projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
anv,vulkan: Implement VK_EXT_private_data
[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", NULL);
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_try_create(struct anv_instance *instance,
333 drmDevicePtr drm_device,
334 struct anv_physical_device **device_out)
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 struct gen_device_info devinfo;
349 if (!gen_get_device_info_from_fd(fd, &devinfo)) {
350 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
351 goto fail_fd;
352 }
353
354 const char *device_name = gen_get_device_name(devinfo.chipset_id);
355
356 if (devinfo.is_haswell) {
357 intel_logw("Haswell Vulkan support is incomplete");
358 } else if (devinfo.gen == 7 && !devinfo.is_baytrail) {
359 intel_logw("Ivy Bridge Vulkan support is incomplete");
360 } else if (devinfo.gen == 7 && devinfo.is_baytrail) {
361 intel_logw("Bay Trail Vulkan support is incomplete");
362 } else if (devinfo.gen >= 8 && devinfo.gen <= 11) {
363 /* Gen8-11 fully supported */
364 } else if (devinfo.gen == 12) {
365 intel_logw("Vulkan is not yet fully supported on gen12");
366 } else {
367 result = vk_errorfi(instance, NULL, VK_ERROR_INCOMPATIBLE_DRIVER,
368 "Vulkan not yet supported on %s", device_name);
369 goto fail_fd;
370 }
371
372 struct anv_physical_device *device =
373 vk_alloc(&instance->alloc, sizeof(*device), 8,
374 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
375 if (device == NULL) {
376 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
377 goto fail_fd;
378 }
379
380 vk_object_base_init(NULL, &device->base, VK_OBJECT_TYPE_PHYSICAL_DEVICE);
381 device->instance = instance;
382
383 assert(strlen(path) < ARRAY_SIZE(device->path));
384 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
385
386 device->info = devinfo;
387 device->name = device_name;
388
389 device->no_hw = device->info.no_hw;
390 if (getenv("INTEL_NO_HW") != NULL)
391 device->no_hw = true;
392
393 device->pci_info.domain = drm_device->businfo.pci->domain;
394 device->pci_info.bus = drm_device->businfo.pci->bus;
395 device->pci_info.device = drm_device->businfo.pci->dev;
396 device->pci_info.function = drm_device->businfo.pci->func;
397
398 device->cmd_parser_version = -1;
399 if (device->info.gen == 7) {
400 device->cmd_parser_version =
401 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
402 if (device->cmd_parser_version == -1) {
403 result = vk_errorfi(device->instance, NULL,
404 VK_ERROR_INITIALIZATION_FAILED,
405 "failed to get command parser version");
406 goto fail_alloc;
407 }
408 }
409
410 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
411 result = vk_errorfi(device->instance, NULL,
412 VK_ERROR_INITIALIZATION_FAILED,
413 "kernel missing gem wait");
414 goto fail_alloc;
415 }
416
417 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
418 result = vk_errorfi(device->instance, NULL,
419 VK_ERROR_INITIALIZATION_FAILED,
420 "kernel missing execbuf2");
421 goto fail_alloc;
422 }
423
424 if (!device->info.has_llc &&
425 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
426 result = vk_errorfi(device->instance, NULL,
427 VK_ERROR_INITIALIZATION_FAILED,
428 "kernel missing wc mmap");
429 goto fail_alloc;
430 }
431
432 device->has_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN);
433 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
434 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
435 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
436 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
437 device->has_syncobj_wait = device->has_syncobj &&
438 anv_gem_supports_syncobj_wait(fd);
439 device->has_context_priority = anv_gem_has_context_priority(fd);
440
441 result = anv_physical_device_init_heaps(device, fd);
442 if (result != VK_SUCCESS)
443 goto fail_alloc;
444
445 device->use_softpin = device->has_softpin &&
446 device->supports_48bit_addresses;
447
448 device->has_context_isolation =
449 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
450
451 device->always_use_bindless =
452 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
453
454 /* We first got the A64 messages on broadwell and we can only use them if
455 * we can pass addresses directly into the shader which requires softpin.
456 */
457 device->has_a64_buffer_access = device->info.gen >= 8 &&
458 device->use_softpin;
459
460 /* We first get bindless image access on Skylake and we can only really do
461 * it if we don't have any relocations so we need softpin.
462 */
463 device->has_bindless_images = device->info.gen >= 9 &&
464 device->use_softpin;
465
466 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
467 * because it's just a matter of setting the sampler address in the sample
468 * message header. However, we've not bothered to wire it up for vec4 so
469 * we leave it disabled on gen7.
470 */
471 device->has_bindless_samplers = device->info.gen >= 8;
472
473 device->has_implicit_ccs = device->info.has_aux_map;
474
475 device->has_mem_available = get_available_system_memory() != 0;
476
477 device->always_flush_cache =
478 driQueryOptionb(&instance->dri_options, "always_flush_cache");
479
480 device->has_mmap_offset =
481 anv_gem_get_param(fd, I915_PARAM_MMAP_GTT_VERSION) >= 4;
482
483 /* GENs prior to 8 do not support EU/Subslice info */
484 if (device->info.gen >= 8) {
485 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
486 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
487
488 /* Without this information, we cannot get the right Braswell
489 * brandstrings, and we have to use conservative numbers for GPGPU on
490 * many platforms, but otherwise, things will just work.
491 */
492 if (device->subslice_total < 1 || device->eu_total < 1) {
493 intel_logw("Kernel 4.1 required to properly query GPU properties");
494 }
495 } else if (device->info.gen == 7) {
496 device->subslice_total = 1 << (device->info.gt - 1);
497 }
498
499 if (device->info.is_cherryview &&
500 device->subslice_total > 0 && device->eu_total > 0) {
501 /* Logical CS threads = EUs per subslice * num threads per EU */
502 uint32_t max_cs_threads =
503 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
504
505 /* Fuse configurations may give more threads than expected, never less. */
506 if (max_cs_threads > device->info.max_cs_threads)
507 device->info.max_cs_threads = max_cs_threads;
508 }
509
510 device->compiler = brw_compiler_create(NULL, &device->info);
511 if (device->compiler == NULL) {
512 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
513 goto fail_alloc;
514 }
515 device->compiler->shader_debug_log = compiler_debug_log;
516 device->compiler->shader_perf_log = compiler_perf_log;
517 device->compiler->supports_pull_constants = false;
518 device->compiler->constant_buffer_0_is_relative =
519 device->info.gen < 8 || !device->has_context_isolation;
520 device->compiler->supports_shader_constants = true;
521 device->compiler->compact_params = false;
522
523 /* Broadwell PRM says:
524 *
525 * "Before Gen8, there was a historical configuration control field to
526 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
527 * different places: TILECTL[1:0], ARB_MODE[5:4], and
528 * DISP_ARB_CTL[14:13].
529 *
530 * For Gen8 and subsequent generations, the swizzle fields are all
531 * reserved, and the CPU's memory controller performs all address
532 * swizzling modifications."
533 */
534 bool swizzled =
535 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
536
537 isl_device_init(&device->isl_dev, &device->info, swizzled);
538
539 result = anv_physical_device_init_uuids(device);
540 if (result != VK_SUCCESS)
541 goto fail_compiler;
542
543 anv_physical_device_init_disk_cache(device);
544
545 if (instance->enabled_extensions.KHR_display) {
546 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
547 if (master_fd >= 0) {
548 /* prod the device with a GETPARAM call which will fail if
549 * we don't have permission to even render on this device
550 */
551 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
552 close(master_fd);
553 master_fd = -1;
554 }
555 }
556 }
557 device->master_fd = master_fd;
558
559 result = anv_init_wsi(device);
560 if (result != VK_SUCCESS)
561 goto fail_disk_cache;
562
563 device->perf = anv_get_perf(&device->info, fd);
564
565 anv_physical_device_get_supported_extensions(device,
566 &device->supported_extensions);
567
568
569 device->local_fd = fd;
570
571 *device_out = device;
572
573 return VK_SUCCESS;
574
575 fail_disk_cache:
576 anv_physical_device_free_disk_cache(device);
577 fail_compiler:
578 ralloc_free(device->compiler);
579 fail_alloc:
580 vk_free(&instance->alloc, device);
581 fail_fd:
582 close(fd);
583 if (master_fd != -1)
584 close(master_fd);
585 return result;
586 }
587
588 static void
589 anv_physical_device_destroy(struct anv_physical_device *device)
590 {
591 anv_finish_wsi(device);
592 anv_physical_device_free_disk_cache(device);
593 ralloc_free(device->compiler);
594 ralloc_free(device->perf);
595 close(device->local_fd);
596 if (device->master_fd >= 0)
597 close(device->master_fd);
598 vk_object_base_finish(&device->base);
599 vk_free(&device->instance->alloc, device);
600 }
601
602 static void *
603 default_alloc_func(void *pUserData, size_t size, size_t align,
604 VkSystemAllocationScope allocationScope)
605 {
606 return malloc(size);
607 }
608
609 static void *
610 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
611 size_t align, VkSystemAllocationScope allocationScope)
612 {
613 return realloc(pOriginal, size);
614 }
615
616 static void
617 default_free_func(void *pUserData, void *pMemory)
618 {
619 free(pMemory);
620 }
621
622 static const VkAllocationCallbacks default_alloc = {
623 .pUserData = NULL,
624 .pfnAllocation = default_alloc_func,
625 .pfnReallocation = default_realloc_func,
626 .pfnFree = default_free_func,
627 };
628
629 VkResult anv_EnumerateInstanceExtensionProperties(
630 const char* pLayerName,
631 uint32_t* pPropertyCount,
632 VkExtensionProperties* pProperties)
633 {
634 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
635
636 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
637 if (anv_instance_extensions_supported.extensions[i]) {
638 vk_outarray_append(&out, prop) {
639 *prop = anv_instance_extensions[i];
640 }
641 }
642 }
643
644 return vk_outarray_status(&out);
645 }
646
647 VkResult anv_CreateInstance(
648 const VkInstanceCreateInfo* pCreateInfo,
649 const VkAllocationCallbacks* pAllocator,
650 VkInstance* pInstance)
651 {
652 struct anv_instance *instance;
653 VkResult result;
654
655 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
656
657 struct anv_instance_extension_table enabled_extensions = {};
658 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
659 int idx;
660 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
661 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
662 anv_instance_extensions[idx].extensionName) == 0)
663 break;
664 }
665
666 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
667 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
668
669 if (!anv_instance_extensions_supported.extensions[idx])
670 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
671
672 enabled_extensions.extensions[idx] = true;
673 }
674
675 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
676 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
677 if (!instance)
678 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
679
680 vk_object_base_init(NULL, &instance->base, VK_OBJECT_TYPE_INSTANCE);
681
682 if (pAllocator)
683 instance->alloc = *pAllocator;
684 else
685 instance->alloc = default_alloc;
686
687 instance->app_info = (struct anv_app_info) { .api_version = 0 };
688 if (pCreateInfo->pApplicationInfo) {
689 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
690
691 instance->app_info.app_name =
692 vk_strdup(&instance->alloc, app->pApplicationName,
693 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
694 instance->app_info.app_version = app->applicationVersion;
695
696 instance->app_info.engine_name =
697 vk_strdup(&instance->alloc, app->pEngineName,
698 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
699 instance->app_info.engine_version = app->engineVersion;
700
701 instance->app_info.api_version = app->apiVersion;
702 }
703
704 if (instance->app_info.api_version == 0)
705 instance->app_info.api_version = VK_API_VERSION_1_0;
706
707 instance->enabled_extensions = enabled_extensions;
708
709 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
710 /* Vulkan requires that entrypoints for extensions which have not been
711 * enabled must not be advertised.
712 */
713 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
714 &instance->enabled_extensions)) {
715 instance->dispatch.entrypoints[i] = NULL;
716 } else {
717 instance->dispatch.entrypoints[i] =
718 anv_instance_dispatch_table.entrypoints[i];
719 }
720 }
721
722 for (unsigned i = 0; i < ARRAY_SIZE(instance->physical_device_dispatch.entrypoints); i++) {
723 /* Vulkan requires that entrypoints for extensions which have not been
724 * enabled must not be advertised.
725 */
726 if (!anv_physical_device_entrypoint_is_enabled(i, instance->app_info.api_version,
727 &instance->enabled_extensions)) {
728 instance->physical_device_dispatch.entrypoints[i] = NULL;
729 } else {
730 instance->physical_device_dispatch.entrypoints[i] =
731 anv_physical_device_dispatch_table.entrypoints[i];
732 }
733 }
734
735 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
736 /* Vulkan requires that entrypoints for extensions which have not been
737 * enabled must not be advertised.
738 */
739 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
740 &instance->enabled_extensions, NULL)) {
741 instance->device_dispatch.entrypoints[i] = NULL;
742 } else {
743 instance->device_dispatch.entrypoints[i] =
744 anv_device_dispatch_table.entrypoints[i];
745 }
746 }
747
748 instance->physical_devices_enumerated = false;
749 list_inithead(&instance->physical_devices);
750
751 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
752 if (result != VK_SUCCESS) {
753 vk_free2(&default_alloc, pAllocator, instance);
754 return vk_error(result);
755 }
756
757 instance->pipeline_cache_enabled =
758 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
759
760 glsl_type_singleton_init_or_ref();
761
762 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
763
764 driParseOptionInfo(&instance->available_dri_options, anv_dri_options_xml);
765 driParseConfigFiles(&instance->dri_options, &instance->available_dri_options,
766 0, "anv", NULL,
767 instance->app_info.engine_name,
768 instance->app_info.engine_version);
769
770 *pInstance = anv_instance_to_handle(instance);
771
772 return VK_SUCCESS;
773 }
774
775 void anv_DestroyInstance(
776 VkInstance _instance,
777 const VkAllocationCallbacks* pAllocator)
778 {
779 ANV_FROM_HANDLE(anv_instance, instance, _instance);
780
781 if (!instance)
782 return;
783
784 list_for_each_entry_safe(struct anv_physical_device, pdevice,
785 &instance->physical_devices, link)
786 anv_physical_device_destroy(pdevice);
787
788 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
789 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
790
791 VG(VALGRIND_DESTROY_MEMPOOL(instance));
792
793 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
794
795 glsl_type_singleton_decref();
796
797 driDestroyOptionCache(&instance->dri_options);
798 driDestroyOptionInfo(&instance->available_dri_options);
799
800 vk_object_base_finish(&instance->base);
801 vk_free(&instance->alloc, instance);
802 }
803
804 static VkResult
805 anv_enumerate_physical_devices(struct anv_instance *instance)
806 {
807 if (instance->physical_devices_enumerated)
808 return VK_SUCCESS;
809
810 instance->physical_devices_enumerated = true;
811
812 /* TODO: Check for more devices ? */
813 drmDevicePtr devices[8];
814 int max_devices;
815
816 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
817 if (max_devices < 1)
818 return VK_SUCCESS;
819
820 VkResult result = VK_SUCCESS;
821 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
822 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
823 devices[i]->bustype == DRM_BUS_PCI &&
824 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
825
826 struct anv_physical_device *pdevice;
827 result = anv_physical_device_try_create(instance, devices[i],
828 &pdevice);
829 /* Incompatible DRM device, skip. */
830 if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
831 result = VK_SUCCESS;
832 continue;
833 }
834
835 /* Error creating the physical device, report the error. */
836 if (result != VK_SUCCESS)
837 break;
838
839 list_addtail(&pdevice->link, &instance->physical_devices);
840 }
841 }
842 drmFreeDevices(devices, max_devices);
843
844 /* If we successfully enumerated any devices, call it success */
845 return result;
846 }
847
848 VkResult anv_EnumeratePhysicalDevices(
849 VkInstance _instance,
850 uint32_t* pPhysicalDeviceCount,
851 VkPhysicalDevice* pPhysicalDevices)
852 {
853 ANV_FROM_HANDLE(anv_instance, instance, _instance);
854 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
855
856 VkResult result = anv_enumerate_physical_devices(instance);
857 if (result != VK_SUCCESS)
858 return result;
859
860 list_for_each_entry(struct anv_physical_device, pdevice,
861 &instance->physical_devices, link) {
862 vk_outarray_append(&out, i) {
863 *i = anv_physical_device_to_handle(pdevice);
864 }
865 }
866
867 return vk_outarray_status(&out);
868 }
869
870 VkResult anv_EnumeratePhysicalDeviceGroups(
871 VkInstance _instance,
872 uint32_t* pPhysicalDeviceGroupCount,
873 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
874 {
875 ANV_FROM_HANDLE(anv_instance, instance, _instance);
876 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
877 pPhysicalDeviceGroupCount);
878
879 VkResult result = anv_enumerate_physical_devices(instance);
880 if (result != VK_SUCCESS)
881 return result;
882
883 list_for_each_entry(struct anv_physical_device, pdevice,
884 &instance->physical_devices, link) {
885 vk_outarray_append(&out, p) {
886 p->physicalDeviceCount = 1;
887 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
888 p->physicalDevices[0] = anv_physical_device_to_handle(pdevice);
889 p->subsetAllocation = false;
890
891 vk_foreach_struct(ext, p->pNext)
892 anv_debug_ignored_stype(ext->sType);
893 }
894 }
895
896 return vk_outarray_status(&out);
897 }
898
899 void anv_GetPhysicalDeviceFeatures(
900 VkPhysicalDevice physicalDevice,
901 VkPhysicalDeviceFeatures* pFeatures)
902 {
903 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
904
905 *pFeatures = (VkPhysicalDeviceFeatures) {
906 .robustBufferAccess = true,
907 .fullDrawIndexUint32 = true,
908 .imageCubeArray = true,
909 .independentBlend = true,
910 .geometryShader = true,
911 .tessellationShader = true,
912 .sampleRateShading = true,
913 .dualSrcBlend = true,
914 .logicOp = true,
915 .multiDrawIndirect = true,
916 .drawIndirectFirstInstance = true,
917 .depthClamp = true,
918 .depthBiasClamp = true,
919 .fillModeNonSolid = true,
920 .depthBounds = pdevice->info.gen >= 12,
921 .wideLines = true,
922 .largePoints = true,
923 .alphaToOne = true,
924 .multiViewport = true,
925 .samplerAnisotropy = true,
926 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
927 pdevice->info.is_baytrail,
928 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
929 .textureCompressionBC = true,
930 .occlusionQueryPrecise = true,
931 .pipelineStatisticsQuery = true,
932 .fragmentStoresAndAtomics = true,
933 .shaderTessellationAndGeometryPointSize = true,
934 .shaderImageGatherExtended = true,
935 .shaderStorageImageExtendedFormats = true,
936 .shaderStorageImageMultisample = false,
937 .shaderStorageImageReadWithoutFormat = false,
938 .shaderStorageImageWriteWithoutFormat = true,
939 .shaderUniformBufferArrayDynamicIndexing = true,
940 .shaderSampledImageArrayDynamicIndexing = true,
941 .shaderStorageBufferArrayDynamicIndexing = true,
942 .shaderStorageImageArrayDynamicIndexing = true,
943 .shaderClipDistance = true,
944 .shaderCullDistance = true,
945 .shaderFloat64 = pdevice->info.gen >= 8 &&
946 pdevice->info.has_64bit_float,
947 .shaderInt64 = pdevice->info.gen >= 8 &&
948 pdevice->info.has_64bit_int,
949 .shaderInt16 = pdevice->info.gen >= 8,
950 .shaderResourceMinLod = pdevice->info.gen >= 9,
951 .variableMultisampleRate = true,
952 .inheritedQueries = true,
953 };
954
955 /* We can't do image stores in vec4 shaders */
956 pFeatures->vertexPipelineStoresAndAtomics =
957 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
958 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
959
960 struct anv_app_info *app_info = &pdevice->instance->app_info;
961
962 /* The new DOOM and Wolfenstein games require depthBounds without
963 * checking for it. They seem to run fine without it so just claim it's
964 * there and accept the consequences.
