it's
978 * there and accept the consequences.
979 */
980 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
981 pFeatures->depthBounds = true;
982 }
983
984 void anv_GetPhysicalDeviceFeatures2(
985 VkPhysicalDevice physicalDevice,
986 VkPhysicalDeviceFeatures2* pFeatures)
987 {
988 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
989 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
990
991 vk_foreach_struct(ext, pFeatures->pNext) {
992 switch (ext->sType) {
993 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
994 VkPhysicalDevice8BitStorageFeaturesKHR *features =
995 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
996 features->storageBuffer8BitAccess = pdevice->info.gen >= 8;
997 features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
998 features->storagePushConstant8 = pdevice->info.gen >= 8;
999 break;
1000 }
1001
1002 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
1003 VkPhysicalDevice16BitStorageFeatures *features =
1004 (VkPhysicalDevice16BitStorageFeatures *)ext;
1005 features->storageBuffer16BitAccess = pdevice->info.gen >= 8;
1006 features->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
1007 features->storagePushConstant16 = pdevice->info.gen >= 8;
1008 features->storageInputOutput16 = false;
1009 break;
1010 }
1011
1012 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
1013 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
1014 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1015 features->bufferDeviceAddressCaptureReplay = false;
1016 features->bufferDeviceAddressMultiDevice = false;
1017 break;
1018 }
1019
1020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1021 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1022 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1023 features->computeDerivativeGroupQuads = true;
1024 features->computeDerivativeGroupLinear = true;
1025 break;
1026 }
1027
1028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1029 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1030 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1031 features->conditionalRendering = pdevice->info.gen >= 8 ||
1032 pdevice->info.is_haswell;
1033 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1034 pdevice->info.is_haswell;
1035 break;
1036 }
1037
1038 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1039 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1040 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1041 features->depthClipEnable = true;
1042 break;
1043 }
1044
1045 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1046 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1047 features->shaderFloat16 = pdevice->info.gen >= 8;
1048 features->shaderInt8 = pdevice->info.gen >= 8;
1049 break;
1050 }
1051
1052 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1053 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1054 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1055 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1056 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1057 features->fragmentShaderShadingRateInterlock = false;
1058 break;
1059 }
1060
1061 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1062 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1063 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1064 features->hostQueryReset = true;
1065 break;
1066 }
1067
1068 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1069 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1070 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1071 features->shaderInputAttachmentArrayDynamicIndexing = false;
1072 features->shaderUniformTexelBufferArrayDynamicIndexing = true;
1073 features->shaderStorageTexelBufferArrayDynamicIndexing = true;
1074 features->shaderUniformBufferArrayNonUniformIndexing = false;
1075 features->shaderSampledImageArrayNonUniformIndexing = true;
1076 features->shaderStorageBufferArrayNonUniformIndexing = true;
1077 features->shaderStorageImageArrayNonUniformIndexing = true;
1078 features->shaderInputAttachmentArrayNonUniformIndexing = false;
1079 features->shaderUniformTexelBufferArrayNonUniformIndexing = true;
1080 features->shaderStorageTexelBufferArrayNonUniformIndexing = true;
1081 features->descriptorBindingUniformBufferUpdateAfterBind = false;
1082 features->descriptorBindingSampledImageUpdateAfterBind = true;
1083 features->descriptorBindingStorageImageUpdateAfterBind = true;
1084 features->descriptorBindingStorageBufferUpdateAfterBind = true;
1085 features->descriptorBindingUniformTexelBufferUpdateAfterBind = true;
1086 features->descriptorBindingStorageTexelBufferUpdateAfterBind = true;
1087 features->descriptorBindingUpdateUnusedWhilePending = true;
1088 features->descriptorBindingPartiallyBound = true;
1089 features->descriptorBindingVariableDescriptorCount = false;
1090 features->runtimeDescriptorArray = true;
1091 break;
1092 }
1093
1094 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1095 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1096 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1097 features->indexTypeUint8 = true;
1098 break;
1099 }
1100
1101 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1102 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1103 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1104 features->inlineUniformBlock = true;
1105 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1106 break;
1107 }
1108
1109 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1110 VkPhysicalDeviceMultiviewFeatures *features =
1111 (VkPhysicalDeviceMultiviewFeatures *)ext;
1112 features->multiview = true;
1113 features->multiviewGeometryShader = true;
1114 features->multiviewTessellationShader = true;
1115 break;
1116 }
1117
1118 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1119 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1120 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1121 features->imagelessFramebuffer = true;
1122 break;
1123 }
1124
1125 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1126 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1127 features->protectedMemory = false;
1128 break;
1129 }
1130
1131 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1132 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1133 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1134 features->samplerYcbcrConversion = true;
1135 break;
1136 }
1137
1138 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1139 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1140 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1141 features->scalarBlockLayout = true;
1142 break;
1143 }
1144
1145 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1146 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1147 features->shaderBufferInt64Atomics =
1148 pdevice->info.gen >= 9 && pdevice->use_softpin;
1149 features->shaderSharedInt64Atomics = VK_FALSE;
1150 break;
1151 }
1152
1153 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1154 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1155 features->shaderDemoteToHelperInvocation = true;
1156 break;
1157 }
1158
1159 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1160 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1161 features->shaderDrawParameters = true;
1162 break;
1163 }
1164
1165 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1166 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1167 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1168 features->texelBufferAlignment = true;
1169 break;
1170 }
1171
1172 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1173 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1174 features->variablePointersStorageBuffer = true;
1175 features->variablePointers = true;
1176 break;
1177 }
1178
1179 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1180 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1181 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1182 features->transformFeedback = true;
1183 features->geometryStreams = true;
1184 break;
1185 }
1186
1187 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1188 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1189 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1190 features->uniformBufferStandardLayout = true;
1191 break;
1192 }
1193
1194 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1195 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1196 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1197 features->vertexAttributeInstanceRateDivisor = true;
1198 features->vertexAttributeInstanceRateZeroDivisor = true;
1199 break;
1200 }
1201
1202 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1203 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1204 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1205 features->ycbcrImageArrays = true;
1206 break;
1207 }
1208
1209 default:
1210 anv_debug_ignored_stype(ext->sType);
1211 break;
1212 }
1213 }
1214 }
1215
1216 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1217
1218 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1219 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1220
1221 void anv_GetPhysicalDeviceProperties(
1222 VkPhysicalDevice physicalDevice,
1223 VkPhysicalDeviceProperties* pProperties)
1224 {
1225 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1226 const struct gen_device_info *devinfo = &pdevice->info;
1227
1228 /* See assertions made when programming the buffer surface state. */
1229 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1230 (1ul << 30) : (1ul << 27);
1231
1232 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1233 const uint32_t max_textures =
1234 pdevice->has_bindless_images ? UINT16_MAX : 128;
1235 const uint32_t max_samplers =
1236 pdevice->has_bindless_samplers ? UINT16_MAX :
1237 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1238 const uint32_t max_images =
1239 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1240
1241 /* The moment we have anything bindless, claim a high per-stage limit */
1242 const uint32_t max_per_stage =
1243 pdevice->has_a64_buffer_access ? UINT32_MAX :
1244 MAX_BINDING_TABLE_SIZE - MAX_RTS;
1245
1246 VkSampleCountFlags sample_counts =
1247 isl_device_get_sample_counts(&pdevice->isl_dev);
1248
1249
1250 VkPhysicalDeviceLimits limits = {
1251 .maxImageDimension1D = (1 << 14),
1252 .maxImageDimension2D = (1 << 14),
1253 .maxImageDimension3D = (1 << 11),
1254 .maxImageDimensionCube = (1 << 14),
1255 .maxImageArrayLayers = (1 << 11),
1256 .maxTexelBufferElements = 128 * 1024 * 1024,
1257 .maxUniformBufferRange = (1ul << 27),
1258 .maxStorageBufferRange = max_raw_buffer_sz,
1259 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1260 .maxMemoryAllocationCount = UINT32_MAX,
1261 .maxSamplerAllocationCount = 64 * 1024,
1262 .bufferImageGranularity = 64, /* A cache line */
1263 .sparseAddressSpaceSize = 0,
1264 .maxBoundDescriptorSets = MAX_SETS,
1265 .maxPerStageDescriptorSamplers = max_samplers,
1266 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1267 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1268 .maxPerStageDescriptorSampledImages = max_textures,
1269 .maxPerStageDescriptorStorageImages = max_images,
1270 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1271 .maxPerStageResources = max_per_stage,
1272 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1273 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1274 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1275 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1276 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1277 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1278 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1279 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1280 .maxVertexInputAttributes = MAX_VBS,
1281 .maxVertexInputBindings = MAX_VBS,
1282 .maxVertexInputAttributeOffset = 2047,
1283 .maxVertexInputBindingStride = 2048,
1284 .maxVertexOutputComponents = 128,
1285 .maxTessellationGenerationLevel = 64,
1286 .maxTessellationPatchSize = 32,
1287 .maxTessellationControlPerVertexInputComponents = 128,
1288 .maxTessellationControlPerVertexOutputComponents = 128,
1289 .maxTessellationControlPerPatchOutputComponents = 128,
1290 .maxTessellationControlTotalOutputComponents = 2048,
1291 .maxTessellationEvaluationInputComponents = 128,
1292 .maxTessellationEvaluationOutputComponents = 128,
1293 .maxGeometryShaderInvocations = 32,
1294 .maxGeometryInputComponents = 64,
1295 .maxGeometryOutputComponents = 128,
1296 .maxGeometryOutputVertices = 256,
1297 .maxGeometryTotalOutputComponents = 1024,
1298 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1299 .maxFragmentOutputAttachments = 8,
1300 .maxFragmentDualSrcAttachments = 1,
1301 .maxFragmentCombinedOutputResources = 8,
1302 .maxComputeSharedMemorySize = 64 * 1024,
1303 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1304 .maxComputeWorkGroupInvocations = 32 * devinfo->max_cs_threads,
1305 .maxComputeWorkGroupSize = {
1306 16 * devinfo->max_cs_threads,
1307 16 * devinfo->max_cs_threads,
1308 16 * devinfo->max_cs_threads,
1309 },
1310 .subPixelPrecisionBits = 8,
1311 .subTexelPrecisionBits = 8,
1312 .mipmapPrecisionBits = 8,
1313 .maxDrawIndexedIndexValue = UINT32_MAX,
1314 .maxDrawIndirectCount = UINT32_MAX,
1315 .maxSamplerLodBias = 16,
1316 .maxSamplerAnisotropy = 16,
1317 .maxViewports = MAX_VIEWPORTS,
1318 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1319 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1320 .viewportSubPixelBits = 13, /* We take a float? */
1321 .minMemoryMapAlignment = 4096, /* A page */
1322 /* The dataport requires texel alignment so we need to assume a worst
1323 * case of R32G32B32A32 which is 16 bytes.
