projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
anv: Drop the instance pointer from anv_device
[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/debug.h"
36 #include "util/build_id.h"
37 #include "util/disk_cache.h"
38 #include "util/mesa-sha1.h"
39 #include "util/os_file.h"
40 #include "util/u_atomic.h"
41 #include "util/u_string.h"
42 #include "util/xmlpool.h"
43 #include "git_sha1.h"
44 #include "vk_util.h"
45 #include "common/gen_aux_map.h"
46 #include "common/gen_defines.h"
47 #include "compiler/glsl_types.h"
48
49 #include "genxml/gen7_pack.h"
50
51 static const char anv_dri_options_xml[] =
52 DRI_CONF_BEGIN
53 DRI_CONF_SECTION_PERFORMANCE
54 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
55 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
56 DRI_CONF_SECTION_END
57
58 DRI_CONF_SECTION_DEBUG
59 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
60 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST("false")
61 DRI_CONF_SECTION_END
62 DRI_CONF_END;
63
64 /* This is probably far to big but it reflects the max size used for messages
65 * in OpenGLs KHR_debug.
66 */
67 #define MAX_DEBUG_MESSAGE_LENGTH 4096
68
69 static void
70 compiler_debug_log(void *data, const char *fmt, ...)
71 {
72 char str[MAX_DEBUG_MESSAGE_LENGTH];
73 struct anv_device *device = (struct anv_device *)data;
74 struct anv_instance *instance = device->physical->instance;
75
76 if (list_is_empty(&instance->debug_report_callbacks.callbacks))
77 return;
78
79 va_list args;
80 va_start(args, fmt);
81 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
82 va_end(args);
83
84 vk_debug_report(&instance->debug_report_callbacks,
85 VK_DEBUG_REPORT_DEBUG_BIT_EXT,
86 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,
87 0, 0, 0, "anv", str);
88 }
89
90 static void
91 compiler_perf_log(void *data, const char *fmt, ...)
92 {
93 va_list args;
94 va_start(args, fmt);
95
96 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
97 intel_logd_v(fmt, args);
98
99 va_end(args);
100 }
101
102 static uint64_t
103 anv_compute_heap_size(int fd, uint64_t gtt_size)
104 {
105 /* Query the total ram from the system */
106 struct sysinfo info;
107 sysinfo(&info);
108
109 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
110
111 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
112 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
113 */
114 uint64_t available_ram;
115 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
116 available_ram = total_ram / 2;
117 else
118 available_ram = total_ram * 3 / 4;
119
120 /* We also want to leave some padding for things we allocate in the driver,
121 * so don't go over 3/4 of the GTT either.
122 */
123 uint64_t available_gtt = gtt_size * 3 / 4;
124
125 return MIN2(available_ram, available_gtt);
126 }
127
128 static VkResult
129 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
130 {
131 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
132 &device->gtt_size) == -1) {
133 /* If, for whatever reason, we can't actually get the GTT size from the
134 * kernel (too old?) fall back to the aperture size.
135 */
136 anv_perf_warn(NULL, NULL,
137 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
138
139 if (anv_gem_get_aperture(fd, &device->gtt_size) == -1) {
140 return vk_errorfi(device->instance, NULL,
141 VK_ERROR_INITIALIZATION_FAILED,
142 "failed to get aperture size: %m");
143 }
144 }
145
146 /* We only allow 48-bit addresses with softpin because knowing the actual
147 * address is required for the vertex cache flush workaround.
148 */
149 device->supports_48bit_addresses = (device->info.gen >= 8) &&
150 device->has_softpin &&
151 device->gtt_size > (4ULL << 30 /* GiB */);
152
153 uint64_t heap_size = anv_compute_heap_size(fd, device->gtt_size);
154
155 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
156 /* When running with an overridden PCI ID, we may get a GTT size from
157 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
158 * address support can still fail. Just clamp the address space size to
159 * 2 GiB if we don't have 48-bit support.
160 */
161 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
162 "not support for 48-bit addresses",
163 __FILE__, __LINE__);
164 heap_size = 2ull << 30;
165 }
166
167 device->memory.heap_count = 1;
168 device->memory.heaps[0] = (struct anv_memory_heap) {
169 .size = heap_size,
170 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
171 };
172
173 uint32_t type_count = 0;
174 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
175 if (device->info.has_llc) {
176 /* Big core GPUs share LLC with the CPU and thus one memory type can be
177 * both cached and coherent at the same time.
178 */
179 device->memory.types[type_count++] = (struct anv_memory_type) {
180 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
181 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
182 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
183 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
184 .heapIndex = heap,
185 };
186 } else {
187 /* The spec requires that we expose a host-visible, coherent memory
188 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
189 * to give the application a choice between cached, but not coherent and
190 * coherent but uncached (WC though).
191 */
192 device->memory.types[type_count++] = (struct anv_memory_type) {
193 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
194 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
195 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
196 .heapIndex = heap,
197 };
198 device->memory.types[type_count++] = (struct anv_memory_type) {
199 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
200 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
201 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
202 .heapIndex = heap,
203 };
204 }
205 }
206 device->memory.type_count = type_count;
207
208 return VK_SUCCESS;
209 }
210
211 static VkResult
212 anv_physical_device_init_uuids(struct anv_physical_device *device)
213 {
214 const struct build_id_note *note =
215 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
216 if (!note) {
217 return vk_errorfi(device->instance, NULL,
218 VK_ERROR_INITIALIZATION_FAILED,
219 "Failed to find build-id");
220 }
221
222 unsigned build_id_len = build_id_length(note);
223 if (build_id_len < 20) {
224 return vk_errorfi(device->instance, NULL,
225 VK_ERROR_INITIALIZATION_FAILED,
226 "build-id too short. It needs to be a SHA");
227 }
228
229 memcpy(device->driver_build_sha1, build_id_data(note), 20);
230
231 struct mesa_sha1 sha1_ctx;
232 uint8_t sha1[20];
233 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
234
235 /* The pipeline cache UUID is used for determining when a pipeline cache is
236 * invalid. It needs both a driver build and the PCI ID of the device.
237 */
238 _mesa_sha1_init(&sha1_ctx);
239 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
240 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
241 sizeof(device->chipset_id));
242 _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless,
243 sizeof(device->always_use_bindless));
244 _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
245 sizeof(device->has_a64_buffer_access));
246 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images,
247 sizeof(device->has_bindless_images));
248 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers,
249 sizeof(device->has_bindless_samplers));
250 _mesa_sha1_final(&sha1_ctx, sha1);
251 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
252
253 /* The driver UUID is used for determining sharability of images and memory
254 * between two Vulkan instances in separate processes. People who want to
255 * share memory need to also check the device UUID (below) so all this
256 * needs to be is the build-id.
257 */
258 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
259
260 /* The device UUID uniquely identifies the given device within the machine.
261 * Since we never have more than one device, this doesn't need to be a real
262 * UUID. However, on the off-chance that someone tries to use this to
263 * cache pre-tiled images or something of the like, we use the PCI ID and
264 * some bits of ISL info to ensure that this is safe.
265 */
266 _mesa_sha1_init(&sha1_ctx);
267 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
268 sizeof(device->chipset_id));
269 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
270 sizeof(device->isl_dev.has_bit6_swizzling));
271 _mesa_sha1_final(&sha1_ctx, sha1);
272 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
273
274 return VK_SUCCESS;
275 }
276
277 static void
278 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
279 {
280 #ifdef ENABLE_SHADER_CACHE
281 char renderer[10];
282 ASSERTED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
283 device->chipset_id);
284 assert(len == sizeof(renderer) - 2);
285
286 char timestamp[41];
287 _mesa_sha1_format(timestamp, device->driver_build_sha1);
288
289 const uint64_t driver_flags =
290 brw_get_compiler_config_value(device->compiler);
291 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
292 #else
293 device->disk_cache = NULL;
294 #endif
295 }
296
297 static void
298 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
299 {
300 #ifdef ENABLE_SHADER_CACHE
301 if (device->disk_cache)
302 disk_cache_destroy(device->disk_cache);
303 #else
304 assert(device->disk_cache == NULL);
305 #endif
306 }
307
308 static uint64_t
309 get_available_system_memory()
310 {
311 char *meminfo = os_read_file("/proc/meminfo");
312 if (!meminfo)
313 return 0;
314
315 char *str = strstr(meminfo, "MemAvailable:");
316 if (!str) {
317 free(meminfo);
318 return 0;
319 }
320
321 uint64_t kb_mem_available;
322 if (sscanf(str, "MemAvailable: %" PRIx64, &kb_mem_available) == 1) {
323 free(meminfo);
324 return kb_mem_available << 10;
325 }
326
327 free(meminfo);
328 return 0;
329 }
330
331 static VkResult
332 anv_physical_device_init(struct anv_physical_device *device,
333 struct anv_instance *instance,
334 drmDevicePtr drm_device)
335 {
336 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
337 const char *path = drm_device->nodes[DRM_NODE_RENDER];
338 VkResult result;
339 int fd;
340 int master_fd = -1;
341
342 brw_process_intel_debug_variable();
343
344 fd = open(path, O_RDWR | O_CLOEXEC);
345 if (fd < 0)
346 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
347
348 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
349 device->instance = instance;
350
351 assert(strlen(path) < ARRAY_SIZE(device->path));
352 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
353
354 if (!gen_get_device_info_from_fd(fd, &device->info)) {
355 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
356 goto fail;
357 }
358 device->chipset_id = device->info.chipset_id;
359 device->no_hw = device->info.no_hw;
360
361 if (getenv("INTEL_NO_HW") != NULL)
362 device->no_hw = true;
363
364 device->pci_info.domain = drm_device->businfo.pci->domain;
365 device->pci_info.bus = drm_device->businfo.pci->bus;
366 device->pci_info.device = drm_device->businfo.pci->dev;
367 device->pci_info.function = drm_device->businfo.pci->func;
368
369 device->name = gen_get_device_name(device->chipset_id);
370
371 if (device->info.is_haswell) {
372 intel_logw("Haswell Vulkan support is incomplete");
373 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
374 intel_logw("Ivy Bridge Vulkan support is incomplete");
375 } else if (device->info.gen == 7 && device->info.is_baytrail) {
376 intel_logw("Bay Trail Vulkan support is incomplete");
377 } else if (device->info.gen >= 8 && device->info.gen <= 11) {
378 /* Gen8-11 fully supported */
379 } else if (device->info.gen == 12) {
380 intel_logw("Vulkan is not yet fully supported on gen12");
381 } else {
382 result = vk_errorfi(instance, NULL, VK_ERROR_INCOMPATIBLE_DRIVER,
383 "Vulkan not yet supported on %s", device->name);
384 goto fail;
385 }
386
387 device->cmd_parser_version = -1;
388 if (device->info.gen == 7) {
389 device->cmd_parser_version =
390 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
391 if (device->cmd_parser_version == -1) {
392 result = vk_errorfi(device->instance, NULL,
393 VK_ERROR_INITIALIZATION_FAILED,
394 "failed to get command parser version");
395 goto fail;
396 }
397 }
398
399 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
400 result = vk_errorfi(device->instance, NULL,
401 VK_ERROR_INITIALIZATION_FAILED,
402 "kernel missing gem wait");
403 goto fail;
404 }
405
406 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
407 result = vk_errorfi(device->instance, NULL,
408 VK_ERROR_INITIALIZATION_FAILED,
409 "kernel missing execbuf2");
410 goto fail;
411 }
412
413 if (!device->info.has_llc &&
414 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
415 result = vk_errorfi(device->instance, NULL,
416 VK_ERROR_INITIALIZATION_FAILED,
417 "kernel missing wc mmap");
418 goto fail;
419 }
420
421 device->has_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN);
422 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
423 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
424 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
425 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
426 device->has_syncobj_wait = device->has_syncobj &&
427 anv_gem_supports_syncobj_wait(fd);
428 device->has_context_priority = anv_gem_has_context_priority(fd);
429
430 result = anv_physical_device_init_heaps(device, fd);
431 if (result != VK_SUCCESS)
432 goto fail;
433
434 device->use_softpin = device->has_softpin &&
435 device->supports_48bit_addresses;
436
437 device->has_context_isolation =
438 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
439
440 device->always_use_bindless =
441 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
442
443 /* We first got the A64 messages on broadwell and we can only use them if
444 * we can pass addresses directly into the shader which requires softpin.
445 */
446 device->has_a64_buffer_access = device->info.gen >= 8 &&
447 device->use_softpin;
448
449 /* We first get bindless image access on Skylake and we can only really do
450 * it if we don't have any relocations so we need softpin.
451 */
452 device->has_bindless_images = device->info.gen >= 9 &&
453 device->use_softpin;
454
455 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
456 * because it's just a matter of setting the sampler address in the sample
457 * message header. However, we've not bothered to wire it up for vec4 so
458 * we leave it disabled on gen7.
459 */
460 device->has_bindless_samplers = device->info.gen >= 8;
461
462 device->has_mem_available = get_available_system_memory() != 0;
463
464 device->always_flush_cache =
465 driQueryOptionb(&instance->dri_options, "always_flush_cache");
466
467 /* Starting with Gen10, the timestamp frequency of the command streamer may
468 * vary from one part to another. We can query the value from the kernel.
469 */
470 if (device->info.gen >= 10) {
471 int timestamp_frequency =
472 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
473
474 if (timestamp_frequency < 0)
475 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
476 else
477 device->info.timestamp_frequency = timestamp_frequency;
478 }
479
480 /* GENs prior to 8 do not support EU/Subslice info */
481 if (device->info.gen >= 8) {
482 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
483 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
484
485 /* Without this information, we cannot get the right Braswell
486 * brandstrings, and we have to use conservative numbers for GPGPU on
487 * many platforms, but otherwise, things will just work.
488 */
489 if (device->subslice_total < 1 || device->eu_total < 1) {
490 intel_logw("Kernel 4.1 required to properly query GPU properties");
491 }
492 } else if (device->info.gen == 7) {
493 device->subslice_total = 1 << (device->info.gt - 1);
494 }
495
496 if (device->info.is_cherryview &&
497 device->subslice_total > 0 && device->eu_total > 0) {
498 /* Logical CS threads = EUs per subslice * num threads per EU */
499 uint32_t max_cs_threads =
500 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
501
502 /* Fuse configurations may give more threads than expected, never less. */
503 if (max_cs_threads > device->info.max_cs_threads)
504 device->info.max_cs_threads = max_cs_threads;
505 }
506
507 device->compiler = brw_compiler_create(NULL, &device->info);
508 if (device->compiler == NULL) {
509 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
510 goto fail;
511 }
512 device->compiler->shader_debug_log = compiler_debug_log;
513 device->compiler->shader_perf_log = compiler_perf_log;
514 device->compiler->supports_pull_constants = false;
515 device->compiler->constant_buffer_0_is_relative =
516 device->info.gen < 8 || !device->has_context_isolation;
517 device->compiler->supports_shader_constants = true;
518 device->compiler->compact_params = false;
519
520 /* Broadwell PRM says:
521 *
522 * "Before Gen8, there was a historical configuration control field to
523 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
524 * different places: TILECTL[1:0], ARB_MODE[5:4], and
525 * DISP_ARB_CTL[14:13].
526 *
527 * For Gen8 and subsequent generations, the swizzle fields are all
528 * reserved, and the CPU's memory controller performs all address
529 * swizzling modifications."