965 */
966 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
967 pFeatures->depthBounds = true;
968 }
969
970 static void
971 anv_get_physical_device_features_1_1(struct anv_physical_device *pdevice,
972 VkPhysicalDeviceVulkan11Features *f)
973 {
974 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES);
975
976 f->storageBuffer16BitAccess = pdevice->info.gen >= 8;
977 f->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
978 f->storagePushConstant16 = pdevice->info.gen >= 8;
979 f->storageInputOutput16 = false;
980 f->multiview = true;
981 f->multiviewGeometryShader = true;
982 f->multiviewTessellationShader = true;
983 f->variablePointersStorageBuffer = true;
984 f->variablePointers = true;
985 f->protectedMemory = false;
986 f->samplerYcbcrConversion = true;
987 f->shaderDrawParameters = true;
988 }
989
990 static void
991 anv_get_physical_device_features_1_2(struct anv_physical_device *pdevice,
992 VkPhysicalDeviceVulkan12Features *f)
993 {
994 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES);
995
996 f->samplerMirrorClampToEdge = true;
997 f->drawIndirectCount = true;
998 f->storageBuffer8BitAccess = pdevice->info.gen >= 8;
999 f->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
1000 f->storagePushConstant8 = pdevice->info.gen >= 8;
1001 f->shaderBufferInt64Atomics = pdevice->info.gen >= 9 &&
1002 pdevice->use_softpin;
1003 f->shaderSharedInt64Atomics = false;
1004 f->shaderFloat16 = pdevice->info.gen >= 8;
1005 f->shaderInt8 = pdevice->info.gen >= 8;
1006
1007 bool descIndexing = pdevice->has_a64_buffer_access &&
1008 pdevice->has_bindless_images;
1009 f->descriptorIndexing = descIndexing;
1010 f->shaderInputAttachmentArrayDynamicIndexing = false;
1011 f->shaderUniformTexelBufferArrayDynamicIndexing = descIndexing;
1012 f->shaderStorageTexelBufferArrayDynamicIndexing = descIndexing;
1013 f->shaderUniformBufferArrayNonUniformIndexing = false;
1014 f->shaderSampledImageArrayNonUniformIndexing = descIndexing;
1015 f->shaderStorageBufferArrayNonUniformIndexing = descIndexing;
1016 f->shaderStorageImageArrayNonUniformIndexing = descIndexing;
1017 f->shaderInputAttachmentArrayNonUniformIndexing = false;
1018 f->shaderUniformTexelBufferArrayNonUniformIndexing = descIndexing;
1019 f->shaderStorageTexelBufferArrayNonUniformIndexing = descIndexing;
1020 f->descriptorBindingUniformBufferUpdateAfterBind = false;
1021 f->descriptorBindingSampledImageUpdateAfterBind = descIndexing;
1022 f->descriptorBindingStorageImageUpdateAfterBind = descIndexing;
1023 f->descriptorBindingStorageBufferUpdateAfterBind = descIndexing;
1024 f->descriptorBindingUniformTexelBufferUpdateAfterBind = descIndexing;
1025 f->descriptorBindingStorageTexelBufferUpdateAfterBind = descIndexing;
1026 f->descriptorBindingUpdateUnusedWhilePending = descIndexing;
1027 f->descriptorBindingPartiallyBound = descIndexing;
1028 f->descriptorBindingVariableDescriptorCount = false;
1029 f->runtimeDescriptorArray = descIndexing;
1030
1031 f->samplerFilterMinmax = pdevice->info.gen >= 9;
1032 f->scalarBlockLayout = true;
1033 f->imagelessFramebuffer = true;
1034 f->uniformBufferStandardLayout = true;
1035 f->shaderSubgroupExtendedTypes = true;
1036 f->separateDepthStencilLayouts = true;
1037 f->hostQueryReset = true;
1038 f->timelineSemaphore = true;
1039 f->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1040 f->bufferDeviceAddressCaptureReplay = pdevice->has_a64_buffer_access;
1041 f->bufferDeviceAddressMultiDevice = false;
1042 f->vulkanMemoryModel = true;
1043 f->vulkanMemoryModelDeviceScope = true;
1044 f->vulkanMemoryModelAvailabilityVisibilityChains = true;
1045 f->shaderOutputViewportIndex = true;
1046 f->shaderOutputLayer = true;
1047 f->subgroupBroadcastDynamicId = true;
1048 }
1049
1050 void anv_GetPhysicalDeviceFeatures2(
1051 VkPhysicalDevice physicalDevice,
1052 VkPhysicalDeviceFeatures2* pFeatures)
1053 {
1054 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1055 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
1056
1057 VkPhysicalDeviceVulkan11Features core_1_1 = {
1058 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES,
1059 };
1060 anv_get_physical_device_features_1_1(pdevice, &core_1_1);
1061
1062 VkPhysicalDeviceVulkan12Features core_1_2 = {
1063 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
1064 };
1065 anv_get_physical_device_features_1_2(pdevice, &core_1_2);
1066
1067 #define CORE_FEATURE(major, minor, feature) \
1068 features->feature = core_##major##_##minor.feature
1069
1070
1071 vk_foreach_struct(ext, pFeatures->pNext) {
1072 switch (ext->sType) {
1073 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
1074 VkPhysicalDevice8BitStorageFeaturesKHR *features =
1075 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
1076 CORE_FEATURE(1, 2, storageBuffer8BitAccess);
1077 CORE_FEATURE(1, 2, uniformAndStorageBuffer8BitAccess);
1078 CORE_FEATURE(1, 2, storagePushConstant8);
1079 break;
1080 }
1081
1082 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
1083 VkPhysicalDevice16BitStorageFeatures *features =
1084 (VkPhysicalDevice16BitStorageFeatures *)ext;
1085 CORE_FEATURE(1, 1, storageBuffer16BitAccess);
1086 CORE_FEATURE(1, 1, uniformAndStorageBuffer16BitAccess);
1087 CORE_FEATURE(1, 1, storagePushConstant16);
1088 CORE_FEATURE(1, 1, storageInputOutput16);
1089 break;
1090 }
1091
1092 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
1093 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
1094 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1095 features->bufferDeviceAddressCaptureReplay = false;
1096 features->bufferDeviceAddressMultiDevice = false;
1097 break;
1098 }
1099
1100 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR: {
1101 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR *features = (void *)ext;
1102 CORE_FEATURE(1, 2, bufferDeviceAddress);
1103 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay);
1104 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice);
1105 break;
1106 }
1107
1108 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1109 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1110 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1111 features->computeDerivativeGroupQuads = true;
1112 features->computeDerivativeGroupLinear = true;
1113 break;
1114 }
1115
1116 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1117 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1118 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1119 features->conditionalRendering = pdevice->info.gen >= 8 ||
1120 pdevice->info.is_haswell;
1121 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1122 pdevice->info.is_haswell;
1123 break;
1124 }
1125
1126 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1127 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1128 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1129 features->depthClipEnable = true;
1130 break;
1131 }
1132
1133 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1134 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1135 CORE_FEATURE(1, 2, shaderFloat16);
1136 CORE_FEATURE(1, 2, shaderInt8);
1137 break;
1138 }
1139
1140 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1141 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1142 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1143 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1144 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1145 features->fragmentShaderShadingRateInterlock = false;
1146 break;
1147 }
1148
1149 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1150 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1151 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1152 CORE_FEATURE(1, 2, hostQueryReset);
1153 break;
1154 }
1155
1156 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1157 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1158 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1159 CORE_FEATURE(1, 2, shaderInputAttachmentArrayDynamicIndexing);
1160 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayDynamicIndexing);
1161 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayDynamicIndexing);
1162 CORE_FEATURE(1, 2, shaderUniformBufferArrayNonUniformIndexing);
1163 CORE_FEATURE(1, 2, shaderSampledImageArrayNonUniformIndexing);
1164 CORE_FEATURE(1, 2, shaderStorageBufferArrayNonUniformIndexing);
1165 CORE_FEATURE(1, 2, shaderStorageImageArrayNonUniformIndexing);
1166 CORE_FEATURE(1, 2, shaderInputAttachmentArrayNonUniformIndexing);
1167 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayNonUniformIndexing);
1168 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayNonUniformIndexing);
1169 CORE_FEATURE(1, 2, descriptorBindingUniformBufferUpdateAfterBind);
1170 CORE_FEATURE(1, 2, descriptorBindingSampledImageUpdateAfterBind);
1171 CORE_FEATURE(1, 2, descriptorBindingStorageImageUpdateAfterBind);
1172 CORE_FEATURE(1, 2, descriptorBindingStorageBufferUpdateAfterBind);
1173 CORE_FEATURE(1, 2, descriptorBindingUniformTexelBufferUpdateAfterBind);
1174 CORE_FEATURE(1, 2, descriptorBindingStorageTexelBufferUpdateAfterBind);
1175 CORE_FEATURE(1, 2, descriptorBindingUpdateUnusedWhilePending);
1176 CORE_FEATURE(1, 2, descriptorBindingPartiallyBound);
1177 CORE_FEATURE(1, 2, descriptorBindingVariableDescriptorCount);
1178 CORE_FEATURE(1, 2, runtimeDescriptorArray);
1179 break;
1180 }
1181
1182 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1183 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1184 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1185 features->indexTypeUint8 = true;
1186 break;
1187 }
1188
1189 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1190 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1191 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1192 features->inlineUniformBlock = true;
1193 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1194 break;
1195 }
1196
1197 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1198 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1199 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1200 features->rectangularLines = true;
1201 features->bresenhamLines = true;
1202 /* Support for Smooth lines with MSAA was removed on gen11. From the
1203 * BSpec section "Multisample ModesState" table for "AA Line Support
1204 * Requirements":
1205 *
1206 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1207 *
1208 * Fortunately, this isn't a case most people care about.