1324 */
1325 .minTexelBufferOffsetAlignment = 16,
1326 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1327 .minUniformBufferOffsetAlignment = 32,
1328 .minStorageBufferOffsetAlignment = 4,
1329 .minTexelOffset = -8,
1330 .maxTexelOffset = 7,
1331 .minTexelGatherOffset = -32,
1332 .maxTexelGatherOffset = 31,
1333 .minInterpolationOffset = -0.5,
1334 .maxInterpolationOffset = 0.4375,
1335 .subPixelInterpolationOffsetBits = 4,
1336 .maxFramebufferWidth = (1 << 14),
1337 .maxFramebufferHeight = (1 << 14),
1338 .maxFramebufferLayers = (1 << 11),
1339 .framebufferColorSampleCounts = sample_counts,
1340 .framebufferDepthSampleCounts = sample_counts,
1341 .framebufferStencilSampleCounts = sample_counts,
1342 .framebufferNoAttachmentsSampleCounts = sample_counts,
1343 .maxColorAttachments = MAX_RTS,
1344 .sampledImageColorSampleCounts = sample_counts,
1345 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1346 .sampledImageDepthSampleCounts = sample_counts,
1347 .sampledImageStencilSampleCounts = sample_counts,
1348 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1349 .maxSampleMaskWords = 1,
1350 .timestampComputeAndGraphics = true,
1351 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1352 .maxClipDistances = 8,
1353 .maxCullDistances = 8,
1354 .maxCombinedClipAndCullDistances = 8,
1355 .discreteQueuePriorities = 2,
1356 .pointSizeRange = { 0.125, 255.875 },
1357 .lineWidthRange = { 0.0, 7.9921875 },
1358 .pointSizeGranularity = (1.0 / 8.0),
1359 .lineWidthGranularity = (1.0 / 128.0),
1360 .strictLines = false, /* FINISHME */
1361 .standardSampleLocations = true,
1362 .optimalBufferCopyOffsetAlignment = 128,
1363 .optimalBufferCopyRowPitchAlignment = 128,
1364 .nonCoherentAtomSize = 64,
1365 };
1366
1367 *pProperties = (VkPhysicalDeviceProperties) {
1368 .apiVersion = anv_physical_device_api_version(pdevice),
1369 .driverVersion = vk_get_driver_version(),
1370 .vendorID = 0x8086,
1371 .deviceID = pdevice->chipset_id,
1372 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1373 .limits = limits,
1374 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1375 };
1376
1377 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1378 "%s", pdevice->name);
1379 memcpy(pProperties->pipelineCacheUUID,
1380 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1381 }
1382
1383 void anv_GetPhysicalDeviceProperties2(
1384 VkPhysicalDevice physicalDevice,
1385 VkPhysicalDeviceProperties2* pProperties)
1386 {
1387 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1388
1389 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1390
1391 vk_foreach_struct(ext, pProperties->pNext) {
1392 switch (ext->sType) {
1393 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1394 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *props =
1395 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1396
1397 /* We support all of the depth resolve modes */
1398 props->supportedDepthResolveModes =
1399 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1400 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1401 VK_RESOLVE_MODE_MIN_BIT_KHR |
1402 VK_RESOLVE_MODE_MAX_BIT_KHR;
1403
1404 /* Average doesn't make sense for stencil so we don't support that */
1405 props->supportedStencilResolveModes =
1406 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1407 if (pdevice->info.gen >= 8) {
1408 /* The advanced stencil resolve modes currently require stencil
1409 * sampling be supported by the hardware.
1410 */
1411 props->supportedStencilResolveModes |=
1412 VK_RESOLVE_MODE_MIN_BIT_KHR |
1413 VK_RESOLVE_MODE_MAX_BIT_KHR;
1414 }
1415
1416 props->independentResolveNone = true;
1417 props->independentResolve = true;
1418 break;
1419 }
1420
1421 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1422 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *props =
1423 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1424
1425 /* It's a bit hard to exactly map our implementation to the limits
1426 * described here. The bindless surface handle in the extended
1427 * message descriptors is 20 bits and it's an index into the table of
1428 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1429 * address. Given that most things consume two surface states per
1430 * view (general/sampled for textures and write-only/read-write for
1431 * images), we claim 2^19 things.
1432 *
1433 * For SSBOs, we just use A64 messages so there is no real limit
1434 * there beyond the limit on the total size of a descriptor set.
1435 */
1436 const unsigned max_bindless_views = 1 << 19;
1437
1438 props->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1439 props->shaderUniformBufferArrayNonUniformIndexingNative = false;
1440 props->shaderSampledImageArrayNonUniformIndexingNative = false;
1441 props->shaderStorageBufferArrayNonUniformIndexingNative = true;
1442 props->shaderStorageImageArrayNonUniformIndexingNative = false;
1443 props->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1444 props->robustBufferAccessUpdateAfterBind = true;
1445 props->quadDivergentImplicitLod = false;
1446 props->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1447 props->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1448 props->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1449 props->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1450 props->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1451 props->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1452 props->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1453 props->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1454 props->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1455 props->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1456 props->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1457 props->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1458 props->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1459 props->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1460 props->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1461 break;
1462 }
1463
1464 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1465 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1466 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1467
1468 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1469 snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1470 "Intel open-source Mesa driver");
1471
1472 snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1473 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1474
1475 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1476 .major = 1,
1477 .minor = 1,
1478 .subminor = 2,
1479 .patch = 0,
1480 };
1481 break;
1482 }
1483
1484 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1485 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1486 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1487 /* Userptr needs page aligned memory. */
1488 props->minImportedHostPointerAlignment = 4096;
1489 break;
1490 }
1491
1492 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1493 VkPhysicalDeviceIDProperties *id_props =
1494 (VkPhysicalDeviceIDProperties *)ext;
1495 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1496 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1497 /* The LUID is for Windows. */
1498 id_props->deviceLUIDValid = false;
1499 break;
1500 }
1501
1502 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1503 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1504 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1505 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1506 props->maxPerStageDescriptorInlineUniformBlocks =
1507 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1508 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1509 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1510 props->maxDescriptorSetInlineUniformBlocks =
1511 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1512 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1513 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1514 break;
1515 }
1516
1517 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1518 VkPhysicalDeviceMaintenance3Properties *props =
1519 (VkPhysicalDeviceMaintenance3Properties *)ext;
1520 /* This value doesn't matter for us today as our per-stage
1521 * descriptors are the real limit.
1522 */
1523 props->maxPerSetDescriptors = 1024;
1524 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1525 break;
1526 }
1527
1528 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1529 VkPhysicalDeviceMultiviewProperties *properties =
1530 (VkPhysicalDeviceMultiviewProperties *)ext;
1531 properties->maxMultiviewViewCount = 16;
1532 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1533 break;
1534 }
1535
1536 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1537 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1538 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1539 properties->pciDomain = pdevice->pci_info.domain;
1540 properties->pciBus = pdevice->pci_info.bus;
1541 properties->pciDevice = pdevice->pci_info.device;
1542 properties->pciFunction = pdevice->pci_info.function;
1543 break;
1544 }
1545
1546 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1547 VkPhysicalDevicePointClippingProperties *properties =
1548 (VkPhysicalDevicePointClippingProperties *) ext;
1549 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1550 break;
1551 }
1552
1553 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1554 VkPhysicalDeviceProtectedMemoryProperties *props =
1555 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1556 props->protectedNoFault = false;
1557 break;
1558 }
1559
1560 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1561 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1562 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1563
1564 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1565 break;
1566 }
1567
1568 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1569 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1570 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1571 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1572 properties->filterMinmaxSingleComponentFormats = true;
1573 break;
1574 }
1575
1576 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1577 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1578
1579 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1580
1581 VkShaderStageFlags scalar_stages = 0;
1582 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1583 if (pdevice->compiler->scalar_stage[stage])
1584 scalar_stages |= mesa_to_vk_shader_stage(stage);
1585 }
1586 properties->supportedStages = scalar_stages;
1587
1588 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1589 VK_SUBGROUP_FEATURE_VOTE_BIT |
1590 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1591 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1592 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1593 VK_SUBGROUP_FEATURE_QUAD_BIT;
1594 if (pdevice->info.gen >= 8) {
1595 /* TODO: There's no technical reason why these can't be made to
1596 * work on gen7 but they don't at the moment so it's best to leave
1597 * the feature disabled than enabled and broken.
1598 */
1599 properties->supportedOperations |=
1600 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1601 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1602 }
1603 properties->quadOperationsInAllStages = pdevice->info.gen >= 8;
1604 break;
1605 }
1606
1607 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1608 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1609 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1610 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1611 props->minSubgroupSize = 8;
1612 props->maxSubgroupSize = 32;
1613 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1614 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1615 break;
1616 }
1617
1618 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1619 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1620 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1621
1622 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1623 * Base Address:
1624 *
1625 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1626 * specifies the base address of the first element of the surface,
1627 * computed in software by adding the surface base address to the
1628 * byte offset of the element in the buffer. The base address must
1629 * be aligned to element size."
1630 *
1631 * The typed dataport messages require that things be texel aligned.
1632 * Otherwise, we may just load/store the wrong data or, in the worst
1633 * case, there may be hangs.
1634 */
1635 props->storageTexelBufferOffsetAlignmentBytes = 16;
1636 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1637
1638 /* The sampler, however, is much more forgiving and it can handle
1639 * arbitrary byte alignment for linear and buffer surfaces. It's
1640 * hard to find a good PRM citation for this but years of empirical
1641 * experience demonstrate that this is true.
1642 */
1643 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1644 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1645 break;
1646 }
1647
1648 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1649 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1650 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1651
1652 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1653 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1654 props->maxTransformFeedbackBufferSize = (1ull << 32);
1655 props->maxTransformFeedbackStreamDataSize = 128 * 4;
1656 props->maxTransformFeedbackBufferDataSize = 128 * 4;
1657 props->maxTransformFeedbackBufferDataStride = 2048;
1658 props->transformFeedbackQueries = true;
1659 props->transformFeedbackStreamsLinesTriangles = false;
1660 props->transformFeedbackRasterizationStreamSelect = false;
1661 props->transformFeedbackDraw = true;
1662 break;
1663 }
1664
1665 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1666 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1667 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1668 /* We have to restrict this a bit for multiview */
1669 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1670 break;
1671 }
1672
1673 default:
1674 anv_debug_ignored_stype(ext->sType);
1675 break;
1676 }
1677 }
1678 }
1679
1680 /* We support exactly one queue family. */
1681 static const VkQueueFamilyProperties
1682 anv_queue_family_properties = {
1683 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1684 VK_QUEUE_COMPUTE_BIT |
1685 VK_QUEUE_TRANSFER_BIT,
1686 .queueCount = 1,
1687 .timestampValidBits = 36, /* XXX: Real value here */
1688 .minImageTransferGranularity = { 1, 1, 1 },
1689 };
1690
1691 void anv_GetPhysicalDeviceQueueFamilyProperties(
1692 VkPhysicalDevice physicalDevice,
1693 uint32_t* pCount,
1694 VkQueueFamilyProperties* pQueueFamilyProperties)
1695 {
1696 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1697
1698 vk_outarray_append(&out, p) {
1699 *p = anv_queue_family_properties;
1700 }
1701 }
1702
1703 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1704 VkPhysicalDevice physicalDevice,
1705 uint32_t* pQueueFamilyPropertyCount,
1706 VkQueueFamilyProperties2* pQueueFamilyProperties)
1707 {
1708
1709 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1710
1711 vk_outarray_append(&out, p) {
1712 p->queueFamilyProperties = anv_queue_family_properties;
1713
1714 vk_foreach_struct(s, p->pNext) {
1715 anv_debug_ignored_stype(s->sType);
1716 }
1717 }
1718 }
1719
1720 void anv_GetPhysicalDeviceMemoryProperties(
1721 VkPhysicalDevice physicalDevice,
1722 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1723 {
1724 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1725
1726 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1727 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1728 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1729 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1730 .heapIndex = physical_device->memory.types[i].heapIndex,
1731 };
1732 }
1733
1734 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1735 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1736 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1737 .size = physical_device->memory.heaps[i].size,
1738 .flags = physical_device->memory.heaps[i].flags,
1739 };
1740 }
1741 }
1742
1743 static void
1744 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
1745 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
1746 {
1747 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1748 uint64_t sys_available = get_available_system_memory();
1749 assert(sys_available > 0);
1750
1751 VkDeviceSize total_heaps_size = 0;
1752 for (size_t i = 0; i < device->memory.heap_count; i++)
1753 total_heaps_size += device->memory.heaps[i].size;
1754
1755 for (size_t i = 0; i < device->memory.heap_count; i++) {
1756 VkDeviceSize heap_size = device->memory.heaps[i].size;
1757 VkDeviceSize heap_used = device->memory.heaps[i].used;
1758 VkDeviceSize heap_budget;
1759
1760 double heap_proportion = (double) heap_size / total_heaps_size;
1761 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
1762
1763 /*
1764 * Let's not incite the app to starve the system: report at most 90% of
1765 * available system memory.