530 */
531 bool swizzled =
532 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
533
534 isl_device_init(&device->isl_dev, &device->info, swizzled);
535
536 result = anv_physical_device_init_uuids(device);
537 if (result != VK_SUCCESS)
538 goto fail;
539
540 anv_physical_device_init_disk_cache(device);
541
542 if (instance->enabled_extensions.KHR_display) {
543 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
544 if (master_fd >= 0) {
545 /* prod the device with a GETPARAM call which will fail if
546 * we don't have permission to even render on this device
547 */
548 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
549 close(master_fd);
550 master_fd = -1;
551 }
552 }
553 }
554 device->master_fd = master_fd;
555
556 result = anv_init_wsi(device);
557 if (result != VK_SUCCESS) {
558 ralloc_free(device->compiler);
559 anv_physical_device_free_disk_cache(device);
560 goto fail;
561 }
562
563 device->perf = anv_get_perf(&device->info, fd);
564
565 anv_physical_device_get_supported_extensions(device,
566 &device->supported_extensions);
567
568
569 device->local_fd = fd;
570
571 return VK_SUCCESS;
572
573 fail:
574 close(fd);
575 if (master_fd != -1)
576 close(master_fd);
577 return result;
578 }
579
580 static void
581 anv_physical_device_finish(struct anv_physical_device *device)
582 {
583 anv_finish_wsi(device);
584 anv_physical_device_free_disk_cache(device);
585 ralloc_free(device->compiler);
586 ralloc_free(device->perf);
587 close(device->local_fd);
588 if (device->master_fd >= 0)
589 close(device->master_fd);
590 }
591
592 static void *
593 default_alloc_func(void *pUserData, size_t size, size_t align,
594 VkSystemAllocationScope allocationScope)
595 {
596 return malloc(size);
597 }
598
599 static void *
600 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
601 size_t align, VkSystemAllocationScope allocationScope)
602 {
603 return realloc(pOriginal, size);
604 }
605
606 static void
607 default_free_func(void *pUserData, void *pMemory)
608 {
609 free(pMemory);
610 }
611
612 static const VkAllocationCallbacks default_alloc = {
613 .pUserData = NULL,
614 .pfnAllocation = default_alloc_func,
615 .pfnReallocation = default_realloc_func,
616 .pfnFree = default_free_func,
617 };
618
619 VkResult anv_EnumerateInstanceExtensionProperties(
620 const char* pLayerName,
621 uint32_t* pPropertyCount,
622 VkExtensionProperties* pProperties)
623 {
624 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
625
626 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
627 if (anv_instance_extensions_supported.extensions[i]) {
628 vk_outarray_append(&out, prop) {
629 *prop = anv_instance_extensions[i];
630 }
631 }
632 }
633
634 return vk_outarray_status(&out);
635 }
636
637 VkResult anv_CreateInstance(
638 const VkInstanceCreateInfo* pCreateInfo,
639 const VkAllocationCallbacks* pAllocator,
640 VkInstance* pInstance)
641 {
642 struct anv_instance *instance;
643 VkResult result;
644
645 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
646
647 struct anv_instance_extension_table enabled_extensions = {};
648 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
649 int idx;
650 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
651 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
652 anv_instance_extensions[idx].extensionName) == 0)
653 break;
654 }
655
656 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
657 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
658
659 if (!anv_instance_extensions_supported.extensions[idx])
660 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
661
662 enabled_extensions.extensions[idx] = true;
663 }
664
665 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
666 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
667 if (!instance)
668 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
669
670 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
671
672 if (pAllocator)
673 instance->alloc = *pAllocator;
674 else
675 instance->alloc = default_alloc;
676
677 instance->app_info = (struct anv_app_info) { .api_version = 0 };
678 if (pCreateInfo->pApplicationInfo) {
679 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
680
681 instance->app_info.app_name =
682 vk_strdup(&instance->alloc, app->pApplicationName,
683 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
684 instance->app_info.app_version = app->applicationVersion;
685
686 instance->app_info.engine_name =
687 vk_strdup(&instance->alloc, app->pEngineName,
688 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
689 instance->app_info.engine_version = app->engineVersion;
690
691 instance->app_info.api_version = app->apiVersion;
692 }
693
694 if (instance->app_info.api_version == 0)
695 instance->app_info.api_version = VK_API_VERSION_1_0;
696
697 instance->enabled_extensions = enabled_extensions;
698
699 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
700 /* Vulkan requires that entrypoints for extensions which have not been
701 * enabled must not be advertised.
702 */
703 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
704 &instance->enabled_extensions)) {
705 instance->dispatch.entrypoints[i] = NULL;
706 } else {
707 instance->dispatch.entrypoints[i] =
708 anv_instance_dispatch_table.entrypoints[i];
709 }
710 }
711
712 struct anv_physical_device *pdevice = &instance->physicalDevice;
713 for (unsigned i = 0; i < ARRAY_SIZE(pdevice->dispatch.entrypoints); i++) {
714 /* Vulkan requires that entrypoints for extensions which have not been
715 * enabled must not be advertised.
716 */
717 if (!anv_physical_device_entrypoint_is_enabled(i, instance->app_info.api_version,
718 &instance->enabled_extensions)) {
719 pdevice->dispatch.entrypoints[i] = NULL;
720 } else {
721 pdevice->dispatch.entrypoints[i] =
722 anv_physical_device_dispatch_table.entrypoints[i];
723 }
724 }
725
726 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
727 /* Vulkan requires that entrypoints for extensions which have not been
728 * enabled must not be advertised.
729 */
730 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
731 &instance->enabled_extensions, NULL)) {
732 instance->device_dispatch.entrypoints[i] = NULL;
733 } else {
734 instance->device_dispatch.entrypoints[i] =
735 anv_device_dispatch_table.entrypoints[i];
736 }
737 }
738
739 instance->physicalDeviceCount = -1;
740
741 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
742 if (result != VK_SUCCESS) {
743 vk_free2(&default_alloc, pAllocator, instance);
744 return vk_error(result);
745 }
746
747 instance->pipeline_cache_enabled =
748 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
749
750 glsl_type_singleton_init_or_ref();
751
752 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
753
754 driParseOptionInfo(&instance->available_dri_options, anv_dri_options_xml);
755 driParseConfigFiles(&instance->dri_options, &instance->available_dri_options,
756 0, "anv", NULL,
757 instance->app_info.engine_name,
758 instance->app_info.engine_version);
759
760 *pInstance = anv_instance_to_handle(instance);
761
762 return VK_SUCCESS;
763 }
764
765 void anv_DestroyInstance(
766 VkInstance _instance,
767 const VkAllocationCallbacks* pAllocator)
768 {
769 ANV_FROM_HANDLE(anv_instance, instance, _instance);
770
771 if (!instance)
772 return;
773
774 if (instance->physicalDeviceCount > 0) {
775 /* We support at most one physical device. */
776 assert(instance->physicalDeviceCount == 1);
777 anv_physical_device_finish(&instance->physicalDevice);
778 }
779
780 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
781 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
782
783 VG(VALGRIND_DESTROY_MEMPOOL(instance));
784
785 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
786
787 glsl_type_singleton_decref();
788
789 driDestroyOptionCache(&instance->dri_options);
790 driDestroyOptionInfo(&instance->available_dri_options);
791
792 vk_free(&instance->alloc, instance);
793 }
794
795 static VkResult
796 anv_enumerate_devices(struct anv_instance *instance)
797 {
798 /* TODO: Check for more devices ? */
799 drmDevicePtr devices[8];
800 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
801 int max_devices;
802
803 instance->physicalDeviceCount = 0;
804
805 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
806 if (max_devices < 1)
807 return VK_ERROR_INCOMPATIBLE_DRIVER;
808
809 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
810 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
811 devices[i]->bustype == DRM_BUS_PCI &&
812 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
813
814 result = anv_physical_device_init(&instance->physicalDevice,
815 instance, devices[i]);
816 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
817 break;
818 }
819 }
820 drmFreeDevices(devices, max_devices);
821
822 if (result == VK_SUCCESS)
823 instance->physicalDeviceCount = 1;
824
825 return result;
826 }
827
828 static VkResult
829 anv_instance_ensure_physical_device(struct anv_instance *instance)
830 {
831 if (instance->physicalDeviceCount < 0) {
832 VkResult result = anv_enumerate_devices(instance);
833 if (result != VK_SUCCESS &&
834 result != VK_ERROR_INCOMPATIBLE_DRIVER)
835 return result;
836 }
837
838 return VK_SUCCESS;
839 }
840
841 VkResult anv_EnumeratePhysicalDevices(
842 VkInstance _instance,
843 uint32_t* pPhysicalDeviceCount,
844 VkPhysicalDevice* pPhysicalDevices)
845 {
846 ANV_FROM_HANDLE(anv_instance, instance, _instance);
847 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
848
849 VkResult result = anv_instance_ensure_physical_device(instance);
850 if (result != VK_SUCCESS)
851 return result;
852
853 if (instance->physicalDeviceCount == 0)
854 return VK_SUCCESS;
855
856 assert(instance->physicalDeviceCount == 1);
857 vk_outarray_append(&out, i) {
858 *i = anv_physical_device_to_handle(&instance->physicalDevice);
859 }
860
861 return vk_outarray_status(&out);
862 }
863
864 VkResult anv_EnumeratePhysicalDeviceGroups(
865 VkInstance _instance,
866 uint32_t* pPhysicalDeviceGroupCount,
867 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
868 {
869 ANV_FROM_HANDLE(anv_instance, instance, _instance);
870 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
871 pPhysicalDeviceGroupCount);
872
873 VkResult result = anv_instance_ensure_physical_device(instance);
874 if (result != VK_SUCCESS)
875 return result;
876
877 if (instance->physicalDeviceCount == 0)
878 return VK_SUCCESS;
879
880 assert(instance->physicalDeviceCount == 1);
881
882 vk_outarray_append(&out, p) {
883 p->physicalDeviceCount = 1;
884 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
885 p->physicalDevices[0] =
886 anv_physical_device_to_handle(&instance->physicalDevice);
887 p->subsetAllocation = false;
888
889 vk_foreach_struct(ext, p->pNext)
890 anv_debug_ignored_stype(ext->sType);
891 }
892
893 return vk_outarray_status(&out);
894 }
895
896 void anv_GetPhysicalDeviceFeatures(
897 VkPhysicalDevice physicalDevice,
898 VkPhysicalDeviceFeatures* pFeatures)
899 {
900 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
901
902 *pFeatures = (VkPhysicalDeviceFeatures) {
903 .robustBufferAccess = true,
904 .fullDrawIndexUint32 = true,
905 .imageCubeArray = true,
906 .independentBlend = true,
907 .geometryShader = true,
908 .tessellationShader = true,
909 .sampleRateShading = true,
910 .dualSrcBlend = true,
911 .logicOp = true,
912 .multiDrawIndirect = true,
913 .drawIndirectFirstInstance = true,
914 .depthClamp = true,
915 .depthBiasClamp = true,
916 .fillModeNonSolid = true,
917 .depthBounds = pdevice->info.gen >= 12,
918 .wideLines = true,
919 .largePoints = true,
920 .alphaToOne = true,
921 .multiViewport = true,
922 .samplerAnisotropy = true,
923 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
924 pdevice->info.is_baytrail,
925 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
926 .textureCompressionBC = true,
927 .occlusionQueryPrecise = true,
928 .pipelineStatisticsQuery = true,
929 .fragmentStoresAndAtomics = true,
930 .shaderTessellationAndGeometryPointSize = true,
931 .shaderImageGatherExtended = true,
932 .shaderStorageImageExtendedFormats = true,
933 .shaderStorageImageMultisample = false,
934 .shaderStorageImageReadWithoutFormat = false,
935 .shaderStorageImageWriteWithoutFormat = true,
936 .shaderUniformBufferArrayDynamicIndexing = true,
937 .shaderSampledImageArrayDynamicIndexing = true,
938 .shaderStorageBufferArrayDynamicIndexing = true,
939 .shaderStorageImageArrayDynamicIndexing = true,
940 .shaderClipDistance = true,
941 .shaderCullDistance = true,
942 .shaderFloat64 = pdevice->info.gen >= 8 &&
943 pdevice->info.has_64bit_types,
944 .shaderInt64 = pdevice->info.gen >= 8 &&
945 pdevice->info.has_64bit_types,
946 .shaderInt16 = pdevice->info.gen >= 8,
947 .shaderResourceMinLod = pdevice->info.gen >= 9,
948 .variableMultisampleRate = true,
949 .inheritedQueries = true,
950 };
951
952 /* We can't do image stores in vec4 shaders */
953 pFeatures->vertexPipelineStoresAndAtomics =
954 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
955 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
956
957 struct anv_app_info *app_info = &pdevice->instance->app_info;
958
959 /* The new DOOM and Wolfenstein games require depthBounds without
960 * checking for it. They seem to run fine without it so just claim it's
961 * there and accept the consequences.
962 */
963 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
964 pFeatures->depthBounds = true;
965 }
966
967 static void
968 anv_get_physical_device_features_1_1(struct anv_physical_device *pdevice,
969 VkPhysicalDeviceVulkan11Features *f)
970 {
971 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES);
972
973 f->storageBuffer16BitAccess = pdevice->info.gen >= 8;
974 f->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
975 f->storagePushConstant16 = pdevice->info.gen >= 8;
976 f->storageInputOutput16 = false;
977 f->multiview = true;
978 f->multiviewGeometryShader = true;
979 f->multiviewTessellationShader = true;
980 f->variablePointersStorageBuffer = true;
981 f->variablePointers = true;
982 f->protectedMemory = false;
983 f->samplerYcbcrConversion = true;
984 f->shaderDrawParameters = true;
985 }
986
987 static void
988 anv_get_physical_device_features_1_2(struct anv_physical_device *pdevice,
989 VkPhysicalDeviceVulkan12Features *f)
990 {
991 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES);
992
993 f->samplerMirrorClampToEdge = true;
994 f->drawIndirectCount = true;
995 f->storageBuffer8BitAccess = pdevice->info.gen >= 8;
996 f->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
997 f->storagePushConstant8 = pdevice->info.gen >= 8;
998 f->shaderBufferInt64Atomics = pdevice->info.gen >= 9 &&
999 pdevice->use_softpin;
1000 f->shaderSharedInt64Atomics = false;
1001 f->shaderFloat16 = pdevice->info.gen >= 8;
1002 f->shaderInt8 = pdevice->info.gen >= 8;
1003
1004 bool descIndexing = pdevice->has_a64_buffer_access &&
1005 pdevice->has_bindless_images;
1006 f->descriptorIndexing = descIndexing;
1007 f->shaderInputAttachmentArrayDynamicIndexing = false;
1008 f->shaderUniformTexelBufferArrayDynamicIndexing = descIndexing;
1009 f->shaderStorageTexelBufferArrayDynamicIndexing = descIndexing;
1010 f->shaderUniformBufferArrayNonUniformIndexing = false;
1011 f->shaderSampledImageArrayNonUniformIndexing = descIndexing;
1012 f->shaderStorageBufferArrayNonUniformIndexing = descIndexing;
1013 f->shaderStorageImageArrayNonUniformIndexing = descIndexing;
1014 f->shaderInputAttachmentArrayNonUniformIndexing = false;
1015 f->shaderUniformTexelBufferArrayNonUniformIndexing = descIndexing;
1016 f->shaderStorageTexelBufferArrayNonUniformIndexing = descIndexing;
1017 f->descriptorBindingUniformBufferUpdateAfterBind = false;
1018 f->descriptorBindingSampledImageUpdateAfterBind = descIndexing;
1019 f->descriptorBindingStorageImageUpdateAfterBind = descIndexing;
1020 f->descriptorBindingStorageBufferUpdateAfterBind = descIndexing;
1021 f->descriptorBindingUniformTexelBufferUpdateAfterBind = descIndexing;
1022 f->descriptorBindingStorageTexelBufferUpdateAfterBind = descIndexing;
1023 f->descriptorBindingUpdateUnusedWhilePending = descIndexing;
1024 f->descriptorBindingPartiallyBound = descIndexing;
1025 f->descriptorBindingVariableDescriptorCount = false;
1026 f->runtimeDescriptorArray = descIndexing;
1027
1028 f->samplerFilterMinmax = pdevice->info.gen >= 9;
1029 f->scalarBlockLayout = true;
1030 f->imagelessFramebuffer = true;
1031 f->uniformBufferStandardLayout = true;
1032 f->shaderSubgroupExtendedTypes = true;
1033 f->separateDepthStencilLayouts = true;
1034 f->hostQueryReset = true;
1035 f->timelineSemaphore = true;
1036 f->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1037 f->bufferDeviceAddressCaptureReplay = pdevice->has_a64_buffer_access;
1038 f->bufferDeviceAddressMultiDevice = false;
1039 f->vulkanMemoryModel = true;
1040 f->vulkanMemoryModelDeviceScope = true;
1041 f->vulkanMemoryModelAvailabilityVisibilityChains = true;
1042 f->shaderOutputViewportIndex = true;
1043 f->shaderOutputLayer = true;
1044 f->subgroupBroadcastDynamicId = true;
1045 }
1046
1047 void anv_GetPhysicalDeviceFeatures2(
1048 VkPhysicalDevice physicalDevice,
1049 VkPhysicalDeviceFeatures2* pFeatures)
1050 {
1051 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1052 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
1053
1054 VkPhysicalDeviceVulkan11Features core_1_1 = {
1055 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES,
1056 };
1057 anv_get_physical_device_features_1_1(pdevice, &core_1_1);
1058
1059 VkPhysicalDeviceVulkan12Features core_1_2 = {
1060 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
1061 };
1062 anv_get_physical_device_features_1_2(pdevice, &core_1_2);
1063
1064 #define CORE_FEATURE(major, minor, feature) \
1065 features->feature = core_##major##_##minor.feature
1066
1067
1068 vk_foreach_struct(ext, pFeatures->pNext) {
1069 switch (ext->sType) {
1070 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
1071 VkPhysicalDevice8BitStorageFeaturesKHR *features =
1072 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
1073 CORE_FEATURE(1, 2, storageBuffer8BitAccess);
1074 CORE_FEATURE(1, 2, uniformAndStorageBuffer8BitAccess);
1075 CORE_FEATURE(1, 2, storagePushConstant8);
1076 break;
1077 }
1078
1079 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
1080 VkPhysicalDevice16BitStorageFeatures *features =
1081 (VkPhysicalDevice16BitStorageFeatures *)ext;
1082 CORE_FEATURE(1, 1, storageBuffer16BitAccess);
1083 CORE_FEATURE(1, 1, uniformAndStorageBuffer16BitAccess);
1084 CORE_FEATURE(1, 1, storagePushConstant16);
1085 CORE_FEATURE(1, 1, storageInputOutput16);
1086 break;
1087 }
1088
1089 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
1090 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
1091 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1092 features->bufferDeviceAddressCaptureReplay = false;
1093 features->bufferDeviceAddressMultiDevice = false;
1094 break;
1095 }
1096
1097 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR: {
1098 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR *features = (void *)ext;
1099 CORE_FEATURE(1, 2, bufferDeviceAddress);
1100 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay);
1101 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice);
1102 break;
1103 }
1104
1105 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1106 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1107 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1108 features->computeDerivativeGroupQuads = true;
1109 features->computeDerivativeGroupLinear = true;
1110 break;
1111 }
1112
1113 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1114 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1115 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1116 features->conditionalRendering = pdevice->info.gen >= 8 ||
1117 pdevice->info.is_haswell;
1118 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1119 pdevice->info.is_haswell;
1120 break;
1121 }
1122
1123 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1124 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1125 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1126 features->depthClipEnable = true;
1127 break;
1128 }
1129
1130 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1131 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1132 CORE_FEATURE(1, 2, shaderFloat16);
1133 CORE_FEATURE(1, 2, shaderInt8);
1134 break;
1135 }
1136
1137 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1138 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1139 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1140 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1141 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1142 features->fragmentShaderShadingRateInterlock = false;
1143 break;
1144 }
1145
1146 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1147 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1148 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1149 CORE_FEATURE(1, 2, hostQueryReset);
1150 break;
1151 }
1152
1153 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1154 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1155 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1156 CORE_FEATURE(1, 2, shaderInputAttachmentArrayDynamicIndexing);
1157 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayDynamicIndexing);
1158 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayDynamicIndexing);
1159 CORE_FEATURE(1, 2, shaderUniformBufferArrayNonUniformIndexing);
1160 CORE_FEATURE(1, 2, shaderSampledImageArrayNonUniformIndexing);
1161 CORE_FEATURE(1, 2, shaderStorageBufferArrayNonUniformIndexing);
1162 CORE_FEATURE(1, 2, shaderStorageImageArrayNonUniformIndexing);
1163 CORE_FEATURE(1, 2, shaderInputAttachmentArrayNonUniformIndexing);
1164 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayNonUniformIndexing);
1165 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayNonUniformIndexing);
1166 CORE_FEATURE(1, 2, descriptorBindingUniformBufferUpdateAfterBind);
1167 CORE_FEATURE(1, 2, descriptorBindingSampledImageUpdateAfterBind);
1168 CORE_FEATURE(1, 2, descriptorBindingStorageImageUpdateAfterBind);
1169 CORE_FEATURE(1, 2, descriptorBindingStorageBufferUpdateAfterBind);
1170 CORE_FEATURE(1, 2, descriptorBindingUniformTexelBufferUpdateAfterBind);
1171 CORE_FEATURE(1, 2, descriptorBindingStorageTexelBufferUpdateAfterBind);
1172 CORE_FEATURE(1, 2, descriptorBindingUpdateUnusedWhilePending);
1173 CORE_FEATURE(1, 2, descriptorBindingPartiallyBound);
1174 CORE_FEATURE(1, 2, descriptorBindingVariableDescriptorCount);
1175 CORE_FEATURE(1, 2, runtimeDescriptorArray);
1176 break;
1177 }
1178
1179 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1180 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1181 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1182 features->indexTypeUint8 = true;
1183 break;
1184 }
1185
1186 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1187 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1188 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1189 features->inlineUniformBlock = true;
1190 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1191 break;
1192 }
1193
1194 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1195 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1196 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1197 features->rectangularLines = true;
1198 features->bresenhamLines = true;
1199 /* Support for Smooth lines with MSAA was removed on gen11. From the
1200 * BSpec section "Multisample ModesState" table for "AA Line Support
1201 * Requirements":
1202 *
1203 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1204 *
1205 * Fortunately, this isn't a case most people care about.
