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