1766 */
1767 uint64_t heap_available = sys_available_prop * 9 / 10;
1768 heap_budget = MIN2(heap_size, heap_used + heap_available);
1769
1770 /*
1771 * Round down to the nearest MB
1772 */
1773 heap_budget &= ~((1ull << 20) - 1);
1774
1775 /*
1776 * The heapBudget value must be non-zero for array elements less than
1777 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1778 * value must be less than or equal to VkMemoryHeap::size for each heap.
1779 */
1780 assert(0 < heap_budget && heap_budget <= heap_size);
1781
1782 memoryBudget->heapUsage[i] = heap_used;
1783 memoryBudget->heapBudget[i] = heap_budget;
1784 }
1785
1786 /* The heapBudget and heapUsage values must be zero for array elements
1787 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1788 */
1789 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
1790 memoryBudget->heapBudget[i] = 0;
1791 memoryBudget->heapUsage[i] = 0;
1792 }
1793 }
1794
1795 void anv_GetPhysicalDeviceMemoryProperties2(
1796 VkPhysicalDevice physicalDevice,
1797 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1798 {
1799 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1800 &pMemoryProperties->memoryProperties);
1801
1802 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1803 switch (ext->sType) {
1804 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
1805 anv_get_memory_budget(physicalDevice, (void*)ext);
1806 break;
1807 default:
1808 anv_debug_ignored_stype(ext->sType);
1809 break;
1810 }
1811 }
1812 }
1813
1814 void
1815 anv_GetDeviceGroupPeerMemoryFeatures(
1816 VkDevice device,
1817 uint32_t heapIndex,
1818 uint32_t localDeviceIndex,
1819 uint32_t remoteDeviceIndex,
1820 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1821 {
1822 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1823 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1824 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1825 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1826 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1827 }
1828
1829 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1830 VkInstance _instance,
1831 const char* pName)
1832 {
1833 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1834
1835 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1836 * when we have to return valid function pointers, NULL, or it's left
1837 * undefined. See the table for exact details.
1838 */
1839 if (pName == NULL)
1840 return NULL;
1841
1842 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1843 if (strcmp(pName, "vk" #entrypoint) == 0) \
1844 return (PFN_vkVoidFunction)anv_##entrypoint
1845
1846 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1847 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1848 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1849 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1850
1851 #undef LOOKUP_ANV_ENTRYPOINT
1852
1853 if (instance == NULL)
1854 return NULL;
1855
1856 int idx = anv_get_instance_entrypoint_index(pName);
1857 if (idx >= 0)
1858 return instance->dispatch.entrypoints[idx];
1859
1860 idx = anv_get_device_entrypoint_index(pName);
1861 if (idx >= 0)
1862 return instance->device_dispatch.entrypoints[idx];
1863
1864 return NULL;
1865 }
1866
1867 /* With version 1+ of the loader interface the ICD should expose
1868 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1869 */
1870 PUBLIC
1871 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1872 VkInstance instance,
1873 const char* pName);
1874
1875 PUBLIC
1876 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1877 VkInstance instance,
1878 const char* pName)
1879 {
1880 return anv_GetInstanceProcAddr(instance, pName);
1881 }
1882
1883 PFN_vkVoidFunction anv_GetDeviceProcAddr(
1884 VkDevice _device,
1885 const char* pName)
1886 {
1887 ANV_FROM_HANDLE(anv_device, device, _device);
1888
1889 if (!device || !pName)
1890 return NULL;
1891
1892 int idx = anv_get_device_entrypoint_index(pName);
1893 if (idx < 0)
1894 return NULL;
1895
1896 return device->dispatch.entrypoints[idx];
1897 }
1898
1899 VkResult
1900 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
1901 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
1902 const VkAllocationCallbacks* pAllocator,
1903 VkDebugReportCallbackEXT* pCallback)
1904 {
1905 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1906 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
1907 pCreateInfo, pAllocator, &instance->alloc,
1908 pCallback);
1909 }
1910
1911 void
1912 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
1913 VkDebugReportCallbackEXT _callback,
1914 const VkAllocationCallbacks* pAllocator)
1915 {
1916 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1917 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
1918 _callback, pAllocator, &instance->alloc);
1919 }
1920
1921 void
1922 anv_DebugReportMessageEXT(VkInstance _instance,
1923 VkDebugReportFlagsEXT flags,
1924 VkDebugReportObjectTypeEXT objectType,
1925 uint64_t object,
1926 size_t location,
1927 int32_t messageCode,
1928 const char* pLayerPrefix,
1929 const char* pMessage)
1930 {
1931 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1932 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
1933 object, location, messageCode, pLayerPrefix, pMessage);
1934 }
1935
1936 static void
1937 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
1938 {
1939 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1940 queue->device = device;
1941 queue->flags = 0;
1942 }
1943
1944 static void
1945 anv_queue_finish(struct anv_queue *queue)
1946 {
1947 }
1948
1949 static struct anv_state
1950 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
1951 {
1952 struct anv_state state;
1953
1954 state = anv_state_pool_alloc(pool, size, align);
1955 memcpy(state.map, p, size);
1956
1957 return state;
1958 }
1959
1960 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
1961 * straightforward 32-bit float color in the first 64 bytes. Instead of using
1962 * a nice float/integer union like Gen8+, Haswell specifies the integer border
1963 * color as a separate entry /after/ the float color. The layout of this entry
1964 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
1965 *
1966 * Since we don't know the format/bpp, we can't make any of the border colors
1967 * containing '1' work for all formats, as it would be in the wrong place for
1968 * some of them. We opt to make 32-bit integers work as this seems like the
1969 * most common option. Fortunately, transparent black works regardless, as
1970 * all zeroes is the same in every bit-size.
1971 */
1972 struct hsw_border_color {
1973 float float32[4];
1974 uint32_t _pad0[12];
1975 uint32_t uint32[4];
1976 uint32_t _pad1[108];
1977 };
1978
1979 struct gen8_border_color {
1980 union {
1981 float float32[4];
1982 uint32_t uint32[4];
1983 };
1984 /* Pad out to 64 bytes */
1985 uint32_t _pad[12];
1986 };
1987
1988 static void
1989 anv_device_init_border_colors(struct anv_device *device)
1990 {
1991 if (device->info.is_haswell) {
1992 static const struct hsw_border_color border_colors[] = {
1993 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
1994 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
1995 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
1996 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
1997 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
1998 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
1999 };
2000
2001 device->border_colors =
2002 anv_state_pool_emit_data(&device->dynamic_state_pool,
2003 sizeof(border_colors), 512, border_colors);
2004 } else {
2005 static const struct gen8_border_color border_colors[] = {
2006 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2007 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2008 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2009 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2010 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2011 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2012 };
2013
2014 device->border_colors =
2015 anv_state_pool_emit_data(&device->dynamic_state_pool,
2016 sizeof(border_colors), 64, border_colors);
2017 }
2018 }
2019
2020 static void
2021 anv_device_init_trivial_batch(struct anv_device *device)
2022 {
2023 anv_bo_init_new(&device->trivial_batch_bo, device, 4096);
2024
2025 if (device->instance->physicalDevice.has_exec_async)
2026 device->trivial_batch_bo.flags |= EXEC_OBJECT_ASYNC;
2027
2028 if (device->instance->physicalDevice.use_softpin)
2029 device->trivial_batch_bo.flags |= EXEC_OBJECT_PINNED;
2030
2031 anv_vma_alloc(device, &device->trivial_batch_bo);
2032
2033 void *map = anv_gem_mmap(device, device->trivial_batch_bo.gem_handle,
2034 0, 4096, 0);
2035
2036 struct anv_batch batch = {
2037 .start = map,
2038 .next = map,
2039 .end = map + 4096,
2040 };
2041
2042 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2043 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2044
2045 if (!device->info.has_llc)
2046 gen_clflush_range(map, batch.next - map);
2047
2048 anv_gem_munmap(map, device->trivial_batch_bo.size);
2049 }
2050
2051 VkResult anv_EnumerateDeviceExtensionProperties(
2052 VkPhysicalDevice physicalDevice,
2053 const char* pLayerName,
2054 uint32_t* pPropertyCount,
2055 VkExtensionProperties* pProperties)
2056 {
2057 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2058 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2059
2060 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2061 if (device->supported_extensions.extensions[i]) {
2062 vk_outarray_append(&out, prop) {
2063 *prop = anv_device_extensions[i];
2064 }
2065 }
2066 }
2067
2068 return vk_outarray_status(&out);
2069 }
2070
2071 static void
2072 anv_device_init_dispatch(struct anv_device *device)
2073 {
2074 const struct anv_device_dispatch_table *genX_table;
2075 switch (device->info.gen) {
2076 case 11:
2077 genX_table = &gen11_device_dispatch_table;
2078 break;
2079 case 10:
2080 genX_table = &gen10_device_dispatch_table;
2081 break;
2082 case 9:
2083 genX_table = &gen9_device_dispatch_table;
2084 break;
2085 case 8:
2086 genX_table = &gen8_device_dispatch_table;
2087 break;
2088 case 7:
2089 if (device->info.is_haswell)
2090 genX_table = &gen75_device_dispatch_table;
2091 else
2092 genX_table = &gen7_device_dispatch_table;
2093 break;
2094 default:
2095 unreachable("unsupported gen\n");
2096 }
2097
2098 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2099 /* Vulkan requires that entrypoints for extensions which have not been
2100 * enabled must not be advertised.