1206 */
1207 features->smoothLines = pdevice->info.gen < 10;
1208 features->stippledRectangularLines = false;
1209 features->stippledBresenhamLines = true;
1210 features->stippledSmoothLines = false;
1211 break;
1212 }
1213
1214 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1215 VkPhysicalDeviceMultiviewFeatures *features =
1216 (VkPhysicalDeviceMultiviewFeatures *)ext;
1217 CORE_FEATURE(1, 1, multiview);
1218 CORE_FEATURE(1, 1, multiviewGeometryShader);
1219 CORE_FEATURE(1, 1, multiviewTessellationShader);
1220 break;
1221 }
1222
1223 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1224 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1225 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1226 CORE_FEATURE(1, 2, imagelessFramebuffer);
1227 break;
1228 }
1229
1230 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1231 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1232 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1233 features->pipelineExecutableInfo = true;
1234 break;
1235 }
1236
1237 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1238 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1239 CORE_FEATURE(1, 1, protectedMemory);
1240 break;
1241 }
1242
1243 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1244 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1245 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1246 CORE_FEATURE(1, 1, samplerYcbcrConversion);
1247 break;
1248 }
1249
1250 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1251 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1252 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1253 CORE_FEATURE(1, 2, scalarBlockLayout);
1254 break;
1255 }
1256
1257 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR: {
1258 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *features =
1259 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *)ext;
1260 CORE_FEATURE(1, 2, separateDepthStencilLayouts);
1261 break;
1262 }
1263
1264 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1265 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1266 CORE_FEATURE(1, 2, shaderBufferInt64Atomics);
1267 CORE_FEATURE(1, 2, shaderSharedInt64Atomics);
1268 break;
1269 }
1270
1271 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1272 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1273 features->shaderDemoteToHelperInvocation = true;
1274 break;
1275 }
1276
1277 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1278 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1279 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1280 features->shaderSubgroupClock = true;
1281 features->shaderDeviceClock = false;
1282 break;
1283 }
1284
1285 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1286 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1287 CORE_FEATURE(1, 1, shaderDrawParameters);
1288 break;
1289 }
1290
1291 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
1292 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
1293 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
1294 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes);
1295 break;
1296 }
1297
1298 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1299 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1300 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1301 features->subgroupSizeControl = true;
1302 features->computeFullSubgroups = true;
1303 break;
1304 }
1305
1306 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1307 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1308 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1309 features->texelBufferAlignment = true;
1310 break;
1311 }
1312
1313 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR: {
1314 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *features =
1315 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *) ext;
1316 CORE_FEATURE(1, 2, timelineSemaphore);
1317 break;
1318 }
1319
1320 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1321 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1322 CORE_FEATURE(1, 1, variablePointersStorageBuffer);
1323 CORE_FEATURE(1, 1, variablePointers);
1324 break;
1325 }
1326
1327 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1328 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1329 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1330 features->transformFeedback = true;
1331 features->geometryStreams = true;
1332 break;
1333 }
1334
1335 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1336 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1337 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1338 CORE_FEATURE(1, 2, uniformBufferStandardLayout);
1339 break;
1340 }
1341
1342 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1343 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1344 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1345 features->vertexAttributeInstanceRateDivisor = true;
1346 features->vertexAttributeInstanceRateZeroDivisor = true;
1347 break;
1348 }
1349
1350 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES:
1351 anv_get_physical_device_features_1_1(pdevice, (void *)ext);
1352 break;
1353
1354 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES:
1355 anv_get_physical_device_features_1_2(pdevice, (void *)ext);
1356 break;
1357
1358 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
1359 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
1360 CORE_FEATURE(1, 2, vulkanMemoryModel);
1361 CORE_FEATURE(1, 2, vulkanMemoryModelDeviceScope);
1362 CORE_FEATURE(1, 2, vulkanMemoryModelAvailabilityVisibilityChains);
1363 break;
1364 }
1365
1366 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1367 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1368 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1369 features->ycbcrImageArrays = true;
1370 break;
1371 }
1372
1373 default:
1374 anv_debug_ignored_stype(ext->sType);
1375 break;
1376 }
1377 }
1378
1379 #undef CORE_FEATURE
1380 }
1381
1382 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1383
1384 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1385 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1386
1387 void anv_GetPhysicalDeviceProperties(
1388 VkPhysicalDevice physicalDevice,
1389 VkPhysicalDeviceProperties* pProperties)
1390 {
1391 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1392 const struct gen_device_info *devinfo = &pdevice->info;
1393
1394 /* See assertions made when programming the buffer surface state. */
1395 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1396 (1ul << 30) : (1ul << 27);
1397
1398 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1399 const uint32_t max_textures =
1400 pdevice->has_bindless_images ? UINT16_MAX : 128;
1401 const uint32_t max_samplers =
1402 pdevice->has_bindless_samplers ? UINT16_MAX :
1403 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1404 const uint32_t max_images =
1405 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1406
1407 /* If we can use bindless for everything, claim a high per-stage limit,
1408 * otherwise use the binding table size, minus the slots reserved for
1409 * render targets and one slot for the descriptor buffer. */
1410 const uint32_t max_per_stage =
1411 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1412 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1413
1414 const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
1415
1416 VkSampleCountFlags sample_counts =
1417 isl_device_get_sample_counts(&pdevice->isl_dev);
1418
1419
1420 VkPhysicalDeviceLimits limits = {
1421 .maxImageDimension1D = (1 << 14),
1422 .maxImageDimension2D = (1 << 14),
1423 .maxImageDimension3D = (1 << 11),
1424 .maxImageDimensionCube = (1 << 14),
1425 .maxImageArrayLayers = (1 << 11),
1426 .maxTexelBufferElements = 128 * 1024 * 1024,
1427 .maxUniformBufferRange = (1ul << 27),
1428 .maxStorageBufferRange = max_raw_buffer_sz,
1429 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1430 .maxMemoryAllocationCount = UINT32_MAX,
1431 .maxSamplerAllocationCount = 64 * 1024,
1432 .bufferImageGranularity = 64, /* A cache line */
1433 .sparseAddressSpaceSize = 0,
1434 .maxBoundDescriptorSets = MAX_SETS,
1435 .maxPerStageDescriptorSamplers = max_samplers,
1436 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1437 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1438 .maxPerStageDescriptorSampledImages = max_textures,
1439 .maxPerStageDescriptorStorageImages = max_images,
1440 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1441 .maxPerStageResources = max_per_stage,
1442 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1443 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1444 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1445 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1446 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1447 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1448 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1449 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1450 .maxVertexInputAttributes = MAX_VBS,
1451 .maxVertexInputBindings = MAX_VBS,
1452 .maxVertexInputAttributeOffset = 2047,
1453 .maxVertexInputBindingStride = 2048,
1454 .maxVertexOutputComponents = 128,
1455 .maxTessellationGenerationLevel = 64,
1456 .maxTessellationPatchSize = 32,
1457 .maxTessellationControlPerVertexInputComponents = 128,
1458 .maxTessellationControlPerVertexOutputComponents = 128,
1459 .maxTessellationControlPerPatchOutputComponents = 128,
1460 .maxTessellationControlTotalOutputComponents = 2048,
1461 .maxTessellationEvaluationInputComponents = 128,
1462 .maxTessellationEvaluationOutputComponents = 128,
1463 .maxGeometryShaderInvocations = 32,
1464 .maxGeometryInputComponents = 64,
1465 .maxGeometryOutputComponents = 128,
1466 .maxGeometryOutputVertices = 256,
1467 .maxGeometryTotalOutputComponents = 1024,
1468 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1469 .maxFragmentOutputAttachments = 8,
1470 .maxFragmentDualSrcAttachments = 1,
1471 .maxFragmentCombinedOutputResources = 8,
1472 .maxComputeSharedMemorySize = 64 * 1024,
1473 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1474 .maxComputeWorkGroupInvocations = max_workgroup_size,
1475 .maxComputeWorkGroupSize = {
1476 max_workgroup_size,
1477 max_workgroup_size,
1478 max_workgroup_size,
1479 },
1480 .subPixelPrecisionBits = 8,
1481 .subTexelPrecisionBits = 8,
1482 .mipmapPrecisionBits = 8,
1483 .maxDrawIndexedIndexValue = UINT32_MAX,
1484 .maxDrawIndirectCount = UINT32_MAX,
1485 .maxSamplerLodBias = 16,
1486 .maxSamplerAnisotropy = 16,
1487 .maxViewports = MAX_VIEWPORTS,
1488 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1489 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1490 .viewportSubPixelBits = 13, /* We take a float? */
1491 .minMemoryMapAlignment = 4096, /* A page */
1492 /* The dataport requires texel alignment so we need to assume a worst
1493 * case of R32G32B32A32 which is 16 bytes.
1494 */
1495 .minTexelBufferOffsetAlignment = 16,
1496 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1497 .minUniformBufferOffsetAlignment = 32,
1498 .minStorageBufferOffsetAlignment = 4,
1499 .minTexelOffset = -8,
1500 .maxTexelOffset = 7,
1501 .minTexelGatherOffset = -32,
1502 .maxTexelGatherOffset = 31,
1503 .minInterpolationOffset = -0.5,
1504 .maxInterpolationOffset = 0.4375,
1505 .subPixelInterpolationOffsetBits = 4,
1506 .maxFramebufferWidth = (1 << 14),
1507 .maxFramebufferHeight = (1 << 14),
1508 .maxFramebufferLayers = (1 << 11),
1509 .framebufferColorSampleCounts = sample_counts,
1510 .framebufferDepthSampleCounts = sample_counts,
1511 .framebufferStencilSampleCounts = sample_counts,
1512 .framebufferNoAttachmentsSampleCounts = sample_counts,
1513 .maxColorAttachments = MAX_RTS,
1514 .sampledImageColorSampleCounts = sample_counts,
1515 .sampledImageIntegerSampleCounts = sample_counts,
1516 .sampledImageDepthSampleCounts = sample_counts,
1517 .sampledImageStencilSampleCounts = sample_counts,
1518 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1519 .maxSampleMaskWords = 1,
1520 .timestampComputeAndGraphics = true,
1521 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1522 .maxClipDistances = 8,
1523 .maxCullDistances = 8,
1524 .maxCombinedClipAndCullDistances = 8,
1525 .discreteQueuePriorities = 2,
1526 .pointSizeRange = { 0.125, 255.875 },
1527 .lineWidthRange = {
1528 0.0,
1529 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1530 2047.9921875 : 7.9921875,
1531 },
1532 .pointSizeGranularity = (1.0 / 8.0),
1533 .lineWidthGranularity = (1.0 / 128.0),
1534 .strictLines = false,
1535 .standardSampleLocations = true,
1536 .optimalBufferCopyOffsetAlignment = 128,
1537 .optimalBufferCopyRowPitchAlignment = 128,
1538 .nonCoherentAtomSize = 64,
1539 };
1540
1541 *pProperties = (VkPhysicalDeviceProperties) {
1542 .apiVersion = anv_physical_device_api_version(pdevice),
1543 .driverVersion = vk_get_driver_version(),
1544 .vendorID = 0x8086,
1545 .deviceID = pdevice->chipset_id,
1546 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1547 .limits = limits,
1548 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1549 };
1550
1551 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1552 "%s", pdevice->name);
1553 memcpy(pProperties->pipelineCacheUUID,
1554 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1555 }
1556
1557 static void
1558 anv_get_physical_device_properties_1_1(struct anv_physical_device *pdevice,
1559 VkPhysicalDeviceVulkan11Properties *p)
1560 {
1561 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES);
1562
1563 memcpy(p->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1564 memcpy(p->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1565 memset(p->deviceLUID, 0, VK_LUID_SIZE);
1566 p->deviceNodeMask = 0;
1567 p->deviceLUIDValid = false;
1568
1569 p->subgroupSize = BRW_SUBGROUP_SIZE;
1570 VkShaderStageFlags scalar_stages = 0;
1571 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1572 if (pdevice->compiler->scalar_stage[stage])
1573 scalar_stages |= mesa_to_vk_shader_stage(stage);
1574 }
1575 p->subgroupSupportedStages = scalar_stages;
1576 p->subgroupSupportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1577 VK_SUBGROUP_FEATURE_VOTE_BIT |
1578 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1579 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1580 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1581 VK_SUBGROUP_FEATURE_QUAD_BIT;
1582 if (pdevice->info.gen >= 8) {
1583 /* TODO: There's no technical reason why these can't be made to
1584 * work on gen7 but they don't at the moment so it's best to leave
1585 * the feature disabled than enabled and broken.
1586 */
1587 p->subgroupSupportedOperations |= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1588 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1589 }
1590 p->subgroupQuadOperationsInAllStages = pdevice->info.gen >= 8;
1591
1592 p->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1593 p->maxMultiviewViewCount = 16;
1594 p->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1595 p->protectedNoFault = false;
1596 /* This value doesn't matter for us today as our per-stage descriptors are
1597 * the real limit.
1598 */
1599 p->maxPerSetDescriptors = 1024;
1600 p->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1601 }
1602
1603 static void
1604 anv_get_physical_device_properties_1_2(struct anv_physical_device *pdevice,
1605 VkPhysicalDeviceVulkan12Properties *p)
1606 {
1607 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES);
1608
1609 p->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1610 memset(p->driverName, 0, sizeof(p->driverName));
1611 snprintf(p->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1612 "Intel open-source Mesa driver");
1613 memset(p->driverInfo, 0, sizeof(p->driverInfo));
1614 snprintf(p->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1615 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1616 p->conformanceVersion = (VkConformanceVersionKHR) {
1617 .major = 1,
1618 .minor = 2,
1619 .subminor = 0,
1620 .patch = 0,
1621 };
1622
1623 p->denormBehaviorIndependence =
1624 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
1625 p->roundingModeIndependence =
1626 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
1627
1628 /* Broadwell does not support HF denorms and there are restrictions
1629 * other gens. According to Kabylake's PRM:
1630 *
1631 * "math - Extended Math Function
1632 * [...]