1209 */
1210 features->smoothLines = pdevice->info.gen < 10;
1211 features->stippledRectangularLines = false;
1212 features->stippledBresenhamLines = true;
1213 features->stippledSmoothLines = false;
1214 break;
1215 }
1216
1217 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1218 VkPhysicalDeviceMultiviewFeatures *features =
1219 (VkPhysicalDeviceMultiviewFeatures *)ext;
1220 CORE_FEATURE(1, 1, multiview);
1221 CORE_FEATURE(1, 1, multiviewGeometryShader);
1222 CORE_FEATURE(1, 1, multiviewTessellationShader);
1223 break;
1224 }
1225
1226 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1227 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1228 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1229 CORE_FEATURE(1, 2, imagelessFramebuffer);
1230 break;
1231 }
1232
1233 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1234 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1235 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1236 features->pipelineExecutableInfo = true;
1237 break;
1238 }
1239
1240 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRIVATE_DATA_FEATURES_EXT: {
1241 VkPhysicalDevicePrivateDataFeaturesEXT *features = (void *)ext;
1242 features->privateData = true;
1243 break;
1244 }
1245
1246 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1247 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1248 CORE_FEATURE(1, 1, protectedMemory);
1249 break;
1250 }
1251
1252 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT: {
1253 VkPhysicalDeviceRobustness2FeaturesEXT *features = (void *)ext;
1254 features->robustBufferAccess2 = true;
1255 features->robustImageAccess2 = true;
1256 features->nullDescriptor = true;
1257 break;
1258 }
1259
1260 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1261 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1262 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1263 CORE_FEATURE(1, 1, samplerYcbcrConversion);
1264 break;
1265 }
1266
1267 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1268 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1269 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1270 CORE_FEATURE(1, 2, scalarBlockLayout);
1271 break;
1272 }
1273
1274 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR: {
1275 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *features =
1276 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *)ext;
1277 CORE_FEATURE(1, 2, separateDepthStencilLayouts);
1278 break;
1279 }
1280
1281 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1282 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1283 CORE_FEATURE(1, 2, shaderBufferInt64Atomics);
1284 CORE_FEATURE(1, 2, shaderSharedInt64Atomics);
1285 break;
1286 }
1287
1288 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1289 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1290 features->shaderDemoteToHelperInvocation = true;
1291 break;
1292 }
1293
1294 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1295 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1296 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1297 features->shaderSubgroupClock = true;
1298 features->shaderDeviceClock = false;
1299 break;
1300 }
1301
1302 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1303 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1304 CORE_FEATURE(1, 1, shaderDrawParameters);
1305 break;
1306 }
1307
1308 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
1309 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
1310 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
1311 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes);
1312 break;
1313 }
1314
1315 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1316 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1317 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1318 features->subgroupSizeControl = true;
1319 features->computeFullSubgroups = true;
1320 break;
1321 }
1322
1323 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1324 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1325 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1326 features->texelBufferAlignment = true;
1327 break;
1328 }
1329
1330 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR: {
1331 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *features =
1332 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *) ext;
1333 CORE_FEATURE(1, 2, timelineSemaphore);
1334 break;
1335 }
1336
1337 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1338 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1339 CORE_FEATURE(1, 1, variablePointersStorageBuffer);
1340 CORE_FEATURE(1, 1, variablePointers);
1341 break;
1342 }
1343
1344 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1345 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1346 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1347 features->transformFeedback = true;
1348 features->geometryStreams = true;
1349 break;
1350 }
1351
1352 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1353 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1354 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1355 CORE_FEATURE(1, 2, uniformBufferStandardLayout);
1356 break;
1357 }
1358
1359 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1360 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1361 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1362 features->vertexAttributeInstanceRateDivisor = true;
1363 features->vertexAttributeInstanceRateZeroDivisor = true;
1364 break;
1365 }
1366
1367 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES:
1368 anv_get_physical_device_features_1_1(pdevice, (void *)ext);
1369 break;
1370
1371 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES:
1372 anv_get_physical_device_features_1_2(pdevice, (void *)ext);
1373 break;
1374
1375 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
1376 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
1377 CORE_FEATURE(1, 2, vulkanMemoryModel);
1378 CORE_FEATURE(1, 2, vulkanMemoryModelDeviceScope);
1379 CORE_FEATURE(1, 2, vulkanMemoryModelAvailabilityVisibilityChains);
1380 break;
1381 }
1382
1383 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1384 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1385 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1386 features->ycbcrImageArrays = true;
1387 break;
1388 }
1389
1390 default:
1391 anv_debug_ignored_stype(ext->sType);
1392 break;
1393 }
1394 }
1395
1396 #undef CORE_FEATURE
1397 }
1398
1399 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1400
1401 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1402 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1403
1404 void anv_GetPhysicalDeviceProperties(
1405 VkPhysicalDevice physicalDevice,
1406 VkPhysicalDeviceProperties* pProperties)
1407 {
1408 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1409 const struct gen_device_info *devinfo = &pdevice->info;
1410
1411 /* See assertions made when programming the buffer surface state. */
1412 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1413 (1ul << 30) : (1ul << 27);
1414
1415 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1416 const uint32_t max_textures =
1417 pdevice->has_bindless_images ? UINT16_MAX : 128;
1418 const uint32_t max_samplers =
1419 pdevice->has_bindless_samplers ? UINT16_MAX :
1420 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1421 const uint32_t max_images =
1422 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1423
1424 /* If we can use bindless for everything, claim a high per-stage limit,
1425 * otherwise use the binding table size, minus the slots reserved for
1426 * render targets and one slot for the descriptor buffer. */
1427 const uint32_t max_per_stage =
1428 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1429 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1430
1431 /* Limit max_threads to 64 for the GPGPU_WALKER command */
1432 const uint32_t max_workgroup_size = 32 * MIN2(64, devinfo->max_cs_threads);
1433
1434 VkSampleCountFlags sample_counts =
1435 isl_device_get_sample_counts(&pdevice->isl_dev);
1436
1437
1438 VkPhysicalDeviceLimits limits = {
1439 .maxImageDimension1D = (1 << 14),
1440 .maxImageDimension2D = (1 << 14),
1441 .maxImageDimension3D = (1 << 11),
1442 .maxImageDimensionCube = (1 << 14),
1443 .maxImageArrayLayers = (1 << 11),
1444 .maxTexelBufferElements = 128 * 1024 * 1024,
1445 .maxUniformBufferRange = (1ul << 27),
1446 .maxStorageBufferRange = max_raw_buffer_sz,
1447 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1448 .maxMemoryAllocationCount = UINT32_MAX,
1449 .maxSamplerAllocationCount = 64 * 1024,
1450 .bufferImageGranularity = 64, /* A cache line */
1451 .sparseAddressSpaceSize = 0,
1452 .maxBoundDescriptorSets = MAX_SETS,
1453 .maxPerStageDescriptorSamplers = max_samplers,
1454 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1455 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1456 .maxPerStageDescriptorSampledImages = max_textures,
1457 .maxPerStageDescriptorStorageImages = max_images,
1458 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1459 .maxPerStageResources = max_per_stage,
1460 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1461 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1462 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1463 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1464 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1465 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1466 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1467 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1468 .maxVertexInputAttributes = MAX_VBS,
1469 .maxVertexInputBindings = MAX_VBS,
1470 .maxVertexInputAttributeOffset = 2047,
1471 .maxVertexInputBindingStride = 2048,
1472 .maxVertexOutputComponents = 128,
1473 .maxTessellationGenerationLevel = 64,
1474 .maxTessellationPatchSize = 32,
1475 .maxTessellationControlPerVertexInputComponents = 128,
1476 .maxTessellationControlPerVertexOutputComponents = 128,
1477 .maxTessellationControlPerPatchOutputComponents = 128,
1478 .maxTessellationControlTotalOutputComponents = 2048,
1479 .maxTessellationEvaluationInputComponents = 128,
1480 .maxTessellationEvaluationOutputComponents = 128,
1481 .maxGeometryShaderInvocations = 32,
1482 .maxGeometryInputComponents = 64,
1483 .maxGeometryOutputComponents = 128,
1484 .maxGeometryOutputVertices = 256,
1485 .maxGeometryTotalOutputComponents = 1024,
1486 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1487 .maxFragmentOutputAttachments = 8,
1488 .maxFragmentDualSrcAttachments = 1,
1489 .maxFragmentCombinedOutputResources = 8,
1490 .maxComputeSharedMemorySize = 64 * 1024,
1491 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1492 .maxComputeWorkGroupInvocations = max_workgroup_size,
1493 .maxComputeWorkGroupSize = {
1494 max_workgroup_size,
1495 max_workgroup_size,
1496 max_workgroup_size,
1497 },
1498 .subPixelPrecisionBits = 8,
1499 .subTexelPrecisionBits = 8,
1500 .mipmapPrecisionBits = 8,
1501 .maxDrawIndexedIndexValue = UINT32_MAX,
1502 .maxDrawIndirectCount = UINT32_MAX,
1503 .maxSamplerLodBias = 16,
1504 .maxSamplerAnisotropy = 16,
1505 .maxViewports = MAX_VIEWPORTS,
1506 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1507 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1508 .viewportSubPixelBits = 13, /* We take a float? */
1509 .minMemoryMapAlignment = 4096, /* A page */
1510 /* The dataport requires texel alignment so we need to assume a worst
1511 * case of R32G32B32A32 which is 16 bytes.
1512 */
1513 .minTexelBufferOffsetAlignment = 16,
1514 .minUniformBufferOffsetAlignment = ANV_UBO_ALIGNMENT,
1515 .minStorageBufferOffsetAlignment = 4,
1516 .minTexelOffset = -8,
1517 .maxTexelOffset = 7,
1518 .minTexelGatherOffset = -32,
1519 .maxTexelGatherOffset = 31,
1520 .minInterpolationOffset = -0.5,
1521 .maxInterpolationOffset = 0.4375,
1522 .subPixelInterpolationOffsetBits = 4,
1523 .maxFramebufferWidth = (1 << 14),
1524 .maxFramebufferHeight = (1 << 14),
1525 .maxFramebufferLayers = (1 << 11),
1526 .framebufferColorSampleCounts = sample_counts,
1527 .framebufferDepthSampleCounts = sample_counts,
1528 .framebufferStencilSampleCounts = sample_counts,
1529 .framebufferNoAttachmentsSampleCounts = sample_counts,
1530 .maxColorAttachments = MAX_RTS,
1531 .sampledImageColorSampleCounts = sample_counts,
1532 .sampledImageIntegerSampleCounts = sample_counts,
1533 .sampledImageDepthSampleCounts = sample_counts,
1534 .sampledImageStencilSampleCounts = sample_counts,
1535 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1536 .maxSampleMaskWords = 1,
1537 .timestampComputeAndGraphics = true,
1538 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1539 .maxClipDistances = 8,
1540 .maxCullDistances = 8,
1541 .maxCombinedClipAndCullDistances = 8,
1542 .discreteQueuePriorities = 2,
1543 .pointSizeRange = { 0.125, 255.875 },
1544 .lineWidthRange = {
1545 0.0,
1546 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1547 2047.9921875 : 7.9921875,
1548 },
1549 .pointSizeGranularity = (1.0 / 8.0),
1550 .lineWidthGranularity = (1.0 / 128.0),
1551 .strictLines = false,
1552 .standardSampleLocations = true,
1553 .optimalBufferCopyOffsetAlignment = 128,
1554 .optimalBufferCopyRowPitchAlignment = 128,
1555 .nonCoherentAtomSize = 64,
1556 };
1557
1558 *pProperties = (VkPhysicalDeviceProperties) {
1559 .apiVersion = anv_physical_device_api_version(pdevice),
1560 .driverVersion = vk_get_driver_version(),
1561 .vendorID = 0x8086,
1562 .deviceID = pdevice->info.chipset_id,
1563 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1564 .limits = limits,
1565 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1566 };
1567
1568 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1569 "%s", pdevice->name);
1570 memcpy(pProperties->pipelineCacheUUID,
1571 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1572 }
1573
1574 static void
1575 anv_get_physical_device_properties_1_1(struct anv_physical_device *pdevice,
1576 VkPhysicalDeviceVulkan11Properties *p)
1577 {
1578 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES);
1579
1580 memcpy(p->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1581 memcpy(p->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1582 memset(p->deviceLUID, 0, VK_LUID_SIZE);
1583 p->deviceNodeMask = 0;
1584 p->deviceLUIDValid = false;
1585
1586 p->subgroupSize = BRW_SUBGROUP_SIZE;
1587 VkShaderStageFlags scalar_stages = 0;
1588 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1589 if (pdevice->compiler->scalar_stage[stage])
1590 scalar_stages |= mesa_to_vk_shader_stage(stage);
1591 }
1592 p->subgroupSupportedStages = scalar_stages;
1593 p->subgroupSupportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1594 VK_SUBGROUP_FEATURE_VOTE_BIT |
1595 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1596 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1597 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1598 VK_SUBGROUP_FEATURE_QUAD_BIT;
1599 if (pdevice->info.gen >= 8) {
1600 /* TODO: There's no technical reason why these can't be made to
1601 * work on gen7 but they don't at the moment so it's best to leave
1602 * the feature disabled than enabled and broken.
1603 */
1604 p->subgroupSupportedOperations |= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1605 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1606 }
1607 p->subgroupQuadOperationsInAllStages = pdevice->info.gen >= 8;
1608
1609 p->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1610 p->maxMultiviewViewCount = 16;
1611 p->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1612 p->protectedNoFault = false;
1613 /* This value doesn't matter for us today as our per-stage descriptors are
1614 * the real limit.
1615 */
1616 p->maxPerSetDescriptors = 1024;
1617 p->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1618 }
1619
1620 static void
1621 anv_get_physical_device_properties_1_2(struct anv_physical_device *pdevice,
1622 VkPhysicalDeviceVulkan12Properties *p)
1623 {
1624 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES);
1625
1626 p->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1627 memset(p->driverName, 0, sizeof(p->driverName));
1628 snprintf(p->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1629 "Intel open-source Mesa driver");
1630 memset(p->driverInfo, 0, sizeof(p->driverInfo));
1631 snprintf(p->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1632 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1633 p->conformanceVersion = (VkConformanceVersionKHR) {
1634 .major = 1,
1635 .minor = 2,
1636 .subminor = 0,
1637 .patch = 0,
1638 };
1639
1640 p->denormBehaviorIndependence =
1641 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
1642 p->roundingModeIndependence =
1643 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
1644
1645 /* Broadwell does not support HF denorms and there are restrictions
1646 * other gens. According to Kabylake's PRM:
1647 *
1648 * "math - Extended Math Function
1649 * [...]
1650 * Restriction : Half-float denorms are always retained."
1651 */
1652 p->shaderDenormFlushToZeroFloat16 = false;
1653 p->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
1654 p->shaderRoundingModeRTEFloat16 = true;
1655 p->shaderRoundingModeRTZFloat16 = true;
1656 p->shaderSignedZeroInfNanPreserveFloat16 = true;
1657
1658 p->shaderDenormFlushToZeroFloat32 = true;
1659 p->shaderDenormPreserveFloat32 = true;
1660 p->shaderRoundingModeRTEFloat32 = true;
1661 p->shaderRoundingModeRTZFloat32 = true;
1662 p->shaderSignedZeroInfNanPreserveFloat32 = true;
1663
1664 p->shaderDenormFlushToZeroFloat64 = true;
1665 p->shaderDenormPreserveFloat64 = true;
1666 p->shaderRoundingModeRTEFloat64 = true;
1667 p->shaderRoundingModeRTZFloat64 = true;
1668 p->shaderSignedZeroInfNanPreserveFloat64 = true;
1669
1670 /* It's a bit hard to exactly map our implementation to the limits
1671 * described here. The bindless surface handle in the extended
1672 * message descriptors is 20 bits and it's an index into the table of
1673 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1674 * address. Given that most things consume two surface states per
1675 * view (general/sampled for textures and write-only/read-write for
1676 * images), we claim 2^19 things.
1677 *
1678 * For SSBOs, we just use A64 messages so there is no real limit
1679 * there beyond the limit on the total size of a descriptor set.
1680 */
1681 const unsigned max_bindless_views = 1 << 19;
1682 p->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1683 p->shaderUniformBufferArrayNonUniformIndexingNative = false;
1684 p->shaderSampledImageArrayNonUniformIndexingNative = false;
1685 p->shaderStorageBufferArrayNonUniformIndexingNative = true;
1686 p->shaderStorageImageArrayNonUniformIndexingNative = false;
1687 p->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1688 p->robustBufferAccessUpdateAfterBind = true;
1689 p->quadDivergentImplicitLod = false;
1690 p->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1691 p->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1692 p->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1693 p->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1694 p->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1695 p->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1696 p->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1697 p->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1698 p->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1699 p->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1700 p->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1701 p->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1702 p->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1703 p->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1704 p->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1705
1706 /* We support all of the depth resolve modes */
1707 p->supportedDepthResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1708 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1709 VK_RESOLVE_MODE_MIN_BIT_KHR |
1710 VK_RESOLVE_MODE_MAX_BIT_KHR;
1711 /* Average doesn't make sense for stencil so we don't support that */
1712 p->supportedStencilResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1713 if (pdevice->info.gen >= 8) {
1714 /* The advanced stencil resolve modes currently require stencil
1715 * sampling be supported by the hardware.
1716 */
1717 p->supportedStencilResolveModes |= VK_RESOLVE_MODE_MIN_BIT_KHR |
1718 VK_RESOLVE_MODE_MAX_BIT_KHR;
1719 }
1720 p->independentResolveNone = true;
1721 p->independentResolve = true;
1722
1723 p->filterMinmaxSingleComponentFormats = pdevice->info.gen >= 9;
1724 p->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1725
1726 p->maxTimelineSemaphoreValueDifference = UINT64_MAX;
1727
1728 p->framebufferIntegerColorSampleCounts =
1729 isl_device_get_sample_counts(&pdevice->isl_dev);
1730 }
1731
1732 void anv_GetPhysicalDeviceProperties2(
1733 VkPhysicalDevice physicalDevice,
1734 VkPhysicalDeviceProperties2* pProperties)
1735 {
1736 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1737
1738 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1739
1740 VkPhysicalDeviceVulkan11Properties core_1_1 = {
1741 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES,
1742 };
1743 anv_get_physical_device_properties_1_1(pdevice, &core_1_1);
1744
1745 VkPhysicalDeviceVulkan12Properties core_1_2 = {
1746 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES,
1747 };
1748 anv_get_physical_device_properties_1_2(pdevice, &core_1_2);
1749
1750 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1751 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1752 sizeof(core_##major##_##minor.core_property))
1753
1754 #define CORE_PROPERTY(major, minor, property) \
1755 CORE_RENAMED_PROPERTY(major, minor, property, property)
1756
1757 vk_foreach_struct(ext, pProperties->pNext) {
1758 switch (ext->sType) {
1759 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1760 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *properties =
1761 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1762 CORE_PROPERTY(1, 2, supportedDepthResolveModes);
1763 CORE_PROPERTY(1, 2, supportedStencilResolveModes);
1764 CORE_PROPERTY(1, 2, independentResolveNone);
1765 CORE_PROPERTY(1, 2, independentResolve);
1766 break;
1767 }
1768
1769 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1770 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *properties =
1771 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1772 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools);
1773 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative);
1774 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative);
1775 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative);
1776 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative);
1777 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative);
1778 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind);
1779 CORE_PROPERTY(1, 2, quadDivergentImplicitLod);
1780 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers);
1781 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers);
1782 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers);
1783 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages);
1784 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages);
1785 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments);
1786 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources);
1787 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers);
1788 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers);
1789 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic);
1790 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers);
1791 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic);
1792 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages);
1793 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages);
1794 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments);
1795 break;
1796 }
1797
1798 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1799 VkPhysicalDeviceDriverPropertiesKHR *properties =
1800 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1801 CORE_PROPERTY(1, 2, driverID);
1802 CORE_PROPERTY(1, 2, driverName);
1803 CORE_PROPERTY(1, 2, driverInfo);
1804 CORE_PROPERTY(1, 2, conformanceVersion);
1805 break;
1806 }
1807
1808 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1809 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1810 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1811 /* Userptr needs page aligned memory. */
1812 props->minImportedHostPointerAlignment = 4096;
1813 break;
1814 }
1815
1816 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1817 VkPhysicalDeviceIDProperties *properties =
1818 (VkPhysicalDeviceIDProperties *)ext;
1819 CORE_PROPERTY(1, 1, deviceUUID);
1820 CORE_PROPERTY(1, 1, driverUUID);
1821 CORE_PROPERTY(1, 1, deviceLUID);
1822 CORE_PROPERTY(1, 1, deviceLUIDValid);
1823 break;
1824 }
1825
1826 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1827 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1828 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1829 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1830 props->maxPerStageDescriptorInlineUniformBlocks =
1831 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1832 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1833 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1834 props->maxDescriptorSetInlineUniformBlocks =
1835 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1836 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1837 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1838 break;
1839 }
1840
1841 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1842 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1843 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1844 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1845 * Sampling Rules - Legacy Mode", it says the following:
1846 *
1847 * "Note that the device divides a pixel into a 16x16 array of
1848 * subpixels, referenced by their upper left corners."