2101 */
2102 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
2103 &device->instance->enabled_extensions,
2104 &device->enabled_extensions)) {
2105 device->dispatch.entrypoints[i] = NULL;
2106 } else if (genX_table->entrypoints[i]) {
2107 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2108 } else {
2109 device->dispatch.entrypoints[i] =
2110 anv_device_dispatch_table.entrypoints[i];
2111 }
2112 }
2113 }
2114
2115 static int
2116 vk_priority_to_gen(int priority)
2117 {
2118 switch (priority) {
2119 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2120 return GEN_CONTEXT_LOW_PRIORITY;
2121 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2122 return GEN_CONTEXT_MEDIUM_PRIORITY;
2123 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2124 return GEN_CONTEXT_HIGH_PRIORITY;
2125 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2126 return GEN_CONTEXT_REALTIME_PRIORITY;
2127 default:
2128 unreachable("Invalid priority");
2129 }
2130 }
2131
2132 static void
2133 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2134 {
2135 anv_bo_init_new(&device->hiz_clear_bo, device, 4096);
2136
2137 if (device->instance->physicalDevice.has_exec_async)
2138 device->hiz_clear_bo.flags |= EXEC_OBJECT_ASYNC;
2139
2140 if (device->instance->physicalDevice.use_softpin)
2141 device->hiz_clear_bo.flags |= EXEC_OBJECT_PINNED;
2142
2143 anv_vma_alloc(device, &device->hiz_clear_bo);
2144
2145 uint32_t *map = anv_gem_mmap(device, device->hiz_clear_bo.gem_handle,
2146 0, 4096, 0);
2147
2148 union isl_color_value hiz_clear = { .u32 = { 0, } };
2149 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2150
2151 memcpy(map, hiz_clear.u32, sizeof(hiz_clear.u32));
2152 anv_gem_munmap(map, device->hiz_clear_bo.size);
2153 }
2154
2155 static bool
2156 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2157 struct anv_block_pool *pool,
2158 uint64_t address)
2159 {
2160 for (uint32_t i = 0; i < pool->nbos; i++) {
2161 uint64_t bo_address = pool->bos[i].offset & (~0ull >> 16);
2162 uint32_t bo_size = pool->bos[i].size;
2163 if (address >= bo_address && address < (bo_address + bo_size)) {
2164 *ret = (struct gen_batch_decode_bo) {
2165 .addr = bo_address,
2166 .size = bo_size,
2167 .map = pool->bos[i].map,
2168 };
2169 return true;
2170 }
2171 }
2172 return false;
2173 }
2174
2175 /* Finding a buffer for batch decoding */
2176 static struct gen_batch_decode_bo
2177 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2178 {
2179 struct anv_device *device = v_batch;
2180 struct gen_batch_decode_bo ret_bo = {};
2181
2182 assert(ppgtt);
2183
2184 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2185 return ret_bo;
2186 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2187 return ret_bo;
2188 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2189 return ret_bo;
2190 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2191 return ret_bo;
2192
2193 if (!device->cmd_buffer_being_decoded)
2194 return (struct gen_batch_decode_bo) { };
2195
2196 struct anv_batch_bo **bo;
2197
2198 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2199 /* The decoder zeroes out the top 16 bits, so we need to as well */
2200 uint64_t bo_address = (*bo)->bo.offset & (~0ull >> 16);
2201
2202 if (address >= bo_address && address < bo_address + (*bo)->bo.size) {
2203 return (struct gen_batch_decode_bo) {
2204 .addr = bo_address,
2205 .size = (*bo)->bo.size,
2206 .map = (*bo)->bo.map,
2207 };
2208 }
2209 }
2210
2211 return (struct gen_batch_decode_bo) { };
2212 }
2213
2214 VkResult anv_CreateDevice(
2215 VkPhysicalDevice physicalDevice,
2216 const VkDeviceCreateInfo* pCreateInfo,
2217 const VkAllocationCallbacks* pAllocator,
2218 VkDevice* pDevice)
2219 {
2220 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2221 VkResult result;
2222 struct anv_device *device;
2223
2224 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2225
2226 struct anv_device_extension_table enabled_extensions = { };
2227 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2228 int idx;
2229 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2230 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2231 anv_device_extensions[idx].extensionName) == 0)
2232 break;
2233 }
2234
2235 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2236 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2237
2238 if (!physical_device->supported_extensions.extensions[idx])
2239 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2240
2241 enabled_extensions.extensions[idx] = true;
2242 }
2243
2244 /* Check enabled features */
2245 if (pCreateInfo->pEnabledFeatures) {
2246 VkPhysicalDeviceFeatures supported_features;
2247 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2248 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2249 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
2250 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2251 for (uint32_t i = 0; i < num_features; i++) {
2252 if (enabled_feature[i] && !supported_feature[i])
2253 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2254 }
2255 }
2256
2257 /* Check requested queues and fail if we are requested to create any
2258 * queues with flags we don't support.
2259 */
2260 assert(pCreateInfo->queueCreateInfoCount > 0);
2261 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2262 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2263 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2264 }
2265
2266 /* Check if client specified queue priority. */
2267 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2268 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2269 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2270
2271 VkQueueGlobalPriorityEXT priority =
2272 queue_priority ? queue_priority->globalPriority :
2273 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2274
2275 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2276 sizeof(*device), 8,
2277 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2278 if (!device)
2279 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2280
2281 if (INTEL_DEBUG & DEBUG_BATCH) {
2282 const unsigned decode_flags =
2283 GEN_BATCH_DECODE_FULL |
2284 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2285 GEN_BATCH_DECODE_OFFSETS |
2286 GEN_BATCH_DECODE_FLOATS;
2287
2288 gen_batch_decode_ctx_init(&device->decoder_ctx,
2289 &physical_device->info,
2290 stderr, decode_flags, NULL,
2291 decode_get_bo, NULL, device);
2292 }
2293
2294 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2295 device->instance = physical_device->instance;
2296 device->chipset_id = physical_device->chipset_id;
2297 device->no_hw = physical_device->no_hw;
2298 device->_lost = false;
2299
2300 if (pAllocator)
2301 device->alloc = *pAllocator;
2302 else
2303 device->alloc = physical_device->instance->alloc;
2304
2305 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2306 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2307 if (device->fd == -1) {
2308 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2309 goto fail_device;
2310 }
2311
2312 device->context_id = anv_gem_create_context(device);
2313 if (device->context_id == -1) {
2314 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2315 goto fail_fd;
2316 }
2317
2318 if (physical_device->use_softpin) {
2319 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2320 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2321 goto fail_fd;
2322 }
2323
2324 /* keep the page with address zero out of the allocator */
2325 struct anv_memory_heap *low_heap =
2326 &physical_device->memory.heaps[physical_device->memory.heap_count - 1];
2327 util_vma_heap_init(&device->vma_lo, low_heap->vma_start, low_heap->vma_size);
2328 device->vma_lo_available = low_heap->size;
2329
2330 struct anv_memory_heap *high_heap =
2331 &physical_device->memory.heaps[0];
2332 util_vma_heap_init(&device->vma_hi, high_heap->vma_start, high_heap->vma_size);
2333 device->vma_hi_available = physical_device->memory.heap_count == 1 ? 0 :
2334 high_heap->size;
2335 }
2336
2337 list_inithead(&device->memory_objects);
2338
2339 /* As per spec, the driver implementation may deny requests to acquire
2340 * a priority above the default priority (MEDIUM) if the caller does not
2341 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2342 * is returned.
2343 */
2344 if (physical_device->has_context_priority) {
2345 int err = anv_gem_set_context_param(device->fd, device->context_id,
2346 I915_CONTEXT_PARAM_PRIORITY,
2347 vk_priority_to_gen(priority));
2348 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2349 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2350 goto fail_fd;
2351 }
2352 }
2353
2354 device->info = physical_device->info;
2355 device->isl_dev = physical_device->isl_dev;
2356
2357 /* On Broadwell and later, we can use batch chaining to more efficiently
2358 * implement growing command buffers. Prior to Haswell, the kernel
2359 * command parser gets in the way and we have to fall back to growing
2360 * the batch.
2361 */
2362 device->can_chain_batches = device->info.gen >= 8;
2363
2364 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
2365 pCreateInfo->pEnabledFeatures->robustBufferAccess;
2366 device->enabled_extensions = enabled_extensions;
2367
2368 anv_device_init_dispatch(device);
2369
2370 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2371 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2372 goto fail_context_id;
2373 }
2374
2375 pthread_condattr_t condattr;
2376 if (pthread_condattr_init(&condattr) != 0) {
2377 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2378 goto fail_mutex;
2379 }
2380 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2381 pthread_condattr_destroy(&condattr);
2382 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2383 goto fail_mutex;
2384 }
2385 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2386 pthread_condattr_destroy(&condattr);
2387 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2388 goto fail_mutex;
2389 }
2390 pthread_condattr_destroy(&condattr);
2391
2392 uint64_t bo_flags =
2393 (physical_device->supports_48bit_addresses ? EXEC_OBJECT_SUPPORTS_48B_ADDRESS : 0) |
2394 (physical_device->has_exec_async ? EXEC_OBJECT_ASYNC : 0) |
2395 (physical_device->has_exec_capture ? EXEC_OBJECT_CAPTURE : 0) |
2396 (physical_device->use_softpin ? EXEC_OBJECT_PINNED : 0);
2397
2398 anv_bo_pool_init(&device->batch_bo_pool, device, bo_flags);
2399
2400 result = anv_bo_cache_init(&device->bo_cache);
2401 if (result != VK_SUCCESS)
2402 goto fail_batch_bo_pool;
2403
2404 if (!physical_device->use_softpin)
2405 bo_flags &= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2406
2407 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2408 DYNAMIC_STATE_POOL_MIN_ADDRESS,
2409 16384,
2410 bo_flags);
2411 if (result != VK_SUCCESS)
2412 goto fail_bo_cache;
2413
2414 result = anv_state_pool_init(&device->instruction_state_pool, device,
2415 INSTRUCTION_STATE_POOL_MIN_ADDRESS,
2416 16384,
2417 bo_flags);
2418 if (result != VK_SUCCESS)
2419 goto fail_dynamic_state_pool;
2420
2421 result = anv_state_pool_init(&device->surface_state_pool, device,
2422 SURFACE_STATE_POOL_MIN_ADDRESS,
2423 4096,
2424 bo_flags);
2425 if (result != VK_SUCCESS)
2426 goto fail_instruction_state_pool;
2427
2428 if (physical_device->use_softpin) {
2429 result = anv_state_pool_init(&device->binding_table_pool, device,
2430 BINDING_TABLE_POOL_MIN_ADDRESS,
2431 4096,
2432 bo_flags);
2433 if (result != VK_SUCCESS)
2434 goto fail_surface_state_pool;
2435 }
2436
2437 result = anv_bo_init_new(&device->workaround_bo, device, 4096);
2438 if (result != VK_SUCCESS)
2439 goto fail_binding_table_pool;
2440
2441 if (physical_device->use_softpin)
2442 device->workaround_bo.flags |= EXEC_OBJECT_PINNED;
2443
2444 if (!anv_vma_alloc(device, &device->workaround_bo))
2445 goto fail_workaround_bo;
2446
2447 anv_device_init_trivial_batch(device);
2448
2449 if (device->info.gen >= 10)
2450 anv_device_init_hiz_clear_value_bo(device);
2451
2452 anv_scratch_pool_init(device, &device->scratch_pool);
2453
2454 anv_queue_init(device, &device->queue);
2455
2456 switch (device->info.gen) {
2457 case 7:
2458 if (!device->info.is_haswell)
2459 result = gen7_init_device_state(device);
2460 else
2461 result = gen75_init_device_state(device);
2462 break;
2463 case 8:
2464 result = gen8_init_device_state(device);
2465 break;
2466 case 9:
2467 result = gen9_init_device_state(device);
2468 break;
2469 case 10:
2470 result = gen10_init_device_state(device);
2471 break;
2472 case 11:
2473 result = gen11_init_device_state(device);
2474 break;
2475 default:
2476 /* Shouldn't get here as we don't create physical devices for any other
2477 * gens. */
2478 unreachable("unhandled gen");
2479 }
2480 if (result != VK_SUCCESS)
2481 goto fail_workaround_bo;
2482
2483 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2484
2485 anv_device_init_blorp(device);
2486
2487 anv_device_init_border_colors(device);
2488
2489 *pDevice = anv_device_to_handle(device);
2490
2491 return VK_SUCCESS;
2492
2493 fail_workaround_bo:
2494 anv_queue_finish(&device->queue);
2495 anv_scratch_pool_finish(device, &device->scratch_pool);
2496 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
2497 anv_gem_close(device, device->workaround_bo.gem_handle);
2498 fail_binding_table_pool:
2499 if (physical_device->use_softpin)
2500 anv_state_pool_finish(&device->binding_table_pool);
2501 fail_surface_state_pool:
2502 anv_state_pool_finish(&device->surface_state_pool);
2503 fail_instruction_state_pool:
2504 anv_state_pool_finish(&device->instruction_state_pool);
2505 fail_dynamic_state_pool:
2506 anv_state_pool_finish(&device->dynamic_state_pool);
2507 fail_bo_cache:
2508 anv_bo_cache_finish(&device->bo_cache);
2509 fail_batch_bo_pool:
2510 anv_bo_pool_finish(&device->batch_bo_pool);
2511 pthread_cond_destroy(&device->queue_submit);
2512 fail_mutex:
2513 pthread_mutex_destroy(&device->mutex);
2514 fail_context_id:
2515 anv_gem_destroy_context(device, device->context_id);
2516 fail_fd:
2517 close(device->fd);
2518 fail_device:
2519 vk_free(&device->alloc, device);
2520
2521 return result;
2522 }
2523
2524 void anv_DestroyDevice(
2525 VkDevice _device,
2526 const VkAllocationCallbacks* pAllocator)
2527 {
2528 ANV_FROM_HANDLE(anv_device, device, _device);
2529 struct anv_physical_device *physical_device;
2530
2531 if (!device)
2532 return;
2533
2534 physical_device = &device->instance->physicalDevice;
2535
2536 anv_device_finish_blorp(device);
2537
2538 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2539
2540 anv_queue_finish(&device->queue);
2541
2542 #ifdef HAVE_VALGRIND
2543 /* We only need to free these to prevent valgrind errors. The backing
2544 * BO will go away in a couple of lines so we don't actually leak.