1633 * Restriction : Half-float denorms are always retained."
1634 */
1635 p->shaderDenormFlushToZeroFloat16 = false;
1636 p->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
1637 p->shaderRoundingModeRTEFloat16 = true;
1638 p->shaderRoundingModeRTZFloat16 = true;
1639 p->shaderSignedZeroInfNanPreserveFloat16 = true;
1640
1641 p->shaderDenormFlushToZeroFloat32 = true;
1642 p->shaderDenormPreserveFloat32 = true;
1643 p->shaderRoundingModeRTEFloat32 = true;
1644 p->shaderRoundingModeRTZFloat32 = true;
1645 p->shaderSignedZeroInfNanPreserveFloat32 = true;
1646
1647 p->shaderDenormFlushToZeroFloat64 = true;
1648 p->shaderDenormPreserveFloat64 = true;
1649 p->shaderRoundingModeRTEFloat64 = true;
1650 p->shaderRoundingModeRTZFloat64 = true;
1651 p->shaderSignedZeroInfNanPreserveFloat64 = true;
1652
1653 /* It's a bit hard to exactly map our implementation to the limits
1654 * described here. The bindless surface handle in the extended
1655 * message descriptors is 20 bits and it's an index into the table of
1656 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1657 * address. Given that most things consume two surface states per
1658 * view (general/sampled for textures and write-only/read-write for
1659 * images), we claim 2^19 things.
1660 *
1661 * For SSBOs, we just use A64 messages so there is no real limit
1662 * there beyond the limit on the total size of a descriptor set.
1663 */
1664 const unsigned max_bindless_views = 1 << 19;
1665 p->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1666 p->shaderUniformBufferArrayNonUniformIndexingNative = false;
1667 p->shaderSampledImageArrayNonUniformIndexingNative = false;
1668 p->shaderStorageBufferArrayNonUniformIndexingNative = true;
1669 p->shaderStorageImageArrayNonUniformIndexingNative = false;
1670 p->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1671 p->robustBufferAccessUpdateAfterBind = true;
1672 p->quadDivergentImplicitLod = false;
1673 p->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1674 p->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1675 p->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1676 p->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1677 p->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1678 p->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1679 p->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1680 p->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1681 p->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1682 p->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1683 p->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1684 p->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1685 p->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1686 p->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1687 p->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1688
1689 /* We support all of the depth resolve modes */
1690 p->supportedDepthResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1691 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1692 VK_RESOLVE_MODE_MIN_BIT_KHR |
1693 VK_RESOLVE_MODE_MAX_BIT_KHR;
1694 /* Average doesn't make sense for stencil so we don't support that */
1695 p->supportedStencilResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1696 if (pdevice->info.gen >= 8) {
1697 /* The advanced stencil resolve modes currently require stencil
1698 * sampling be supported by the hardware.
1699 */
1700 p->supportedStencilResolveModes |= VK_RESOLVE_MODE_MIN_BIT_KHR |
1701 VK_RESOLVE_MODE_MAX_BIT_KHR;
1702 }
1703 p->independentResolveNone = true;
1704 p->independentResolve = true;
1705
1706 p->filterMinmaxSingleComponentFormats = pdevice->info.gen >= 9;
1707 p->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1708
1709 p->maxTimelineSemaphoreValueDifference = UINT64_MAX;
1710
1711 p->framebufferIntegerColorSampleCounts =
1712 isl_device_get_sample_counts(&pdevice->isl_dev);
1713 }
1714
1715 void anv_GetPhysicalDeviceProperties2(
1716 VkPhysicalDevice physicalDevice,
1717 VkPhysicalDeviceProperties2* pProperties)
1718 {
1719 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1720
1721 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1722
1723 VkPhysicalDeviceVulkan11Properties core_1_1 = {
1724 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES,
1725 };
1726 anv_get_physical_device_properties_1_1(pdevice, &core_1_1);
1727
1728 VkPhysicalDeviceVulkan12Properties core_1_2 = {
1729 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES,
1730 };
1731 anv_get_physical_device_properties_1_2(pdevice, &core_1_2);
1732
1733 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1734 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1735 sizeof(core_##major##_##minor.core_property))
1736
1737 #define CORE_PROPERTY(major, minor, property) \
1738 CORE_RENAMED_PROPERTY(major, minor, property, property)
1739
1740 vk_foreach_struct(ext, pProperties->pNext) {
1741 switch (ext->sType) {
1742 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1743 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *properties =
1744 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1745 CORE_PROPERTY(1, 2, supportedDepthResolveModes);
1746 CORE_PROPERTY(1, 2, supportedStencilResolveModes);
1747 CORE_PROPERTY(1, 2, independentResolveNone);
1748 CORE_PROPERTY(1, 2, independentResolve);
1749 break;
1750 }
1751
1752 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1753 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *properties =
1754 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1755 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools);
1756 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative);
1757 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative);
1758 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative);
1759 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative);
1760 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative);
1761 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind);
1762 CORE_PROPERTY(1, 2, quadDivergentImplicitLod);
1763 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers);
1764 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers);
1765 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers);
1766 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages);
1767 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages);
1768 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments);
1769 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources);
1770 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers);
1771 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers);
1772 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic);
1773 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers);
1774 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic);
1775 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages);
1776 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages);
1777 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments);
1778 break;
1779 }
1780
1781 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1782 VkPhysicalDeviceDriverPropertiesKHR *properties =
1783 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1784 CORE_PROPERTY(1, 2, driverID);
1785 CORE_PROPERTY(1, 2, driverName);
1786 CORE_PROPERTY(1, 2, driverInfo);
1787 CORE_PROPERTY(1, 2, conformanceVersion);
1788 break;
1789 }
1790
1791 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1792 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1793 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1794 /* Userptr needs page aligned memory. */
1795 props->minImportedHostPointerAlignment = 4096;
1796 break;
1797 }
1798
1799 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1800 VkPhysicalDeviceIDProperties *properties =
1801 (VkPhysicalDeviceIDProperties *)ext;
1802 CORE_PROPERTY(1, 1, deviceUUID);
1803 CORE_PROPERTY(1, 1, driverUUID);
1804 CORE_PROPERTY(1, 1, deviceLUID);
1805 CORE_PROPERTY(1, 1, deviceLUIDValid);
1806 break;
1807 }
1808
1809 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1810 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1811 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1812 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1813 props->maxPerStageDescriptorInlineUniformBlocks =
1814 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1815 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1816 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1817 props->maxDescriptorSetInlineUniformBlocks =
1818 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1819 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1820 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1821 break;
1822 }
1823
1824 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1825 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1826 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1827 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1828 * Sampling Rules - Legacy Mode", it says the following:
1829 *
1830 * "Note that the device divides a pixel into a 16x16 array of
1831 * subpixels, referenced by their upper left corners."
1832 *
1833 * This is the only known reference in the PRMs to the subpixel
1834 * precision of line rasterization and a "16x16 array of subpixels"
1835 * implies 4 subpixel precision bits. Empirical testing has shown
1836 * that 4 subpixel precision bits applies to all line rasterization
1837 * types.
1838 */
1839 props->lineSubPixelPrecisionBits = 4;
1840 break;
1841 }
1842
1843 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1844 VkPhysicalDeviceMaintenance3Properties *properties =
1845 (VkPhysicalDeviceMaintenance3Properties *)ext;
1846 /* This value doesn't matter for us today as our per-stage
1847 * descriptors are the real limit.
1848 */
1849 CORE_PROPERTY(1, 1, maxPerSetDescriptors);
1850 CORE_PROPERTY(1, 1, maxMemoryAllocationSize);
1851 break;
1852 }
1853
1854 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1855 VkPhysicalDeviceMultiviewProperties *properties =
1856 (VkPhysicalDeviceMultiviewProperties *)ext;
1857 CORE_PROPERTY(1, 1, maxMultiviewViewCount);
1858 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex);
1859 break;
1860 }
1861
1862 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1863 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1864 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1865 properties->pciDomain = pdevice->pci_info.domain;
1866 properties->pciBus = pdevice->pci_info.bus;
1867 properties->pciDevice = pdevice->pci_info.device;
1868 properties->pciFunction = pdevice->pci_info.function;
1869 break;
1870 }
1871
1872 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1873 VkPhysicalDevicePointClippingProperties *properties =
1874 (VkPhysicalDevicePointClippingProperties *) ext;
1875 CORE_PROPERTY(1, 1, pointClippingBehavior);
1876 break;
1877 }
1878
1879 #pragma GCC diagnostic push
1880 #pragma GCC diagnostic ignored "-Wswitch"
1881 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1882 VkPhysicalDevicePresentationPropertiesANDROID *props =
1883 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1884 props->sharedImage = VK_FALSE;
1885 break;
1886 }
1887 #pragma GCC diagnostic pop
1888
1889 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1890 VkPhysicalDeviceProtectedMemoryProperties *properties =
1891 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1892 CORE_PROPERTY(1, 1, protectedNoFault);
1893 break;
1894 }
1895
1896 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1897 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1898 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1899 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1900 break;
1901 }
1902
1903 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1904 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1905 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1906 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping);
1907 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats);
1908 break;
1909 }
1910
1911 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1912 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1913 CORE_PROPERTY(1, 1, subgroupSize);
1914 CORE_RENAMED_PROPERTY(1, 1, supportedStages,
1915 subgroupSupportedStages);
1916 CORE_RENAMED_PROPERTY(1, 1, supportedOperations,
1917 subgroupSupportedOperations);
1918 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages,
1919 subgroupQuadOperationsInAllStages);
1920 break;
1921 }
1922
1923 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1924 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1925 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1926 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1927 props->minSubgroupSize = 8;
1928 props->maxSubgroupSize = 32;
1929 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1930 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1931 break;
1932 }
1933 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
1934 VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
1935 CORE_PROPERTY(1, 2, denormBehaviorIndependence);
1936 CORE_PROPERTY(1, 2, roundingModeIndependence);
1937 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16);
1938 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16);
1939 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16);
1940 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16);
1941 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16);
1942 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32);
1943 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32);
1944 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32);
1945 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32);
1946 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32);
1947 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64);
1948 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64);
1949 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64);
1950 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64);
1951 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64);
1952 break;
1953 }
1954
1955 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1956 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1957 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1958
1959 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1960 * Base Address:
1961 *
1962 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1963 * specifies the base address of the first element of the surface,
1964 * computed in software by adding the surface base address to the
1965 * byte offset of the element in the buffer. The base address must
1966 * be aligned to element size."
1967 *
1968 * The typed dataport messages require that things be texel aligned.
1969 * Otherwise, we may just load/store the wrong data or, in the worst
1970 * case, there may be hangs.
1971 */
1972 props->storageTexelBufferOffsetAlignmentBytes = 16;
1973 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1974
1975 /* The sampler, however, is much more forgiving and it can handle
1976 * arbitrary byte alignment for linear and buffer surfaces. It's
1977 * hard to find a good PRM citation for this but years of empirical
1978 * experience demonstrate that this is true.
1979 */
1980 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1981 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1982 break;
1983 }
1984
1985 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR: {
1986 VkPhysicalDeviceTimelineSemaphorePropertiesKHR *properties =
1987 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR *) ext;
1988 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference);
1989 break;
1990 }
1991
1992 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1993 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1994 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1995
1996 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1997 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1998 props->maxTransformFeedbackBufferSize = (1ull << 32);
1999 props->maxTransformFeedbackStreamDataSize = 128 * 4;
2000 props->maxTransformFeedbackBufferDataSize = 128 * 4;
2001 props->maxTransformFeedbackBufferDataStride = 2048;
2002 props->transformFeedbackQueries = true;
2003 props->transformFeedbackStreamsLinesTriangles = false;
2004 props->transformFeedbackRasterizationStreamSelect = false;
2005 props->transformFeedbackDraw = true;
2006 break;
2007 }
2008
2009 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
2010 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
2011 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
2012 /* We have to restrict this a bit for multiview */
2013 props->maxVertexAttribDivisor = UINT32_MAX / 16;
2014 break;
2015 }
2016
2017 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES:
2018 anv_get_physical_device_properties_1_1(pdevice, (void *)ext);
2019 break;
2020
2021 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES:
2022 anv_get_physical_device_properties_1_2(pdevice, (void *)ext);
2023 break;
2024
2025 default:
2026 anv_debug_ignored_stype(ext->sType);
2027 break;
2028 }
2029 }
2030
2031 #undef CORE_RENAMED_PROPERTY
2032 #undef CORE_PROPERTY
2033 }
2034
2035 /* We support exactly one queue family. */
2036 static const VkQueueFamilyProperties
2037 anv_queue_family_properties = {
2038 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
2039 VK_QUEUE_COMPUTE_BIT |
2040 VK_QUEUE_TRANSFER_BIT,
2041 .queueCount = 1,
2042 .timestampValidBits = 36, /* XXX: Real value here */
2043 .minImageTransferGranularity = { 1, 1, 1 },
2044 };
2045
2046 void anv_GetPhysicalDeviceQueueFamilyProperties(
2047 VkPhysicalDevice physicalDevice,
2048 uint32_t* pCount,
2049 VkQueueFamilyProperties* pQueueFamilyProperties)
2050 {
2051 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
2052
2053 vk_outarray_append(&out, p) {
2054 *p = anv_queue_family_properties;
2055 }
2056 }
2057
2058 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2059 VkPhysicalDevice physicalDevice,
2060 uint32_t* pQueueFamilyPropertyCount,
2061 VkQueueFamilyProperties2* pQueueFamilyProperties)
2062 {
2063
2064 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
2065
2066 vk_outarray_append(&out, p) {
2067 p->queueFamilyProperties = anv_queue_family_properties;
2068
2069 vk_foreach_struct(s, p->pNext) {
2070 anv_debug_ignored_stype(s->sType);
2071 }
2072 }
2073 }
2074
2075 void anv_GetPhysicalDeviceMemoryProperties(
2076 VkPhysicalDevice physicalDevice,
2077 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
2078 {
2079 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2080
2081 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
2082 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
2083 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
2084 .propertyFlags = physical_device->memory.types[i].propertyFlags,
2085 .heapIndex = physical_device->memory.types[i].heapIndex,
2086 };
2087 }
2088
2089 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
2090 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
2091 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
2092 .size = physical_device->memory.heaps[i].size,
2093 .flags = physical_device->memory.heaps[i].flags,
2094 };
2095 }
2096 }
2097
2098 static void
2099 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
2100 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
2101 {
2102 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2103 uint64_t sys_available = get_available_system_memory();
2104 assert(sys_available > 0);
2105
2106 VkDeviceSize total_heaps_size = 0;
2107 for (size_t i = 0; i < device->memory.heap_count; i++)
2108 total_heaps_size += device->memory.heaps[i].size;
2109
2110 for (size_t i = 0; i < device->memory.heap_count; i++) {
2111 VkDeviceSize heap_size = device->memory.heaps[i].size;
2112 VkDeviceSize heap_used = device->memory.heaps[i].used;
2113 VkDeviceSize heap_budget;
2114
2115 double heap_proportion = (double) heap_size / total_heaps_size;
2116 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
2117
2118 /*
2119 * Let's not incite the app to starve the system: report at most 90% of
2120 * available system memory.
2121 */
2122 uint64_t heap_available = sys_available_prop * 9 / 10;
2123 heap_budget = MIN2(heap_size, heap_used + heap_available);
2124
2125 /*
2126 * Round down to the nearest MB
2127 */
2128 heap_budget &= ~((1ull << 20) - 1);
2129
2130 /*
2131 * The heapBudget value must be non-zero for array elements less than
2132 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2133 * value must be less than or equal to VkMemoryHeap::size for each heap.
2134 */
2135 assert(0 < heap_budget && heap_budget <= heap_size);
2136
2137 memoryBudget->heapUsage[i] = heap_used;
2138 memoryBudget->heapBudget[i] = heap_budget;
2139 }
2140
2141 /* The heapBudget and heapUsage values must be zero for array elements
2142 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2143 */
2144 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
2145 memoryBudget->heapBudget[i] = 0;
2146 memoryBudget->heapUsage[i] = 0;
2147 }
2148 }
2149
2150 void anv_GetPhysicalDeviceMemoryProperties2(
2151 VkPhysicalDevice physicalDevice,
2152 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
2153 {
2154 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
2155 &pMemoryProperties->memoryProperties);
2156
2157 vk_foreach_struct(ext, pMemoryProperties->pNext) {
2158 switch (ext->sType) {
2159 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
2160 anv_get_memory_budget(physicalDevice, (void*)ext);
2161 break;
2162 default:
2163 anv_debug_ignored_stype(ext->sType);
2164 break;
2165 }
2166 }
2167 }
2168
2169 void
2170 anv_GetDeviceGroupPeerMemoryFeatures(
2171 VkDevice device,
2172 uint32_t heapIndex,
2173 uint32_t localDeviceIndex,
2174 uint32_t remoteDeviceIndex,
2175 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
2176 {
2177 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
2178 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2179 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2180 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2181 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2182 }
2183
2184 PFN_vkVoidFunction anv_GetInstanceProcAddr(
2185 VkInstance _instance,
2186 const char* pName)
2187 {
2188 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2189
2190 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2191 * when we have to return valid function pointers, NULL, or it's left
2192 * undefined. See the table for exact details.
