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