2545 */
2546 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2547 #endif
2548
2549 anv_scratch_pool_finish(device, &device->scratch_pool);
2550
2551 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
2552 anv_vma_free(device, &device->workaround_bo);
2553 anv_gem_close(device, device->workaround_bo.gem_handle);
2554
2555 anv_vma_free(device, &device->trivial_batch_bo);
2556 anv_gem_close(device, device->trivial_batch_bo.gem_handle);
2557 if (device->info.gen >= 10)
2558 anv_gem_close(device, device->hiz_clear_bo.gem_handle);
2559
2560 if (physical_device->use_softpin)
2561 anv_state_pool_finish(&device->binding_table_pool);
2562 anv_state_pool_finish(&device->surface_state_pool);
2563 anv_state_pool_finish(&device->instruction_state_pool);
2564 anv_state_pool_finish(&device->dynamic_state_pool);
2565
2566 anv_bo_cache_finish(&device->bo_cache);
2567
2568 anv_bo_pool_finish(&device->batch_bo_pool);
2569
2570 pthread_cond_destroy(&device->queue_submit);
2571 pthread_mutex_destroy(&device->mutex);
2572
2573 anv_gem_destroy_context(device, device->context_id);
2574
2575 if (INTEL_DEBUG & DEBUG_BATCH)
2576 gen_batch_decode_ctx_finish(&device->decoder_ctx);
2577
2578 close(device->fd);
2579
2580 vk_free(&device->alloc, device);
2581 }
2582
2583 VkResult anv_EnumerateInstanceLayerProperties(
2584 uint32_t* pPropertyCount,
2585 VkLayerProperties* pProperties)
2586 {
2587 if (pProperties == NULL) {
2588 *pPropertyCount = 0;
2589 return VK_SUCCESS;
2590 }
2591
2592 /* None supported at this time */
2593 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2594 }
2595
2596 VkResult anv_EnumerateDeviceLayerProperties(
2597 VkPhysicalDevice physicalDevice,
2598 uint32_t* pPropertyCount,
2599 VkLayerProperties* pProperties)
2600 {
2601 if (pProperties == NULL) {
2602 *pPropertyCount = 0;
2603 return VK_SUCCESS;
2604 }
2605
2606 /* None supported at this time */
2607 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2608 }
2609
2610 void anv_GetDeviceQueue(
2611 VkDevice _device,
2612 uint32_t queueNodeIndex,
2613 uint32_t queueIndex,
2614 VkQueue* pQueue)
2615 {
2616 ANV_FROM_HANDLE(anv_device, device, _device);
2617
2618 assert(queueIndex == 0);
2619
2620 *pQueue = anv_queue_to_handle(&device->queue);
2621 }
2622
2623 void anv_GetDeviceQueue2(
2624 VkDevice _device,
2625 const VkDeviceQueueInfo2* pQueueInfo,
2626 VkQueue* pQueue)
2627 {
2628 ANV_FROM_HANDLE(anv_device, device, _device);
2629
2630 assert(pQueueInfo->queueIndex == 0);
2631
2632 if (pQueueInfo->flags == device->queue.flags)
2633 *pQueue = anv_queue_to_handle(&device->queue);
2634 else
2635 *pQueue = NULL;
2636 }
2637
2638 VkResult
2639 _anv_device_set_lost(struct anv_device *device,
2640 const char *file, int line,
2641 const char *msg, ...)
2642 {
2643 VkResult err;
2644 va_list ap;
2645
2646 device->_lost = true;
2647
2648 va_start(ap, msg);
2649 err = __vk_errorv(device->instance, device,
2650 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2651 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2652 va_end(ap);
2653
2654 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2655 abort();
2656
2657 return err;
2658 }
2659
2660 VkResult
2661 anv_device_query_status(struct anv_device *device)
2662 {
2663 /* This isn't likely as most of the callers of this function already check
2664 * for it. However, it doesn't hurt to check and it potentially lets us
2665 * avoid an ioctl.
2666 */
2667 if (anv_device_is_lost(device))
2668 return VK_ERROR_DEVICE_LOST;
2669
2670 uint32_t active, pending;
2671 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2672 if (ret == -1) {
2673 /* We don't know the real error. */
2674 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2675 }
2676
2677 if (active) {
2678 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2679 } else if (pending) {
2680 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2681 }
2682
2683 return VK_SUCCESS;
2684 }
2685
2686 VkResult
2687 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2688 {
2689 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2690 * Other usages of the BO (such as on different hardware) will not be
2691 * flagged as "busy" by this ioctl. Use with care.
2692 */
2693 int ret = anv_gem_busy(device, bo->gem_handle);
2694 if (ret == 1) {
2695 return VK_NOT_READY;
2696 } else if (ret == -1) {
2697 /* We don't know the real error. */
2698 return anv_device_set_lost(device, "gem wait failed: %m");
2699 }
2700
2701 /* Query for device status after the busy call. If the BO we're checking
2702 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2703 * client because it clearly doesn't have valid data. Yes, this most
2704 * likely means an ioctl, but we just did an ioctl to query the busy status
2705 * so it's no great loss.
2706 */
2707 return anv_device_query_status(device);
2708 }
2709
2710 VkResult
2711 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2712 int64_t timeout)
2713 {
2714 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2715 if (ret == -1 && errno == ETIME) {
2716 return VK_TIMEOUT;
2717 } else if (ret == -1) {
2718 /* We don't know the real error. */
2719 return anv_device_set_lost(device, "gem wait failed: %m");
2720 }
2721
2722 /* Query for device status after the wait. If the BO we're waiting on got
2723 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2724 * because it clearly doesn't have valid data. Yes, this most likely means
2725 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2726 */
2727 return anv_device_query_status(device);
2728 }
2729
2730 VkResult anv_DeviceWaitIdle(
2731 VkDevice _device)
2732 {
2733 ANV_FROM_HANDLE(anv_device, device, _device);
2734 if (anv_device_is_lost(device))
2735 return VK_ERROR_DEVICE_LOST;
2736
2737 struct anv_batch batch;
2738
2739 uint32_t cmds[8];
2740 batch.start = batch.next = cmds;
2741 batch.end = (void *) cmds + sizeof(cmds);
2742
2743 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2744 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2745
2746 return anv_device_submit_simple_batch(device, &batch);
2747 }
2748
2749 bool
2750 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo)
2751 {
2752 if (!(bo->flags & EXEC_OBJECT_PINNED))
2753 return true;
2754
2755 pthread_mutex_lock(&device->vma_mutex);
2756
2757 bo->offset = 0;
2758
2759 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS &&
2760 device->vma_hi_available >= bo->size) {
2761 uint64_t addr = util_vma_heap_alloc(&device->vma_hi, bo->size, 4096);
2762 if (addr) {
2763 bo->offset = gen_canonical_address(addr);
2764 assert(addr == gen_48b_address(bo->offset));
2765 device->vma_hi_available -= bo->size;
2766 }
2767 }
2768
2769 if (bo->offset == 0 && device->vma_lo_available >= bo->size) {
2770 uint64_t addr = util_vma_heap_alloc(&device->vma_lo, bo->size, 4096);
2771 if (addr) {
2772 bo->offset = gen_canonical_address(addr);
2773 assert(addr == gen_48b_address(bo->offset));
2774 device->vma_lo_available -= bo->size;
2775 }
2776 }
2777
2778 pthread_mutex_unlock(&device->vma_mutex);
2779
2780 return bo->offset != 0;
2781 }
2782
2783 void
2784 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
2785 {
2786 if (!(bo->flags & EXEC_OBJECT_PINNED))
2787 return;
2788
2789 const uint64_t addr_48b = gen_48b_address(bo->offset);
2790
2791 pthread_mutex_lock(&device->vma_mutex);
2792
2793 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
2794 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
2795 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
2796 device->vma_lo_available += bo->size;
2797 } else {
2798 MAYBE_UNUSED const struct anv_physical_device *physical_device =
2799 &device->instance->physicalDevice;
2800 assert(addr_48b >= physical_device->memory.heaps[0].vma_start &&
2801 addr_48b < (physical_device->memory.heaps[0].vma_start +
2802 physical_device->memory.heaps[0].vma_size));
2803 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
2804 device->vma_hi_available += bo->size;
2805 }
2806
2807 pthread_mutex_unlock(&device->vma_mutex);
2808
2809 bo->offset = 0;
2810 }
2811
2812 VkResult
2813 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
2814 {
2815 uint32_t gem_handle = anv_gem_create(device, size);
2816 if (!gem_handle)
2817 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
2818
2819 anv_bo_init(bo, gem_handle, size);
2820
2821 return VK_SUCCESS;
2822 }
2823
2824 VkResult anv_AllocateMemory(
2825 VkDevice _device,
2826 const VkMemoryAllocateInfo* pAllocateInfo,
2827 const VkAllocationCallbacks* pAllocator,
2828 VkDeviceMemory* pMem)
2829 {
2830 ANV_FROM_HANDLE(anv_device, device, _device);
2831 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2832 struct anv_device_memory *mem;
2833 VkResult result = VK_SUCCESS;
2834
2835 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
2836
2837 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2838 assert(pAllocateInfo->allocationSize > 0);
2839
2840 if (pAllocateInfo->allocationSize > MAX_MEMORY_ALLOCATION_SIZE)
2841 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
2842
2843 /* FINISHME: Fail if allocation request exceeds heap size. */
2844
2845 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
2846 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2847 if (mem == NULL)
2848 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2849
2850 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
2851 mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
2852 mem->map = NULL;
2853 mem->map_size = 0;
2854 mem->ahw = NULL;
2855 mem->host_ptr = NULL;
2856
2857 uint64_t bo_flags = 0;
2858
2859 assert(mem->type->heapIndex < pdevice->memory.heap_count);
2860 if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
2861 bo_flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2862
2863 const struct wsi_memory_allocate_info *wsi_info =
2864 vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
2865 if (wsi_info && wsi_info->implicit_sync) {
2866 /* We need to set the WRITE flag on window system buffers so that GEM
2867 * will know we're writing to them and synchronize uses on other rings
2868 * (eg if the display server uses the blitter ring).