2193 */
2194 if (pName == NULL)
2195 return NULL;
2196
2197 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2198 if (strcmp(pName, "vk" #entrypoint) == 0) \
2199 return (PFN_vkVoidFunction)anv_##entrypoint
2200
2201 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
2202 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
2203 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
2204 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
2205
2206 #undef LOOKUP_ANV_ENTRYPOINT
2207
2208 if (instance == NULL)
2209 return NULL;
2210
2211 int idx = anv_get_instance_entrypoint_index(pName);
2212 if (idx >= 0)
2213 return instance->dispatch.entrypoints[idx];
2214
2215 idx = anv_get_physical_device_entrypoint_index(pName);
2216 if (idx >= 0)
2217 return instance->physicalDevice.dispatch.entrypoints[idx];
2218
2219 idx = anv_get_device_entrypoint_index(pName);
2220 if (idx >= 0)
2221 return instance->device_dispatch.entrypoints[idx];
2222
2223 return NULL;
2224 }
2225
2226 /* With version 1+ of the loader interface the ICD should expose
2227 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2228 */
2229 PUBLIC
2230 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2231 VkInstance instance,
2232 const char* pName);
2233
2234 PUBLIC
2235 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2236 VkInstance instance,
2237 const char* pName)
2238 {
2239 return anv_GetInstanceProcAddr(instance, pName);
2240 }
2241
2242 PFN_vkVoidFunction anv_GetDeviceProcAddr(
2243 VkDevice _device,
2244 const char* pName)
2245 {
2246 ANV_FROM_HANDLE(anv_device, device, _device);
2247
2248 if (!device || !pName)
2249 return NULL;
2250
2251 int idx = anv_get_device_entrypoint_index(pName);
2252 if (idx < 0)
2253 return NULL;
2254
2255 return device->dispatch.entrypoints[idx];
2256 }
2257
2258 /* With version 4+ of the loader interface the ICD should expose
2259 * vk_icdGetPhysicalDeviceProcAddr()
2260 */
2261 PUBLIC
2262 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2263 VkInstance _instance,
2264 const char* pName);
2265
2266 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2267 VkInstance _instance,
2268 const char* pName)
2269 {
2270 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2271
2272 if (!pName || !instance)
2273 return NULL;
2274
2275 int idx = anv_get_physical_device_entrypoint_index(pName);
2276 if (idx < 0)
2277 return NULL;
2278
2279 return instance->physicalDevice.dispatch.entrypoints[idx];
2280 }
2281
2282
2283 VkResult
2284 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
2285 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
2286 const VkAllocationCallbacks* pAllocator,
2287 VkDebugReportCallbackEXT* pCallback)
2288 {
2289 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2290 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2291 pCreateInfo, pAllocator, &instance->alloc,
2292 pCallback);
2293 }
2294
2295 void
2296 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
2297 VkDebugReportCallbackEXT _callback,
2298 const VkAllocationCallbacks* pAllocator)
2299 {
2300 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2301 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2302 _callback, pAllocator, &instance->alloc);
2303 }
2304
2305 void
2306 anv_DebugReportMessageEXT(VkInstance _instance,
2307 VkDebugReportFlagsEXT flags,
2308 VkDebugReportObjectTypeEXT objectType,
2309 uint64_t object,
2310 size_t location,
2311 int32_t messageCode,
2312 const char* pLayerPrefix,
2313 const char* pMessage)
2314 {
2315 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2316 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2317 object, location, messageCode, pLayerPrefix, pMessage);
2318 }
2319
2320 static struct anv_state
2321 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2322 {
2323 struct anv_state state;
2324
2325 state = anv_state_pool_alloc(pool, size, align);
2326 memcpy(state.map, p, size);
2327
2328 return state;
2329 }
2330
2331 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2332 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2333 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2334 * color as a separate entry /after/ the float color. The layout of this entry
2335 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2336 *
2337 * Since we don't know the format/bpp, we can't make any of the border colors
2338 * containing '1' work for all formats, as it would be in the wrong place for
2339 * some of them. We opt to make 32-bit integers work as this seems like the
2340 * most common option. Fortunately, transparent black works regardless, as
2341 * all zeroes is the same in every bit-size.
2342 */
2343 struct hsw_border_color {
2344 float float32[4];
2345 uint32_t _pad0[12];
2346 uint32_t uint32[4];
2347 uint32_t _pad1[108];
2348 };
2349
2350 struct gen8_border_color {
2351 union {
2352 float float32[4];
2353 uint32_t uint32[4];
2354 };
2355 /* Pad out to 64 bytes */
2356 uint32_t _pad[12];
2357 };
2358
2359 static void
2360 anv_device_init_border_colors(struct anv_device *device)
2361 {
2362 if (device->info.is_haswell) {
2363 static const struct hsw_border_color border_colors[] = {
2364 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2365 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2366 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2367 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2368 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2369 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2370 };
2371
2372 device->border_colors =
2373 anv_state_pool_emit_data(&device->dynamic_state_pool,
2374 sizeof(border_colors), 512, border_colors);
2375 } else {
2376 static const struct gen8_border_color border_colors[] = {
2377 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2378 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2379 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2380 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2381 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2382 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2383 };
2384
2385 device->border_colors =
2386 anv_state_pool_emit_data(&device->dynamic_state_pool,
2387 sizeof(border_colors), 64, border_colors);
2388 }
2389 }
2390
2391 static VkResult
2392 anv_device_init_trivial_batch(struct anv_device *device)
2393 {
2394 VkResult result = anv_device_alloc_bo(device, 4096,
2395 ANV_BO_ALLOC_MAPPED,
2396 0 /* explicit_address */,
2397 &device->trivial_batch_bo);
2398 if (result != VK_SUCCESS)
2399 return result;
2400
2401 struct anv_batch batch = {
2402 .start = device->trivial_batch_bo->map,
2403 .next = device->trivial_batch_bo->map,
2404 .end = device->trivial_batch_bo->map + 4096,
2405 };
2406
2407 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2408 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2409
2410 if (!device->info.has_llc)
2411 gen_clflush_range(batch.start, batch.next - batch.start);
2412
2413 return VK_SUCCESS;
2414 }
2415
2416 VkResult anv_EnumerateDeviceExtensionProperties(
2417 VkPhysicalDevice physicalDevice,
2418 const char* pLayerName,
2419 uint32_t* pPropertyCount,
2420 VkExtensionProperties* pProperties)
2421 {
2422 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2423 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2424
2425 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2426 if (device->supported_extensions.extensions[i]) {
2427 vk_outarray_append(&out, prop) {
2428 *prop = anv_device_extensions[i];
2429 }
2430 }
2431 }
2432
2433 return vk_outarray_status(&out);
2434 }
2435
2436 static void
2437 anv_device_init_dispatch(struct anv_device *device)
2438 {
2439 const struct anv_instance *instance = device->physical->instance;
2440
2441 const struct anv_device_dispatch_table *genX_table;
2442 switch (device->info.gen) {
2443 case 12:
2444 genX_table = &gen12_device_dispatch_table;
2445 break;
2446 case 11:
2447 genX_table = &gen11_device_dispatch_table;
2448 break;
2449 case 10:
2450 genX_table = &gen10_device_dispatch_table;
2451 break;
2452 case 9:
2453 genX_table = &gen9_device_dispatch_table;
2454 break;
2455 case 8:
2456 genX_table = &gen8_device_dispatch_table;
2457 break;
2458 case 7:
2459 if (device->info.is_haswell)
2460 genX_table = &gen75_device_dispatch_table;
2461 else
2462 genX_table = &gen7_device_dispatch_table;
2463 break;
2464 default:
2465 unreachable("unsupported gen\n");
2466 }
2467
2468 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2469 /* Vulkan requires that entrypoints for extensions which have not been
2470 * enabled must not be advertised.
2471 */
2472 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
2473 &instance->enabled_extensions,
2474 &device->enabled_extensions)) {
2475 device->dispatch.entrypoints[i] = NULL;
2476 } else if (genX_table->entrypoints[i]) {
2477 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2478 } else {
2479 device->dispatch.entrypoints[i] =
2480 anv_device_dispatch_table.entrypoints[i];
2481 }
2482 }
2483 }
2484
2485 static int
2486 vk_priority_to_gen(int priority)
2487 {
2488 switch (priority) {
2489 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2490 return GEN_CONTEXT_LOW_PRIORITY;
2491 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2492 return GEN_CONTEXT_MEDIUM_PRIORITY;
2493 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2494 return GEN_CONTEXT_HIGH_PRIORITY;
2495 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2496 return GEN_CONTEXT_REALTIME_PRIORITY;
2497 default:
2498 unreachable("Invalid priority");
2499 }
2500 }
2501
2502 static VkResult
2503 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2504 {
2505 VkResult result = anv_device_alloc_bo(device, 4096,
2506 ANV_BO_ALLOC_MAPPED,
2507 0 /* explicit_address */,
2508 &device->hiz_clear_bo);
2509 if (result != VK_SUCCESS)
2510 return result;
2511
2512 union isl_color_value hiz_clear = { .u32 = { 0, } };
2513 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2514
2515 memcpy(device->hiz_clear_bo->map, hiz_clear.u32, sizeof(hiz_clear.u32));
2516
2517 if (!device->info.has_llc)
2518 gen_clflush_range(device->hiz_clear_bo->map, sizeof(hiz_clear.u32));
2519
2520 return VK_SUCCESS;
2521 }
2522
2523 static bool
2524 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2525 struct anv_block_pool *pool,
2526 uint64_t address)
2527 {
2528 anv_block_pool_foreach_bo(bo, pool) {
2529 uint64_t bo_address = gen_48b_address(bo->offset);
2530 if (address >= bo_address && address < (bo_address + bo->size)) {
2531 *ret = (struct gen_batch_decode_bo) {
2532 .addr = bo_address,
2533 .size = bo->size,
2534 .map = bo->map,
2535 };
2536 return true;
2537 }
2538 }
2539 return false;
2540 }
2541
2542 /* Finding a buffer for batch decoding */
2543 static struct gen_batch_decode_bo
2544 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2545 {
2546 struct anv_device *device = v_batch;
2547 struct gen_batch_decode_bo ret_bo = {};
2548
2549 assert(ppgtt);
2550
2551 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2552 return ret_bo;
2553 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2554 return ret_bo;
2555 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2556 return ret_bo;
2557 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2558 return ret_bo;
2559
2560 if (!device->cmd_buffer_being_decoded)
2561 return (struct gen_batch_decode_bo) { };
2562
2563 struct anv_batch_bo **bo;
2564
2565 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2566 /* The decoder zeroes out the top 16 bits, so we need to as well */
2567 uint64_t bo_address = (*bo)->bo->offset & (~0ull >> 16);
2568
2569 if (address >= bo_address && address < bo_address + (*bo)->bo->size) {
2570 return (struct gen_batch_decode_bo) {
2571 .addr = bo_address,
2572 .size = (*bo)->bo->size,
2573 .map = (*bo)->bo->map,
2574 };
2575 }
2576 }
2577
2578 return (struct gen_batch_decode_bo) { };
2579 }
2580
2581 struct gen_aux_map_buffer {
2582 struct gen_buffer base;
2583 struct anv_state state;
2584 };
2585
2586 static struct gen_buffer *
2587 gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2588 {
2589 struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
2590 if (!buf)
2591 return NULL;
2592
2593 struct anv_device *device = (struct anv_device*)driver_ctx;
2594 assert(device->physical->supports_48bit_addresses &&
2595 device->physical->use_softpin);
2596
2597 struct anv_state_pool *pool = &device->dynamic_state_pool;
2598 buf->state = anv_state_pool_alloc(pool, size, size);
2599
2600 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2601 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2602 buf->base.map = buf->state.map;
2603 buf->base.driver_bo = &buf->state;
2604 return &buf->base;
2605 }
2606
2607 static void
2608 gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
2609 {
2610 struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
2611 struct anv_device *device = (struct anv_device*)driver_ctx;
2612 struct anv_state_pool *pool = &device->dynamic_state_pool;
2613 anv_state_pool_free(pool, buf->state);
2614 free(buf);
2615 }
2616
2617 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
2618 .alloc = gen_aux_map_buffer_alloc,
2619 .free = gen_aux_map_buffer_free,
2620 };
2621
2622 VkResult anv_CreateDevice(
2623 VkPhysicalDevice physicalDevice,
2624 const VkDeviceCreateInfo* pCreateInfo,
2625 const VkAllocationCallbacks* pAllocator,
2626 VkDevice* pDevice)
2627 {
2628 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2629 VkResult result;
2630 struct anv_device *device;
2631
2632 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2633
2634 struct anv_device_extension_table enabled_extensions = { };
2635 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2636 int idx;
2637 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2638 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2639 anv_device_extensions[idx].extensionName) == 0)
2640 break;
2641 }
2642
2643 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2644 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2645
2646 if (!physical_device->supported_extensions.extensions[idx])
2647 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2648
2649 enabled_extensions.extensions[idx] = true;
2650 }
2651
2652 /* Check enabled features */
2653 if (pCreateInfo->pEnabledFeatures) {
2654 VkPhysicalDeviceFeatures supported_features;
2655 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2656 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2657 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
2658 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2659 for (uint32_t i = 0; i < num_features; i++) {
2660 if (enabled_feature[i] && !supported_feature[i])
2661 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2662 }
2663 }
2664
2665 /* Check requested queues and fail if we are requested to create any
2666 * queues with flags we don't support.
2667 */
2668 assert(pCreateInfo->queueCreateInfoCount > 0);
2669 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2670 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2671 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2672 }
2673
2674 /* Check if client specified queue priority. */
2675 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2676 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2677 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2678
2679 VkQueueGlobalPriorityEXT priority =
2680 queue_priority ? queue_priority->globalPriority :
2681 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2682
2683 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2684 sizeof(*device), 8,
2685 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2686 if (!device)
2687 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2688
2689 if (INTEL_DEBUG & DEBUG_BATCH) {
2690 const unsigned decode_flags =
2691 GEN_BATCH_DECODE_FULL |
2692 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2693 GEN_BATCH_DECODE_OFFSETS |
2694 GEN_BATCH_DECODE_FLOATS;
2695
2696 gen_batch_decode_ctx_init(&device->decoder_ctx,
2697 &physical_device->info,
2698 stderr, decode_flags, NULL,
2699 decode_get_bo, NULL, device);
2700 }
2701
2702 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2703 device->physical = physical_device;
2704 device->chipset_id = physical_device->chipset_id;
2705 device->no_hw = physical_device->no_hw;
2706 device->_lost = false;
2707
2708 if (pAllocator)
2709 device->alloc = *pAllocator;
2710 else
2711 device->alloc = physical_device->instance->alloc;
2712
2713 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2714 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2715 if (device->fd == -1) {
2716 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2717 goto fail_device;
2718 }
2719
2720 device->context_id = anv_gem_create_context(device);
2721 if (device->context_id == -1) {
2722 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2723 goto fail_fd;
2724 }
2725
2726 result = anv_queue_init(device, &device->queue);
2727 if (result != VK_SUCCESS)
2728 goto fail_context_id;
2729
2730 if (physical_device->use_softpin) {
2731 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2732 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2733 goto fail_queue;
2734 }
2735
2736 /* keep the page with address zero out of the allocator */
2737 util_vma_heap_init(&device->vma_lo,
2738 LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
2739
2740 util_vma_heap_init(&device->vma_cva, CLIENT_VISIBLE_HEAP_MIN_ADDRESS,
2741 CLIENT_VISIBLE_HEAP_SIZE);
2742
2743 /* Leave the last 4GiB out of the high vma range, so that no state
2744 * base address + size can overflow 48 bits. For more information see
2745 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2746 */
2747 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
2748 physical_device->gtt_size - (1ull << 32) -
2749 HIGH_HEAP_MIN_ADDRESS);
2750 }
2751
2752 list_inithead(&device->memory_objects);
2753
2754 /* As per spec, the driver implementation may deny requests to acquire
2755 * a priority above the default priority (MEDIUM) if the caller does not
2756 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2757 * is returned.
2758 */
2759 if (physical_device->has_context_priority) {
2760 int err = anv_gem_set_context_param(device->fd, device->context_id,
2761 I915_CONTEXT_PARAM_PRIORITY,
2762 vk_priority_to_gen(priority));
2763 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2764 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2765 goto fail_vmas;
2766 }
2767 }
2768
2769 device->info = physical_device->info;
2770 device->isl_dev = physical_device->isl_dev;
2771
2772 /* On Broadwell and later, we can use batch chaining to more efficiently
2773 * implement growing command buffers. Prior to Haswell, the kernel
2774 * command parser gets in the way and we have to fall back to growing
2775 * the batch.