1849 *
1850 * This is the only known reference in the PRMs to the subpixel
1851 * precision of line rasterization and a "16x16 array of subpixels"
1852 * implies 4 subpixel precision bits. Empirical testing has shown
1853 * that 4 subpixel precision bits applies to all line rasterization
1854 * types.
1855 */
1856 props->lineSubPixelPrecisionBits = 4;
1857 break;
1858 }
1859
1860 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1861 VkPhysicalDeviceMaintenance3Properties *properties =
1862 (VkPhysicalDeviceMaintenance3Properties *)ext;
1863 /* This value doesn't matter for us today as our per-stage
1864 * descriptors are the real limit.
1865 */
1866 CORE_PROPERTY(1, 1, maxPerSetDescriptors);
1867 CORE_PROPERTY(1, 1, maxMemoryAllocationSize);
1868 break;
1869 }
1870
1871 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1872 VkPhysicalDeviceMultiviewProperties *properties =
1873 (VkPhysicalDeviceMultiviewProperties *)ext;
1874 CORE_PROPERTY(1, 1, maxMultiviewViewCount);
1875 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex);
1876 break;
1877 }
1878
1879 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1880 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1881 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1882 properties->pciDomain = pdevice->pci_info.domain;
1883 properties->pciBus = pdevice->pci_info.bus;
1884 properties->pciDevice = pdevice->pci_info.device;
1885 properties->pciFunction = pdevice->pci_info.function;
1886 break;
1887 }
1888
1889 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1890 VkPhysicalDevicePointClippingProperties *properties =
1891 (VkPhysicalDevicePointClippingProperties *) ext;
1892 CORE_PROPERTY(1, 1, pointClippingBehavior);
1893 break;
1894 }
1895
1896 #pragma GCC diagnostic push
1897 #pragma GCC diagnostic ignored "-Wswitch"
1898 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1899 VkPhysicalDevicePresentationPropertiesANDROID *props =
1900 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1901 props->sharedImage = VK_FALSE;
1902 break;
1903 }
1904 #pragma GCC diagnostic pop
1905
1906 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1907 VkPhysicalDeviceProtectedMemoryProperties *properties =
1908 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1909 CORE_PROPERTY(1, 1, protectedNoFault);
1910 break;
1911 }
1912
1913 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1914 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1915 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1916 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1917 break;
1918 }
1919
1920 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_PROPERTIES_EXT: {
1921 VkPhysicalDeviceRobustness2PropertiesEXT *properties = (void *)ext;
1922 properties->robustStorageBufferAccessSizeAlignment =
1923 ANV_SSBO_BOUNDS_CHECK_ALIGNMENT;
1924 properties->robustUniformBufferAccessSizeAlignment =
1925 ANV_UBO_ALIGNMENT;
1926 break;
1927 }
1928
1929 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1930 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1931 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1932 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping);
1933 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats);
1934 break;
1935 }
1936
1937 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1938 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1939 CORE_PROPERTY(1, 1, subgroupSize);
1940 CORE_RENAMED_PROPERTY(1, 1, supportedStages,
1941 subgroupSupportedStages);
1942 CORE_RENAMED_PROPERTY(1, 1, supportedOperations,
1943 subgroupSupportedOperations);
1944 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages,
1945 subgroupQuadOperationsInAllStages);
1946 break;
1947 }
1948
1949 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1950 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1951 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1952 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1953 props->minSubgroupSize = 8;
1954 props->maxSubgroupSize = 32;
1955 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1956 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1957 break;
1958 }
1959 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
1960 VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
1961 CORE_PROPERTY(1, 2, denormBehaviorIndependence);
1962 CORE_PROPERTY(1, 2, roundingModeIndependence);
1963 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16);
1964 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16);
1965 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16);
1966 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16);
1967 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16);
1968 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32);
1969 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32);
1970 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32);
1971 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32);
1972 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32);
1973 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64);
1974 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64);
1975 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64);
1976 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64);
1977 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64);
1978 break;
1979 }
1980
1981 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1982 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1983 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1984
1985 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1986 * Base Address:
1987 *
1988 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1989 * specifies the base address of the first element of the surface,
1990 * computed in software by adding the surface base address to the
1991 * byte offset of the element in the buffer. The base address must
1992 * be aligned to element size."
1993 *
1994 * The typed dataport messages require that things be texel aligned.
1995 * Otherwise, we may just load/store the wrong data or, in the worst
1996 * case, there may be hangs.
1997 */
1998 props->storageTexelBufferOffsetAlignmentBytes = 16;
1999 props->storageTexelBufferOffsetSingleTexelAlignment = true;
2000
2001 /* The sampler, however, is much more forgiving and it can handle
2002 * arbitrary byte alignment for linear and buffer surfaces. It's
2003 * hard to find a good PRM citation for this but years of empirical
2004 * experience demonstrate that this is true.
2005 */
2006 props->uniformTexelBufferOffsetAlignmentBytes = 1;
2007 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
2008 break;
2009 }
2010
2011 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR: {
2012 VkPhysicalDeviceTimelineSemaphorePropertiesKHR *properties =
2013 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR *) ext;
2014 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference);
2015 break;
2016 }
2017
2018 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
2019 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
2020 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
2021
2022 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
2023 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
2024 props->maxTransformFeedbackBufferSize = (1ull << 32);
2025 props->maxTransformFeedbackStreamDataSize = 128 * 4;
2026 props->maxTransformFeedbackBufferDataSize = 128 * 4;
2027 props->maxTransformFeedbackBufferDataStride = 2048;
2028 props->transformFeedbackQueries = true;
2029 props->transformFeedbackStreamsLinesTriangles = false;
2030 props->transformFeedbackRasterizationStreamSelect = false;
2031 props->transformFeedbackDraw = true;
2032 break;
2033 }
2034
2035 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
2036 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
2037 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
2038 /* We have to restrict this a bit for multiview */
2039 props->maxVertexAttribDivisor = UINT32_MAX / 16;
2040 break;
2041 }
2042
2043 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES:
2044 anv_get_physical_device_properties_1_1(pdevice, (void *)ext);
2045 break;
2046
2047 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES:
2048 anv_get_physical_device_properties_1_2(pdevice, (void *)ext);
2049 break;
2050
2051 default:
2052 anv_debug_ignored_stype(ext->sType);
2053 break;
2054 }
2055 }
2056
2057 #undef CORE_RENAMED_PROPERTY
2058 #undef CORE_PROPERTY
2059 }
2060
2061 /* We support exactly one queue family. */
2062 static const VkQueueFamilyProperties
2063 anv_queue_family_properties = {
2064 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
2065 VK_QUEUE_COMPUTE_BIT |
2066 VK_QUEUE_TRANSFER_BIT,
2067 .queueCount = 1,
2068 .timestampValidBits = 36, /* XXX: Real value here */
2069 .minImageTransferGranularity = { 1, 1, 1 },
2070 };
2071
2072 void anv_GetPhysicalDeviceQueueFamilyProperties(
2073 VkPhysicalDevice physicalDevice,
2074 uint32_t* pCount,
2075 VkQueueFamilyProperties* pQueueFamilyProperties)
2076 {
2077 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
2078
2079 vk_outarray_append(&out, p) {
2080 *p = anv_queue_family_properties;
2081 }
2082 }
2083
2084 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2085 VkPhysicalDevice physicalDevice,
2086 uint32_t* pQueueFamilyPropertyCount,
2087 VkQueueFamilyProperties2* pQueueFamilyProperties)
2088 {
2089
2090 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
2091
2092 vk_outarray_append(&out, p) {
2093 p->queueFamilyProperties = anv_queue_family_properties;
2094
2095 vk_foreach_struct(s, p->pNext) {
2096 anv_debug_ignored_stype(s->sType);
2097 }
2098 }
2099 }
2100
2101 void anv_GetPhysicalDeviceMemoryProperties(
2102 VkPhysicalDevice physicalDevice,
2103 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
2104 {
2105 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2106
2107 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
2108 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
2109 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
2110 .propertyFlags = physical_device->memory.types[i].propertyFlags,
2111 .heapIndex = physical_device->memory.types[i].heapIndex,
2112 };
2113 }
2114
2115 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
2116 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
2117 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
2118 .size = physical_device->memory.heaps[i].size,
2119 .flags = physical_device->memory.heaps[i].flags,
2120 };
2121 }
2122 }
2123
2124 static void
2125 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
2126 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
2127 {
2128 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2129 uint64_t sys_available = get_available_system_memory();
2130 assert(sys_available > 0);
2131
2132 VkDeviceSize total_heaps_size = 0;
2133 for (size_t i = 0; i < device->memory.heap_count; i++)
2134 total_heaps_size += device->memory.heaps[i].size;
2135
2136 for (size_t i = 0; i < device->memory.heap_count; i++) {
2137 VkDeviceSize heap_size = device->memory.heaps[i].size;
2138 VkDeviceSize heap_used = device->memory.heaps[i].used;
2139 VkDeviceSize heap_budget;
2140
2141 double heap_proportion = (double) heap_size / total_heaps_size;
2142 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
2143
2144 /*
2145 * Let's not incite the app to starve the system: report at most 90% of
2146 * available system memory.
2147 */
2148 uint64_t heap_available = sys_available_prop * 9 / 10;
2149 heap_budget = MIN2(heap_size, heap_used + heap_available);
2150
2151 /*
2152 * Round down to the nearest MB
2153 */
2154 heap_budget &= ~((1ull << 20) - 1);
2155
2156 /*
2157 * The heapBudget value must be non-zero for array elements less than
2158 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2159 * value must be less than or equal to VkMemoryHeap::size for each heap.
2160 */
2161 assert(0 < heap_budget && heap_budget <= heap_size);
2162
2163 memoryBudget->heapUsage[i] = heap_used;
2164 memoryBudget->heapBudget[i] = heap_budget;
2165 }
2166
2167 /* The heapBudget and heapUsage values must be zero for array elements
2168 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2169 */
2170 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
2171 memoryBudget->heapBudget[i] = 0;
2172 memoryBudget->heapUsage[i] = 0;
2173 }
2174 }
2175
2176 void anv_GetPhysicalDeviceMemoryProperties2(
2177 VkPhysicalDevice physicalDevice,
2178 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
2179 {
2180 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
2181 &pMemoryProperties->memoryProperties);
2182
2183 vk_foreach_struct(ext, pMemoryProperties->pNext) {
2184 switch (ext->sType) {
2185 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
2186 anv_get_memory_budget(physicalDevice, (void*)ext);
2187 break;
2188 default:
2189 anv_debug_ignored_stype(ext->sType);
2190 break;
2191 }
2192 }
2193 }
2194
2195 void
2196 anv_GetDeviceGroupPeerMemoryFeatures(
2197 VkDevice device,
2198 uint32_t heapIndex,
2199 uint32_t localDeviceIndex,
2200 uint32_t remoteDeviceIndex,
2201 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
2202 {
2203 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
2204 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2205 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2206 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2207 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2208 }
2209
2210 PFN_vkVoidFunction anv_GetInstanceProcAddr(
2211 VkInstance _instance,
2212 const char* pName)
2213 {
2214 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2215
2216 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2217 * when we have to return valid function pointers, NULL, or it's left
2218 * undefined. See the table for exact details.
2219 */
2220 if (pName == NULL)
2221 return NULL;
2222
2223 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2224 if (strcmp(pName, "vk" #entrypoint) == 0) \
2225 return (PFN_vkVoidFunction)anv_##entrypoint
2226
2227 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
2228 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
2229 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
2230 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
2231
2232 /* GetInstanceProcAddr() can also be called with a NULL instance.
2233 * See https://gitlab.khronos.org/vulkan/vulkan/issues/2057
2234 */
2235 LOOKUP_ANV_ENTRYPOINT(GetInstanceProcAddr);
2236
2237 #undef LOOKUP_ANV_ENTRYPOINT
2238
2239 if (instance == NULL)
2240 return NULL;
2241
2242 int idx = anv_get_instance_entrypoint_index(pName);
2243 if (idx >= 0)
2244 return instance->dispatch.entrypoints[idx];
2245
2246 idx = anv_get_physical_device_entrypoint_index(pName);
2247 if (idx >= 0)
2248 return instance->physical_device_dispatch.entrypoints[idx];
2249
2250 idx = anv_get_device_entrypoint_index(pName);
2251 if (idx >= 0)
2252 return instance->device_dispatch.entrypoints[idx];
2253
2254 return NULL;
2255 }
2256
2257 /* With version 1+ of the loader interface the ICD should expose
2258 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2259 */
2260 PUBLIC
2261 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2262 VkInstance instance,
2263 const char* pName);
2264
2265 PUBLIC
2266 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2267 VkInstance instance,
2268 const char* pName)
2269 {
2270 return anv_GetInstanceProcAddr(instance, pName);
2271 }
2272
2273 PFN_vkVoidFunction anv_GetDeviceProcAddr(
2274 VkDevice _device,
2275 const char* pName)
2276 {
2277 ANV_FROM_HANDLE(anv_device, device, _device);
2278
2279 if (!device || !pName)
2280 return NULL;
2281
2282 int idx = anv_get_device_entrypoint_index(pName);
2283 if (idx < 0)
2284 return NULL;
2285
2286 return device->dispatch.entrypoints[idx];
2287 }
2288
2289 /* With version 4+ of the loader interface the ICD should expose
2290 * vk_icdGetPhysicalDeviceProcAddr()
2291 */
2292 PUBLIC
2293 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2294 VkInstance _instance,
2295 const char* pName);
2296
2297 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2298 VkInstance _instance,
2299 const char* pName)
2300 {
2301 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2302
2303 if (!pName || !instance)
2304 return NULL;
2305
2306 int idx = anv_get_physical_device_entrypoint_index(pName);
2307 if (idx < 0)
2308 return NULL;
2309
2310 return instance->physical_device_dispatch.entrypoints[idx];
2311 }
2312
2313
2314 VkResult
2315 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
2316 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
2317 const VkAllocationCallbacks* pAllocator,
2318 VkDebugReportCallbackEXT* pCallback)
2319 {
2320 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2321 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2322 pCreateInfo, pAllocator, &instance->alloc,
2323 pCallback);
2324 }
2325
2326 void
2327 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
2328 VkDebugReportCallbackEXT _callback,
2329 const VkAllocationCallbacks* pAllocator)
2330 {
2331 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2332 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2333 _callback, pAllocator, &instance->alloc);
2334 }
2335
2336 void
2337 anv_DebugReportMessageEXT(VkInstance _instance,
2338 VkDebugReportFlagsEXT flags,
2339 VkDebugReportObjectTypeEXT objectType,
2340 uint64_t object,
2341 size_t location,
2342 int32_t messageCode,
2343 const char* pLayerPrefix,
2344 const char* pMessage)
2345 {
2346 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2347 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2348 object, location, messageCode, pLayerPrefix, pMessage);
2349 }
2350
2351 static struct anv_state
2352 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2353 {
2354 struct anv_state state;
2355
2356 state = anv_state_pool_alloc(pool, size, align);
2357 memcpy(state.map, p, size);
2358
2359 return state;
2360 }
2361
2362 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2363 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2364 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2365 * color as a separate entry /after/ the float color. The layout of this entry
2366 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2367 *
2368 * Since we don't know the format/bpp, we can't make any of the border colors
2369 * containing '1' work for all formats, as it would be in the wrong place for
2370 * some of them. We opt to make 32-bit integers work as this seems like the
2371 * most common option. Fortunately, transparent black works regardless, as
2372 * all zeroes is the same in every bit-size.