2869 */
2870 bo_flags |= EXEC_OBJECT_WRITE;
2871 } else if (pdevice->has_exec_async) {
2872 bo_flags |= EXEC_OBJECT_ASYNC;
2873 }
2874
2875 if (pdevice->use_softpin)
2876 bo_flags |= EXEC_OBJECT_PINNED;
2877
2878 const VkExportMemoryAllocateInfo *export_info =
2879 vk_find_struct_const(pAllocateInfo->pNext, EXPORT_MEMORY_ALLOCATE_INFO);
2880
2881 /* Check if we need to support Android HW buffer export. If so,
2882 * create AHardwareBuffer and import memory from it.
2883 */
2884 bool android_export = false;
2885 if (export_info && export_info->handleTypes &
2886 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
2887 android_export = true;
2888
2889 /* Android memory import. */
2890 const struct VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info =
2891 vk_find_struct_const(pAllocateInfo->pNext,
2892 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID);
2893
2894 if (ahw_import_info) {
2895 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
2896 if (result != VK_SUCCESS)
2897 goto fail;
2898
2899 goto success;
2900 } else if (android_export) {
2901 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
2902 if (result != VK_SUCCESS)
2903 goto fail;
2904
2905 const struct VkImportAndroidHardwareBufferInfoANDROID import_info = {
2906 .buffer = mem->ahw,
2907 };
2908 result = anv_import_ahw_memory(_device, mem, &import_info);
2909 if (result != VK_SUCCESS)
2910 goto fail;
2911
2912 goto success;
2913 }
2914
2915 const VkImportMemoryFdInfoKHR *fd_info =
2916 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
2917
2918 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2919 * ignored.
2920 */
2921 if (fd_info && fd_info->handleType) {
2922 /* At the moment, we support only the below handle types. */
2923 assert(fd_info->handleType ==
2924 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2925 fd_info->handleType ==
2926 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2927
2928 result = anv_bo_cache_import(device, &device->bo_cache, fd_info->fd,
2929 bo_flags | ANV_BO_EXTERNAL, &mem->bo);
2930 if (result != VK_SUCCESS)
2931 goto fail;
2932
2933 VkDeviceSize aligned_alloc_size =
2934 align_u64(pAllocateInfo->allocationSize, 4096);
2935
2936 /* For security purposes, we reject importing the bo if it's smaller
2937 * than the requested allocation size. This prevents a malicious client
2938 * from passing a buffer to a trusted client, lying about the size, and
2939 * telling the trusted client to try and texture from an image that goes
2940 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2941 * in the trusted client. The trusted client can protect itself against
2942 * this sort of attack but only if it can trust the buffer size.
2943 */
2944 if (mem->bo->size < aligned_alloc_size) {
2945 result = vk_errorf(device->instance, device,
2946 VK_ERROR_INVALID_EXTERNAL_HANDLE,
2947 "aligned allocationSize too large for "
2948 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2949 "%"PRIu64"B > %"PRIu64"B",
2950 aligned_alloc_size, mem->bo->size);
2951 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2952 goto fail;
2953 }
2954
2955 /* From the Vulkan spec:
2956 *
2957 * "Importing memory from a file descriptor transfers ownership of
2958 * the file descriptor from the application to the Vulkan
2959 * implementation. The application must not perform any operations on
2960 * the file descriptor after a successful import."
2961 *
2962 * If the import fails, we leave the file descriptor open.
2963 */
2964 close(fd_info->fd);
2965 goto success;
2966 }
2967
2968 const VkImportMemoryHostPointerInfoEXT *host_ptr_info =
2969 vk_find_struct_const(pAllocateInfo->pNext,
2970 IMPORT_MEMORY_HOST_POINTER_INFO_EXT);
2971 if (host_ptr_info && host_ptr_info->handleType) {
2972 if (host_ptr_info->handleType ==
2973 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
2974 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
2975 goto fail;
2976 }
2977
2978 assert(host_ptr_info->handleType ==
2979 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
2980
2981 result = anv_bo_cache_import_host_ptr(
2982 device, &device->bo_cache, host_ptr_info->pHostPointer,
2983 pAllocateInfo->allocationSize, bo_flags, &mem->bo);
2984
2985 if (result != VK_SUCCESS)
2986 goto fail;
2987
2988 mem->host_ptr = host_ptr_info->pHostPointer;
2989 goto success;
2990 }
2991
2992 /* Regular allocate (not importing memory). */
2993
2994 if (export_info && export_info->handleTypes)
2995 bo_flags |= ANV_BO_EXTERNAL;
2996
2997 result = anv_bo_cache_alloc(device, &device->bo_cache,
2998 pAllocateInfo->allocationSize, bo_flags,
2999 &mem->bo);
3000 if (result != VK_SUCCESS)
3001 goto fail;
3002
3003 const VkMemoryDedicatedAllocateInfo *dedicated_info =
3004 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_DEDICATED_ALLOCATE_INFO);
3005 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3006 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3007
3008 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3009 * the BO. In this case, we have a dedicated allocation.
3010 */
3011 if (image->needs_set_tiling) {
3012 const uint32_t i915_tiling =
3013 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3014 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3015 image->planes[0].surface.isl.row_pitch_B,
3016 i915_tiling);
3017 if (ret) {
3018 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
3019 return vk_errorf(device->instance, NULL,
3020 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3021 "failed to set BO tiling: %m");
3022 }
3023 }
3024 }
3025
3026 success:
3027 pthread_mutex_lock(&device->mutex);
3028 list_addtail(&mem->link, &device->memory_objects);
3029 pthread_mutex_unlock(&device->mutex);
3030
3031 *pMem = anv_device_memory_to_handle(mem);
3032
3033 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3034 mem->bo->size);
3035
3036 return VK_SUCCESS;
3037
3038 fail:
3039 vk_free2(&device->alloc, pAllocator, mem);
3040
3041 return result;
3042 }
3043
3044 VkResult anv_GetMemoryFdKHR(
3045 VkDevice device_h,
3046 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3047 int* pFd)
3048 {
3049 ANV_FROM_HANDLE(anv_device, dev, device_h);
3050 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3051
3052 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3053
3054 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3055 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3056
3057 return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
3058 }
3059
3060 VkResult anv_GetMemoryFdPropertiesKHR(
3061 VkDevice _device,
3062 VkExternalMemoryHandleTypeFlagBits handleType,
3063 int fd,
3064 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3065 {
3066 ANV_FROM_HANDLE(anv_device, device, _device);
3067 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3068
3069 switch (handleType) {
3070 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3071 /* dma-buf can be imported as any memory type */
3072 pMemoryFdProperties->memoryTypeBits =
3073 (1 << pdevice->memory.type_count) - 1;
3074 return VK_SUCCESS;
3075
3076 default:
3077 /* The valid usage section for this function says:
3078 *
3079 * "handleType must not be one of the handle types defined as
3080 * opaque."
3081 *
3082 * So opaque handle types fall into the default "unsupported" case.
3083 */
3084 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3085 }
3086 }
3087
3088 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3089 VkDevice _device,
3090 VkExternalMemoryHandleTypeFlagBits handleType,
3091 const void* pHostPointer,
3092 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3093 {
3094 ANV_FROM_HANDLE(anv_device, device, _device);
3095
3096 assert(pMemoryHostPointerProperties->sType ==
3097 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3098
3099 switch (handleType) {
3100 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT: {
3101 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3102
3103 /* Host memory can be imported as any memory type. */
3104 pMemoryHostPointerProperties->memoryTypeBits =
3105 (1ull << pdevice->memory.type_count) - 1;
3106
3107 return VK_SUCCESS;
3108 }
3109 default:
3110 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3111 }
3112 }
3113
3114 void anv_FreeMemory(
3115 VkDevice _device,
3116 VkDeviceMemory _mem,
3117 const VkAllocationCallbacks* pAllocator)
3118 {
3119 ANV_FROM_HANDLE(anv_device, device, _device);
3120 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3121 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3122
3123 if (mem == NULL)
3124 return;
3125
3126 pthread_mutex_lock(&device->mutex);
3127 list_del(&mem->link);
3128 pthread_mutex_unlock(&device->mutex);
3129
3130 if (mem->map)
3131 anv_UnmapMemory(_device, _mem);
3132
3133 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3134 -mem->bo->size);
3135
3136 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
3137
3138 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3139 if (mem->ahw)
3140 AHardwareBuffer_release(mem->ahw);
3141 #endif
3142
3143 vk_free2(&device->alloc, pAllocator, mem);
3144 }
3145
3146 VkResult anv_MapMemory(
3147 VkDevice _device,
3148 VkDeviceMemory _memory,
3149 VkDeviceSize offset,
3150 VkDeviceSize size,
3151 VkMemoryMapFlags flags,
3152 void** ppData)
3153 {
3154 ANV_FROM_HANDLE(anv_device, device, _device);
3155 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3156
3157 if (mem == NULL) {
3158 *ppData = NULL;
3159 return VK_SUCCESS;
3160 }
3161
3162 if (mem->host_ptr) {
3163 *ppData = mem->host_ptr + offset;
3164 return VK_SUCCESS;
3165 }
3166
3167 if (size == VK_WHOLE_SIZE)
3168 size = mem->bo->size - offset;
3169
3170 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3171 *
3172 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3173 * assert(size != 0);
3174 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3175 * equal to the size of the memory minus offset
3176 */
3177 assert(size > 0);
3178 assert(offset + size <= mem->bo->size);
3179
3180 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3181 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3182 * at a time is valid. We could just mmap up front and return an offset
3183 * pointer here, but that may exhaust virtual memory on 32 bit
3184 * userspace. */
3185
3186 uint32_t gem_flags = 0;
3187
3188 if (!device->info.has_llc &&
3189 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3190 gem_flags |= I915_MMAP_WC;
3191
3192 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3193 uint64_t map_offset = offset & ~4095ull;
3194 assert(offset >= map_offset);
3195 uint64_t map_size = (offset + size) - map_offset;
3196
3197 /* Let's map whole pages */
3198 map_size = align_u64(map_size, 4096);
3199
3200 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3201 map_offset, map_size, gem_flags);
3202 if (map == MAP_FAILED)
3203 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3204
3205 mem->map = map;
3206 mem->map_size = map_size;
3207
3208 *ppData = mem->map + (offset - map_offset);
3209
3210 return VK_SUCCESS;
3211 }
3212
3213 void anv_UnmapMemory(
3214 VkDevice _device,
3215 VkDeviceMemory _memory)
3216 {
3217 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3218
3219 if (mem == NULL || mem->host_ptr)
3220 return;
3221
3222 anv_gem_munmap(mem->map, mem->map_size);
3223
3224 mem->map = NULL;
3225 mem->map_size = 0;
3226 }
3227
3228 static void
3229 clflush_mapped_ranges(struct anv_device *device,
3230 uint32_t count,
3231 const VkMappedMemoryRange *ranges)
3232 {
3233 for (uint32_t i = 0; i < count; i++) {
3234 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3235 if (ranges[i].offset >= mem->map_size)
3236 continue;
3237
3238 gen_clflush_range(mem->map + ranges[i].offset,
3239 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3240 }
3241 }
3242
3243 VkResult anv_FlushMappedMemoryRanges(
3244 VkDevice _device,
3245 uint32_t memoryRangeCount,
3246 const VkMappedMemoryRange* pMemoryRanges)
3247 {
3248 ANV_FROM_HANDLE(anv_device, device, _device);
3249
3250 if (device->info.has_llc)
3251 return VK_SUCCESS;
3252
3253 /* Make sure the writes we're flushing have landed. */
3254 __builtin_ia32_mfence();
3255
3256 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3257
3258 return VK_SUCCESS;
3259 }
3260
3261 VkResult anv_InvalidateMappedMemoryRanges(
3262 VkDevice _device,
3263 uint32_t memoryRangeCount,
3264 const VkMappedMemoryRange* pMemoryRanges)
3265 {
3266 ANV_FROM_HANDLE(anv_device, device, _device);
3267
3268 if (device->info.has_llc)
3269 return VK_SUCCESS;
3270
3271 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3272
3273 /* Make sure no reads get moved up above the invalidate. */
3274 __builtin_ia32_mfence();
3275
3276 return VK_SUCCESS;
3277 }
3278
3279 void anv_GetBufferMemoryRequirements(
3280 VkDevice _device,
3281 VkBuffer _buffer,
3282 VkMemoryRequirements* pMemoryRequirements)
3283 {
3284 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3285 ANV_FROM_HANDLE(anv_device, device, _device);
3286 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3287
3288 /* The Vulkan spec (git aaed022) says:
3289 *
3290 * memoryTypeBits is a bitfield and contains one bit set for every
3291 * supported memory type for the resource. The bit `1<<i` is set if and
3292 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3293 * structure for the physical device is supported.