2776 */
2777 device->can_chain_batches = device->info.gen >= 8;
2778
2779 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
2780 pCreateInfo->pEnabledFeatures->robustBufferAccess;
2781 device->enabled_extensions = enabled_extensions;
2782
2783 anv_device_init_dispatch(device);
2784
2785 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2786 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2787 goto fail_queue;
2788 }
2789
2790 pthread_condattr_t condattr;
2791 if (pthread_condattr_init(&condattr) != 0) {
2792 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2793 goto fail_mutex;
2794 }
2795 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2796 pthread_condattr_destroy(&condattr);
2797 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2798 goto fail_mutex;
2799 }
2800 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2801 pthread_condattr_destroy(&condattr);
2802 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2803 goto fail_mutex;
2804 }
2805 pthread_condattr_destroy(&condattr);
2806
2807 result = anv_bo_cache_init(&device->bo_cache);
2808 if (result != VK_SUCCESS)
2809 goto fail_queue_cond;
2810
2811 anv_bo_pool_init(&device->batch_bo_pool, device);
2812
2813 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2814 DYNAMIC_STATE_POOL_MIN_ADDRESS, 16384);
2815 if (result != VK_SUCCESS)
2816 goto fail_batch_bo_pool;
2817
2818 result = anv_state_pool_init(&device->instruction_state_pool, device,
2819 INSTRUCTION_STATE_POOL_MIN_ADDRESS, 16384);
2820 if (result != VK_SUCCESS)
2821 goto fail_dynamic_state_pool;
2822
2823 result = anv_state_pool_init(&device->surface_state_pool, device,
2824 SURFACE_STATE_POOL_MIN_ADDRESS, 4096);
2825 if (result != VK_SUCCESS)
2826 goto fail_instruction_state_pool;
2827
2828 if (physical_device->use_softpin) {
2829 result = anv_state_pool_init(&device->binding_table_pool, device,
2830 BINDING_TABLE_POOL_MIN_ADDRESS, 4096);
2831 if (result != VK_SUCCESS)
2832 goto fail_surface_state_pool;
2833 }
2834
2835 if (device->info.gen >= 12) {
2836 device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
2837 &physical_device->info);
2838 if (!device->aux_map_ctx)
2839 goto fail_binding_table_pool;
2840 }
2841
2842 result = anv_device_alloc_bo(device, 4096, 0 /* flags */,
2843 0 /* explicit_address */,
2844 &device->workaround_bo);
2845 if (result != VK_SUCCESS)
2846 goto fail_surface_aux_map_pool;
2847
2848 result = anv_device_init_trivial_batch(device);
2849 if (result != VK_SUCCESS)
2850 goto fail_workaround_bo;
2851
2852 if (device->info.gen >= 10) {
2853 result = anv_device_init_hiz_clear_value_bo(device);
2854 if (result != VK_SUCCESS)
2855 goto fail_trivial_batch_bo;
2856 }
2857
2858 anv_scratch_pool_init(device, &device->scratch_pool);
2859
2860 switch (device->info.gen) {
2861 case 7:
2862 if (!device->info.is_haswell)
2863 result = gen7_init_device_state(device);
2864 else
2865 result = gen75_init_device_state(device);
2866 break;
2867 case 8:
2868 result = gen8_init_device_state(device);
2869 break;
2870 case 9:
2871 result = gen9_init_device_state(device);
2872 break;
2873 case 10:
2874 result = gen10_init_device_state(device);
2875 break;
2876 case 11:
2877 result = gen11_init_device_state(device);
2878 break;
2879 case 12:
2880 result = gen12_init_device_state(device);
2881 break;
2882 default:
2883 /* Shouldn't get here as we don't create physical devices for any other
2884 * gens. */
2885 unreachable("unhandled gen");
2886 }
2887 if (result != VK_SUCCESS)
2888 goto fail_workaround_bo;
2889
2890 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2891
2892 anv_device_init_blorp(device);
2893
2894 anv_device_init_border_colors(device);
2895
2896 anv_device_perf_init(device);
2897
2898 *pDevice = anv_device_to_handle(device);
2899
2900 return VK_SUCCESS;
2901
2902 fail_workaround_bo:
2903 anv_scratch_pool_finish(device, &device->scratch_pool);
2904 if (device->info.gen >= 10)
2905 anv_device_release_bo(device, device->hiz_clear_bo);
2906 anv_device_release_bo(device, device->workaround_bo);
2907 fail_trivial_batch_bo:
2908 anv_device_release_bo(device, device->trivial_batch_bo);
2909 fail_surface_aux_map_pool:
2910 if (device->info.gen >= 12) {
2911 gen_aux_map_finish(device->aux_map_ctx);
2912 device->aux_map_ctx = NULL;
2913 }
2914 fail_binding_table_pool:
2915 if (physical_device->use_softpin)
2916 anv_state_pool_finish(&device->binding_table_pool);
2917 fail_surface_state_pool:
2918 anv_state_pool_finish(&device->surface_state_pool);
2919 fail_instruction_state_pool:
2920 anv_state_pool_finish(&device->instruction_state_pool);
2921 fail_dynamic_state_pool:
2922 anv_state_pool_finish(&device->dynamic_state_pool);
2923 fail_batch_bo_pool:
2924 anv_bo_pool_finish(&device->batch_bo_pool);
2925 anv_bo_cache_finish(&device->bo_cache);
2926 fail_queue_cond:
2927 pthread_cond_destroy(&device->queue_submit);
2928 fail_mutex:
2929 pthread_mutex_destroy(&device->mutex);
2930 fail_vmas:
2931 if (physical_device->use_softpin) {
2932 util_vma_heap_finish(&device->vma_hi);
2933 util_vma_heap_finish(&device->vma_cva);
2934 util_vma_heap_finish(&device->vma_lo);
2935 }
2936 fail_queue:
2937 anv_queue_finish(&device->queue);
2938 fail_context_id:
2939 anv_gem_destroy_context(device, device->context_id);
2940 fail_fd:
2941 close(device->fd);
2942 fail_device:
2943 vk_free(&device->alloc, device);
2944
2945 return result;
2946 }
2947
2948 void anv_DestroyDevice(
2949 VkDevice _device,
2950 const VkAllocationCallbacks* pAllocator)
2951 {
2952 ANV_FROM_HANDLE(anv_device, device, _device);
2953
2954 if (!device)
2955 return;
2956
2957 anv_device_finish_blorp(device);
2958
2959 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2960
2961 anv_queue_finish(&device->queue);
2962
2963 #ifdef HAVE_VALGRIND
2964 /* We only need to free these to prevent valgrind errors. The backing
2965 * BO will go away in a couple of lines so we don't actually leak.
2966 */
2967 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2968 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
2969 #endif
2970
2971 anv_scratch_pool_finish(device, &device->scratch_pool);
2972
2973 anv_device_release_bo(device, device->workaround_bo);
2974 anv_device_release_bo(device, device->trivial_batch_bo);
2975 if (device->info.gen >= 10)
2976 anv_device_release_bo(device, device->hiz_clear_bo);
2977
2978 if (device->info.gen >= 12) {
2979 gen_aux_map_finish(device->aux_map_ctx);
2980 device->aux_map_ctx = NULL;
2981 }
2982
2983 if (device->physical->use_softpin)
2984 anv_state_pool_finish(&device->binding_table_pool);
2985 anv_state_pool_finish(&device->surface_state_pool);
2986 anv_state_pool_finish(&device->instruction_state_pool);
2987 anv_state_pool_finish(&device->dynamic_state_pool);
2988
2989 anv_bo_pool_finish(&device->batch_bo_pool);
2990
2991 anv_bo_cache_finish(&device->bo_cache);
2992
2993 if (device->physical->use_softpin) {
2994 util_vma_heap_finish(&device->vma_hi);
2995 util_vma_heap_finish(&device->vma_cva);
2996 util_vma_heap_finish(&device->vma_lo);
2997 }
2998
2999 pthread_cond_destroy(&device->queue_submit);
3000 pthread_mutex_destroy(&device->mutex);
3001
3002 anv_gem_destroy_context(device, device->context_id);
3003
3004 if (INTEL_DEBUG & DEBUG_BATCH)
3005 gen_batch_decode_ctx_finish(&device->decoder_ctx);
3006
3007 close(device->fd);
3008
3009 vk_free(&device->alloc, device);
3010 }
3011
3012 VkResult anv_EnumerateInstanceLayerProperties(
3013 uint32_t* pPropertyCount,
3014 VkLayerProperties* pProperties)
3015 {
3016 if (pProperties == NULL) {
3017 *pPropertyCount = 0;
3018 return VK_SUCCESS;
3019 }
3020
3021 /* None supported at this time */
3022 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
3023 }
3024
3025 VkResult anv_EnumerateDeviceLayerProperties(
3026 VkPhysicalDevice physicalDevice,
3027 uint32_t* pPropertyCount,
3028 VkLayerProperties* pProperties)
3029 {
3030 if (pProperties == NULL) {
3031 *pPropertyCount = 0;
3032 return VK_SUCCESS;
3033 }
3034
3035 /* None supported at this time */
3036 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
3037 }
3038
3039 void anv_GetDeviceQueue(
3040 VkDevice _device,
3041 uint32_t queueNodeIndex,
3042 uint32_t queueIndex,
3043 VkQueue* pQueue)
3044 {
3045 const VkDeviceQueueInfo2 info = {
3046 .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
3047 .pNext = NULL,
3048 .flags = 0,
3049 .queueFamilyIndex = queueNodeIndex,
3050 .queueIndex = queueIndex,
3051 };
3052
3053 anv_GetDeviceQueue2(_device, &info, pQueue);
3054 }
3055
3056 void anv_GetDeviceQueue2(
3057 VkDevice _device,
3058 const VkDeviceQueueInfo2* pQueueInfo,
3059 VkQueue* pQueue)
3060 {
3061 ANV_FROM_HANDLE(anv_device, device, _device);
3062
3063 assert(pQueueInfo->queueIndex == 0);
3064
3065 if (pQueueInfo->flags == device->queue.flags)
3066 *pQueue = anv_queue_to_handle(&device->queue);
3067 else
3068 *pQueue = NULL;
3069 }
3070
3071 VkResult
3072 _anv_device_set_lost(struct anv_device *device,
3073 const char *file, int line,
3074 const char *msg, ...)
3075 {
3076 VkResult err;
3077 va_list ap;
3078
3079 p_atomic_inc(&device->_lost);
3080
3081 va_start(ap, msg);
3082 err = __vk_errorv(device->physical->instance, device,
3083 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
3084 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3085 va_end(ap);
3086
3087 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3088 abort();
3089
3090 return err;
3091 }
3092
3093 VkResult
3094 _anv_queue_set_lost(struct anv_queue *queue,
3095 const char *file, int line,
3096 const char *msg, ...)
3097 {
3098 VkResult err;
3099 va_list ap;
3100
3101 p_atomic_inc(&queue->device->_lost);
3102
3103 va_start(ap, msg);
3104 err = __vk_errorv(queue->device->physical->instance, queue->device,
3105 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
3106 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3107 va_end(ap);
3108
3109 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3110 abort();
3111
3112 return err;
3113 }
3114
3115 VkResult
3116 anv_device_query_status(struct anv_device *device)
3117 {
3118 /* This isn't likely as most of the callers of this function already check
3119 * for it. However, it doesn't hurt to check and it potentially lets us
3120 * avoid an ioctl.
3121 */
3122 if (anv_device_is_lost(device))
3123 return VK_ERROR_DEVICE_LOST;
3124
3125 uint32_t active, pending;
3126 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
3127 if (ret == -1) {
3128 /* We don't know the real error. */
3129 return anv_device_set_lost(device, "get_reset_stats failed: %m");
3130 }
3131
3132 if (active) {
3133 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
3134 } else if (pending) {
3135 return anv_device_set_lost(device, "GPU hung with commands in-flight");
3136 }
3137
3138 return VK_SUCCESS;
3139 }
3140
3141 VkResult
3142 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
3143 {
3144 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3145 * Other usages of the BO (such as on different hardware) will not be
3146 * flagged as "busy" by this ioctl. Use with care.
3147 */
3148 int ret = anv_gem_busy(device, bo->gem_handle);
3149 if (ret == 1) {
3150 return VK_NOT_READY;
3151 } else if (ret == -1) {
3152 /* We don't know the real error. */
3153 return anv_device_set_lost(device, "gem wait failed: %m");
3154 }
3155
3156 /* Query for device status after the busy call. If the BO we're checking
3157 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3158 * client because it clearly doesn't have valid data. Yes, this most
3159 * likely means an ioctl, but we just did an ioctl to query the busy status
3160 * so it's no great loss.
3161 */
3162 return anv_device_query_status(device);
3163 }
3164
3165 VkResult
3166 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
3167 int64_t timeout)
3168 {
3169 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
3170 if (ret == -1 && errno == ETIME) {
3171 return VK_TIMEOUT;
3172 } else if (ret == -1) {
3173 /* We don't know the real error. */
3174 return anv_device_set_lost(device, "gem wait failed: %m");
3175 }
3176
3177 /* Query for device status after the wait. If the BO we're waiting on got
3178 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3179 * because it clearly doesn't have valid data. Yes, this most likely means
3180 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3181 */
3182 return anv_device_query_status(device);
3183 }
3184
3185 VkResult anv_DeviceWaitIdle(
3186 VkDevice _device)
3187 {
3188 ANV_FROM_HANDLE(anv_device, device, _device);
3189
3190 if (anv_device_is_lost(device))
3191 return VK_ERROR_DEVICE_LOST;
3192
3193 return anv_queue_submit_simple_batch(&device->queue, NULL);
3194 }
3195
3196 bool
3197 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo,
3198 uint64_t client_address)
3199 {
3200 const struct gen_device_info *devinfo = &device->info;
3201 /* Gen12 CCS surface addresses need to be 64K aligned. We have no way of
3202 * telling what this allocation is for so pick the largest alignment.
3203 */
3204 const uint32_t vma_alignment =
3205 devinfo->gen >= 12 ? (64 * 1024) : (4 * 1024);
3206
3207 if (!(bo->flags & EXEC_OBJECT_PINNED)) {
3208 assert(!(bo->has_client_visible_address));
3209 return true;
3210 }
3211
3212 pthread_mutex_lock(&device->vma_mutex);
3213
3214 bo->offset = 0;
3215
3216 if (bo->has_client_visible_address) {
3217 assert(bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS);
3218 if (client_address) {
3219 if (util_vma_heap_alloc_addr(&device->vma_cva,
3220 client_address, bo->size)) {
3221 bo->offset = gen_canonical_address(client_address);
3222 }
3223 } else {
3224 uint64_t addr =
3225 util_vma_heap_alloc(&device->vma_cva, bo->size, vma_alignment);
3226 if (addr) {
3227 bo->offset = gen_canonical_address(addr);
3228 assert(addr == gen_48b_address(bo->offset));
3229 }
3230 }
3231 /* We don't want to fall back to other heaps */
3232 goto done;
3233 }
3234
3235 assert(client_address == 0);
3236
3237 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) {
3238 uint64_t addr =
3239 util_vma_heap_alloc(&device->vma_hi, bo->size, vma_alignment);
3240 if (addr) {
3241 bo->offset = gen_canonical_address(addr);
3242 assert(addr == gen_48b_address(bo->offset));
3243 }
3244 }
3245
3246 if (bo->offset == 0) {
3247 uint64_t addr =
3248 util_vma_heap_alloc(&device->vma_lo, bo->size, vma_alignment);
3249 if (addr) {
3250 bo->offset = gen_canonical_address(addr);
3251 assert(addr == gen_48b_address(bo->offset));
3252 }
3253 }
3254
3255 done:
3256 pthread_mutex_unlock(&device->vma_mutex);
3257
3258 return bo->offset != 0;
3259 }
3260
3261 void
3262 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
3263 {
3264 if (!(bo->flags & EXEC_OBJECT_PINNED))
3265 return;
3266
3267 const uint64_t addr_48b = gen_48b_address(bo->offset);
3268
3269 pthread_mutex_lock(&device->vma_mutex);
3270
3271 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3272 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3273 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
3274 } else if (addr_48b >= CLIENT_VISIBLE_HEAP_MIN_ADDRESS &&
3275 addr_48b <= CLIENT_VISIBLE_HEAP_MAX_ADDRESS) {
3276 util_vma_heap_free(&device->vma_cva, addr_48b, bo->size);
3277 } else {
3278 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS);
3279 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
3280 }
3281
3282 pthread_mutex_unlock(&device->vma_mutex);
3283
3284 bo->offset = 0;
3285 }
3286
3287 VkResult anv_AllocateMemory(
3288 VkDevice _device,
3289 const VkMemoryAllocateInfo* pAllocateInfo,
3290 const VkAllocationCallbacks* pAllocator,
3291 VkDeviceMemory* pMem)
3292 {
3293 ANV_FROM_HANDLE(anv_device, device, _device);
3294 struct anv_physical_device *pdevice = device->physical;
3295 struct anv_device_memory *mem;
3296 VkResult result = VK_SUCCESS;
3297
3298 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3299
3300 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3301 assert(pAllocateInfo->allocationSize > 0);
3302
3303 VkDeviceSize aligned_alloc_size =
3304 align_u64(pAllocateInfo->allocationSize, 4096);
3305
3306 if (aligned_alloc_size > MAX_MEMORY_ALLOCATION_SIZE)
3307 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3308
3309 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3310 struct anv_memory_type *mem_type =
3311 &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3312 assert(mem_type->heapIndex < pdevice->memory.heap_count);
3313 struct anv_memory_heap *mem_heap =
3314 &pdevice->memory.heaps[mem_type->heapIndex];
3315
3316 uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
3317 if (mem_heap_used + aligned_alloc_size > mem_heap->size)
3318 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3319
3320 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
3321 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3322 if (mem == NULL)
3323 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3324
3325 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3326 mem->type = mem_type;
3327 mem->map = NULL;
3328 mem->map_size = 0;
3329 mem->ahw = NULL;
3330 mem->host_ptr = NULL;
3331
3332 enum anv_bo_alloc_flags alloc_flags = 0;
3333
3334 const VkExportMemoryAllocateInfo *export_info = NULL;
3335 const VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info = NULL;
3336 const VkImportMemoryFdInfoKHR *fd_info = NULL;
3337 const VkImportMemoryHostPointerInfoEXT *host_ptr_info = NULL;
3338 const VkMemoryDedicatedAllocateInfo *dedicated_info = NULL;
3339 VkMemoryAllocateFlags vk_flags = 0;
3340 uint64_t client_address = 0;
3341
3342 vk_foreach_struct_const(ext, pAllocateInfo->pNext) {
3343 switch (ext->sType) {
3344 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO:
3345 export_info = (void *)ext;
3346 break;
3347
3348 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID:
3349 ahw_import_info = (void *)ext;
3350 break;
3351
3352 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR:
3353 fd_info = (void *)ext;
3354 break;
3355
3356 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT:
3357 host_ptr_info = (void *)ext;
3358 break;
3359
3360 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO: {
3361 const VkMemoryAllocateFlagsInfo *flags_info = (void *)ext;
3362 vk_flags = flags_info->flags;
3363 break;
3364 }
3365
3366 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO:
3367 dedicated_info = (void *)ext;
3368 break;
3369
3370 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR: {
3371 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *addr_info =
3372 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *)ext;
3373 client_address = addr_info->opaqueCaptureAddress;
3374 break;
3375 }
3376
3377 default:
3378 anv_debug_ignored_stype(ext->sType);
3379 break;
3380 }
3381 }
3382
3383 if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR)
3384 alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS;
3385
3386 /* Check if we need to support Android HW buffer export. If so,
3387 * create AHardwareBuffer and import memory from it.