2373 */
2374 struct hsw_border_color {
2375 float float32[4];
2376 uint32_t _pad0[12];
2377 uint32_t uint32[4];
2378 uint32_t _pad1[108];
2379 };
2380
2381 struct gen8_border_color {
2382 union {
2383 float float32[4];
2384 uint32_t uint32[4];
2385 };
2386 /* Pad out to 64 bytes */
2387 uint32_t _pad[12];
2388 };
2389
2390 static void
2391 anv_device_init_border_colors(struct anv_device *device)
2392 {
2393 if (device->info.is_haswell) {
2394 static const struct hsw_border_color border_colors[] = {
2395 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2396 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2397 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2398 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2399 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2400 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2401 };
2402
2403 device->border_colors =
2404 anv_state_pool_emit_data(&device->dynamic_state_pool,
2405 sizeof(border_colors), 512, border_colors);
2406 } else {
2407 static const struct gen8_border_color border_colors[] = {
2408 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2409 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2410 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2411 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2412 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2413 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2414 };
2415
2416 device->border_colors =
2417 anv_state_pool_emit_data(&device->dynamic_state_pool,
2418 sizeof(border_colors), 64, border_colors);
2419 }
2420 }
2421
2422 static VkResult
2423 anv_device_init_trivial_batch(struct anv_device *device)
2424 {
2425 VkResult result = anv_device_alloc_bo(device, 4096,
2426 ANV_BO_ALLOC_MAPPED,
2427 0 /* explicit_address */,
2428 &device->trivial_batch_bo);
2429 if (result != VK_SUCCESS)
2430 return result;
2431
2432 struct anv_batch batch = {
2433 .start = device->trivial_batch_bo->map,
2434 .next = device->trivial_batch_bo->map,
2435 .end = device->trivial_batch_bo->map + 4096,
2436 };
2437
2438 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2439 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2440
2441 if (!device->info.has_llc)
2442 gen_clflush_range(batch.start, batch.next - batch.start);
2443
2444 return VK_SUCCESS;
2445 }
2446
2447 VkResult anv_EnumerateDeviceExtensionProperties(
2448 VkPhysicalDevice physicalDevice,
2449 const char* pLayerName,
2450 uint32_t* pPropertyCount,
2451 VkExtensionProperties* pProperties)
2452 {
2453 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2454 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2455
2456 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2457 if (device->supported_extensions.extensions[i]) {
2458 vk_outarray_append(&out, prop) {
2459 *prop = anv_device_extensions[i];
2460 }
2461 }
2462 }
2463
2464 return vk_outarray_status(&out);
2465 }
2466
2467 static void
2468 anv_device_init_dispatch(struct anv_device *device)
2469 {
2470 const struct anv_instance *instance = device->physical->instance;
2471
2472 const struct anv_device_dispatch_table *genX_table;
2473 switch (device->info.gen) {
2474 case 12:
2475 genX_table = &gen12_device_dispatch_table;
2476 break;
2477 case 11:
2478 genX_table = &gen11_device_dispatch_table;
2479 break;
2480 case 10:
2481 genX_table = &gen10_device_dispatch_table;
2482 break;
2483 case 9:
2484 genX_table = &gen9_device_dispatch_table;
2485 break;
2486 case 8:
2487 genX_table = &gen8_device_dispatch_table;
2488 break;
2489 case 7:
2490 if (device->info.is_haswell)
2491 genX_table = &gen75_device_dispatch_table;
2492 else
2493 genX_table = &gen7_device_dispatch_table;
2494 break;
2495 default:
2496 unreachable("unsupported gen\n");
2497 }
2498
2499 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2500 /* Vulkan requires that entrypoints for extensions which have not been
2501 * enabled must not be advertised.
2502 */
2503 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
2504 &instance->enabled_extensions,
2505 &device->enabled_extensions)) {
2506 device->dispatch.entrypoints[i] = NULL;
2507 } else if (genX_table->entrypoints[i]) {
2508 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2509 } else {
2510 device->dispatch.entrypoints[i] =
2511 anv_device_dispatch_table.entrypoints[i];
2512 }
2513 }
2514 }
2515
2516 static int
2517 vk_priority_to_gen(int priority)
2518 {
2519 switch (priority) {
2520 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2521 return GEN_CONTEXT_LOW_PRIORITY;
2522 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2523 return GEN_CONTEXT_MEDIUM_PRIORITY;
2524 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2525 return GEN_CONTEXT_HIGH_PRIORITY;
2526 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2527 return GEN_CONTEXT_REALTIME_PRIORITY;
2528 default:
2529 unreachable("Invalid priority");
2530 }
2531 }
2532
2533 static VkResult
2534 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2535 {
2536 VkResult result = anv_device_alloc_bo(device, 4096,
2537 ANV_BO_ALLOC_MAPPED,
2538 0 /* explicit_address */,
2539 &device->hiz_clear_bo);
2540 if (result != VK_SUCCESS)
2541 return result;
2542
2543 union isl_color_value hiz_clear = { .u32 = { 0, } };
2544 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2545
2546 memcpy(device->hiz_clear_bo->map, hiz_clear.u32, sizeof(hiz_clear.u32));
2547
2548 if (!device->info.has_llc)
2549 gen_clflush_range(device->hiz_clear_bo->map, sizeof(hiz_clear.u32));
2550
2551 return VK_SUCCESS;
2552 }
2553
2554 static bool
2555 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2556 struct anv_block_pool *pool,
2557 uint64_t address)
2558 {
2559 anv_block_pool_foreach_bo(bo, pool) {
2560 uint64_t bo_address = gen_48b_address(bo->offset);
2561 if (address >= bo_address && address < (bo_address + bo->size)) {
2562 *ret = (struct gen_batch_decode_bo) {
2563 .addr = bo_address,
2564 .size = bo->size,
2565 .map = bo->map,
2566 };
2567 return true;
2568 }
2569 }
2570 return false;
2571 }
2572
2573 /* Finding a buffer for batch decoding */
2574 static struct gen_batch_decode_bo
2575 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2576 {
2577 struct anv_device *device = v_batch;
2578 struct gen_batch_decode_bo ret_bo = {};
2579
2580 assert(ppgtt);
2581
2582 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2583 return ret_bo;
2584 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2585 return ret_bo;
2586 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2587 return ret_bo;
2588 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2589 return ret_bo;
2590
2591 if (!device->cmd_buffer_being_decoded)
2592 return (struct gen_batch_decode_bo) { };
2593
2594 struct anv_batch_bo **bo;
2595
2596 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2597 /* The decoder zeroes out the top 16 bits, so we need to as well */
2598 uint64_t bo_address = (*bo)->bo->offset & (~0ull >> 16);
2599
2600 if (address >= bo_address && address < bo_address + (*bo)->bo->size) {
2601 return (struct gen_batch_decode_bo) {
2602 .addr = bo_address,
2603 .size = (*bo)->bo->size,
2604 .map = (*bo)->bo->map,
2605 };
2606 }
2607 }
2608
2609 return (struct gen_batch_decode_bo) { };
2610 }
2611
2612 struct gen_aux_map_buffer {
2613 struct gen_buffer base;
2614 struct anv_state state;
2615 };
2616
2617 static struct gen_buffer *
2618 gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2619 {
2620 struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
2621 if (!buf)
2622 return NULL;
2623
2624 struct anv_device *device = (struct anv_device*)driver_ctx;
2625 assert(device->physical->supports_48bit_addresses &&
2626 device->physical->use_softpin);
2627
2628 struct anv_state_pool *pool = &device->dynamic_state_pool;
2629 buf->state = anv_state_pool_alloc(pool, size, size);
2630
2631 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2632 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2633 buf->base.map = buf->state.map;
2634 buf->base.driver_bo = &buf->state;
2635 return &buf->base;
2636 }
2637
2638 static void
2639 gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
2640 {
2641 struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
2642 struct anv_device *device = (struct anv_device*)driver_ctx;
2643 struct anv_state_pool *pool = &device->dynamic_state_pool;
2644 anv_state_pool_free(pool, buf->state);
2645 free(buf);
2646 }
2647
2648 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
2649 .alloc = gen_aux_map_buffer_alloc,
2650 .free = gen_aux_map_buffer_free,
2651 };
2652
2653 static VkResult
2654 check_physical_device_features(VkPhysicalDevice physicalDevice,
2655 const VkPhysicalDeviceFeatures *features)
2656 {
2657 VkPhysicalDeviceFeatures supported_features;
2658 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2659 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2660 VkBool32 *enabled_feature = (VkBool32 *)features;
2661 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2662 for (uint32_t i = 0; i < num_features; i++) {
2663 if (enabled_feature[i] && !supported_feature[i])
2664 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2665 }
2666
2667 return VK_SUCCESS;
2668 }
2669
2670 VkResult anv_CreateDevice(
2671 VkPhysicalDevice physicalDevice,
2672 const VkDeviceCreateInfo* pCreateInfo,
2673 const VkAllocationCallbacks* pAllocator,
2674 VkDevice* pDevice)
2675 {
2676 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2677 VkResult result;
2678 struct anv_device *device;
2679
2680 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2681
2682 struct anv_device_extension_table enabled_extensions = { };
2683 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2684 int idx;
2685 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2686 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2687 anv_device_extensions[idx].extensionName) == 0)
2688 break;
2689 }
2690
2691 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2692 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2693
2694 if (!physical_device->supported_extensions.extensions[idx])
2695 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2696
2697 enabled_extensions.extensions[idx] = true;
2698 }
2699
2700 /* Check enabled features */
2701 bool robust_buffer_access = false;
2702 if (pCreateInfo->pEnabledFeatures) {
2703 result = check_physical_device_features(physicalDevice,
2704 pCreateInfo->pEnabledFeatures);
2705 if (result != VK_SUCCESS)
2706 return result;
2707
2708 if (pCreateInfo->pEnabledFeatures->robustBufferAccess)
2709 robust_buffer_access = true;
2710 }
2711
2712 vk_foreach_struct_const(ext, pCreateInfo->pNext) {
2713 switch (ext->sType) {
2714 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2: {
2715 const VkPhysicalDeviceFeatures2 *features = (const void *)ext;
2716 result = check_physical_device_features(physicalDevice,
2717 &features->features);
2718 if (result != VK_SUCCESS)
2719 return result;
2720
2721 if (features->features.robustBufferAccess)
2722 robust_buffer_access = true;
2723 break;
2724 }
2725
2726 default:
2727 /* Don't warn */
2728 break;
2729 }
2730 }
2731
2732 /* Check requested queues and fail if we are requested to create any
2733 * queues with flags we don't support.
2734 */
2735 assert(pCreateInfo->queueCreateInfoCount > 0);
2736 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2737 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2738 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2739 }
2740
2741 /* Check if client specified queue priority. */
2742 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2743 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2744 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2745
2746 VkQueueGlobalPriorityEXT priority =
2747 queue_priority ? queue_priority->globalPriority :
2748 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2749
2750 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2751 sizeof(*device), 8,
2752 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2753 if (!device)
2754 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2755
2756 vk_device_init(&device->vk, pCreateInfo,
2757 &physical_device->instance->alloc, pAllocator);
2758
2759 if (INTEL_DEBUG & DEBUG_BATCH) {
2760 const unsigned decode_flags =
2761 GEN_BATCH_DECODE_FULL |
2762 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2763 GEN_BATCH_DECODE_OFFSETS |
2764 GEN_BATCH_DECODE_FLOATS;
2765
2766 gen_batch_decode_ctx_init(&device->decoder_ctx,
2767 &physical_device->info,
2768 stderr, decode_flags, NULL,
2769 decode_get_bo, NULL, device);
2770 }
2771
2772 device->physical = physical_device;
2773 device->no_hw = physical_device->no_hw;
2774 device->_lost = false;
2775
2776 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2777 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2778 if (device->fd == -1) {
2779 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2780 goto fail_device;
2781 }
2782
2783 device->context_id = anv_gem_create_context(device);
2784 if (device->context_id == -1) {
2785 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2786 goto fail_fd;
2787 }
2788
2789 result = anv_queue_init(device, &device->queue);
2790 if (result != VK_SUCCESS)
2791 goto fail_context_id;
2792
2793 if (physical_device->use_softpin) {
2794 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2795 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2796 goto fail_queue;
2797 }
2798
2799 /* keep the page with address zero out of the allocator */
2800 util_vma_heap_init(&device->vma_lo,
2801 LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
2802
2803 util_vma_heap_init(&device->vma_cva, CLIENT_VISIBLE_HEAP_MIN_ADDRESS,
2804 CLIENT_VISIBLE_HEAP_SIZE);
2805
2806 /* Leave the last 4GiB out of the high vma range, so that no state
2807 * base address + size can overflow 48 bits. For more information see
2808 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2809 */
2810 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
2811 physical_device->gtt_size - (1ull << 32) -
2812 HIGH_HEAP_MIN_ADDRESS);
2813 }
2814
2815 list_inithead(&device->memory_objects);
2816
2817 /* As per spec, the driver implementation may deny requests to acquire
2818 * a priority above the default priority (MEDIUM) if the caller does not
2819 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2820 * is returned.
2821 */
2822 if (physical_device->has_context_priority) {
2823 int err = anv_gem_set_context_param(device->fd, device->context_id,
2824 I915_CONTEXT_PARAM_PRIORITY,
2825 vk_priority_to_gen(priority));
2826 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2827 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2828 goto fail_vmas;
2829 }
2830 }
2831
2832 device->info = physical_device->info;
2833 device->isl_dev = physical_device->isl_dev;
2834
2835 /* On Broadwell and later, we can use batch chaining to more efficiently
2836 * implement growing command buffers. Prior to Haswell, the kernel
2837 * command parser gets in the way and we have to fall back to growing
2838 * the batch.
2839 */
2840 device->can_chain_batches = device->info.gen >= 8;
2841
2842 device->robust_buffer_access = robust_buffer_access;
2843 device->enabled_extensions = enabled_extensions;
2844
2845 anv_device_init_dispatch(device);
2846
2847 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2848 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2849 goto fail_queue;
2850 }
2851
2852 pthread_condattr_t condattr;
2853 if (pthread_condattr_init(&condattr) != 0) {
2854 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2855 goto fail_mutex;
2856 }
2857 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2858 pthread_condattr_destroy(&condattr);
2859 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2860 goto fail_mutex;
2861 }
2862 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2863 pthread_condattr_destroy(&condattr);
2864 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2865 goto fail_mutex;
2866 }
2867 pthread_condattr_destroy(&condattr);
2868
2869 result = anv_bo_cache_init(&device->bo_cache);
2870 if (result != VK_SUCCESS)
2871 goto fail_queue_cond;
2872
2873 anv_bo_pool_init(&device->batch_bo_pool, device);
2874
2875 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2876 DYNAMIC_STATE_POOL_MIN_ADDRESS, 0, 16384);
2877 if (result != VK_SUCCESS)
2878 goto fail_batch_bo_pool;
2879
2880 result = anv_state_pool_init(&device->instruction_state_pool, device,
2881 INSTRUCTION_STATE_POOL_MIN_ADDRESS, 0, 16384);
2882 if (result != VK_SUCCESS)
2883 goto fail_dynamic_state_pool;
2884
2885 result = anv_state_pool_init(&device->surface_state_pool, device,
2886 SURFACE_STATE_POOL_MIN_ADDRESS, 0, 4096);
2887 if (result != VK_SUCCESS)
2888 goto fail_instruction_state_pool;
2889
2890 if (physical_device->use_softpin) {
2891 int64_t bt_pool_offset = (int64_t)BINDING_TABLE_POOL_MIN_ADDRESS -
2892 (int64_t)SURFACE_STATE_POOL_MIN_ADDRESS;
2893 assert(INT32_MIN < bt_pool_offset && bt_pool_offset < 0);
2894 result = anv_state_pool_init(&device->binding_table_pool, device,
2895 SURFACE_STATE_POOL_MIN_ADDRESS,
2896 bt_pool_offset, 4096);
2897 if (result != VK_SUCCESS)
2898 goto fail_surface_state_pool;
2899 }
2900
2901 if (device->info.gen >= 12) {
2902 device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
2903 &physical_device->info);
2904 if (!device->aux_map_ctx)
2905 goto fail_binding_table_pool;
2906 }
2907
2908 result = anv_device_alloc_bo(device, 4096, 0 /* flags */,
2909 0 /* explicit_address */,
2910 &device->workaround_bo);
2911 if (result != VK_SUCCESS)
2912 goto fail_surface_aux_map_pool;
2913
2914 result = anv_device_init_trivial_batch(device);
2915 if (result != VK_SUCCESS)
2916 goto fail_workaround_bo;
2917
2918 /* Allocate a null surface state at surface state offset 0. This makes
2919 * NULL descriptor handling trivial because we can just memset structures
2920 * to zero and they have a valid descriptor.