3294 */
3295 uint32_t memory_types = 0;
3296 for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
3297 uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
3298 if ((valid_usage & buffer->usage) == buffer->usage)
3299 memory_types |= (1u << i);
3300 }
3301
3302 /* Base alignment requirement of a cache line */
3303 uint32_t alignment = 16;
3304
3305 /* We need an alignment of 32 for pushing UBOs */
3306 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3307 alignment = MAX2(alignment, 32);
3308
3309 pMemoryRequirements->size = buffer->size;
3310 pMemoryRequirements->alignment = alignment;
3311
3312 /* Storage and Uniform buffers should have their size aligned to
3313 * 32-bits to avoid boundary checks when last DWord is not complete.
3314 * This would ensure that not internal padding would be needed for
3315 * 16-bit types.
3316 */
3317 if (device->robust_buffer_access &&
3318 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3319 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3320 pMemoryRequirements->size = align_u64(buffer->size, 4);
3321
3322 pMemoryRequirements->memoryTypeBits = memory_types;
3323 }
3324
3325 void anv_GetBufferMemoryRequirements2(
3326 VkDevice _device,
3327 const VkBufferMemoryRequirementsInfo2* pInfo,
3328 VkMemoryRequirements2* pMemoryRequirements)
3329 {
3330 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3331 &pMemoryRequirements->memoryRequirements);
3332
3333 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3334 switch (ext->sType) {
3335 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3336 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3337 requirements->prefersDedicatedAllocation = false;
3338 requirements->requiresDedicatedAllocation = false;
3339 break;
3340 }
3341
3342 default:
3343 anv_debug_ignored_stype(ext->sType);
3344 break;
3345 }
3346 }
3347 }
3348
3349 void anv_GetImageMemoryRequirements(
3350 VkDevice _device,
3351 VkImage _image,
3352 VkMemoryRequirements* pMemoryRequirements)
3353 {
3354 ANV_FROM_HANDLE(anv_image, image, _image);
3355 ANV_FROM_HANDLE(anv_device, device, _device);
3356 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3357
3358 /* The Vulkan spec (git aaed022) says:
3359 *
3360 * memoryTypeBits is a bitfield and contains one bit set for every
3361 * supported memory type for the resource. The bit `1<<i` is set if and
3362 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3363 * structure for the physical device is supported.
3364 *
3365 * All types are currently supported for images.
3366 */
3367 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3368
3369 /* We must have image allocated or imported at this point. According to the
3370 * specification, external images must have been bound to memory before
3371 * calling GetImageMemoryRequirements.
3372 */
3373 assert(image->size > 0);
3374
3375 pMemoryRequirements->size = image->size;
3376 pMemoryRequirements->alignment = image->alignment;
3377 pMemoryRequirements->memoryTypeBits = memory_types;
3378 }
3379
3380 void anv_GetImageMemoryRequirements2(
3381 VkDevice _device,
3382 const VkImageMemoryRequirementsInfo2* pInfo,
3383 VkMemoryRequirements2* pMemoryRequirements)
3384 {
3385 ANV_FROM_HANDLE(anv_device, device, _device);
3386 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3387
3388 anv_GetImageMemoryRequirements(_device, pInfo->image,
3389 &pMemoryRequirements->memoryRequirements);
3390
3391 vk_foreach_struct_const(ext, pInfo->pNext) {
3392 switch (ext->sType) {
3393 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3394 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3395 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3396 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3397 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3398 plane_reqs->planeAspect);
3399
3400 assert(image->planes[plane].offset == 0);
3401
3402 /* The Vulkan spec (git aaed022) says:
3403 *
3404 * memoryTypeBits is a bitfield and contains one bit set for every
3405 * supported memory type for the resource. The bit `1<<i` is set
3406 * if and only if the memory type `i` in the
3407 * VkPhysicalDeviceMemoryProperties structure for the physical
3408 * device is supported.
3409 *
3410 * All types are currently supported for images.
3411 */
3412 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3413 (1ull << pdevice->memory.type_count) - 1;
3414
3415 /* We must have image allocated or imported at this point. According to the
3416 * specification, external images must have been bound to memory before
3417 * calling GetImageMemoryRequirements.
3418 */
3419 assert(image->planes[plane].size > 0);
3420
3421 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3422 pMemoryRequirements->memoryRequirements.alignment =
3423 image->planes[plane].alignment;
3424 break;
3425 }
3426
3427 default:
3428 anv_debug_ignored_stype(ext->sType);
3429 break;
3430 }
3431 }
3432
3433 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3434 switch (ext->sType) {
3435 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3436 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3437 if (image->needs_set_tiling || image->external_format) {
3438 /* If we need to set the tiling for external consumers, we need a
3439 * dedicated allocation.
3440 *
3441 * See also anv_AllocateMemory.
3442 */
3443 requirements->prefersDedicatedAllocation = true;
3444 requirements->requiresDedicatedAllocation = true;
3445 } else {
3446 requirements->prefersDedicatedAllocation = false;
3447 requirements->requiresDedicatedAllocation = false;
3448 }
3449 break;
3450 }
3451
3452 default:
3453 anv_debug_ignored_stype(ext->sType);
3454 break;
3455 }
3456 }
3457 }
3458
3459 void anv_GetImageSparseMemoryRequirements(
3460 VkDevice device,
3461 VkImage image,
3462 uint32_t* pSparseMemoryRequirementCount,
3463 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3464 {
3465 *pSparseMemoryRequirementCount = 0;
3466 }
3467
3468 void anv_GetImageSparseMemoryRequirements2(
3469 VkDevice device,
3470 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3471 uint32_t* pSparseMemoryRequirementCount,
3472 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3473 {
3474 *pSparseMemoryRequirementCount = 0;
3475 }
3476
3477 void anv_GetDeviceMemoryCommitment(
3478 VkDevice device,
3479 VkDeviceMemory memory,
3480 VkDeviceSize* pCommittedMemoryInBytes)
3481 {
3482 *pCommittedMemoryInBytes = 0;
3483 }
3484
3485 static void
3486 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3487 {
3488 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3489 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3490
3491 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3492
3493 if (mem) {
3494 assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
3495 buffer->address = (struct anv_address) {
3496 .bo = mem->bo,
3497 .offset = pBindInfo->memoryOffset,
3498 };
3499 } else {
3500 buffer->address = ANV_NULL_ADDRESS;
3501 }
3502 }
3503
3504 VkResult anv_BindBufferMemory(
3505 VkDevice device,
3506 VkBuffer buffer,
3507 VkDeviceMemory memory,
3508 VkDeviceSize memoryOffset)
3509 {
3510 anv_bind_buffer_memory(
3511 &(VkBindBufferMemoryInfo) {
3512 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
3513 .buffer = buffer,
3514 .memory = memory,
3515 .memoryOffset = memoryOffset,
3516 });
3517
3518 return VK_SUCCESS;
3519 }
3520
3521 VkResult anv_BindBufferMemory2(
3522 VkDevice device,
3523 uint32_t bindInfoCount,
3524 const VkBindBufferMemoryInfo* pBindInfos)
3525 {
3526 for (uint32_t i = 0; i < bindInfoCount; i++)
3527 anv_bind_buffer_memory(&pBindInfos[i]);
3528
3529 return VK_SUCCESS;
3530 }
3531
3532 VkResult anv_QueueBindSparse(
3533 VkQueue _queue,
3534 uint32_t bindInfoCount,
3535 const VkBindSparseInfo* pBindInfo,
3536 VkFence fence)
3537 {
3538 ANV_FROM_HANDLE(anv_queue, queue, _queue);
3539 if (anv_device_is_lost(queue->device))
3540 return VK_ERROR_DEVICE_LOST;
3541
3542 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
3543 }
3544
3545 // Event functions
3546
3547 VkResult anv_CreateEvent(
3548 VkDevice _device,
3549 const VkEventCreateInfo* pCreateInfo,
3550 const VkAllocationCallbacks* pAllocator,
3551 VkEvent* pEvent)
3552 {
3553 ANV_FROM_HANDLE(anv_device, device, _device);
3554 struct anv_state state;
3555 struct anv_event *event;
3556
3557 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
3558
3559 state = anv_state_pool_alloc(&device->dynamic_state_pool,
3560 sizeof(*event), 8);
3561 event = state.map;
3562 event->state = state;
3563 event->semaphore = VK_EVENT_RESET;
3564
3565 if (!device->info.has_llc) {
3566 /* Make sure the writes we're flushing have landed. */
3567 __builtin_ia32_mfence();
3568 __builtin_ia32_clflush(event);
3569 }
3570
3571 *pEvent = anv_event_to_handle(event);
3572
3573 return VK_SUCCESS;
3574 }
3575
3576 void anv_DestroyEvent(
3577 VkDevice _device,
3578 VkEvent _event,
3579 const VkAllocationCallbacks* pAllocator)
3580 {
3581 ANV_FROM_HANDLE(anv_device, device, _device);
3582 ANV_FROM_HANDLE(anv_event, event, _event);
3583
3584 if (!event)
3585 return;
3586
3587 anv_state_pool_free(&device->dynamic_state_pool, event->state);
3588 }
3589
3590 VkResult anv_GetEventStatus(
3591 VkDevice _device,
3592 VkEvent _event)
3593 {
3594 ANV_FROM_HANDLE(anv_device, device, _device);
3595 ANV_FROM_HANDLE(anv_event, event, _event);
3596
3597 if (anv_device_is_lost(device))
3598 return VK_ERROR_DEVICE_LOST;
3599
3600 if (!device->info.has_llc) {
3601 /* Invalidate read cache before reading event written by GPU. */
3602 __builtin_ia32_clflush(event);
3603 __builtin_ia32_mfence();
3604
3605 }
3606
3607 return event->semaphore;
3608 }
3609
3610 VkResult anv_SetEvent(
3611 VkDevice _device,
3612 VkEvent _event)
3613 {
3614 ANV_FROM_HANDLE(anv_device, device, _device);
3615 ANV_FROM_HANDLE(anv_event, event, _event);
3616
3617 event->semaphore = VK_EVENT_SET;
3618
3619 if (!device->info.has_llc) {
3620 /* Make sure the writes we're flushing have landed. */
3621 __builtin_ia32_mfence();
3622 __builtin_ia32_clflush(event);
3623 }
3624
3625 return VK_SUCCESS;
3626 }
3627
3628 VkResult anv_ResetEvent(
3629 VkDevice _device,
3630 VkEvent _event)
3631 {
3632 ANV_FROM_HANDLE(anv_device, device, _device);
3633 ANV_FROM_HANDLE(anv_event, event, _event);
3634
3635 event->semaphore = VK_EVENT_RESET;
3636
3637 if (!device->info.has_llc) {