3388 */
3389 bool android_export = false;
3390 if (export_info && export_info->handleTypes &
3391 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3392 android_export = true;
3393
3394 if (ahw_import_info) {
3395 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3396 if (result != VK_SUCCESS)
3397 goto fail;
3398
3399 goto success;
3400 } else if (android_export) {
3401 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3402 if (result != VK_SUCCESS)
3403 goto fail;
3404
3405 const VkImportAndroidHardwareBufferInfoANDROID import_info = {
3406 .buffer = mem->ahw,
3407 };
3408 result = anv_import_ahw_memory(_device, mem, &import_info);
3409 if (result != VK_SUCCESS)
3410 goto fail;
3411
3412 goto success;
3413 }
3414
3415 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3416 * ignored.
3417 */
3418 if (fd_info && fd_info->handleType) {
3419 /* At the moment, we support only the below handle types. */
3420 assert(fd_info->handleType ==
3421 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3422 fd_info->handleType ==
3423 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3424
3425 result = anv_device_import_bo(device, fd_info->fd, alloc_flags,
3426 client_address, &mem->bo);
3427 if (result != VK_SUCCESS)
3428 goto fail;
3429
3430 VkDeviceSize aligned_alloc_size =
3431 align_u64(pAllocateInfo->allocationSize, 4096);
3432
3433 /* For security purposes, we reject importing the bo if it's smaller
3434 * than the requested allocation size. This prevents a malicious client
3435 * from passing a buffer to a trusted client, lying about the size, and
3436 * telling the trusted client to try and texture from an image that goes
3437 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3438 * in the trusted client. The trusted client can protect itself against
3439 * this sort of attack but only if it can trust the buffer size.
3440 */
3441 if (mem->bo->size < aligned_alloc_size) {
3442 result = vk_errorf(device, device, VK_ERROR_INVALID_EXTERNAL_HANDLE,
3443 "aligned allocationSize too large for "
3444 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3445 "%"PRIu64"B > %"PRIu64"B",
3446 aligned_alloc_size, mem->bo->size);
3447 anv_device_release_bo(device, mem->bo);
3448 goto fail;
3449 }
3450
3451 /* From the Vulkan spec:
3452 *
3453 * "Importing memory from a file descriptor transfers ownership of
3454 * the file descriptor from the application to the Vulkan
3455 * implementation. The application must not perform any operations on
3456 * the file descriptor after a successful import."
3457 *
3458 * If the import fails, we leave the file descriptor open.
3459 */
3460 close(fd_info->fd);
3461 goto success;
3462 }
3463
3464 if (host_ptr_info && host_ptr_info->handleType) {
3465 if (host_ptr_info->handleType ==
3466 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3467 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3468 goto fail;
3469 }
3470
3471 assert(host_ptr_info->handleType ==
3472 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3473
3474 result = anv_device_import_bo_from_host_ptr(device,
3475 host_ptr_info->pHostPointer,
3476 pAllocateInfo->allocationSize,
3477 alloc_flags,
3478 client_address,
3479 &mem->bo);
3480 if (result != VK_SUCCESS)
3481 goto fail;
3482
3483 mem->host_ptr = host_ptr_info->pHostPointer;
3484 goto success;
3485 }
3486
3487 /* Regular allocate (not importing memory). */
3488
3489 if (export_info && export_info->handleTypes)
3490 alloc_flags |= ANV_BO_ALLOC_EXTERNAL;
3491
3492 result = anv_device_alloc_bo(device, pAllocateInfo->allocationSize,
3493 alloc_flags, client_address, &mem->bo);
3494 if (result != VK_SUCCESS)
3495 goto fail;
3496
3497 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3498 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3499
3500 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3501 * the BO. In this case, we have a dedicated allocation.
3502 */
3503 if (image->needs_set_tiling) {
3504 const uint32_t i915_tiling =
3505 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3506 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3507 image->planes[0].surface.isl.row_pitch_B,
3508 i915_tiling);
3509 if (ret) {
3510 anv_device_release_bo(device, mem->bo);
3511 result = vk_errorf(device, device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
3512 "failed to set BO tiling: %m");
3513 goto fail;
3514 }
3515 }
3516 }
3517
3518 success:
3519 mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
3520 if (mem_heap_used > mem_heap->size) {
3521 p_atomic_add(&mem_heap->used, -mem->bo->size);
3522 anv_device_release_bo(device, mem->bo);
3523 result = vk_errorf(device, device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
3524 "Out of heap memory");
3525 goto fail;
3526 }
3527
3528 pthread_mutex_lock(&device->mutex);
3529 list_addtail(&mem->link, &device->memory_objects);
3530 pthread_mutex_unlock(&device->mutex);
3531
3532 *pMem = anv_device_memory_to_handle(mem);
3533
3534 return VK_SUCCESS;
3535
3536 fail:
3537 vk_free2(&device->alloc, pAllocator, mem);
3538
3539 return result;
3540 }
3541
3542 VkResult anv_GetMemoryFdKHR(
3543 VkDevice device_h,
3544 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3545 int* pFd)
3546 {
3547 ANV_FROM_HANDLE(anv_device, dev, device_h);
3548 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3549
3550 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3551
3552 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3553 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3554
3555 return anv_device_export_bo(dev, mem->bo, pFd);
3556 }
3557
3558 VkResult anv_GetMemoryFdPropertiesKHR(
3559 VkDevice _device,
3560 VkExternalMemoryHandleTypeFlagBits handleType,
3561 int fd,
3562 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3563 {
3564 ANV_FROM_HANDLE(anv_device, device, _device);
3565
3566 switch (handleType) {
3567 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3568 /* dma-buf can be imported as any memory type */
3569 pMemoryFdProperties->memoryTypeBits =
3570 (1 << device->physical->memory.type_count) - 1;
3571 return VK_SUCCESS;
3572
3573 default:
3574 /* The valid usage section for this function says:
3575 *
3576 * "handleType must not be one of the handle types defined as
3577 * opaque."
3578 *
3579 * So opaque handle types fall into the default "unsupported" case.
3580 */
3581 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3582 }
3583 }
3584
3585 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3586 VkDevice _device,
3587 VkExternalMemoryHandleTypeFlagBits handleType,
3588 const void* pHostPointer,
3589 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3590 {
3591 ANV_FROM_HANDLE(anv_device, device, _device);
3592
3593 assert(pMemoryHostPointerProperties->sType ==
3594 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3595
3596 switch (handleType) {
3597 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT:
3598 /* Host memory can be imported as any memory type. */
3599 pMemoryHostPointerProperties->memoryTypeBits =
3600 (1ull << device->physical->memory.type_count) - 1;
3601
3602 return VK_SUCCESS;
3603
3604 default:
3605 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3606 }
3607 }
3608
3609 void anv_FreeMemory(
3610 VkDevice _device,
3611 VkDeviceMemory _mem,
3612 const VkAllocationCallbacks* pAllocator)
3613 {
3614 ANV_FROM_HANDLE(anv_device, device, _device);
3615 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3616
3617 if (mem == NULL)
3618 return;
3619
3620 pthread_mutex_lock(&device->mutex);
3621 list_del(&mem->link);
3622 pthread_mutex_unlock(&device->mutex);
3623
3624 if (mem->map)
3625 anv_UnmapMemory(_device, _mem);
3626
3627 p_atomic_add(&device->physical->memory.heaps[mem->type->heapIndex].used,
3628 -mem->bo->size);
3629
3630 anv_device_release_bo(device, mem->bo);
3631
3632 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3633 if (mem->ahw)
3634 AHardwareBuffer_release(mem->ahw);
3635 #endif
3636
3637 vk_free2(&device->alloc, pAllocator, mem);
3638 }
3639
3640 VkResult anv_MapMemory(
3641 VkDevice _device,
3642 VkDeviceMemory _memory,
3643 VkDeviceSize offset,
3644 VkDeviceSize size,
3645 VkMemoryMapFlags flags,
3646 void** ppData)
3647 {
3648 ANV_FROM_HANDLE(anv_device, device, _device);
3649 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3650
3651 if (mem == NULL) {
3652 *ppData = NULL;
3653 return VK_SUCCESS;
3654 }
3655
3656 if (mem->host_ptr) {
3657 *ppData = mem->host_ptr + offset;
3658 return VK_SUCCESS;
3659 }
3660
3661 if (size == VK_WHOLE_SIZE)
3662 size = mem->bo->size - offset;
3663
3664 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3665 *
3666 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3667 * assert(size != 0);
3668 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3669 * equal to the size of the memory minus offset
3670 */
3671 assert(size > 0);
3672 assert(offset + size <= mem->bo->size);
3673
3674 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3675 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3676 * at a time is valid. We could just mmap up front and return an offset
3677 * pointer here, but that may exhaust virtual memory on 32 bit
3678 * userspace. */
3679
3680 uint32_t gem_flags = 0;
3681
3682 if (!device->info.has_llc &&
3683 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3684 gem_flags |= I915_MMAP_WC;
3685
3686 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3687 uint64_t map_offset = offset & ~4095ull;
3688 assert(offset >= map_offset);
3689 uint64_t map_size = (offset + size) - map_offset;
3690
3691 /* Let's map whole pages */
3692 map_size = align_u64(map_size, 4096);
3693
3694 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3695 map_offset, map_size, gem_flags);
3696 if (map == MAP_FAILED)
3697 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3698
3699 mem->map = map;
3700 mem->map_size = map_size;
3701
3702 *ppData = mem->map + (offset - map_offset);
3703
3704 return VK_SUCCESS;
3705 }
3706
3707 void anv_UnmapMemory(
3708 VkDevice _device,
3709 VkDeviceMemory _memory)
3710 {
3711 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3712
3713 if (mem == NULL || mem->host_ptr)
3714 return;
3715
3716 anv_gem_munmap(mem->map, mem->map_size);
3717
3718 mem->map = NULL;
3719 mem->map_size = 0;
3720 }
3721
3722 static void
3723 clflush_mapped_ranges(struct anv_device *device,
3724 uint32_t count,
3725 const VkMappedMemoryRange *ranges)
3726 {
3727 for (uint32_t i = 0; i < count; i++) {
3728 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3729 if (ranges[i].offset >= mem->map_size)
3730 continue;
3731
3732 gen_clflush_range(mem->map + ranges[i].offset,
3733 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3734 }
3735 }
3736
3737 VkResult anv_FlushMappedMemoryRanges(
3738 VkDevice _device,
3739 uint32_t memoryRangeCount,
3740 const VkMappedMemoryRange* pMemoryRanges)
3741 {
3742 ANV_FROM_HANDLE(anv_device, device, _device);
3743
3744 if (device->info.has_llc)
3745 return VK_SUCCESS;
3746
3747 /* Make sure the writes we're flushing have landed. */
3748 __builtin_ia32_mfence();
3749
3750 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3751
3752 return VK_SUCCESS;
3753 }
3754
3755 VkResult anv_InvalidateMappedMemoryRanges(
3756 VkDevice _device,
3757 uint32_t memoryRangeCount,
3758 const VkMappedMemoryRange* pMemoryRanges)
3759 {
3760 ANV_FROM_HANDLE(anv_device, device, _device);
3761
3762 if (device->info.has_llc)
3763 return VK_SUCCESS;
3764
3765 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3766
3767 /* Make sure no reads get moved up above the invalidate. */
3768 __builtin_ia32_mfence();
3769
3770 return VK_SUCCESS;
3771 }
3772
3773 void anv_GetBufferMemoryRequirements(
3774 VkDevice _device,
3775 VkBuffer _buffer,
3776 VkMemoryRequirements* pMemoryRequirements)
3777 {
3778 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3779 ANV_FROM_HANDLE(anv_device, device, _device);
3780
3781 /* The Vulkan spec (git aaed022) says:
3782 *
3783 * memoryTypeBits is a bitfield and contains one bit set for every
3784 * supported memory type for the resource. The bit `1<<i` is set if and
3785 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3786 * structure for the physical device is supported.
3787 */
3788 uint32_t memory_types = (1ull << device->physical->memory.type_count) - 1;
3789
3790 /* Base alignment requirement of a cache line */
3791 uint32_t alignment = 16;
3792
3793 /* We need an alignment of 32 for pushing UBOs */
3794 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3795 alignment = MAX2(alignment, 32);
3796
3797 pMemoryRequirements->size = buffer->size;
3798 pMemoryRequirements->alignment = alignment;
3799
3800 /* Storage and Uniform buffers should have their size aligned to
3801 * 32-bits to avoid boundary checks when last DWord is not complete.
3802 * This would ensure that not internal padding would be needed for
3803 * 16-bit types.
3804 */
3805 if (device->robust_buffer_access &&
3806 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3807 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3808 pMemoryRequirements->size = align_u64(buffer->size, 4);
3809
3810 pMemoryRequirements->memoryTypeBits = memory_types;
3811 }
3812
3813 void anv_GetBufferMemoryRequirements2(
3814 VkDevice _device,
3815 const VkBufferMemoryRequirementsInfo2* pInfo,
3816 VkMemoryRequirements2* pMemoryRequirements)
3817 {
3818 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3819 &pMemoryRequirements->memoryRequirements);
3820
3821 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3822 switch (ext->sType) {
3823 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3824 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3825 requirements->prefersDedicatedAllocation = false;
3826 requirements->requiresDedicatedAllocation = false;
3827 break;
3828 }
3829
3830 default:
3831 anv_debug_ignored_stype(ext->sType);
3832 break;
3833 }
3834 }
3835 }
3836
3837 void anv_GetImageMemoryRequirements(
3838 VkDevice _device,
3839 VkImage _image,
3840 VkMemoryRequirements* pMemoryRequirements)
3841 {
3842 ANV_FROM_HANDLE(anv_image, image, _image);
3843 ANV_FROM_HANDLE(anv_device, device, _device);
3844
3845 /* The Vulkan spec (git aaed022) says:
3846 *
3847 * memoryTypeBits is a bitfield and contains one bit set for every
3848 * supported memory type for the resource. The bit `1<<i` is set if and
3849 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3850 * structure for the physical device is supported.
3851 *
3852 * All types are currently supported for images.
3853 */
3854 uint32_t memory_types = (1ull << device->physical->memory.type_count) - 1;
3855
3856 /* We must have image allocated or imported at this point. According to the
3857 * specification, external images must have been bound to memory before
3858 * calling GetImageMemoryRequirements.
3859 */
3860 assert(image->size > 0);
3861
3862 pMemoryRequirements->size = image->size;
3863 pMemoryRequirements->alignment = image->alignment;
3864 pMemoryRequirements->memoryTypeBits = memory_types;
3865 }
3866
3867 void anv_GetImageMemoryRequirements2(
3868 VkDevice _device,
3869 const VkImageMemoryRequirementsInfo2* pInfo,
3870 VkMemoryRequirements2* pMemoryRequirements)
3871 {
3872 ANV_FROM_HANDLE(anv_device, device, _device);
3873 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3874
3875 anv_GetImageMemoryRequirements(_device, pInfo->image,
3876 &pMemoryRequirements->memoryRequirements);
3877
3878 vk_foreach_struct_const(ext, pInfo->pNext) {
3879 switch (ext->sType) {
3880 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3881 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3882 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3883 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3884 plane_reqs->planeAspect);
3885
3886 assert(image->planes[plane].offset == 0);
3887
3888 /* The Vulkan spec (git aaed022) says:
3889 *
3890 * memoryTypeBits is a bitfield and contains one bit set for every
3891 * supported memory type for the resource. The bit `1<<i` is set
3892 * if and only if the memory type `i` in the
3893 * VkPhysicalDeviceMemoryProperties structure for the physical
3894 * device is supported.