2921 */
2922 device->null_surface_state =
2923 anv_state_pool_alloc(&device->surface_state_pool,
2924 device->isl_dev.ss.size,
2925 device->isl_dev.ss.align);
2926 isl_null_fill_state(&device->isl_dev, device->null_surface_state.map,
2927 isl_extent3d(1, 1, 1) /* This shouldn't matter */);
2928 assert(device->null_surface_state.offset == 0);
2929
2930 if (device->info.gen >= 10) {
2931 result = anv_device_init_hiz_clear_value_bo(device);
2932 if (result != VK_SUCCESS)
2933 goto fail_trivial_batch_bo;
2934 }
2935
2936 anv_scratch_pool_init(device, &device->scratch_pool);
2937
2938 switch (device->info.gen) {
2939 case 7:
2940 if (!device->info.is_haswell)
2941 result = gen7_init_device_state(device);
2942 else
2943 result = gen75_init_device_state(device);
2944 break;
2945 case 8:
2946 result = gen8_init_device_state(device);
2947 break;
2948 case 9:
2949 result = gen9_init_device_state(device);
2950 break;
2951 case 10:
2952 result = gen10_init_device_state(device);
2953 break;
2954 case 11:
2955 result = gen11_init_device_state(device);
2956 break;
2957 case 12:
2958 result = gen12_init_device_state(device);
2959 break;
2960 default:
2961 /* Shouldn't get here as we don't create physical devices for any other
2962 * gens. */
2963 unreachable("unhandled gen");
2964 }
2965 if (result != VK_SUCCESS)
2966 goto fail_workaround_bo;
2967
2968 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2969
2970 anv_device_init_blorp(device);
2971
2972 anv_device_init_border_colors(device);
2973
2974 anv_device_perf_init(device);
2975
2976 *pDevice = anv_device_to_handle(device);
2977
2978 return VK_SUCCESS;
2979
2980 fail_workaround_bo:
2981 anv_scratch_pool_finish(device, &device->scratch_pool);
2982 if (device->info.gen >= 10)
2983 anv_device_release_bo(device, device->hiz_clear_bo);
2984 anv_device_release_bo(device, device->workaround_bo);
2985 fail_trivial_batch_bo:
2986 anv_device_release_bo(device, device->trivial_batch_bo);
2987 fail_surface_aux_map_pool:
2988 if (device->info.gen >= 12) {
2989 gen_aux_map_finish(device->aux_map_ctx);
2990 device->aux_map_ctx = NULL;
2991 }
2992 fail_binding_table_pool:
2993 if (physical_device->use_softpin)
2994 anv_state_pool_finish(&device->binding_table_pool);
2995 fail_surface_state_pool:
2996 anv_state_pool_finish(&device->surface_state_pool);
2997 fail_instruction_state_pool:
2998 anv_state_pool_finish(&device->instruction_state_pool);
2999 fail_dynamic_state_pool:
3000 anv_state_pool_finish(&device->dynamic_state_pool);
3001 fail_batch_bo_pool:
3002 anv_bo_pool_finish(&device->batch_bo_pool);
3003 anv_bo_cache_finish(&device->bo_cache);
3004 fail_queue_cond:
3005 pthread_cond_destroy(&device->queue_submit);
3006 fail_mutex:
3007 pthread_mutex_destroy(&device->mutex);
3008 fail_vmas:
3009 if (physical_device->use_softpin) {
3010 util_vma_heap_finish(&device->vma_hi);
3011 util_vma_heap_finish(&device->vma_cva);
3012 util_vma_heap_finish(&device->vma_lo);
3013 }
3014 fail_queue:
3015 anv_queue_finish(&device->queue);
3016 fail_context_id:
3017 anv_gem_destroy_context(device, device->context_id);
3018 fail_fd:
3019 close(device->fd);
3020 fail_device:
3021 vk_free(&device->vk.alloc, device);
3022
3023 return result;
3024 }
3025
3026 void anv_DestroyDevice(
3027 VkDevice _device,
3028 const VkAllocationCallbacks* pAllocator)
3029 {
3030 ANV_FROM_HANDLE(anv_device, device, _device);
3031
3032 if (!device)
3033 return;
3034
3035 anv_device_finish_blorp(device);
3036
3037 anv_pipeline_cache_finish(&device->default_pipeline_cache);
3038
3039 anv_queue_finish(&device->queue);
3040
3041 #ifdef HAVE_VALGRIND
3042 /* We only need to free these to prevent valgrind errors. The backing
3043 * BO will go away in a couple of lines so we don't actually leak.
3044 */
3045 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
3046 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
3047 #endif
3048
3049 anv_scratch_pool_finish(device, &device->scratch_pool);
3050
3051 anv_device_release_bo(device, device->workaround_bo);
3052 anv_device_release_bo(device, device->trivial_batch_bo);
3053 if (device->info.gen >= 10)
3054 anv_device_release_bo(device, device->hiz_clear_bo);
3055
3056 if (device->info.gen >= 12) {
3057 gen_aux_map_finish(device->aux_map_ctx);
3058 device->aux_map_ctx = NULL;
3059 }
3060
3061 if (device->physical->use_softpin)
3062 anv_state_pool_finish(&device->binding_table_pool);
3063 anv_state_pool_finish(&device->surface_state_pool);
3064 anv_state_pool_finish(&device->instruction_state_pool);
3065 anv_state_pool_finish(&device->dynamic_state_pool);
3066
3067 anv_bo_pool_finish(&device->batch_bo_pool);
3068
3069 anv_bo_cache_finish(&device->bo_cache);
3070
3071 if (device->physical->use_softpin) {
3072 util_vma_heap_finish(&device->vma_hi);
3073 util_vma_heap_finish(&device->vma_cva);
3074 util_vma_heap_finish(&device->vma_lo);
3075 }
3076
3077 pthread_cond_destroy(&device->queue_submit);
3078 pthread_mutex_destroy(&device->mutex);
3079
3080 anv_gem_destroy_context(device, device->context_id);
3081
3082 if (INTEL_DEBUG & DEBUG_BATCH)
3083 gen_batch_decode_ctx_finish(&device->decoder_ctx);
3084
3085 close(device->fd);
3086
3087 vk_device_finish(&device->vk);
3088 vk_free(&device->vk.alloc, device);
3089 }
3090
3091 VkResult anv_EnumerateInstanceLayerProperties(
3092 uint32_t* pPropertyCount,
3093 VkLayerProperties* pProperties)
3094 {
3095 if (pProperties == NULL) {
3096 *pPropertyCount = 0;
3097 return VK_SUCCESS;
3098 }
3099
3100 /* None supported at this time */
3101 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
3102 }
3103
3104 VkResult anv_EnumerateDeviceLayerProperties(
3105 VkPhysicalDevice physicalDevice,
3106 uint32_t* pPropertyCount,
3107 VkLayerProperties* pProperties)
3108 {
3109 if (pProperties == NULL) {
3110 *pPropertyCount = 0;
3111 return VK_SUCCESS;
3112 }
3113
3114 /* None supported at this time */
3115 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
3116 }
3117
3118 void anv_GetDeviceQueue(
3119 VkDevice _device,
3120 uint32_t queueNodeIndex,
3121 uint32_t queueIndex,
3122 VkQueue* pQueue)
3123 {
3124 const VkDeviceQueueInfo2 info = {
3125 .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
3126 .pNext = NULL,
3127 .flags = 0,
3128 .queueFamilyIndex = queueNodeIndex,
3129 .queueIndex = queueIndex,
3130 };
3131
3132 anv_GetDeviceQueue2(_device, &info, pQueue);
3133 }
3134
3135 void anv_GetDeviceQueue2(
3136 VkDevice _device,
3137 const VkDeviceQueueInfo2* pQueueInfo,
3138 VkQueue* pQueue)
3139 {
3140 ANV_FROM_HANDLE(anv_device, device, _device);
3141
3142 assert(pQueueInfo->queueIndex == 0);
3143
3144 if (pQueueInfo->flags == device->queue.flags)
3145 *pQueue = anv_queue_to_handle(&device->queue);
3146 else
3147 *pQueue = NULL;
3148 }
3149
3150 VkResult
3151 _anv_device_set_lost(struct anv_device *device,
3152 const char *file, int line,
3153 const char *msg, ...)
3154 {
3155 VkResult err;
3156 va_list ap;
3157
3158 p_atomic_inc(&device->_lost);
3159
3160 va_start(ap, msg);
3161 err = __vk_errorv(device->physical->instance, device,
3162 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
3163 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3164 va_end(ap);
3165
3166 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3167 abort();
3168
3169 return err;
3170 }
3171
3172 VkResult
3173 _anv_queue_set_lost(struct anv_queue *queue,
3174 const char *file, int line,
3175 const char *msg, ...)
3176 {
3177 VkResult err;
3178 va_list ap;
3179
3180 p_atomic_inc(&queue->device->_lost);
3181
3182 va_start(ap, msg);
3183 err = __vk_errorv(queue->device->physical->instance, queue->device,
3184 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
3185 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3186 va_end(ap);
3187
3188 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3189 abort();
3190
3191 return err;
3192 }
3193
3194 VkResult
3195 anv_device_query_status(struct anv_device *device)
3196 {
3197 /* This isn't likely as most of the callers of this function already check
3198 * for it. However, it doesn't hurt to check and it potentially lets us
3199 * avoid an ioctl.
3200 */
3201 if (anv_device_is_lost(device))
3202 return VK_ERROR_DEVICE_LOST;
3203
3204 uint32_t active, pending;
3205 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
3206 if (ret == -1) {
3207 /* We don't know the real error. */
3208 return anv_device_set_lost(device, "get_reset_stats failed: %m");
3209 }
3210
3211 if (active) {
3212 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
3213 } else if (pending) {
3214 return anv_device_set_lost(device, "GPU hung with commands in-flight");
3215 }
3216
3217 return VK_SUCCESS;
3218 }
3219
3220 VkResult
3221 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
3222 {
3223 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3224 * Other usages of the BO (such as on different hardware) will not be
3225 * flagged as "busy" by this ioctl. Use with care.
3226 */
3227 int ret = anv_gem_busy(device, bo->gem_handle);
3228 if (ret == 1) {
3229 return VK_NOT_READY;
3230 } else if (ret == -1) {
3231 /* We don't know the real error. */
3232 return anv_device_set_lost(device, "gem wait failed: %m");
3233 }
3234
3235 /* Query for device status after the busy call. If the BO we're checking
3236 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3237 * client because it clearly doesn't have valid data. Yes, this most
3238 * likely means an ioctl, but we just did an ioctl to query the busy status
3239 * so it's no great loss.
3240 */
3241 return anv_device_query_status(device);
3242 }
3243
3244 VkResult
3245 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
3246 int64_t timeout)
3247 {
3248 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
3249 if (ret == -1 && errno == ETIME) {
3250 return VK_TIMEOUT;
3251 } else if (ret == -1) {
3252 /* We don't know the real error. */
3253 return anv_device_set_lost(device, "gem wait failed: %m");
3254 }
3255
3256 /* Query for device status after the wait. If the BO we're waiting on got
3257 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3258 * because it clearly doesn't have valid data. Yes, this most likely means
3259 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3260 */
3261 return anv_device_query_status(device);
3262 }
3263
3264 VkResult anv_DeviceWaitIdle(
3265 VkDevice _device)
3266 {
3267 ANV_FROM_HANDLE(anv_device, device, _device);
3268
3269 if (anv_device_is_lost(device))
3270 return VK_ERROR_DEVICE_LOST;
3271
3272 return anv_queue_submit_simple_batch(&device->queue, NULL);
3273 }
3274
3275 uint64_t
3276 anv_vma_alloc(struct anv_device *device,
3277 uint64_t size, uint64_t align,
3278 enum anv_bo_alloc_flags alloc_flags,
3279 uint64_t client_address)
3280 {
3281 pthread_mutex_lock(&device->vma_mutex);
3282
3283 uint64_t addr = 0;
3284
3285 if (alloc_flags & ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS) {
3286 if (client_address) {
3287 if (util_vma_heap_alloc_addr(&device->vma_cva,
3288 client_address, size)) {
3289 addr = client_address;
3290 }
3291 } else {
3292 addr = util_vma_heap_alloc(&device->vma_cva, size, align);
3293 }
3294 /* We don't want to fall back to other heaps */
3295 goto done;
3296 }
3297
3298 assert(client_address == 0);
3299
3300 if (!(alloc_flags & ANV_BO_ALLOC_32BIT_ADDRESS))
3301 addr = util_vma_heap_alloc(&device->vma_hi, size, align);
3302
3303 if (addr == 0)
3304 addr = util_vma_heap_alloc(&device->vma_lo, size, align);
3305
3306 done:
3307 pthread_mutex_unlock(&device->vma_mutex);
3308
3309 assert(addr == gen_48b_address(addr));
3310 return gen_canonical_address(addr);
3311 }
3312
3313 void
3314 anv_vma_free(struct anv_device *device,
3315 uint64_t address, uint64_t size)
3316 {
3317 const uint64_t addr_48b = gen_48b_address(address);
3318
3319 pthread_mutex_lock(&device->vma_mutex);
3320
3321 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3322 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3323 util_vma_heap_free(&device->vma_lo, addr_48b, size);
3324 } else if (addr_48b >= CLIENT_VISIBLE_HEAP_MIN_ADDRESS &&
3325 addr_48b <= CLIENT_VISIBLE_HEAP_MAX_ADDRESS) {
3326 util_vma_heap_free(&device->vma_cva, addr_48b, size);
3327 } else {
3328 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS);
3329 util_vma_heap_free(&device->vma_hi, addr_48b, size);
3330 }
3331
3332 pthread_mutex_unlock(&device->vma_mutex);
3333 }
3334
3335 VkResult anv_AllocateMemory(
3336 VkDevice _device,
3337 const VkMemoryAllocateInfo* pAllocateInfo,
3338 const VkAllocationCallbacks* pAllocator,
3339 VkDeviceMemory* pMem)
3340 {
3341 ANV_FROM_HANDLE(anv_device, device, _device);
3342 struct anv_physical_device *pdevice = device->physical;
3343 struct anv_device_memory *mem;
3344 VkResult result = VK_SUCCESS;
3345
3346 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3347
3348 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3349 assert(pAllocateInfo->allocationSize > 0);
3350
3351 VkDeviceSize aligned_alloc_size =
3352 align_u64(pAllocateInfo->allocationSize, 4096);
3353
3354 if (aligned_alloc_size > MAX_MEMORY_ALLOCATION_SIZE)
3355 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3356
3357 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3358 struct anv_memory_type *mem_type =
3359 &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3360 assert(mem_type->heapIndex < pdevice->memory.heap_count);
3361 struct anv_memory_heap *mem_heap =
3362 &pdevice->memory.heaps[mem_type->heapIndex];
3363
3364 uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
3365 if (mem_heap_used + aligned_alloc_size > mem_heap->size)
3366 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3367
3368 mem = vk_alloc2(&device->vk.alloc, pAllocator, sizeof(*mem), 8,
3369 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3370 if (mem == NULL)
3371 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3372
3373 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3374 vk_object_base_init(&device->vk, &mem->base, VK_OBJECT_TYPE_DEVICE_MEMORY);
3375 mem->type = mem_type;
3376 mem->map = NULL;
3377 mem->map_size = 0;
3378 mem->ahw = NULL;
3379 mem->host_ptr = NULL;
3380
3381 enum anv_bo_alloc_flags alloc_flags = 0;
3382
3383 const VkExportMemoryAllocateInfo *export_info = NULL;
3384 const VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info = NULL;
3385 const VkImportMemoryFdInfoKHR *fd_info = NULL;
3386 const VkImportMemoryHostPointerInfoEXT *host_ptr_info = NULL;
3387 const VkMemoryDedicatedAllocateInfo *dedicated_info = NULL;
3388 VkMemoryAllocateFlags vk_flags = 0;
3389 uint64_t client_address = 0;
3390
3391 vk_foreach_struct_const(ext, pAllocateInfo->pNext) {
3392 switch (ext->sType) {
3393 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO:
3394 export_info = (void *)ext;
3395 break;
3396
3397 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID:
3398 ahw_import_info = (void *)ext;
3399 break;
3400
3401 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR:
3402 fd_info = (void *)ext;
3403 break;
3404
3405 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT:
3406 host_ptr_info = (void *)ext;
3407 break;
3408
3409 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO: {
3410 const VkMemoryAllocateFlagsInfo *flags_info = (void *)ext;
3411 vk_flags = flags_info->flags;
3412 break;
3413 }
3414
3415 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO:
3416 dedicated_info = (void *)ext;
3417 break;
3418
3419 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR: {
3420 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *addr_info =
3421 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *)ext;
3422 client_address = addr_info->opaqueCaptureAddress;
3423 break;
3424 }
3425
3426 default:
3427 anv_debug_ignored_stype(ext->sType);
3428 break;
3429 }
3430 }
3431
3432 /* By default, we want all VkDeviceMemory objects to support CCS */
3433 if (device->physical->has_implicit_ccs)
3434 alloc_flags |= ANV_BO_ALLOC_IMPLICIT_CCS;
3435
3436 if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR)
3437 alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS;
3438
3439 if ((export_info && export_info->handleTypes) ||
3440 (fd_info && fd_info->handleType) ||
3441 (host_ptr_info && host_ptr_info->handleType)) {
3442 /* Anything imported or exported is EXTERNAL */
3443 alloc_flags |= ANV_BO_ALLOC_EXTERNAL;
3444
3445 /* We can't have implicit CCS on external memory with an AUX-table.
3446 * Doing so would require us to sync the aux tables across processes
3447 * which is impractical.
3448 */
3449 if (device->info.has_aux_map)
3450 alloc_flags &= ~ANV_BO_ALLOC_IMPLICIT_CCS;
3451 }
3452
3453 /* Check if we need to support Android HW buffer export. If so,
3454 * create AHardwareBuffer and import memory from it.