3638 /* Make sure the writes we're flushing have landed. */
3639 __builtin_ia32_mfence();
3640 __builtin_ia32_clflush(event);
3641 }
3642
3643 return VK_SUCCESS;
3644 }
3645
3646 // Buffer functions
3647
3648 VkResult anv_CreateBuffer(
3649 VkDevice _device,
3650 const VkBufferCreateInfo* pCreateInfo,
3651 const VkAllocationCallbacks* pAllocator,
3652 VkBuffer* pBuffer)
3653 {
3654 ANV_FROM_HANDLE(anv_device, device, _device);
3655 struct anv_buffer *buffer;
3656
3657 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
3658
3659 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
3660 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3661 if (buffer == NULL)
3662 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3663
3664 buffer->size = pCreateInfo->size;
3665 buffer->usage = pCreateInfo->usage;
3666 buffer->address = ANV_NULL_ADDRESS;
3667
3668 *pBuffer = anv_buffer_to_handle(buffer);
3669
3670 return VK_SUCCESS;
3671 }
3672
3673 void anv_DestroyBuffer(
3674 VkDevice _device,
3675 VkBuffer _buffer,
3676 const VkAllocationCallbacks* pAllocator)
3677 {
3678 ANV_FROM_HANDLE(anv_device, device, _device);
3679 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3680
3681 if (!buffer)
3682 return;
3683
3684 vk_free2(&device->alloc, pAllocator, buffer);
3685 }
3686
3687 VkDeviceAddress anv_GetBufferDeviceAddressEXT(
3688 VkDevice device,
3689 const VkBufferDeviceAddressInfoEXT* pInfo)
3690 {
3691 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
3692
3693 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
3694
3695 return anv_address_physical(buffer->address);
3696 }
3697
3698 void
3699 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
3700 enum isl_format format,
3701 struct anv_address address,
3702 uint32_t range, uint32_t stride)
3703 {
3704 isl_buffer_fill_state(&device->isl_dev, state.map,
3705 .address = anv_address_physical(address),
3706 .mocs = device->default_mocs,
3707 .size_B = range,
3708 .format = format,
3709 .swizzle = ISL_SWIZZLE_IDENTITY,
3710 .stride_B = stride);
3711 }
3712
3713 void anv_DestroySampler(
3714 VkDevice _device,
3715 VkSampler _sampler,
3716 const VkAllocationCallbacks* pAllocator)
3717 {
3718 ANV_FROM_HANDLE(anv_device, device, _device);
3719 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
3720
3721 if (!sampler)
3722 return;
3723
3724 if (sampler->bindless_state.map) {
3725 anv_state_pool_free(&device->dynamic_state_pool,
3726 sampler->bindless_state);
3727 }
3728
3729 vk_free2(&device->alloc, pAllocator, sampler);
3730 }
3731
3732 VkResult anv_CreateFramebuffer(
3733 VkDevice _device,
3734 const VkFramebufferCreateInfo* pCreateInfo,
3735 const VkAllocationCallbacks* pAllocator,
3736 VkFramebuffer* pFramebuffer)
3737 {
3738 ANV_FROM_HANDLE(anv_device, device, _device);
3739 struct anv_framebuffer *framebuffer;
3740
3741 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
3742
3743 size_t size = sizeof(*framebuffer);
3744
3745 /* VK_KHR_imageless_framebuffer extension says:
3746 *
3747 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3748 * parameter pAttachments is ignored.
3749 */
3750 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
3751 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
3752 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
3753 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3754 if (framebuffer == NULL)
3755 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3756
3757 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
3758 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
3759 framebuffer->attachments[i] = iview;
3760 }
3761 framebuffer->attachment_count = pCreateInfo->attachmentCount;
3762 } else {
3763 assert(device->enabled_extensions.KHR_imageless_framebuffer);
3764 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
3765 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3766 if (framebuffer == NULL)
3767 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3768
3769 framebuffer->attachment_count = 0;
3770 }
3771
3772 framebuffer->width = pCreateInfo->width;
3773 framebuffer->height = pCreateInfo->height;
3774 framebuffer->layers = pCreateInfo->layers;
3775
3776 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
3777
3778 return VK_SUCCESS;
3779 }
3780
3781 void anv_DestroyFramebuffer(
3782 VkDevice _device,
3783 VkFramebuffer _fb,
3784 const VkAllocationCallbacks* pAllocator)
3785 {
3786 ANV_FROM_HANDLE(anv_device, device, _device);
3787 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
3788
3789 if (!fb)
3790 return;
3791
3792 vk_free2(&device->alloc, pAllocator, fb);
3793 }
3794
3795 static const VkTimeDomainEXT anv_time_domains[] = {
3796 VK_TIME_DOMAIN_DEVICE_EXT,
3797 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
3798 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
3799 };
3800
3801 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3802 VkPhysicalDevice physicalDevice,
3803 uint32_t *pTimeDomainCount,
3804 VkTimeDomainEXT *pTimeDomains)
3805 {
3806 int d;
3807 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
3808
3809 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
3810 vk_outarray_append(&out, i) {
3811 *i = anv_time_domains[d];
3812 }
3813 }
3814
3815 return vk_outarray_status(&out);
3816 }
3817
3818 static uint64_t
3819 anv_clock_gettime(clockid_t clock_id)
3820 {
3821 struct timespec current;
3822 int ret;
3823
3824 ret = clock_gettime(clock_id, &current);
3825 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
3826 ret = clock_gettime(CLOCK_MONOTONIC, &current);
3827 if (ret < 0)
3828 return 0;
3829
3830 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
3831 }
3832
3833 #define TIMESTAMP 0x2358
3834
3835 VkResult anv_GetCalibratedTimestampsEXT(
3836 VkDevice _device,
3837 uint32_t timestampCount,
3838 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
3839 uint64_t *pTimestamps,
3840 uint64_t *pMaxDeviation)
3841 {
3842 ANV_FROM_HANDLE(anv_device, device, _device);
3843 uint64_t timestamp_frequency = device->info.timestamp_frequency;
3844 int ret;
3845 int d;
3846 uint64_t begin, end;
3847 uint64_t max_clock_period = 0;
3848
3849 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3850
3851 for (d = 0; d < timestampCount; d++) {
3852 switch (pTimestampInfos[d].timeDomain) {
3853 case VK_TIME_DOMAIN_DEVICE_EXT:
3854 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
3855 &pTimestamps[d]);
3856
3857 if (ret != 0) {
3858 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
3859 "register: %m");
3860 }
3861 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
3862 max_clock_period = MAX2(max_clock_period, device_period);
3863 break;
3864 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
3865 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
3866 max_clock_period = MAX2(max_clock_period, 1);
3867 break;
3868
3869 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
3870 pTimestamps[d] = begin;
3871 break;
3872 default:
3873 pTimestamps[d] = 0;
3874 break;
3875 }
3876 }
3877
3878 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3879
3880 /*
3881 * The maximum deviation is the sum of the interval over which we
3882 * perform the sampling and the maximum period of any sampled
3883 * clock. That's because the maximum skew between any two sampled
3884 * clock edges is when the sampled clock with the largest period is
3885 * sampled at the end of that period but right at the beginning of the
3886 * sampling interval and some other clock is sampled right at the
3887 * begining of its sampling period and right at the end of the
3888 * sampling interval. Let's assume the GPU has the longest clock
3889 * period and that the application is sampling GPU and monotonic:
3890 *
3891 * s e
3892 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3893 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3894 *
3895 * g
3896 * 0 1 2 3
3897 * GPU -----_____-----_____-----_____-----_____
3898 *
3899 * m
3900 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3901 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3902 *
3903 * Interval <----------------->
3904 * Deviation <-------------------------->
3905 *
3906 * s = read(raw) 2
3907 * g = read(GPU) 1
3908 * m = read(monotonic) 2
3909 * e = read(raw) b
3910 *
3911 * We round the sample interval up by one tick to cover sampling error
3912 * in the interval clock
3913 */
3914
3915 uint64_t sample_interval = end - begin + 1;
3916
3917 *pMaxDeviation = sample_interval + max_clock_period;
3918
3919 return VK_SUCCESS;
3920 }
3921
3922 /* vk_icd.h does not declare this function, so we declare it here to
3923 * suppress Wmissing-prototypes.
3924 */
3925 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3926 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
3927
3928 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3929 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
3930 {
3931 /* For the full details on loader interface versioning, see
3932 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3933 * What follows is a condensed summary, to help you navigate the large and
3934 * confusing official doc.
3935 *
3936 * - Loader interface v0 is incompatible with later versions. We don't
3937 * support it.
3938 *
3939 * - In loader interface v1:
3940 * - The first ICD entrypoint called by the loader is
3941 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3942 * entrypoint.
3943 * - The ICD must statically expose no other Vulkan symbol unless it is
3944 * linked with -Bsymbolic.
3945 * - Each dispatchable Vulkan handle created by the ICD must be
3946 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3947 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3948 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3949 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3950 * such loader-managed surfaces.
3951 *
3952 * - Loader interface v2 differs from v1 in:
3953 * - The first ICD entrypoint called by the loader is
3954 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3955 * statically expose this entrypoint.
3956 *
3957 * - Loader interface v3 differs from v2 in:
3958 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3959 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3960 * because the loader no longer does so.
3961 */
3962 *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
3963 return VK_SUCCESS;
3964 }