3895 *
3896 * All types are currently supported for images.
3897 */
3898 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3899 (1ull << device->physical->memory.type_count) - 1;
3900
3901 /* We must have image allocated or imported at this point. According to the
3902 * specification, external images must have been bound to memory before
3903 * calling GetImageMemoryRequirements.
3904 */
3905 assert(image->planes[plane].size > 0);
3906
3907 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3908 pMemoryRequirements->memoryRequirements.alignment =
3909 image->planes[plane].alignment;
3910 break;
3911 }
3912
3913 default:
3914 anv_debug_ignored_stype(ext->sType);
3915 break;
3916 }
3917 }
3918
3919 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3920 switch (ext->sType) {
3921 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3922 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3923 if (image->needs_set_tiling || image->external_format) {
3924 /* If we need to set the tiling for external consumers, we need a
3925 * dedicated allocation.
3926 *
3927 * See also anv_AllocateMemory.
3928 */
3929 requirements->prefersDedicatedAllocation = true;
3930 requirements->requiresDedicatedAllocation = true;
3931 } else {
3932 requirements->prefersDedicatedAllocation = false;
3933 requirements->requiresDedicatedAllocation = false;
3934 }
3935 break;
3936 }
3937
3938 default:
3939 anv_debug_ignored_stype(ext->sType);
3940 break;
3941 }
3942 }
3943 }
3944
3945 void anv_GetImageSparseMemoryRequirements(
3946 VkDevice device,
3947 VkImage image,
3948 uint32_t* pSparseMemoryRequirementCount,
3949 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3950 {
3951 *pSparseMemoryRequirementCount = 0;
3952 }
3953
3954 void anv_GetImageSparseMemoryRequirements2(
3955 VkDevice device,
3956 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3957 uint32_t* pSparseMemoryRequirementCount,
3958 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3959 {
3960 *pSparseMemoryRequirementCount = 0;
3961 }
3962
3963 void anv_GetDeviceMemoryCommitment(
3964 VkDevice device,
3965 VkDeviceMemory memory,
3966 VkDeviceSize* pCommittedMemoryInBytes)
3967 {
3968 *pCommittedMemoryInBytes = 0;
3969 }
3970
3971 static void
3972 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3973 {
3974 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3975 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3976
3977 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3978
3979 if (mem) {
3980 buffer->address = (struct anv_address) {
3981 .bo = mem->bo,
3982 .offset = pBindInfo->memoryOffset,
3983 };
3984 } else {
3985 buffer->address = ANV_NULL_ADDRESS;
3986 }
3987 }
3988
3989 VkResult anv_BindBufferMemory(
3990 VkDevice device,
3991 VkBuffer buffer,
3992 VkDeviceMemory memory,
3993 VkDeviceSize memoryOffset)
3994 {
3995 anv_bind_buffer_memory(
3996 &(VkBindBufferMemoryInfo) {
3997 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
3998 .buffer = buffer,
3999 .memory = memory,
4000 .memoryOffset = memoryOffset,
4001 });
4002
4003 return VK_SUCCESS;
4004 }
4005
4006 VkResult anv_BindBufferMemory2(
4007 VkDevice device,
4008 uint32_t bindInfoCount,
4009 const VkBindBufferMemoryInfo* pBindInfos)
4010 {
4011 for (uint32_t i = 0; i < bindInfoCount; i++)
4012 anv_bind_buffer_memory(&pBindInfos[i]);
4013
4014 return VK_SUCCESS;
4015 }
4016
4017 VkResult anv_QueueBindSparse(
4018 VkQueue _queue,
4019 uint32_t bindInfoCount,
4020 const VkBindSparseInfo* pBindInfo,
4021 VkFence fence)
4022 {
4023 ANV_FROM_HANDLE(anv_queue, queue, _queue);
4024 if (anv_device_is_lost(queue->device))
4025 return VK_ERROR_DEVICE_LOST;
4026
4027 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
4028 }
4029
4030 // Event functions
4031
4032 VkResult anv_CreateEvent(
4033 VkDevice _device,
4034 const VkEventCreateInfo* pCreateInfo,
4035 const VkAllocationCallbacks* pAllocator,
4036 VkEvent* pEvent)
4037 {
4038 ANV_FROM_HANDLE(anv_device, device, _device);
4039 struct anv_state state;
4040 struct anv_event *event;
4041
4042 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
4043
4044 state = anv_state_pool_alloc(&device->dynamic_state_pool,
4045 sizeof(*event), 8);
4046 event = state.map;
4047 event->state = state;
4048 event->semaphore = VK_EVENT_RESET;
4049
4050 if (!device->info.has_llc) {
4051 /* Make sure the writes we're flushing have landed. */
4052 __builtin_ia32_mfence();
4053 __builtin_ia32_clflush(event);
4054 }
4055
4056 *pEvent = anv_event_to_handle(event);
4057
4058 return VK_SUCCESS;
4059 }
4060
4061 void anv_DestroyEvent(
4062 VkDevice _device,
4063 VkEvent _event,
4064 const VkAllocationCallbacks* pAllocator)
4065 {
4066 ANV_FROM_HANDLE(anv_device, device, _device);
4067 ANV_FROM_HANDLE(anv_event, event, _event);
4068
4069 if (!event)
4070 return;
4071
4072 anv_state_pool_free(&device->dynamic_state_pool, event->state);
4073 }
4074
4075 VkResult anv_GetEventStatus(
4076 VkDevice _device,
4077 VkEvent _event)
4078 {
4079 ANV_FROM_HANDLE(anv_device, device, _device);
4080 ANV_FROM_HANDLE(anv_event, event, _event);
4081
4082 if (anv_device_is_lost(device))
4083 return VK_ERROR_DEVICE_LOST;
4084
4085 if (!device->info.has_llc) {
4086 /* Invalidate read cache before reading event written by GPU. */
4087 __builtin_ia32_clflush(event);
4088 __builtin_ia32_mfence();
4089
4090 }
4091
4092 return event->semaphore;
4093 }
4094
4095 VkResult anv_SetEvent(
4096 VkDevice _device,
4097 VkEvent _event)
4098 {
4099 ANV_FROM_HANDLE(anv_device, device, _device);
4100 ANV_FROM_HANDLE(anv_event, event, _event);
4101
4102 event->semaphore = VK_EVENT_SET;
4103
4104 if (!device->info.has_llc) {
4105 /* Make sure the writes we're flushing have landed. */
4106 __builtin_ia32_mfence();
4107 __builtin_ia32_clflush(event);
4108 }
4109
4110 return VK_SUCCESS;
4111 }
4112
4113 VkResult anv_ResetEvent(
4114 VkDevice _device,
4115 VkEvent _event)
4116 {
4117 ANV_FROM_HANDLE(anv_device, device, _device);
4118 ANV_FROM_HANDLE(anv_event, event, _event);
4119
4120 event->semaphore = VK_EVENT_RESET;
4121
4122 if (!device->info.has_llc) {
4123 /* Make sure the writes we're flushing have landed. */
4124 __builtin_ia32_mfence();
4125 __builtin_ia32_clflush(event);
4126 }
4127
4128 return VK_SUCCESS;
4129 }
4130
4131 // Buffer functions
4132
4133 VkResult anv_CreateBuffer(
4134 VkDevice _device,
4135 const VkBufferCreateInfo* pCreateInfo,
4136 const VkAllocationCallbacks* pAllocator,
4137 VkBuffer* pBuffer)
4138 {
4139 ANV_FROM_HANDLE(anv_device, device, _device);
4140 struct anv_buffer *buffer;
4141
4142 /* Don't allow creating buffers bigger than our address space. The real
4143 * issue here is that we may align up the buffer size and we don't want
4144 * doing so to cause roll-over. However, no one has any business
4145 * allocating a buffer larger than our GTT size.
4146 */
4147 if (pCreateInfo->size > device->physical->gtt_size)
4148 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
4149
4150 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
4151
4152 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
4153 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4154 if (buffer == NULL)
4155 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4156
4157 buffer->size = pCreateInfo->size;
4158 buffer->usage = pCreateInfo->usage;
4159 buffer->address = ANV_NULL_ADDRESS;
4160
4161 *pBuffer = anv_buffer_to_handle(buffer);
4162
4163 return VK_SUCCESS;
4164 }
4165
4166 void anv_DestroyBuffer(
4167 VkDevice _device,
4168 VkBuffer _buffer,
4169 const VkAllocationCallbacks* pAllocator)
4170 {
4171 ANV_FROM_HANDLE(anv_device, device, _device);
4172 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
4173
4174 if (!buffer)
4175 return;
4176
4177 vk_free2(&device->alloc, pAllocator, buffer);
4178 }
4179
4180 VkDeviceAddress anv_GetBufferDeviceAddress(
4181 VkDevice device,
4182 const VkBufferDeviceAddressInfoKHR* pInfo)
4183 {
4184 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
4185
4186 assert(!anv_address_is_null(buffer->address));
4187 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
4188
4189 return anv_address_physical(buffer->address);
4190 }
4191
4192 uint64_t anv_GetBufferOpaqueCaptureAddress(
4193 VkDevice device,
4194 const VkBufferDeviceAddressInfoKHR* pInfo)
4195 {
4196 return 0;
4197 }
4198
4199 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4200 VkDevice device,
4201 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR* pInfo)
4202 {
4203 ANV_FROM_HANDLE(anv_device_memory, memory, pInfo->memory);
4204
4205 assert(memory->bo->flags & EXEC_OBJECT_PINNED);
4206 assert(memory->bo->has_client_visible_address);
4207
4208 return gen_48b_address(memory->bo->offset);
4209 }
4210
4211 void
4212 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
4213 enum isl_format format,
4214 struct anv_address address,
4215 uint32_t range, uint32_t stride)
4216 {
4217 isl_buffer_fill_state(&device->isl_dev, state.map,
4218 .address = anv_address_physical(address),
4219 .mocs = device->isl_dev.mocs.internal,
4220 .size_B = range,
4221 .format = format,
4222 .swizzle = ISL_SWIZZLE_IDENTITY,
4223 .stride_B = stride);
4224 }
4225
4226 void anv_DestroySampler(
4227 VkDevice _device,
4228 VkSampler _sampler,
4229 const VkAllocationCallbacks* pAllocator)
4230 {
4231 ANV_FROM_HANDLE(anv_device, device, _device);
4232 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
4233
4234 if (!sampler)
4235 return;
4236
4237 if (sampler->bindless_state.map) {
4238 anv_state_pool_free(&device->dynamic_state_pool,
4239 sampler->bindless_state);
4240 }
4241
4242 vk_free2(&device->alloc, pAllocator, sampler);
4243 }
4244
4245 VkResult anv_CreateFramebuffer(
4246 VkDevice _device,
4247 const VkFramebufferCreateInfo* pCreateInfo,
4248 const VkAllocationCallbacks* pAllocator,
4249 VkFramebuffer* pFramebuffer)
4250 {
4251 ANV_FROM_HANDLE(anv_device, device, _device);
4252 struct anv_framebuffer *framebuffer;
4253
4254 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
4255
4256 size_t size = sizeof(*framebuffer);
4257
4258 /* VK_KHR_imageless_framebuffer extension says:
4259 *
4260 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4261 * parameter pAttachments is ignored.
4262 */
4263 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
4264 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4265 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4266 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4267 if (framebuffer == NULL)
4268 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4269
4270 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4271 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4272 framebuffer->attachments[i] = iview;
4273 }
4274 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4275 } else {
4276 assert(device->enabled_extensions.KHR_imageless_framebuffer);
4277 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4278 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4279 if (framebuffer == NULL)
4280 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4281
4282 framebuffer->attachment_count = 0;
4283 }
4284
4285 framebuffer->width = pCreateInfo->width;
4286 framebuffer->height = pCreateInfo->height;
4287 framebuffer->layers = pCreateInfo->layers;
4288
4289 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4290
4291 return VK_SUCCESS;
4292 }
4293
4294 void anv_DestroyFramebuffer(
4295 VkDevice _device,
4296 VkFramebuffer _fb,
4297 const VkAllocationCallbacks* pAllocator)
4298 {
4299 ANV_FROM_HANDLE(anv_device, device, _device);
4300 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4301
4302 if (!fb)
4303 return;
4304
4305 vk_free2(&device->alloc, pAllocator, fb);
4306 }
4307
4308 static const VkTimeDomainEXT anv_time_domains[] = {
4309 VK_TIME_DOMAIN_DEVICE_EXT,
4310 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4311 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4312 };
4313
4314 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4315 VkPhysicalDevice physicalDevice,
4316 uint32_t *pTimeDomainCount,
4317 VkTimeDomainEXT *pTimeDomains)
4318 {
4319 int d;
4320 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4321
4322 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4323 vk_outarray_append(&out, i) {
4324 *i = anv_time_domains[d];
4325 }
4326 }
4327
4328 return vk_outarray_status(&out);
4329 }
4330
4331 static uint64_t
4332 anv_clock_gettime(clockid_t clock_id)
4333 {
4334 struct timespec current;
4335 int ret;
4336
4337 ret = clock_gettime(clock_id, &current);
4338 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4339 ret = clock_gettime(CLOCK_MONOTONIC, &current);
4340 if (ret < 0)
4341 return 0;
4342
4343 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4344 }
4345
4346 #define TIMESTAMP 0x2358
4347
4348 VkResult anv_GetCalibratedTimestampsEXT(
4349 VkDevice _device,
4350 uint32_t timestampCount,
4351 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4352 uint64_t *pTimestamps,
4353 uint64_t *pMaxDeviation)
4354 {
4355 ANV_FROM_HANDLE(anv_device, device, _device);
4356 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4357 int ret;
4358 int d;
4359 uint64_t begin, end;
4360 uint64_t max_clock_period = 0;
4361
4362 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4363
4364 for (d = 0; d < timestampCount; d++) {
4365 switch (pTimestampInfos[d].timeDomain) {
4366 case VK_TIME_DOMAIN_DEVICE_EXT:
4367 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
4368 &pTimestamps[d]);
4369
4370 if (ret != 0) {
4371 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4372 "register: %m");
4373 }
4374 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4375 max_clock_period = MAX2(max_clock_period, device_period);
4376 break;
4377 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4378 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4379 max_clock_period = MAX2(max_clock_period, 1);
4380 break;
4381
4382 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4383 pTimestamps[d] = begin;
4384 break;
4385 default:
4386 pTimestamps[d] = 0;
4387 break;
4388 }
4389 }
4390
4391 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4392
4393 /*
4394 * The maximum deviation is the sum of the interval over which we
4395 * perform the sampling and the maximum period of any sampled
4396 * clock. That's because the maximum skew between any two sampled
4397 * clock edges is when the sampled clock with the largest period is
4398 * sampled at the end of that period but right at the beginning of the
4399 * sampling interval and some other clock is sampled right at the
4400 * begining of its sampling period and right at the end of the
4401 * sampling interval. Let's assume the GPU has the longest clock
4402 * period and that the application is sampling GPU and monotonic:
4403 *
4404 * s e
4405 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4406 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4407 *
4408 * g
4409 * 0 1 2 3
4410 * GPU -----_____-----_____-----_____-----_____
4411 *
4412 * m
4413 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4414 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4415 *
4416 * Interval <----------------->
4417 * Deviation <-------------------------->
4418 *
4419 * s = read(raw) 2
4420 * g = read(GPU) 1
4421 * m = read(monotonic) 2
4422 * e = read(raw) b
4423 *
4424 * We round the sample interval up by one tick to cover sampling error
4425 * in the interval clock
4426 */
4427
4428 uint64_t sample_interval = end - begin + 1;
4429
4430 *pMaxDeviation = sample_interval + max_clock_period;
4431
4432 return VK_SUCCESS;
4433 }
4434
4435 /* vk_icd.h does not declare this function, so we declare it here to
4436 * suppress Wmissing-prototypes.
4437 */
4438 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4439 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4440
4441 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4442 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4443 {
4444 /* For the full details on loader interface versioning, see
4445 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4446 * What follows is a condensed summary, to help you navigate the large and
4447 * confusing official doc.
4448 *
4449 * - Loader interface v0 is incompatible with later versions. We don't
4450 * support it.
4451 *
4452 * - In loader interface v1:
4453 * - The first ICD entrypoint called by the loader is
4454 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4455 * entrypoint.
4456 * - The ICD must statically expose no other Vulkan symbol unless it is
4457 * linked with -Bsymbolic.
4458 * - Each dispatchable Vulkan handle created by the ICD must be
4459 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4460 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4461 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4462 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4463 * such loader-managed surfaces.
4464 *
4465 * - Loader interface v2 differs from v1 in:
4466 * - The first ICD entrypoint called by the loader is
4467 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4468 * statically expose this entrypoint.
4469 *
4470 * - Loader interface v3 differs from v2 in:
4471 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4472 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4473 * because the loader no longer does so.
4474 *
4475 * - Loader interface v4 differs from v3 in:
4476 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4477 */
4478 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4479 return VK_SUCCESS;
4480 }