3455 */
3456 bool android_export = false;
3457 if (export_info && export_info->handleTypes &
3458 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3459 android_export = true;
3460
3461 if (ahw_import_info) {
3462 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3463 if (result != VK_SUCCESS)
3464 goto fail;
3465
3466 goto success;
3467 } else if (android_export) {
3468 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3469 if (result != VK_SUCCESS)
3470 goto fail;
3471
3472 const VkImportAndroidHardwareBufferInfoANDROID import_info = {
3473 .buffer = mem->ahw,
3474 };
3475 result = anv_import_ahw_memory(_device, mem, &import_info);
3476 if (result != VK_SUCCESS)
3477 goto fail;
3478
3479 goto success;
3480 }
3481
3482 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3483 * ignored.
3484 */
3485 if (fd_info && fd_info->handleType) {
3486 /* At the moment, we support only the below handle types. */
3487 assert(fd_info->handleType ==
3488 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3489 fd_info->handleType ==
3490 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3491
3492 result = anv_device_import_bo(device, fd_info->fd, alloc_flags,
3493 client_address, &mem->bo);
3494 if (result != VK_SUCCESS)
3495 goto fail;
3496
3497 /* For security purposes, we reject importing the bo if it's smaller
3498 * than the requested allocation size. This prevents a malicious client
3499 * from passing a buffer to a trusted client, lying about the size, and
3500 * telling the trusted client to try and texture from an image that goes
3501 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3502 * in the trusted client. The trusted client can protect itself against
3503 * this sort of attack but only if it can trust the buffer size.
3504 */
3505 if (mem->bo->size < aligned_alloc_size) {
3506 result = vk_errorf(device, device, VK_ERROR_INVALID_EXTERNAL_HANDLE,
3507 "aligned allocationSize too large for "
3508 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3509 "%"PRIu64"B > %"PRIu64"B",
3510 aligned_alloc_size, mem->bo->size);
3511 anv_device_release_bo(device, mem->bo);
3512 goto fail;
3513 }
3514
3515 /* From the Vulkan spec:
3516 *
3517 * "Importing memory from a file descriptor transfers ownership of
3518 * the file descriptor from the application to the Vulkan
3519 * implementation. The application must not perform any operations on
3520 * the file descriptor after a successful import."
3521 *
3522 * If the import fails, we leave the file descriptor open.
3523 */
3524 close(fd_info->fd);
3525 goto success;
3526 }
3527
3528 if (host_ptr_info && host_ptr_info->handleType) {
3529 if (host_ptr_info->handleType ==
3530 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3531 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3532 goto fail;
3533 }
3534
3535 assert(host_ptr_info->handleType ==
3536 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3537
3538 result = anv_device_import_bo_from_host_ptr(device,
3539 host_ptr_info->pHostPointer,
3540 pAllocateInfo->allocationSize,
3541 alloc_flags,
3542 client_address,
3543 &mem->bo);
3544 if (result != VK_SUCCESS)
3545 goto fail;
3546
3547 mem->host_ptr = host_ptr_info->pHostPointer;
3548 goto success;
3549 }
3550
3551 /* Regular allocate (not importing memory). */
3552
3553 result = anv_device_alloc_bo(device, pAllocateInfo->allocationSize,
3554 alloc_flags, client_address, &mem->bo);
3555 if (result != VK_SUCCESS)
3556 goto fail;
3557
3558 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3559 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3560
3561 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3562 * the BO. In this case, we have a dedicated allocation.
3563 */
3564 if (image->needs_set_tiling) {
3565 const uint32_t i915_tiling =
3566 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3567 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3568 image->planes[0].surface.isl.row_pitch_B,
3569 i915_tiling);
3570 if (ret) {
3571 anv_device_release_bo(device, mem->bo);
3572 result = vk_errorf(device, device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
3573 "failed to set BO tiling: %m");
3574 goto fail;
3575 }
3576 }
3577 }
3578
3579 success:
3580 mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
3581 if (mem_heap_used > mem_heap->size) {
3582 p_atomic_add(&mem_heap->used, -mem->bo->size);
3583 anv_device_release_bo(device, mem->bo);
3584 result = vk_errorf(device, device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
3585 "Out of heap memory");
3586 goto fail;
3587 }
3588
3589 pthread_mutex_lock(&device->mutex);
3590 list_addtail(&mem->link, &device->memory_objects);
3591 pthread_mutex_unlock(&device->mutex);
3592
3593 *pMem = anv_device_memory_to_handle(mem);
3594
3595 return VK_SUCCESS;
3596
3597 fail:
3598 vk_free2(&device->vk.alloc, pAllocator, mem);
3599
3600 return result;
3601 }
3602
3603 VkResult anv_GetMemoryFdKHR(
3604 VkDevice device_h,
3605 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3606 int* pFd)
3607 {
3608 ANV_FROM_HANDLE(anv_device, dev, device_h);
3609 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3610
3611 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3612
3613 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3614 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3615
3616 return anv_device_export_bo(dev, mem->bo, pFd);
3617 }
3618
3619 VkResult anv_GetMemoryFdPropertiesKHR(
3620 VkDevice _device,
3621 VkExternalMemoryHandleTypeFlagBits handleType,
3622 int fd,
3623 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3624 {
3625 ANV_FROM_HANDLE(anv_device, device, _device);
3626
3627 switch (handleType) {
3628 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3629 /* dma-buf can be imported as any memory type */
3630 pMemoryFdProperties->memoryTypeBits =
3631 (1 << device->physical->memory.type_count) - 1;
3632 return VK_SUCCESS;
3633
3634 default:
3635 /* The valid usage section for this function says:
3636 *
3637 * "handleType must not be one of the handle types defined as
3638 * opaque."
3639 *
3640 * So opaque handle types fall into the default "unsupported" case.
3641 */
3642 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3643 }
3644 }
3645
3646 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3647 VkDevice _device,
3648 VkExternalMemoryHandleTypeFlagBits handleType,
3649 const void* pHostPointer,
3650 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3651 {
3652 ANV_FROM_HANDLE(anv_device, device, _device);
3653
3654 assert(pMemoryHostPointerProperties->sType ==
3655 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3656
3657 switch (handleType) {
3658 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT:
3659 /* Host memory can be imported as any memory type. */
3660 pMemoryHostPointerProperties->memoryTypeBits =
3661 (1ull << device->physical->memory.type_count) - 1;
3662
3663 return VK_SUCCESS;
3664
3665 default:
3666 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3667 }
3668 }
3669
3670 void anv_FreeMemory(
3671 VkDevice _device,
3672 VkDeviceMemory _mem,
3673 const VkAllocationCallbacks* pAllocator)
3674 {
3675 ANV_FROM_HANDLE(anv_device, device, _device);
3676 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3677
3678 if (mem == NULL)
3679 return;
3680
3681 pthread_mutex_lock(&device->mutex);
3682 list_del(&mem->link);
3683 pthread_mutex_unlock(&device->mutex);
3684
3685 if (mem->map)
3686 anv_UnmapMemory(_device, _mem);
3687
3688 p_atomic_add(&device->physical->memory.heaps[mem->type->heapIndex].used,
3689 -mem->bo->size);
3690
3691 anv_device_release_bo(device, mem->bo);
3692
3693 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3694 if (mem->ahw)
3695 AHardwareBuffer_release(mem->ahw);
3696 #endif
3697
3698 vk_object_base_finish(&mem->base);
3699 vk_free2(&device->vk.alloc, pAllocator, mem);
3700 }
3701
3702 VkResult anv_MapMemory(
3703 VkDevice _device,
3704 VkDeviceMemory _memory,
3705 VkDeviceSize offset,
3706 VkDeviceSize size,
3707 VkMemoryMapFlags flags,
3708 void** ppData)
3709 {
3710 ANV_FROM_HANDLE(anv_device, device, _device);
3711 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3712
3713 if (mem == NULL) {
3714 *ppData = NULL;
3715 return VK_SUCCESS;
3716 }
3717
3718 if (mem->host_ptr) {
3719 *ppData = mem->host_ptr + offset;
3720 return VK_SUCCESS;
3721 }
3722
3723 if (size == VK_WHOLE_SIZE)
3724 size = mem->bo->size - offset;
3725
3726 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3727 *
3728 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3729 * assert(size != 0);
3730 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3731 * equal to the size of the memory minus offset
3732 */
3733 assert(size > 0);
3734 assert(offset + size <= mem->bo->size);
3735
3736 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3737 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3738 * at a time is valid. We could just mmap up front and return an offset
3739 * pointer here, but that may exhaust virtual memory on 32 bit
3740 * userspace. */
3741
3742 uint32_t gem_flags = 0;
3743
3744 if (!device->info.has_llc &&
3745 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3746 gem_flags |= I915_MMAP_WC;
3747
3748 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3749 uint64_t map_offset;
3750 if (!device->physical->has_mmap_offset)
3751 map_offset = offset & ~4095ull;
3752 else
3753 map_offset = 0;
3754 assert(offset >= map_offset);
3755 uint64_t map_size = (offset + size) - map_offset;
3756
3757 /* Let's map whole pages */
3758 map_size = align_u64(map_size, 4096);
3759
3760 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3761 map_offset, map_size, gem_flags);
3762 if (map == MAP_FAILED)
3763 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3764
3765 mem->map = map;
3766 mem->map_size = map_size;
3767
3768 *ppData = mem->map + (offset - map_offset);
3769
3770 return VK_SUCCESS;
3771 }
3772
3773 void anv_UnmapMemory(
3774 VkDevice _device,
3775 VkDeviceMemory _memory)
3776 {
3777 ANV_FROM_HANDLE(anv_device, device, _device);
3778 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3779
3780 if (mem == NULL || mem->host_ptr)
3781 return;
3782
3783 anv_gem_munmap(device, mem->map, mem->map_size);
3784
3785 mem->map = NULL;
3786 mem->map_size = 0;
3787 }
3788
3789 static void
3790 clflush_mapped_ranges(struct anv_device *device,
3791 uint32_t count,
3792 const VkMappedMemoryRange *ranges)
3793 {
3794 for (uint32_t i = 0; i < count; i++) {
3795 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3796 if (ranges[i].offset >= mem->map_size)
3797 continue;
3798
3799 gen_clflush_range(mem->map + ranges[i].offset,
3800 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3801 }
3802 }
3803
3804 VkResult anv_FlushMappedMemoryRanges(
3805 VkDevice _device,
3806 uint32_t memoryRangeCount,
3807 const VkMappedMemoryRange* pMemoryRanges)
3808 {
3809 ANV_FROM_HANDLE(anv_device, device, _device);
3810
3811 if (device->info.has_llc)
3812 return VK_SUCCESS;
3813
3814 /* Make sure the writes we're flushing have landed. */
3815 __builtin_ia32_mfence();
3816
3817 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3818
3819 return VK_SUCCESS;
3820 }
3821
3822 VkResult anv_InvalidateMappedMemoryRanges(
3823 VkDevice _device,
3824 uint32_t memoryRangeCount,
3825 const VkMappedMemoryRange* pMemoryRanges)
3826 {
3827 ANV_FROM_HANDLE(anv_device, device, _device);
3828
3829 if (device->info.has_llc)
3830 return VK_SUCCESS;
3831
3832 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3833
3834 /* Make sure no reads get moved up above the invalidate. */
3835 __builtin_ia32_mfence();
3836
3837 return VK_SUCCESS;
3838 }
3839
3840 void anv_GetBufferMemoryRequirements(
3841 VkDevice _device,
3842 VkBuffer _buffer,
3843 VkMemoryRequirements* pMemoryRequirements)
3844 {
3845 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3846 ANV_FROM_HANDLE(anv_device, device, _device);
3847
3848 /* The Vulkan spec (git aaed022) says:
3849 *
3850 * memoryTypeBits is a bitfield and contains one bit set for every
3851 * supported memory type for the resource. The bit `1<<i` is set if and
3852 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3853 * structure for the physical device is supported.
3854 */
3855 uint32_t memory_types = (1ull << device->physical->memory.type_count) - 1;
3856
3857 /* Base alignment requirement of a cache line */
3858 uint32_t alignment = 16;
3859
3860 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3861 alignment = MAX2(alignment, ANV_UBO_ALIGNMENT);
3862
3863 pMemoryRequirements->size = buffer->size;
3864 pMemoryRequirements->alignment = alignment;
3865
3866 /* Storage and Uniform buffers should have their size aligned to
3867 * 32-bits to avoid boundary checks when last DWord is not complete.
3868 * This would ensure that not internal padding would be needed for
3869 * 16-bit types.
3870 */
3871 if (device->robust_buffer_access &&
3872 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3873 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3874 pMemoryRequirements->size = align_u64(buffer->size, 4);
3875
3876 pMemoryRequirements->memoryTypeBits = memory_types;
3877 }
3878
3879 void anv_GetBufferMemoryRequirements2(
3880 VkDevice _device,
3881 const VkBufferMemoryRequirementsInfo2* pInfo,
3882 VkMemoryRequirements2* pMemoryRequirements)
3883 {
3884 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3885 &pMemoryRequirements->memoryRequirements);
3886
3887 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3888 switch (ext->sType) {
3889 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3890 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3891 requirements->prefersDedicatedAllocation = false;
3892 requirements->requiresDedicatedAllocation = false;
3893 break;
3894 }
3895
3896 default:
3897 anv_debug_ignored_stype(ext->sType);
3898 break;
3899 }
3900 }
3901 }
3902
3903 void anv_GetImageMemoryRequirements(
3904 VkDevice _device,
3905 VkImage _image,
3906 VkMemoryRequirements* pMemoryRequirements)
3907 {
3908 ANV_FROM_HANDLE(anv_image, image, _image);
3909 ANV_FROM_HANDLE(anv_device, device, _device);
3910
3911 /* The Vulkan spec (git aaed022) says:
3912 *
3913 * memoryTypeBits is a bitfield and contains one bit set for every
3914 * supported memory type for the resource. The bit `1<<i` is set if and
3915 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3916 * structure for the physical device is supported.
3917 *
3918 * All types are currently supported for images.
3919 */
3920 uint32_t memory_types = (1ull << device->physical->memory.type_count) - 1;
3921
3922 pMemoryRequirements->size = image->size;
3923 pMemoryRequirements->alignment = image->alignment;
3924 pMemoryRequirements->memoryTypeBits = memory_types;
3925 }
3926
3927 void anv_GetImageMemoryRequirements2(
3928 VkDevice _device,
3929 const VkImageMemoryRequirementsInfo2* pInfo,
3930 VkMemoryRequirements2* pMemoryRequirements)
3931 {
3932 ANV_FROM_HANDLE(anv_device, device, _device);
3933 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3934
3935 anv_GetImageMemoryRequirements(_device, pInfo->image,
3936 &pMemoryRequirements->memoryRequirements);
3937
3938 vk_foreach_struct_const(ext, pInfo->pNext) {
3939 switch (ext->sType) {
3940 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3941 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3942 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3943 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3944 plane_reqs->planeAspect);
3945
3946 assert(image->planes[plane].offset == 0);
3947
3948 /* The Vulkan spec (git aaed022) says:
3949 *
3950 * memoryTypeBits is a bitfield and contains one bit set for every
3951 * supported memory type for the resource. The bit `1<<i` is set
3952 * if and only if the memory type `i` in the
3953 * VkPhysicalDeviceMemoryProperties structure for the physical
3954 * device is supported.
3955 *
3956 * All types are currently supported for images.
3957 */
3958 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3959 (1ull << device->physical->memory.type_count) - 1;
3960
3961 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3962 pMemoryRequirements->memoryRequirements.alignment =
3963 image->planes[plane].alignment;
3964 break;
3965 }
3966
3967 default:
3968 anv_debug_ignored_stype(ext->sType);
3969 break;
3970 }
3971 }
3972
3973 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3974 switch (ext->sType) {
3975 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3976 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3977 if (image->needs_set_tiling || image->external_format) {
3978 /* If we need to set the tiling for external consumers, we need a
3979 * dedicated allocation.
3980 *
3981 * See also anv_AllocateMemory.
3982 */
3983 requirements->prefersDedicatedAllocation = true;
3984 requirements->requiresDedicatedAllocation = true;
3985 } else {
3986 requirements->prefersDedicatedAllocation = false;
3987 requirements->requiresDedicatedAllocation = false;
3988 }
3989 break;
3990 }
3991
3992 default:
3993 anv_debug_ignored_stype(ext->sType);
3994 break;
3995 }
3996 }
3997 }
3998
3999 void anv_GetImageSparseMemoryRequirements(
4000 VkDevice device,
4001 VkImage image,
4002 uint32_t* pSparseMemoryRequirementCount,
4003 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
4004 {
4005 *pSparseMemoryRequirementCount = 0;
4006 }
4007
4008 void anv_GetImageSparseMemoryRequirements2(
4009 VkDevice device,
4010 const VkImageSparseMemoryRequirementsInfo2* pInfo,
4011 uint32_t* pSparseMemoryRequirementCount,
4012 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
4013 {
4014 *pSparseMemoryRequirementCount = 0;
4015 }
4016
4017 void anv_GetDeviceMemoryCommitment(
4018 VkDevice device,
4019 VkDeviceMemory memory,
4020 VkDeviceSize* pCommittedMemoryInBytes)
4021 {
4022 *pCommittedMemoryInBytes = 0;
4023 }
4024
4025 static void
4026 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
4027 {
4028 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
4029 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
4030
4031 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
4032
4033 if (mem) {
4034 buffer->address = (struct anv_address) {
4035 .bo = mem->bo,
4036 .offset = pBindInfo->memoryOffset,
4037 };
4038 } else {
4039 buffer->address = ANV_NULL_ADDRESS;
4040 }
4041 }
4042
4043 VkResult anv_BindBufferMemory(
4044 VkDevice device,
4045 VkBuffer buffer,
4046 VkDeviceMemory memory,
4047 VkDeviceSize memoryOffset)
4048 {
4049 anv_bind_buffer_memory(
4050 &(VkBindBufferMemoryInfo) {
4051 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
4052 .buffer = buffer,
4053 .memory = memory,
4054 .memoryOffset = memoryOffset,
4055 });
4056
4057 return VK_SUCCESS;
4058 }
4059
4060 VkResult anv_BindBufferMemory2(
4061 VkDevice device,
4062 uint32_t bindInfoCount,
4063 const VkBindBufferMemoryInfo* pBindInfos)
4064 {
4065 for (uint32_t i = 0; i < bindInfoCount; i++)
4066 anv_bind_buffer_memory(&pBindInfos[i]);
4067
4068 return VK_SUCCESS;
4069 }
4070
4071 VkResult anv_QueueBindSparse(
4072 VkQueue _queue,
4073 uint32_t bindInfoCount,
4074 const VkBindSparseInfo* pBindInfo,
4075 VkFence fence)
4076 {
4077 ANV_FROM_HANDLE(anv_queue, queue, _queue);
4078 if (anv_device_is_lost(queue->device))
4079 return VK_ERROR_DEVICE_LOST;
4080
4081 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
4082 }
4083
4084 // Event functions
4085
4086 VkResult anv_CreateEvent(
4087 VkDevice _device,
4088 const VkEventCreateInfo* pCreateInfo,
4089 const VkAllocationCallbacks* pAllocator,
4090 VkEvent* pEvent)
4091 {
4092 ANV_FROM_HANDLE(anv_device, device, _device);
4093 struct anv_event *event;
4094
4095 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
4096
4097 event = vk_alloc2(&device->vk.alloc, pAllocator, sizeof(*event), 8,
4098 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4099 if (event == NULL)
4100 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4101
4102 vk_object_base_init(&device->vk, &event->base, VK_OBJECT_TYPE_EVENT);
4103 event->state = anv_state_pool_alloc(&device->dynamic_state_pool,
4104 sizeof(uint64_t), 8);
4105 *(uint64_t *)event->state.map = VK_EVENT_RESET;
4106
4107 *pEvent = anv_event_to_handle(event);
4108
4109 return VK_SUCCESS;
4110 }
4111
4112 void anv_DestroyEvent(
4113 VkDevice _device,
4114 VkEvent _event,
4115 const VkAllocationCallbacks* pAllocator)
4116 {
4117 ANV_FROM_HANDLE(anv_device, device, _device);
4118 ANV_FROM_HANDLE(anv_event, event, _event);
4119
4120 if (!event)
4121 return;
4122
4123 anv_state_pool_free(&device->dynamic_state_pool, event->state);
4124
4125 vk_object_base_finish(&event->base);
4126 vk_free2(&device->vk.alloc, pAllocator, event);
4127 }
4128
4129 VkResult anv_GetEventStatus(
4130 VkDevice _device,
4131 VkEvent _event)
4132 {
4133 ANV_FROM_HANDLE(anv_device, device, _device);
4134 ANV_FROM_HANDLE(anv_event, event, _event);
4135
4136 if (anv_device_is_lost(device))
4137 return VK_ERROR_DEVICE_LOST;
4138
4139 return *(uint64_t *)event->state.map;
4140 }
4141
4142 VkResult anv_SetEvent(
4143 VkDevice _device,
4144 VkEvent _event)
4145 {
4146 ANV_FROM_HANDLE(anv_event, event, _event);
4147
4148 *(uint64_t *)event->state.map = VK_EVENT_SET;
4149
4150 return VK_SUCCESS;
4151 }
4152
4153 VkResult anv_ResetEvent(
4154 VkDevice _device,
4155 VkEvent _event)
4156 {
4157 ANV_FROM_HANDLE(anv_event, event, _event);
4158
4159 *(uint64_t *)event->state.map = VK_EVENT_RESET;
4160
4161 return VK_SUCCESS;
4162 }
4163
4164 // Buffer functions
4165
4166 VkResult anv_CreateBuffer(
4167 VkDevice _device,
4168 const VkBufferCreateInfo* pCreateInfo,
4169 const VkAllocationCallbacks* pAllocator,
4170 VkBuffer* pBuffer)
4171 {
4172 ANV_FROM_HANDLE(anv_device, device, _device);
4173 struct anv_buffer *buffer;
4174
4175 /* Don't allow creating buffers bigger than our address space. The real
4176 * issue here is that we may align up the buffer size and we don't want
4177 * doing so to cause roll-over. However, no one has any business
4178 * allocating a buffer larger than our GTT size.
4179 */
4180 if (pCreateInfo->size > device->physical->gtt_size)
4181 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
4182
4183 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
4184
4185 buffer = vk_alloc2(&device->vk.alloc, pAllocator, sizeof(*buffer), 8,
4186 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4187 if (buffer == NULL)
4188 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4189
4190 vk_object_base_init(&device->vk, &buffer->base, VK_OBJECT_TYPE_BUFFER);
4191 buffer->size = pCreateInfo->size;
4192 buffer->usage = pCreateInfo->usage;
4193 buffer->address = ANV_NULL_ADDRESS;
4194
4195 *pBuffer = anv_buffer_to_handle(buffer);
4196
4197 return VK_SUCCESS;
4198 }
4199
4200 void anv_DestroyBuffer(
4201 VkDevice _device,
4202 VkBuffer _buffer,
4203 const VkAllocationCallbacks* pAllocator)
4204 {
4205 ANV_FROM_HANDLE(anv_device, device, _device);
4206 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
4207
4208 if (!buffer)
4209 return;
4210
4211 vk_object_base_finish(&buffer->base);
4212 vk_free2(&device->vk.alloc, pAllocator, buffer);
4213 }
4214
4215 VkDeviceAddress anv_GetBufferDeviceAddress(
4216 VkDevice device,
4217 const VkBufferDeviceAddressInfoKHR* pInfo)
4218 {
4219 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
4220
4221 assert(!anv_address_is_null(buffer->address));
4222 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
4223
4224 return anv_address_physical(buffer->address);
4225 }
4226
4227 uint64_t anv_GetBufferOpaqueCaptureAddress(
4228 VkDevice device,
4229 const VkBufferDeviceAddressInfoKHR* pInfo)
4230 {
4231 return 0;
4232 }
4233
4234 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4235 VkDevice device,
4236 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR* pInfo)
4237 {
4238 ANV_FROM_HANDLE(anv_device_memory, memory, pInfo->memory);
4239
4240 assert(memory->bo->flags & EXEC_OBJECT_PINNED);
4241 assert(memory->bo->has_client_visible_address);
4242
4243 return gen_48b_address(memory->bo->offset);
4244 }
4245
4246 void
4247 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
4248 enum isl_format format,
4249 struct anv_address address,
4250 uint32_t range, uint32_t stride)
4251 {
4252 isl_buffer_fill_state(&device->isl_dev, state.map,
4253 .address = anv_address_physical(address),
4254 .mocs = device->isl_dev.mocs.internal,
4255 .size_B = range,
4256 .format = format,
4257 .swizzle = ISL_SWIZZLE_IDENTITY,
4258 .stride_B = stride);
4259 }
4260
4261 void anv_DestroySampler(
4262 VkDevice _device,
4263 VkSampler _sampler,
4264 const VkAllocationCallbacks* pAllocator)
4265 {
4266 ANV_FROM_HANDLE(anv_device, device, _device);
4267 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
4268
4269 if (!sampler)
4270 return;
4271
4272 if (sampler->bindless_state.map) {
4273 anv_state_pool_free(&device->dynamic_state_pool,
4274 sampler->bindless_state);
4275 }
4276
4277 vk_object_base_finish(&sampler->base);
4278 vk_free2(&device->vk.alloc, pAllocator, sampler);
4279 }
4280
4281 VkResult anv_CreateFramebuffer(
4282 VkDevice _device,
4283 const VkFramebufferCreateInfo* pCreateInfo,
4284 const VkAllocationCallbacks* pAllocator,
4285 VkFramebuffer* pFramebuffer)
4286 {
4287 ANV_FROM_HANDLE(anv_device, device, _device);
4288 struct anv_framebuffer *framebuffer;
4289
4290 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
4291
4292 size_t size = sizeof(*framebuffer);
4293
4294 /* VK_KHR_imageless_framebuffer extension says:
4295 *
4296 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4297 * parameter pAttachments is ignored.
4298 */
4299 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
4300 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4301 framebuffer = vk_alloc2(&device->vk.alloc, pAllocator, size, 8,
4302 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4303 if (framebuffer == NULL)
4304 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4305
4306 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4307 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4308 framebuffer->attachments[i] = iview;
4309 }
4310 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4311 } else {
4312 framebuffer = vk_alloc2(&device->vk.alloc, pAllocator, size, 8,
4313 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4314 if (framebuffer == NULL)
4315 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4316
4317 framebuffer->attachment_count = 0;
4318 }
4319
4320 vk_object_base_init(&device->vk, &framebuffer->base,
4321 VK_OBJECT_TYPE_FRAMEBUFFER);
4322
4323 framebuffer->width = pCreateInfo->width;
4324 framebuffer->height = pCreateInfo->height;
4325 framebuffer->layers = pCreateInfo->layers;
4326
4327 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4328
4329 return VK_SUCCESS;
4330 }
4331
4332 void anv_DestroyFramebuffer(
4333 VkDevice _device,
4334 VkFramebuffer _fb,
4335 const VkAllocationCallbacks* pAllocator)
4336 {
4337 ANV_FROM_HANDLE(anv_device, device, _device);
4338 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4339
4340 if (!fb)
4341 return;
4342
4343 vk_object_base_finish(&fb->base);
4344 vk_free2(&device->vk.alloc, pAllocator, fb);
4345 }
4346
4347 static const VkTimeDomainEXT anv_time_domains[] = {
4348 VK_TIME_DOMAIN_DEVICE_EXT,
4349 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4350 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4351 };
4352
4353 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4354 VkPhysicalDevice physicalDevice,
4355 uint32_t *pTimeDomainCount,
4356 VkTimeDomainEXT *pTimeDomains)
4357 {
4358 int d;
4359 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4360
4361 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4362 vk_outarray_append(&out, i) {
4363 *i = anv_time_domains[d];
4364 }
4365 }
4366
4367 return vk_outarray_status(&out);
4368 }
4369
4370 static uint64_t
4371 anv_clock_gettime(clockid_t clock_id)
4372 {
4373 struct timespec current;
4374 int ret;
4375
4376 ret = clock_gettime(clock_id, &current);
4377 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4378 ret = clock_gettime(CLOCK_MONOTONIC, &current);
4379 if (ret < 0)
4380 return 0;
4381
4382 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4383 }
4384
4385 #define TIMESTAMP 0x2358
4386
4387 VkResult anv_GetCalibratedTimestampsEXT(
4388 VkDevice _device,
4389 uint32_t timestampCount,
4390 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4391 uint64_t *pTimestamps,
4392 uint64_t *pMaxDeviation)
4393 {
4394 ANV_FROM_HANDLE(anv_device, device, _device);
4395 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4396 int ret;
4397 int d;
4398 uint64_t begin, end;
4399 uint64_t max_clock_period = 0;
4400
4401 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4402
4403 for (d = 0; d < timestampCount; d++) {
4404 switch (pTimestampInfos[d].timeDomain) {
4405 case VK_TIME_DOMAIN_DEVICE_EXT:
4406 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
4407 &pTimestamps[d]);
4408
4409 if (ret != 0) {
4410 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4411 "register: %m");
4412 }
4413 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4414 max_clock_period = MAX2(max_clock_period, device_period);
4415 break;
4416 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4417 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4418 max_clock_period = MAX2(max_clock_period, 1);
4419 break;
4420
4421 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4422 pTimestamps[d] = begin;
4423 break;
4424 default:
4425 pTimestamps[d] = 0;
4426 break;
4427 }
4428 }
4429
4430 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4431
4432 /*
4433 * The maximum deviation is the sum of the interval over which we
4434 * perform the sampling and the maximum period of any sampled
4435 * clock. That's because the maximum skew between any two sampled
4436 * clock edges is when the sampled clock with the largest period is
4437 * sampled at the end of that period but right at the beginning of the
4438 * sampling interval and some other clock is sampled right at the
4439 * begining of its sampling period and right at the end of the
4440 * sampling interval. Let's assume the GPU has the longest clock
4441 * period and that the application is sampling GPU and monotonic:
4442 *
4443 * s e
4444 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4445 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4446 *
4447 * g
4448 * 0 1 2 3
4449 * GPU -----_____-----_____-----_____-----_____
4450 *
4451 * m
4452 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4453 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4454 *
4455 * Interval <----------------->
4456 * Deviation <-------------------------->
4457 *
4458 * s = read(raw) 2
4459 * g = read(GPU) 1
4460 * m = read(monotonic) 2
4461 * e = read(raw) b
4462 *
4463 * We round the sample interval up by one tick to cover sampling error
4464 * in the interval clock
4465 */
4466
4467 uint64_t sample_interval = end - begin + 1;
4468
4469 *pMaxDeviation = sample_interval + max_clock_period;
4470
4471 return VK_SUCCESS;
4472 }
4473
4474 /* vk_icd.h does not declare this function, so we declare it here to
4475 * suppress Wmissing-prototypes.
4476 */
4477 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4478 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4479
4480 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4481 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4482 {
4483 /* For the full details on loader interface versioning, see
4484 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4485 * What follows is a condensed summary, to help you navigate the large and
4486 * confusing official doc.
4487 *
4488 * - Loader interface v0 is incompatible with later versions. We don't
4489 * support it.
4490 *
4491 * - In loader interface v1:
4492 * - The first ICD entrypoint called by the loader is
4493 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4494 * entrypoint.
4495 * - The ICD must statically expose no other Vulkan symbol unless it is
4496 * linked with -Bsymbolic.
4497 * - Each dispatchable Vulkan handle created by the ICD must be
4498 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4499 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4500 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4501 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4502 * such loader-managed surfaces.
4503 *
4504 * - Loader interface v2 differs from v1 in:
4505 * - The first ICD entrypoint called by the loader is
4506 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4507 * statically expose this entrypoint.
4508 *
4509 * - Loader interface v3 differs from v2 in:
4510 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4511 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4512 * because the loader no longer does so.
4513 *
4514 * - Loader interface v4 differs from v3 in:
4515 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4516 */
4517 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4518 return VK_SUCCESS;
4519 }
4520
4521 VkResult anv_CreatePrivateDataSlotEXT(
4522 VkDevice _device,
4523 const VkPrivateDataSlotCreateInfoEXT* pCreateInfo,
4524 const VkAllocationCallbacks* pAllocator,
4525 VkPrivateDataSlotEXT* pPrivateDataSlot)
4526 {
4527 ANV_FROM_HANDLE(anv_device, device, _device);
4528 return vk_private_data_slot_create(&device->vk, pCreateInfo, pAllocator,
4529 pPrivateDataSlot);
4530 }
4531
4532 void anv_DestroyPrivateDataSlotEXT(
4533 VkDevice _device,
4534 VkPrivateDataSlotEXT privateDataSlot,
4535 const VkAllocationCallbacks* pAllocator)
4536 {
4537 ANV_FROM_HANDLE(anv_device, device, _device);
4538 vk_private_data_slot_destroy(&device->vk, privateDataSlot, pAllocator);
4539 }
4540
4541 VkResult anv_SetPrivateDataEXT(
4542 VkDevice _device,
4543 VkObjectType objectType,
4544 uint64_t objectHandle,
4545 VkPrivateDataSlotEXT privateDataSlot,
4546 uint64_t data)
4547 {
4548 ANV_FROM_HANDLE(anv_device, device, _device);
4549 return vk_object_base_set_private_data(&device->vk,
4550 objectType, objectHandle,
4551 privateDataSlot, data);
4552 }
4553
4554 void anv_GetPrivateDataEXT(
4555 VkDevice _device,
4556 VkObjectType objectType,
4557 uint64_t objectHandle,
4558 VkPrivateDataSlotEXT privateDataSlot,
4559 uint64_t* pData)
4560 {
4561 ANV_FROM_HANDLE(anv_device, device, _device);
4562 vk_object_base_get_private_data(&device->vk,
4563 objectType, objectHandle,
4564 privateDataSlot, pData);
4565 }